WO2025259196A1 - A system for and method of predicting bilirubin levels - Google Patents

Abstract

A system for and method of predicting bilirubin levels based on a plurality of images. The method including: pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject; color calibrating the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; generating color-related features based on the plurality of color calibrated patches; and providing the color-related features as a predictor input to a machine learning model to generate a prediction of bilirubin levels of the subject.

Description

A SYSTEM FOR AND METHOD OF PREDICTING BILIRUBIN LEVELS

CROSS-REFERENCE TO RELATED APPLICATION

[0001] This application claims the benefit of priority to the Singapore application no. 10202401708T filed 13 June, 2024, the contents of which are hereby incorporated by reference in their entirety for all purposes.

TECHNICAL FIELD

[0002] This application relates generally to the field of health screening, and more particularly, to a system for predicting bilirubin levels and a method of predicting bilirubin levels.

BACKGROUND

[0003] Neonatal j aundice (NNJ), or hyperbilirubinaemia, affects approximately 60% of term and 80% of preterm infants. Early detection of NNJ is essential for preventing bilirubin-induced neurological dysfunction, which includes kernicterus. The current gold standard for diagnosis, total serum bilirubin (TSB) measurement, is invasive, resource-intensive, and costly. While non-invasive methods such as visual inspection and icterometry have been explored, their accuracy remains variable.

[0004] Transcutaneous bilirubinometry (TcB), which uses optical spectroscopy to estimate bilirubin levels through the skin, is a candidate for non-invasive screening. However, use of TcB is restricted to healthcare facilities and requires regular calibration, maintenance, and trained personnel, leading to increased healthcare costs and inconvenience for families. SUMMARY

[0005] According to an aspect, disclosed herein a system. The system comprises: memory storing instructions; and a processor coupled to the memory and configured to process the stored instructions to implement: a module configured to perform a method of predicting bilirubin levels based on a plurality of images, the method including: pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject; color calibrating the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; generating color-related features based on the plurality of color calibrated patches; and providing the color-related features as a predictor input to a machine learning model to generate a prediction of bilirubin levels of the subject.

[0006] According to another aspect, disclosed herein a method of predicting bilirubin levels based on a plurality of images, the method including: pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject; color calibrating the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; generating color-related features based on the plurality of color calibrated patches; and providing the color-related features as a predictor input to a machine learning model to generate a prediction of bilirubin levels of the subject. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Various embodiments of the present disclosure are described below with reference to the following drawings:

[0008] FIG. 1 is a schematic diagram showing a system for predicting bilirubin levels according to embodiments of the present disclosure;

[0009] FIG. 2 is a schematic diagram showing a calibration card according to various embodiments,

[0010] FIG. 3 A is a schematic diagram illustrating the system for predicting bilirubin levels of FIG. 1 capturing a first image;

[0011] FIG. 3B is a schematic diagram illustrating the system for predicting bilirubin levels of FIG. 1 capturing a second image;

[0012] FIG. 4 is a flow chart illustrating a system for and a method of predicting bilirubin levels according to embodiments of the present disclosure;

[0013] FIG. 5 is a schematic diagram illustrating disposing multiple calibration cards on multiple unique body surfaces of a subject according to embodiments of the present disclosure; [0014] FIG. 6 is a schematic diagram illustrating a system for predicting bilirubin levels capturing an image of a skin surface according to embodiments of the present disclosure;

[0015] FIG. 7 is a schematic diagram illustrating a system for predicting bilirubin levels capturing an image of a sclera according to embodiments of the present disclosure;

[0016] FIG. 8 shows an outcome of a SNAP analysis of a system for predicting bilirubin levels according to embodiments of the present disclosure;

[0017] FIG. 9A is a schematic diagram showing a calibration card with an occluding portion according to various embodiments;

[0018] FIG. 9B is a schematic diagram showing a calibration card with an occluder according to various embodiments; [0019] FIG. 10 is a flow chart illustrating a method of predicting bilirubin levels according to embodiments of the present disclosure;

[0020] FIG. 11A is an image of an exemplary implementation BiliSG of the proposed system;

[0021] FIG. 1 IB illustrates the elimination of shadows using a color sticker of the exemplary implementation BiliSG of FIG. 1 1 A;

[0022] FIG. 12 shows a sternal-abdomen “yellowness” gradient across TSB range;

[0023] FIG. 13 is a consort diagram of patient recruitment;

[0024] FIGs. 14A and 14B show a distribution of Total Serum Bilirubin (TSB) Levels of the exemplary implementation;

[0025] FIG. 15 A is a correlation plot between ground truth and model predictions, with each dot representing a paired observation between the reference standard TSB and the SpB for individual measurements. The scatter of the dots shows the relationship between the actual and predicted values. The solid line represents the line of best fit through the data points, showing the trend in the relationship between reference standard TSB and SpB;

[0026] FIG. 15B is a Bland-Altman plot, with each dot representing the difference between the SpB and reference standard TSB levels for each observation, plotted against the mean of the SpB and reference standard TSB values for that observation;

[0027] FIG. 16 shows Receiver Operating Characteristic (ROC) Curves of Screening Tool Bilirubin (SpB) and Transcutaneous Bilirubin (TcB); and

[0028] FIG. 17 is a schematic diagram of a processor system

DETAILED DESCRIPTION

[0029] The following detailed description is made with reference to the accompanying drawings, showing details and embodiments of the present disclosure for the purposes of illustration. Features that are described in the context of an embodiment may correspondingly be applicable to the same or similar features in the other embodiments, even if not explicitly described in these other embodiments. Additions and/or combinations and/or alternatives as described for a feature in the context of an embodiment may correspondingly be applicable to the same or similar feature in the other embodiments.

[0030] In the context of various embodiments, the articles “a”, “an” and “the” as used with regard to a feature or element include a reference to one or more of the features or elements.

[0031] In the context of various embodiments, the term “about” or “approximately” as applied to a numeric value encompasses the exact value and a reasonable variance as generally understood in the relevant technical field, e.g., within 10% of the specified value.

[0032] As used herein, the term “and/or” includes any and all combinations of one or more of the associated listed items.

[0033] As used herein, terms “concurrently”, “simultaneously”, “at the same time”, or the like, may refer to events or actions that coincide or overlap within a period of time, regardless of whether the events start at the same time instant, and regardless of whether the events end at the same time instant.

[0034] As used herein, the term “skin patch” which may be segmented from an image may be used interchangeably with the terms “surface patch”, “skin region”, “skin surface”, “skin area”, etc., and may generally refer to an image of an external body surface (including skin surface) of a subject or a segmented portion of an image of an external body surface of a subject, as it may be understood.

[0035] As used herein, the term “calibration patch” which may be segmented from an image may be used interchangeably with the tenns “calibration region”, “calibration surface”, “calibration area”, etc., and may generally refer to a predetermined calibration image, a combination of predetermined calibration images, a predetermined calibration color patch, a combination of predetermined calibration color patches, a predetermined imaging pattern, a combination of predetermined imaging patterns, and/or a combination thereof, as it may be understood.

[0036] As used herein, the term “unique body surface(s)” may generally refer to external surfaces (such as skin surfaces) of different body parts of a subject. As an exemplary illustration, a skin surface of a forehead and a skin surface of a chest of the same subject may be deemed “unique body surfaces”. As another exemplary illustration, a first skin portion of a forehead and a second skin portion of the forehead of the same subject may not be deemed “unique body surfaces”, even if the first skin portion does not overlap with the second skin portion. In yet another exemplary illustration, a scleral surface of an eye and a skin surface of an abdomen of the same subject may be deemed “unique body surfaces”.

[0037] Neonatal Jaundice (NNJ), or hyperbilirubinaemia, is a common problem among babies, affecting 60% of term and 80% of preterm babies. Early detection is essential to prevent morbidity and mortality associated with bilirubin toxicity. Conventional approach towards assessing NNJ may either be invasive and/or mainly confined to healthcare facilities, precluding home-based NNJ screening.

[0038] Other approaches, such as smartphone applications (apps), have emerged as alternatives for estimating bilirubin levels using digital images of the skin. However, existing solutions typically result in inconsistent performance across different regions of the body or different skin surfaces.

[0039] Departing from conventional approaches, the proposed system and method utilize images of different regions of interest (body surfaces of the subject) for bilirubin level prediction, which at the same time, corresponds to the cephalocaudal advancement of dennal icterus, which often finds relevance in NNJ evaluation. It is also worth noting that no existing smartphone-based NNJ apps concurrently use digital images from multiple regions of interest arranged in a cephalocaudal fashion, such as from the forehead, sternum, and abdomen, as predictors for a single bilirubin estimate. Further departing from conventional approaches, the proposed system and method steer away from making a prediction based on absolute measurement(s) of a single body surface and/or multiple independent measurements of the body surfaces.

[0040] The present disclosure proposes a system and method to predict or estimate bilirubin levels in newborn infants based on multiple images of external surfaces (such as skin surfaces) of the subject (infant/newbom). The proposed system and method may be applicable to a multiethnic neonatal population. The present disclosure may also include color analysis of skin images and/or scleral images implemented as a machine learning-based system and method. In various embodiments, the proposed system and method may be implemented in a portable device, thus overcoming limitations of mandatory healthcare facilities-based NJJ screening, enabling monitoring of bilirubin levels in newborn infants in the comfort of home-based screening. In addition, continuous monitoring of bilirubin levels may be performed over multiple data points, thus providing a comprehensive trend in the alteration of bilirubin levels and enabling prompt medical intervention or attention

[0041] Departing from conventional methods, the proposed system and method may comprise the steps of pre-processing and color calibrating prior to predicting the bilirubin levels. The proposed system and method may predict bilirubin levels based on multiple images captured of the subject, in which each image may include a calibration card or a calibration device placed adjacent to the skin surface of the subject Provision of the calibration card aids in overcoming the limitations posed by capturing images in a variety of lighting conditions, such as in bright light, in shade, in artificial lighting, etc In various embodiments, the calibration card may act as a reference for calibrating or correcting for color shifts under the influence of different lighting conditions. In addition, for images taken of a skin surface of the subject, the calibration card may aid in determining a skin tone of the subject.

[0042] Further, apart from assisting with color calibration or color correction, the calibration card may also aid in the segmentation of a skin patch or skin surface from the captured images. In exemplary embodiments, the calibration card may define a window or a through opening for exposing the external surface of the subject to an imaging device. In addition, the calibration card may also aid in the segmentation of calibration patch or calibration zones from the captured images.

[0043] Based on one or more skin patches and calibration patches, the proposed system and method may determine one or more color calibrated patches or calibrated skin patches, which correspond to the skin surface of the subject, and further, to generate one or more color-related features based on the one or more color calibrated patches which may act as inputs to a machine learning model for predicting bilirubin levels.

[0044] An exemplary color-related feature includes a yellowness gradient determined based on unique body surfaces or different body surfaces of the subject. The yellowness gradient may be obtained from images captured of multiple skin surfaces of a subject. It may be noted that each of the images captured may be segmented and calibrated prior to determining the yellowness gradient.

[0045] Another exemplary color-related feature may include a skin tone of the subject determining based on a skin tone scale, such as a Monk Skin Test Scale or a Fitzpatrick skin phototypes The skin tone scale may aid in overcoming clinical implication of the reduced reliability of visual estimation of neonatal jaundice in babies with darker skin tone, thus improving the robustness of bilirubin level prediction.

[0046] Referring to FIGs. 1 to 7, in one aspect, the proposed system 100 according to some embodiments of the present disclosure is configurable as a device 110 for predicting bilirubin levels. The device 110 may be configured for predicting bilirubin levels based on a plurality of images of a subject 80 (such as an infant).

[0047] According to various embodiments as shown in FIG. 1, the device 110 may include a processor 900 configured to perform a method of predicting bilirubin levels based on a plurality of images of a subject 80. The device 110 may further include an imaging device 120, such as a color camera, for capturing the plurality of images. The imaging device 120 may be in signal communication 130 with the processor 900 such that the plurality of images are provided to the processor 900 for further processing or computing. In various embodiments, the imaging device 120 and the processor 900 may be in a wired signal communication such that the processor 900 may be located locally relative to the imaging device 120. In other embodiments, the processor 900 may be in a wireless signal communication with the imaging device 120, such that the processor 900 may be located remotely relative to the imaging device 120 and subject 80. For example, the processor 900 may be a server located in a medical facility. [0048] According to various embodiments, referring to FIGs. 1 and 2, the system 100 may further include one or more calibration devices such as calibration cards 200. Referring to FIG. 2, the calibration cards 200 may comprise a frame portion 210 defining a window 220. The window 220 may expose a skin surface 82/83 of the subject 80 during the capturing of the plurality of images.

[0049] In various embodiments, the frame portion 210 may comprise a plurality of colors. In an example, the plurality of colors may comprise at least 20 colors. In another example, the plurality of colors may comprise 20 to 40 colors. In another example, the plurality of colors may comprise 32 to 40 colors.

[0050] In various embodiments, the plurality of colors may be divided into one or more arrays 212/213/214/215 of the plurality of colors, or color arrays 212/213/214/215. The color arrays 212/213/214/215 may act as calibration patches for color calibration or correction as well as borders for image segmentation. In various embodiments, the color arrays 212/213/214/215 may correspond to the four peripheral sides of the window 220. This enables each of the color arrays 212/213/214/215 to be disposed positioned adjacent to the window 220 enabling convenient referencing.

[0051] In various embodiments, selected ones of the plurality of colors may correspond to colors for color calibration/correction under different lighting conditions, such as indoor lighting, dimmed lighting, etc.

[0052] In various embodiments, selected ones of the plurality of colors correspond to skin tone scale, such as a Monk Skin Test scale or a Fitzpatrick skin phototypes. For example, a portion of color array 214 may include colors corresponding to the Monk Skin Tone (MST) scale. The skin tone scale provides multiple skin tone colors to allow a comparison of color with the body surfaces 82/83 under different lighting conditions.

[0053] In various embodiments, the calibration card 220 may comprise a thumb portion 211 protruding from the frame portion 210, for allowing easy handling by a user, such as a parent of the subject 80. The thumb portion 211 enables the calibration card 220 to be positioned on a target body surface of the subject 80 with minimal disturbance to the subject 80 (e g. when the infant is asleep).

[0054] In various embodiments, the calibration card 220 may be generally flexible or conformable to different body surfaces 82/83, such as a forehead 82 or a chest 83 of the subject 80. The calibration card 220 may be configured to be bendable for conforming to the body surfaces 82/83. However, the calibration card 220 may be limited in the extension direction and hence with limited stretchability.

[0055] In addition, the calibration card 220 may comprise an adhesive layer for adhering and conforming to the body surface 82/83. This advantageously aids in reducing or mitigating shadows from forming in the window 220, thus improving the quality of the images captured. [0056] Alternatively, the calibration card 220 may be provided with a non-adhesive backing to provide a reusable or repositionable calibration card 220, which allows multiple usage and may be placed on different body surfaces 82/83 of the subject 80.

[0057] In various embodiments, the frame portion 210 may further comprise at least one visual marker 216. The visual marker(s) 216 may aid in determining an orientation and/or position of the calibration card 220 in the captured images Tn an exemplary embodiment shown in FIG. 2, the frame portion 210 may include four visual markers 216 disposed on the four corners of the frame portion 210. The visual markers 216 may be ArUco markers.

[0058] According to various embodiments of the proposed system 100, as shown in FIGs.

3 A and 3B, the device 110 for predicting bilirubin levels may be configured as a portable device or a portable smart device, such as a mobile phone, operable by a user (or parent of the subject 80). Similarly, the portable device 110 may include an imaging device 120 for capturing a plurality of images 310 of a subject 80. In addition, the imaging device 120 may be in signal communication 130 with a processor 900 configured to perform a method of predicting bilirubin levels based on the plurality of images 310 provided by the imaging device 120. In various embodiments, the imaging device 120 and the processor 900 may be integrated into a single portable device 110. Therefore, the process of image capturing as well as the prediction of bilirubin levels may be collectively integrated to run on a mobile application

[0059] FIGs. 4 to 7 illustrate a method 400 of predicting bilirubin levels according to various embodiments. The method 400 of predicting bilirubin levels may begin with adhering a calibration card 200 to unique body surfaces 82/83 of the subject 80. Tn addition, the method 400 of predicting bilirubin levels may include capturing a plurality of images 310 of the unique body surfaces 82/83. As such, each of the plurality of images 310 may correspond to an image of a unique body surface of the subject 80. In various embodiments, the plurality of images 310 may comprise at least: an image of a skin surface on a first body surface selected from a plurality of unique body surfaces, and an image of a skin surface on a second body surface selected from the plurality of unique body surfaces, wherein the first body surface is superior relative to the second body surface. In other words, the first body surface is closer to the head of the subject in comparison to the second body surface, such that the first body surface and the second body surface are located in a cephalocaudal fashion. It may be understood that the plurality of images 310 may include two or more images of respective unique body surfaces. In various embodiments, the plurality of images 310 may further comprises: an image of a skin surface on a third body surface, an image of a skin surface on a fourth body surface and an image of a skin surface on a fifth body surface, wherein the second body surface is superior to the third body surface, wherein the third body surface is superior to the fourth body surface, and the further body surface is superior to the fifth body surface. In an example, the plurality of images 310 may include five images corresponding to five unique body surfaces, such as a forehead, a sternum, an abdomen, a shin (or a thigh), and a feet of a subject 80.

[0060] FIGs. 3 A and 3B illustrate exemplary images 310 captured and used for the method of predicting bilirubin levels. FIG. 3A shows an exemplary image 310A corresponding to a forehead 82 (a first body surface) of the subject 80. FIG. 3B shows an exemplary image 310B corresponding to a sternum 83 (a second body surface) of the subject 80. Each of the plurality of images 310 may include a picture or a photo of the calibration card 200 disposed on a body surface or skin surface 82/83 of the subject 80. It may be noted that the skin surfaces 82/83 may be captured through the window 220 of the calibration card. Hence, each of the plurality of images 310 may include a calibration patch corresponding to the frame portion 210 of the calibration card 200, and a skin patch corresponding to the window 220 of the calibration card 200.

[0061] Referring to FIG. 4, according to various embodiments, the method 400 of predicting bilirubin levels may also include pre-processing 410 each of the plurality of images 310 by segmenting a skin patch 322 and a calibration patch 324 in each of the plurality of images, to obtain a plurality of pre-processed images 320. The skin patch 322 may correspond to the window 220 of the calibration card 200 and the calibration patch 324 may correspond to the frame portion 210 of the calibration card 200.

[0062] In various embodiments, prior to pre-processing 410 each of the plurality of images 310, the plurality of images 310 may be screened or checked according to various criteria, such as image quality, appropriate lighting, and distance of the calibration card from the imaging device. In various embodiments, images which did not meet the various criteria may be rejected and discarded, while images which satisfy the criteria may be retained and used for the later pre-processing 410 step.

[0063] According to embodiments of the proposed system 100 and method 400, referring to FIG. 5, the plurality of images 310 of the unique body surfaces of the subject may correspond to images taken multiple body surfaces, such as: a skin surface 82 on a forehead of the subject 80, a skin surface 83 on a sternum of the subject 80, a skin surface 84 on an abdomen of the subject 80, a skin surface 85 on a shin of the subject 80, and a skin surface 86 of a feet of the subject 80. As it may be understood, multiple calibration cards 200 may be positioned or adhered to the subject 80 during the process. As such, each of the calibration cards 200 may correspond to the corresponding body surfaces 82/83/84/85 or calibration patches 324 in each of the plurality of images 310.

[0064] Referring to FIG. 6, in various alternative embodiments, a single image 310C of the plurality of images 310 may include a plurality of unique body surfaces 82/83/84 and corresponding calibration cards 200. As such, the processor 900 may capture and pre-process 410 a plurality of skin patches 322 and a plurality of calibration patch 324 from the image 310C. [0065] The method 400 of predicting bilirubin levels may also comprise color calibrating 420 the skin patch 322 in each of the plurality of pre-processed images 320, to obtain a plurality of color calibrated patches 330. The color calibration 420 performed on each skin patch 322 may be based on the calibration patch 324 associated with the skin patch 322. In other words, each of the skin patch 322 may be color calibrated with the corresponding calibration patch 324 present in the same image 320. In various embodiments, the skin patch 322 may be associated with the calibration patch 324 by being adjacent to one another in the same image 310.

[0066] In various embodiments, the color calibration 420 performed on each skin patch 322 may be performed using the calibration patch 324 and a reference calibration data (such as a reference calibration card). As an example, the reference calibration data may be obtained from an image of the calibration card 220 in an ideal lighting situation, such as 5000-6000K color temperature. In various embodiments, a calibration algorithm or calibration model may be trained based on a relationship between the calibration patch 324 (of the calibration card) and the reference calibration data. The calibration algorithm may be applied to the skin patch 322 as well as to the calibration patches 324 at the frame portion, thus generating the plurality of color calibrated patches 330 as well as one or more color calibrated calibration patches.

[0067] In various embodiments, the step of color calibrating 420 the skin patch 322 in each of the plurality of pre-processed images 320 may further comprise the step of color calibrating 420 the skin patch 322 in each of the plurality of pre-processed images 320 with the respective calibration patch 324 (and/or color calibrated calibration patch), based on a polynomial regression. In various embodiments, the polynomial regression is a second order polynomial regression.

[0068] In an exemplary embodiment, the calibration card 200 may comprise a total of 32 colors in the calibration patch 324, in which 8 colors are added to the commercially available 24-color MacBeth color card. The step of color calibrating 420 may be performed using the 32- color calibration card. The metric for evaluating the calibration techniques may be the Delta E (AE) difference of the processed calibration card and a reference calibration card. Delta E is a metric used to quantify the difference between two colors, and represents the distance between two colors in a three-dimensional color space. The CIELAB color space may be used for AE calculations due to the CIELAB color space bring perceptually uniform, in other words, the same amount of numerical change corresponds to about the same amount of visually perceived change. As such, AE may be computed in the CIELAB space for color calibration.

[0069] In another exemplary embodiment, the calibration card 200 may further comprise 8 additional color patches (total of 40 colors) in which 6 color patches represents the Fitzpatrick scale of skin phototypes. Three shortlisted color calibration techniques, namely, i) second order polynomial regression; ii) third order polynomial regression; and iii) Gaussian process regression, were performed on the 40 colors calibration card. The Delta E method was used with average AE was computed and compared across the 3 shortlisted methods.

[0070] It was found with surprise that the second order polynomial regression is a preferred color calibration method over the third order polynomial regression and the Gaussian process regression. While higher order polynomial regression tends to provide a better fitted regression outcome, visual inspection of the regression outcome showed a tendency to overprocess the overall image, resulting in the transformation of skin patches into colors that were not skin-like and often lacking of generalization. Further, the Gaussian process regression often requires the adjustment of multiple parameters, and hence is sensitive to particular dataset that the technique is applied on. Second order polynomial regression provided stability in calibrating images, lower tendency to overprocess and over-calibrate images, and greater capability in terms of generalization.

[0071] According to various embodiments, the step of pre-processing 410 each of the plurality of images 310 may further comprise the step of: segmenting a sclera patch 323 in each of the plurality of images 310D to obtain a plurality of pre-processed images 320. In addition, the step of color calibrating 420 may further comprise the step of: color calibrating the sclera patch 323 in each of the plurality of pre-processed images 320 based on the calibration patch 324 associated with the sclera patch 323 to obtain a plurality of color calibrated patches 330.

[0072] Further, the method 400 may comprise generating 430 one or more color-related features 340 based on the plurality of color calibrated patches 330. The color-related features 340 may be based on various color spaces, such as Cyan-Magenta-Yellow and Key (Black) (CMYK) color space, blue chromaticity, Jaundice Eye Color Index (JECI), LAB color space, Weighted Average Yellowness, Individual Typology Angle (ITA), and CIE L*a*b* color space.

[0073] In addition, the method 400 may comprise providing 440 the color-related features 340 as a predictor input to a machine learning model 300 to generate 460 a prediction of bilirubin levels 360 of the subject 80. According to various exemplary embodiments, the machine learning model 300 may be one or a combination selected from: a Support Vector Regression, a Random Forest Regression, and an Extreme Gradient Boosting Regressor, Gradient-boosted Trees Regressor, Extra Trees Regressor, Light Gradient-boosting Machine.

[0074] In various embodiments, one of the color-related features 340 may include a yellowness gradient of the plurality of color calibrated patches. The yellowness gradient may be computed based on and/or corresponds to a color difference between each of the plurality of color calibrated patches 330. In various embodiments, the color difference may be a difference in the degree of yellowness. In various embodiments, the color difference may be a gradient in the degree of yellowness. In an exemplary embodiment, the yellowness gradient may be computed based on and/or correspond to the color difference between images of two unique body surfaces. In various embodiments, the yellowness gradient may be computed based on the color difference between an image of a skin surface on a first body surface and an image of a skin surface on a second body surface, wherein the first body surface is superior (or nearer to the head of the subject) relative to the second body surface. In an exemplary embodiment, the yellowness gradient may be computed based on and/or correspond to the color difference between an image of a skin surface on a first body surface, such as a sternum (or chest) of the subject 80, and an image of a skin surface on a second body surface, such as an abdomen of the subject 80.

[0075] In various embodiments, the color-related features 340 may further comprise a second yellowness gradient of the plurality of color calibrated patches. Similarly, the second yellowness gradient may be computed based on and/or correspond to a color difference between images of two unique body surfaces, in which one of the body surface is superior relative to another of the body surface. In an exemplary embodiment, the second yellowness gradient may be computed based on and/or correspond to a color difference between an image of a skin surface on a forehead and an image of a skin surface on a sternum (or chest) of the subject 80. [0076] In various embodiments, the color-related features 340 may further comprise a plurality of subsequent yellowness gradient. The plurality of subsequent yellowness gradient may correspond to: a color difference between a skin surface on an abdomen and a skin surface on a shin (or a thigh) of the subject 80; and a color difference between a skin surface on the shin (or the thigh) and a skin surface on a feet of the subject 80 It may be appreciated that the abdomen is superior to the shin (or the thigh), and the shin (or the thigh) is superior to the feet. It may be appreciated that multiple color-related features 340 may be used for the method 400. [0077] In various embodiments, the yellowness gradient and/or the second yellow gradient may comprise markers selected from one of or a combination of: Cyan-Magenta-Yellow and Key (Black) (CMYK) color space’s Yellow channel, blue chromaticity, Jaundice Eye Color Index (JECI), LAB color space’s B channel, and Weighted Average Yellowness.

[0078] In various embodiments, the color-related features 340 may further comprise one of or a combination of: color-space related features, yellowness-related features, clustering features, gradient features and skin tone similarity features. [0079] In various embodiments, the color-space related features may be obtained from one of or a combination of color spaces: Red-Green-Blue (RGB), LAB, Luminance-Chrominance red difference-Chrominance blue difference (YCrCb), Hue Saturation Value (HSV), Cyan- Magenta-Yellow and Key (Black) (CMYK) and Hue Saturation Lightness (HSL). Additional color-space related features may be derived from these color spaces, such as the Individual Typology Angle (IT A) which is a measure derived from theL* (lightness) and B* (yellow/blue) components of the CIE L*a*b* color space.

[0080] In various embodiments, the yellowness-related features may be features that select/pick out the magnitude of yellowness in the skin color. The yellowness-related features may comprise a magnitude of yellowness selected from one of or a combination of: a colorspace related Y channel of CMYK color space, B channel of LAB color space, blue chromaticity values (which is defined as blue pixel value divided by the summation of the red, green, and blue pixel values. It has an inversed relationship with yellowness, JECI value.

[0081] Clustering features may be derived in the following manner. For each skin patch, K- Means clustering may be performed in the RGB color space to get various clusters. The cluster centroid values may be extracted and converted to various color spaces The channels in the respective color spaces which contain yellowness may be selected and the channel values may be used. The average of the channel values (also known as ‘yellowness values’) of all the centroids may be taken to derive the weighted average clustering features.

[0082] In various embodiments, the clustering features correspond to an average of respective yellowness value of each of a plurality of cluster centroids. The plurality of cluster centroids may be obtained via K-means clustering in a RGB color space. In various embodiments, each of the plurality of cluster centroids may comprise yellowness extracted from a corresponding color space, wherein the corresponding color space is converted from the RGB color space. [0083] In various embodiments, the gradient features may correspond to a difference between the respective yellowness-related features of respective ones of the plurality of unique body surfaces. For example, the gradient features may correspond to a difference between yellowness-related features of the forehead and the chest of the subject 80.

[0084] In various embodiments, the skin tone similarity features may be obtained between ones of the plurality of color calibrated patches 330 (or skin patches 322) and respective ones of the plurality of calibration patches 324. In some embodiments, the skin tone similarity features may be obtained between ones of the plurality of color calibrated patches 330 and respective ones of a plurality of color calibrated calibration patches. Skin tone similarity features may be obtained on the assumption that lighting conditions will impact the entire image (including skin patch 322 and calibration patch 324) equally.

[0085] For example, a skin tone similarity feature may be obtained between the skin patch 322 and the portion of color array 214 (FIG. 2) which includes colors corresponding to the Monk Skin Tone (MST) scale. Hence, the respective ones of the plurality of calibration patches corresponds to a Monk Skin Tone (MST) scale. The skin tone similarity features depart from conventional methods and systems in which skin tone of the subject 80 is typically not included in the prediction of bilirubin levels, enabling the proposed system and method to be generalizable across a wide spectrum of skin tones

[0086] In various embodiments, referring again to FIG. 4, the method 400 may further comprise the step of providing user-input features 350 as another predictor input to the machine learning model 300 for predicting bilirubin levels of the subject 80.

[0087] As exemplary embodiments, the user-input features comprise one of or a combination of: a gestational age, an hour of life, a mode of delivery, a compatibility of motherbaby blood type (rh/abo), a type of feeding (breastfeeding/formula/mixed), and an indication of Glucose-6-Phosphate Dehydrogenase (G6PD) deficiency. Further, the user-input features may also include one or a combination of: race, skin tone, gender, cephalohematoma.

[0088] Referring to FIG 7, images taken of the eye of the subject 80 may also be utilized for the proposed system 100 and method 400 for improved accuracy and robustness. In various embodiments, the plurality of images 310 of the unique body surfaces of the subject may correspond to one or more images taken of a sclera 87 (or sclera region) of an eye of the subject 80. The sclera, which turns yellow in jaundiced individuals due to bilirubin accumulation, provides a valuable indicator for assessing jaundice.

[0089] According to various embodiments, a color calibration card 200 may be positioned at a side of an eye of the subject 80 (FIG. 7), prior to capturing the images. The images may undergo color correction based on the calibration card, similar to the process used for the skin surfaces such as the sternum and abdomen. In addition, normalization of colors and reduction of lighting and device discrepancies may also be performed. Thereafter, the region of interest (ROI) of the sclera may be identified, focusing on the eyes of the newborn using the ArUco marker detection, calibration card and center eye patch separation. For the center eye patch, image segmentation techniques such as Mask R-CNN may be applied to isolate the sclera patch 323 or sclera region. As alternatives, other segmentation techniques such as U-Net, SAM, and eye landmark detection may also be used. The sclera patch 323 may undergo feature extraction process similar to that of the skin sites (e g. sternum and abdomen). The color-related features extracted from the sclera patch 323 may be fused with existing color-related features obtained from the skin sites as input to the ML model. In various embodiments, feature scaling and normalization may be applied to ensure compatibility between the sclera patch 323 derived color-related features and skin patch 322 derived color-related features.

[0090] Referring to FIG. 8, in various embodiments, SHAP analysis may be performed in determining the potential predictors/features for use with the model. As the ML model may be an ensemble model, features may be identified as input to the ML model to make a prediction on bilirubin level. According to an exemplary embodiment, the features for use with the ML model may include: the clustering features (weighted average Cb value from YCrCb color space), gradient features (gradient of blue chromaticity value between sternum and abdomen, gradient of Y value between sternum and abdomen), yellowness-related features (V channel value from HSV color space, B value of LAB color space, A channel value from LAB color space, Y value of CMYK color space), skin tone similarity features between the calibration patches and the skin patches, and user-input features (hour of life, temperature value / light value of which the photo was taken at, birth weight, type of feeding), and other color-related features.

[0091] Referring to FIGs. 9A and 9B, in various embodiments, the window 220 of the calibration card 200 may be selectively occluded prior to and during a phototherapy session. Provision of the selectable occlusion of window 220 allows the mitigation of skin bleaching effects on the skin surface(s) due to the phototherapy session, which may cause the skin surface(s) not to be representative of the bilirubin levels in the subject’s blood (due to phototherapy). This enables the proposed system 100 and method 400 to be expanded for use to subject 80 who is undergoing or underwent phototherapy.

[0092] Referring to FIG 9A, according to various embodiments of the present disclosure, the calibration card 200 may further comprise at least one occluding portion 230 movably coupled to the frame portion 210. The at least one occluding portion 230 may be movable to at least partially cover the window 220. The at least one occluding portion 230 may be opaque to a phototherapy wave. In an example shown in FIG. 9A, the occluding portion 230 may be a foldable flap displaceable between a folded position to occlude the window 220 and an unfolded position to reveal the window 220. [0093] Referring to FIG. 9B, according to various embodiments, the calibration card 200 may further be provided with an occluder 230 detachably coupleable to the frame portion 210. The occluder 230 may be attachable to the frame portion 210 to at least partially or fully cover the window 220 or be detachable from the frame portion 210 to reveal the window 220. The occlude 230 may be opaque to a phototherapy wave.

[0094] Tn various embodiments, the subject 80 may undergo phototherapy session with the calibration card 200 attached to the subject 80. Further, during the phototherapy session, the window 220 of the calibration card 200 may be fully occluded or at least partially occluded. This enables at least a portion of the window 220 to be occluded during a prior phototherapy session.

[0095] Hence, in various embodiments, one or more of the unique body surfaces of the subject 80 may be occluded during a prior phototherapy session. Alternatively or additionally, at least a portion of the unique body surface of the subject 80 is occluded during the prior phototherapy session.

[0096] According to another aspect of the present disclosure, disclosed herein is a method of predicting bilirubin levels based on a plurality of images of a subject Tn various embodiments, the method 400 comprises: in 410, pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject, in 420, color calibrating (or color correcting) the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; in 430, generating color-related features based on the plurality of color calibrated patches; and in 440 providing the color-related features as a predictor input to a machine learning model to generate 460 a prediction of bilirubin levels of the subject. 1 [0097] According to various embodiments, the method 400 may further comprise in 450, providing user-input features as another predictor input to the machine learning model. According to various embodiments, the method 400 may further comprise: in 470, segmenting a sclera patch in each of the plurality of images to obtain a plurality of pre-processed images, and wherein the step of color calibrating further comprises the step of: color calibrating the sclera patch in each of the plurality of pre-processed i ages based on the calibration patch associated with the sclera patch to obtain a plurality of color calibrated patches.

[0098] Exemplary Implementation - BiliSG

[0099] BiliSG is an exemplary implementation of the proposed system for and method of predicting bilirubin levels. Referring to FIG. 11 A, the implementation aims to develop and validate a smartphone-based or smart device-based Machine Learning (ML) application to predict or estimate bilirubin levels in newborn infants based on color analysis of skin and/or scleral images (without iris) from the local multi-ethnic neonatal population.

[00100] BiliSG application is accompanied by a color card/sticker (or calibration card) which defines a central aperture (or window). The color card may include multiple color squares, such as 40 colored squares, which included the Fitzpatrick scale of skin tone. The color card may provide color references for pre-processing of the images captured prior to being input to the ML model. The color card may be used as a sticker to prevent shadows between the baby ’ s skin and the color card, see FIG. 1 IB. The BiliSG application is also configured to screen and accept the images/photographs in which satisfactory conditions are met for lighting conditions, sharpness, and the detection of color card/sticker.

[00101] The utility of clinically relevant predictors such as yellowness or skin tone were explored, in view of the clinical implication of the reduced reliability of visual estimation of neonatal jaundice in babies with darker skin tone. It was hypothesized that the collection of a localized multi-ethnic “training set” and development of a customized Machine Learning Regression model that incorporated clinically relevant predictors such as sternum-abdomen “yellowness” gradient and skin tone could improve the performance of the application, with the aim of clinical deployment for purpose of nation-wide NNJ surveillance

[00102] A cross-sectional prospective study was conducted at the neonatal unit in a tertiary teaching hospital and 4 primary care clinics. All babies born at the gestation of 35 weeks and above with clinical jaundice or are recommended for screening of jaundice within 21 days of post-natal age were recruited.

[00103] Technical development of BiliSG Machine Learning (ML) model was done by performing K-Folds Cross-Validation for model selection and prospective validation to further validate the selected model. Model selection was done based on the metric Root Mean Square Error (RMSE), Pearson correlation, and statistical agreement on Bland-Altman plot. The overall best performing model in view of all 3 metrics was selected as the ML model. The ML model included the feature “yellowness” gradient between the sternal and abdominal regions of interest (ROI).

[00104] Using BiliSG, images of the babies’ skin over the forehead, sternum, abdomen, and sclera (optional) were taken against the color card/sticker. Smartphone-predicted bilirubin estimates (SpB) were compared with total serum bilirubin (TSB). Pearson correlation coefficient was assessed to establish the strength of the relationship between SpB versus TSB. Bland Altman plots were used to establish the agreement between SpB and TSB using the clinically acceptable limits of agreement of 35 pmol/L and 50 pmol/L. which were defined a priori Sensitivity, specificity, and likelihood ratios for detection of TSB of >17 mg/dL was assessed. The correlation and statistical agreement between TcB and TSB were also analysed. It was concluded that BiliSG provided good correlation and statistical agreement with the gold standard test of TSB, with high sensitivity, and could be embedded in a decentralised screening model for Neonatal Jaundice (NNJ). [00105] The BiliSG application advantageously enables parents to monitor jaundice levels of newborn in the comfort of their homes. Taking reference to the predicted bilirubin levels, healthcare professionals may provide medical advice and instructions through a teleconsultation service. The combination of a smartphone application and teleconsultation service represents a low-cost and non-invasive solution which facilitates the de-centralisation of care and reduces frequent (and often non-crucial) visits to healthcare facilities, especially during the current COVID-19 pandemic or any future pandemic or disaster. It was also found that parents have shown high acceptance of the BiliSG application.

[00106] Color calibration card/sticker

[00107] The color calibration card/sticker has a central aperture (for skin color assessment) surrounded by squares of different colors The colored squares were printed on matte photo paper using the same commercial photo printer to aid with color consistency. The color calibration card/sticker compensates for variations in light intensity and temperature since color varies under different lighting conditions. The properties of the skin and the calibration card segments may be equally affected by those differing lighting conditions. Color correction techniques were performed using the calibration card as reference, with the aim of adjusting for the confounding effects of different illumination conditions on the color properties of the acquired image i.e. babies’ skin. The sticker function was provided to eliminate the formation of shadows due to gaps between the color card and the infant’s skin.

[00108] Methods: Study Design, Setting and Ethics

[00109] The study was conducted as a cross-sectional prospective study which was initiated in a neonatal unit in a tertiary hospital [Singapore General Hospital (SGH)] and subsequently carried out at four public primary care clinics [SingHealth Polyclinics (SHP)- Bukit Merah, Sengkang, Bedok and Punggol] in the south and eastern part of Singapore. SGH is the largest tertiary hospital in Singapore, providing wide-ranging multi-disciplinary care. The Neonatal Department in SGH delivers a complete range of inpatient and outpatient services to neonates and infants. The hospital cares for about 1500 neonates annually. SHP provide comprehensive primary care to mother and child in an ambulatory setting. This ranges from routine infant care such as developmental screening and vaccination to acute care of an unwell infant and child.

[00110] Participants and Sampling:

[001 11] In SGH, eligible infants were identified using records based on gestational age, postnatal age and indication for a bilirubin check. Parents of the eligible infants were approached by a study team member when the infants were admitted to the hospital or when they came for their regular specialist outpatient clinic visits. In SHP, eligible infants were identified by the healthcare providers in the clinics when the infants visited the clinics for their neonatal follow-up. They were informed of the study and given a participant information sheet detailing the study. Parents who consented to the study were required to sign the informed consent form.

[00112] Inclusion and Exclusion criteria

[00113] Clinically stable term and late preterm infants born at gestation of 35 weeks and above with clinical jaundice or were recommended for screening of jaundice within the postnatal age of 21 days were eligible for recruitment. Both inpatients and outpatients will be eligible for recruitment. Neonates with skin lesions over the forehead, sternum or abdomen that will interfere with image acquisition were excluded.

[00114] Potential risks and limitations

[001 15] This study involves minimal risk in participating infants. SpB measurement is a non- invasive point-of-care test. It did not interfere with the current standard of care for infants with neonatal jaundice Data has been de-identified and stored in a secure platform to minimise the risk of a breach of patient confidentiality.

[00116] Baseline characteristics and medical data [00117] Baseline demographic and medical data including gestational age, birthweight, gender, ethnicity, skin tone, presence of cephalohematoma, ABO and Rhesus incompatibility, presence of glucose-6-phosphate dehydrogenase deficiency, type of feeding, and phototherapy status was collected from patients and medical records as de-identified data. Qualitative information about end user acceptability was also collected. All data was de-identified and stored in a passcode protected encrypted drive accessible to research team. Any hardcopy research data was stored in the department office with lock and key access.

[00118] Ascertainment of bilirubin level

[00119] Transcutaneous bilirubin (TcB) measurements and total serum bilirubin (TSB) was paired with images captured from the mobile application BiliSG, which would generate the Smartphone-predicted Bilirubin estimate (SpB), with the measurements taken within an hour of each other.

[00120] Transcutaneous bilirubin (TcB) measurement

[00121] Enrolled infants recommended for jaundice screening would have TcB assayed using the Drager JM-105 (Drager Medical GmbH) Bilirubinometer. The device has been calibrated on a daily basis as per manufacturer’s guidelines. Measurements were obtained from the sternum with the probe placed perpendicular to the skin. Three measurements were taken in succession, following which the device would reflect the mean of the three measurements. The TcB measurements were performed by trained personnel following departmental protocol and guidelines on care of neonatal jaundice.

[00122] Total serum bilirubin (TSB) measurement

[00123] Where TSB measurement was triggered based on initial TcB screening, as per current clinical practice guidelines, heel prick was done to obtain 0.2 cm³ capillary tube of blood and TSB assayed using the Unistat® analyser which was based on direct spectrophotometry. In babies where TSB measurement has not been triggered based on initial screening via TcB measurement, TSB would also be assayed if parental consent was obtained. In babies where multiple SpB-TSB data pairs were collected at different timepoints, a sub-group analysis using only the first measurement was performed to investigate the possibility of intra-class correlation.

[00124] BiliSG / Smartphone-predicted Bilirubin (SpB) measurement

[00125] Where TSB and/or TcB is measured, Smartphone application bilirubin estimate (SpB) measurement was obtained within 1 hour of these measurements, which was in accordance with other validation studies. The BiliSG application was used on an Apple® iPhone 12 (Apple Inc, Cupertino, California, USA) dedicated for the study purpose. The procedure involved placement of a color calibration card/sticker on the baby’s forehead, sternum, abdomen and at the side of the eye, followed by alignment of the application photograph frame with the color card (see FIGs. HA and 1 IB). The choice of forehead, sternal and abdominal images corresponds to the cephalocaudal progression of jaundice, predictably from the face to the trunk, extremities and finally to the palms and soles that accompany increasing TSB levels.

[00126] The BiliSG application was configured to accept skin or scleral images upon checking for image quality, appropriate lighting, and distance of the calibration card from the camera lens. The BiliSG application was configured to reject skin and/or scleral images that did not meet all three criteria. The images collected were uploaded to a server for further processing. As another criteria, images selected were obtained under ambient light conditions and without the use of the built-in flashlight. To minimise inter-operator variability, screening using the application was limited to selected trained investigators.

[00127] Data preprocessing

[00128] Image Segmentation - After the images were collected by the BiliSG application, the calibration card (or color card) and the skin patches were segmented. The segmenting process or image segmentation process involved identifying and isolating regions of interest within each image. Since the BiliSG application was designed with an image frame within which the calibration card may be aligned, the BiliSG application was configured to focus on the image frame for further segmentation purposes.

[00129] Color Calibration - By applying color correction techniques using the color patches in the calibration card as reference, the colors of the photographed skin/scleral patches were adjusted such that the impact of varying illumination conditions were minimized.

[00130] Predictors and Machine Learning Model Development

[00131] After color calibration has been performed, multiple color-related features from the skin sites were obtained. These features were included as predictors in the model development process. Derived demographic features and other additional information that were deemed clinically relevant and related to bilirubin levels (e g. hour of life of babies) were separately collected and included as predictors as well. In particular, the age of the baby was deemed to be an important predictor due to the natural history of neonatal jaundice, where TSB increases in the first few days of life, and reaches a peak around day 5 of life.

[00132] Machine learning regression techniques were used to identify predictors that were highly correlated Total Serum Bilirubin (TSB) as the ground truth

[00133] Cephalocaudal advancement of dermal icterus i.e. “yellowness” gradient between the forehead, sternal and abdominal regions of interest (ROI) was specifically explored as a predictor. The markers of “yellowness”, namely CMYK color space’s Yellow channel, blue chromaticity, JECI index, and LAB color space’s B channel were, in varying degrees, correlated with TSB. The proportion of cases that demonstrated a “yellowness” gradient between the sternal and abdominal ROI decreased with increasing TSB levels (FIG. 12). It was found that there was no consistent “yellowness” gradient observed between the forehead and sternal ROI, a finding which was consistent with previous observations that demonstrated under-estimation of bilirubin measurements when the transcutaneous bilirubinometer was applied to exposed areas such as the forehead.

[00134] Machine Learning Model Selection and Validation

[00135] Various ML models or regression algorithms were developed and trained for bilirubin level prediction, such as Support Vector Regression, Random Forest Regression, Extreme Gradient Boosting Regression, Light Gradient-Boosting Machine Regressor, Gradient Boosted Trees Regressor, and Extra Trees Regressor. The ML models were developed with patient data collected from November 2023 to mid-April 2024, which were from 416 unique patients. This dataset was used for model training and selection.

[00136] The ML model development and selection was carried out using the K-Folds Cross- Validation technique By using a 5-fold cross-validation procedure, the available study sample was randomly divided into 5 equal- sized subgroups (i.e., folds) that were stratified by TSB value. One-fold was removed from the sample, and data from the other four folds were used in training a regression model for estimating bilirubin levels. This ML model was then used to estimate the SpB values for participants in the removed fold. The process was repeated for 5 iterations, successively removing one-fold, developing a model by using data from the remaining four folds, and applying the model to the removed fold. Thus, data from a particular newborn were not used in developing the model that was used to estimate the bilirubin level for that newborn, providing an unbiased evaluation of BiliSG.

[00137] The performance of the various ML models was serially monitored in terms of RMSE, Pearson correlation, and statistical agreement on Bland-Altman plot. The overall best performing model in view of all 3 metrics was selected as the ML model for the application to predict a single value of bilirubin level (SpB). Further external temporal validation of this ML model was performed on patients who were prospectively recruited between mid-April to June

2024. [00138] Statistical analyses and Results

[00139] Descriptive analysis was used to summarise the characteristics of the infants. Pearson’s correlation coefficient was used to assess the strength of linear relationship between predicted SpB values and TSB (gold standard). Bland-Altman (BA) plots were used to examine the statistical agreement between SpB values and TSB (gold standard).

[00140] Subgroup analysis was performed based on skin tone types, as this factor has been reported to affect the accuracy of jaundice screening. Sensitivity, specificity, likelihood ratios, and positive and negative predictive values for detection of TSB phototherapy thresholds were assessed for the decision rule of SpB >13 mg/dL to identify a newborn with a TSB value >17 mg/dL, for comparability with other studies.

[00141] The utility of BiliSG and TcB as screening tools for identifying neonates with TSB level of >17 mg/dL were compared by constructing receiver operator characteristic curves and comparing the area under the curve (AUC) for each using the above approach. All analyses were performed using Python software version 3.9. 17.

[00142] The sample size was estimated based on clinically acceptable limits of agreement for TSB (±3mg/dL). A conservative mean difference of 0.5 mg/dL and a SD of 1 ,3mg/dL indicated that 463 paired measurements were needed to detect agreement based on a 95% CI for the limits of agreement at 80% power and maximum allowable difference of 3 mg/dL. Accounting for a 15% loss of information due to image quality or participant withdrawals, 545 babies would need to be recruited, contributing to 545 pairs of measurements for model development (70%) and prospective validation (30%). A 2-sided P value less than .001 was considered significant. [00143] Results

[00144] Between November 2023 and June 2024, 633 unique neonates were recruited at SGH and 4 primary care clinics (FIG. 13). A total of 627 neonates had paired TSB and SpB readings for analysis, with 499 Results (79.6%) having 1 reading, 98 (15.6%) 2 readings, 22 (3.5%) 3 readings, 5 (0.8%) 4 readings, and 3 (0.5%) 5 readings.

Table 1. Demographic and Clinical Characteristics [00145] During the iterative development and cross-validation phase, 395 neonates were recruited, but were excluded for poor image quality, leaving 352 for cross-validation. After determining the ML model, 232 additional neonates were recruited, with 37 excluded for image quality and 1 for phototherapy, resulting in 194 for temporal validation.

[00146] Demographic data (see Table 1) showed an ethnic distribution similar to the national population (315[57.7%] Chinese, 35 [6.4%] Indian, 169 [31.0%] Malay, and 27 [4.9%] other ethnicities), with a median (IQR) gestational age of 38.0 (35.0^41.0) weeks and birth weight of 3045 (1975-4514) grams and 285 male neonates (52.4%). Glucose-6-phosphate deficiency was present in 15 neonates (2.7%), and ABO group incompatibility was documented in 149 (27.7%).

[00147] TSB level distribution (FIGs. 14A and 14B) showed comparable patterns between the prospective validation dataset and the training set, particularly for TSB values lOmg/dL or less and 14mg/dL or greater, despite a slight leftward skew in the validation dataset.

[00148] Predictors

[00149] In the ML model, yellowness-related predictors from the forehead, sternal, and abdominal regions were among the top predictors. The discernible yellowness gradient between the sternal and abdominal regions decreased at higher TSB levels, particularly higher than 12mg/dL (FIG. 12), while no consistent gradient was observed between the forehead and other regions. Scleral images were excluded from analysis due to successful capture in only 10 participants.

[00150] To improve interpretability of the model, Shapley additive explainability tools were employed, revealing that the neonate’s hour of life was consistently 1 of the top predictors. This aligns with clinical knowledge that TSB levels typically rise in the first few days of life Other significant features included yellowness-related attributes from skin images, such as the CMYK yellow channel, the LAB B channel, and the jaundice eye color index value, indicating the model’s reliance on the degree of yellowness in the images for predictions.

[00151] Cross-Validation and Temporal Validation of the ML Model

[00152] The gradient boosted trees model was selected as the algorithm due to its consistent performance, achieving an RMSE of 2.41mg/dL and a Pearson correlation of 0.77 (P < .001) between SpB and TSB, with 76% of data pairs within a clinically acceptable difference of 50 pmol/L (approximately 3mg/dL). After determining the ML model, 194 individuals were prospectively recruited for temporal validation, which revealed a strong correlation between SpB and TSB (FIG. 15A), with a Pearson coefficient of 0.84 (95% CI, 0.79-0.88; P < .001). For ethnic groups, the coefficients were 0.86 (95% CI, 0.80-0.90; P < .001) for Chinese, 0.91 (95% CI, 0.73-0.97; P < .001) for Indian, and 0.81 (95% CI, 0.69-0.89; P < .001) for Malay neonates. For Fitzpatrick skin phototype II and III, coefficients were 0.85 (95% CI, 0.79-0.90; P < .001) and 0.92 (95% CI, 0.20-0.99; P = .03), respectively. A sensitivity analysis of 231 babies (194 without artifacts and 7 with significant artifacts affecting >35% of regions of interest) showed a Pearson coefficient of 0.81 (95%CI, 0.76-0.85; P < .001).

[00153] The Bland-Altman plot (FIG. 15B) indicated that 82% of paired measurements were within the maximum acceptable difference, with SpB readings slightly lower than TSB, with a mean difference of -0.18mg/dL (95% limits of agreement [LoA], -4.20 to 3.84mg/dL). The mean differences among the largest ethnic groups were similar: -0.21 (95% LoA, -4.01 to 3.59mg/dL) for Chinese, -0.30 (95% LoA, -3.55 to 2.95mg/dL) for Indian, and -0.46 (95% LoA, -4.72 to 3.79mg/dL) for Malay. The ML model’s RMSE was 2 06mg/dL.

[00154] Calibration of ML Model

[00155] To assess the calibration of the ML model across the TSB range, the prospective validation set (TSB levels from 0.88mg/dL to 20.46mg/dL) was divided into 3 equidistant groups: (1) TSB less than 6.5 mg/dL, (2) TSB between 6.5 mg/dL and 13.0 mg/dL, and (3) TSB greater than 13.0 mg/dL. In each group, 83% (24 of 29), 85% (88 of 104), and 77% (47 of 61) of cases, respectively, fell within the clinically acceptable range of +3mg/dL, suggesting an alignment of predictions and underscoring the ML model’s reliability.

[00156] Diagnostic Accuracy of SpB and TcB Measurements

[00157] Using TSB as the criterion standard, the sensitivity and specificity of SpB were 100% (95% CI, 100%- 100%) and 70% (95% CI, 63%-76%), respectively, while for TcB, they were 100% (95% CI, 100%-100%) and 51% (95% CI, 40%-61%) (see Table 2 and Table 3 below). The positive likelihood ratios were 3.30 for SpB and 2.02 for TcB, with both having a negative likelihood ratio of 0.0. The positive predictive value for SpB was 10% (95%CI, 2%-17%), and for TcB, it was 8% (95% CI, 1 %-16%), while both had a negative predictive value of 100% (95% CI, 100%-100%). The areas under the receiver operating characteristic curve were 0.89 (95% CI, 0.82-0.96) for SpB and 0.95 (95% CI, 0.87-1.00) for TcB, suggesting diagnostic accuracy for both measurements (FIG. 16).

Table 2. Diagnostic accuracy of SpB and TcB Table 3. Cross Tabulation of SpB and TSB results

[00158] The smartphone-based BiliSG app, developed and validated in the Singapore population, demonstrated a strong correlation (Pearson coefficient of 0.84) between SpB and TSB. It achieved 100% sensitivity and 70% specificity based on the decision rule of SpB 13mg/dL or greater to predict TSB 17mgl/dL or greater. Its diagnostic accuracy was comparable with TcB.17 Screening tool-derived SpB demonstrated a stronger correlation with TSB compared with the pooled correlation coefficient of 0.77 previously reported which analyzed data from 1733 neonates across 10 studies.

[00159] Referring to FIG. 12, the proportion of cases with a discernible “yellowness” gradient between the sternal and abdominal ROI decreased with increasing TSB levels and was therefore included as a predictor in the ML model. This departs from other solutions which use single images from a single ROI to generate each SpB and would inherently not be able to incorporate the “yellowness” gradient between two separate regions of interest, which are taken at the same setting, for the purpose of generating a single SpB.

[00160] Further, with the provision of the color sticker rather than a color card aids in eliminating the formation of shadow within the central aperture by minimizing any potential gap between the color card and the skin (see FIG. 1 IB), thereby improving image quality, which was essential for accurate SpB readings. BiliSG was validated on a larger number of datapoints, which includes an extensive validation cohort which helps with improving accuracy by reducing the risk of overfitting and random effects that can occur with smaller sample sizes.

[00161] In addition to validating the BiliSG application on a multi-ethnic population in Singapore, this study was also unique because a significant proportion of babies were of Malay ethnic origin. Data analysis was performed objectively based on the Fitzpatrick skin tone phototype rather than ethnicity, which was helpful in objectively assessing the effects of intra- and inter- ethnic group skin tone variation on the ML model performance.

[00162] Next Generation BiliSG- BiliSG v2.0

[00163] BiliSG v2.0 is a further development of BiliSG which incorporates images from multiple cephalocaudally arranged regions of interest (ROIs) — forehead, sternum, abdomen, shin, and foot — and further integrates the skin tone marker Monk Skin Test scale in addition to the Fitzpatrick scale, as predictive features in the ML model. It was hypothesized that combining skin yellowness data from all five ROIs with skin tone markers, and training the model on a larger, ethnically diverse neonatal dataset, will improve the accuracy and generalizability of BiliSG v2.0. In addition, the effects of phototherapy and the utility of photoopaque occluders on the app’s performance was also implemented and explored.

[00164] In the previous implementation, development and validation, BiliSG demonstrated excellent correlation between smartphone-predicted bilirubin (SpB) and total serum bilirubin TSB, with strong statistical agreement and 100% sensitivity. In addition to forehead, sternum, and abdomen (zones 1 to 3) of the previous implementation, BiliSG v2.0 incorporates the limb regions (zones 4 and 5) to capture the full cephalocaudal progression of dermal icterus.

[00165] In addition, skin tone was identified as a key factor influencing the accuracy of non- invasive NNJ screening across modalities, including visual assessment and TcB measurement. However, in the original BiliSG implementation, skin tone — graded using the Fitzpatrick scale — was not among the top predictors, as revealed by SHAP analysis. This may reflect limitations of the Fitzpatrick scale, which does not capture the full diversity of skin tones, and potential sample size constraints. Nevertheless, no existing solutions incorporate a skin tone marker as a predictor.

[00166] Additionally, phototherapy has also been shown to affect the accuracy of non-invasive NNJ screening methods, including TcB. It was postulated that the proposed system and method may have an improved TcB accuracy when measurements are taken from non-exposed skin, either with or without the use of photo-opaque occluders.

[00167] Other aspects of the implementation also includes: ensuring adequate recruitment of neonates across a range of skin tones to improve generalizability; assessing inter-rater reliability across the multiple study centres in use of BiliSG app; determining the influence of varying ambient lighting conditions on image quality and SpB measurement accuracy; and determining the optimal threshold for smartphone-predicted bilirubin (SpB) levels to detect clinically relevant total serum bilirubin (TSB) thresholds (>100, 150, 175, 200, 225, 250, 275, 300, and 325 umol/L i.e. 5.9, 8.8, 10.2, 11.7, 13.2, 14.6, 16.1, 17.5, and 19 rng/dL), using receiver operating characteristic (ROC) analysis and other appropriate statistical techniques.

[00168] Methods

[00169] Design

[00170] The current implementation is an extension of the earlier BiliSG implementation and adopts a dual-phase prospective design, aimed at the development and validation of an enhanced version of the BiliSG app — BiliSG version 2.0. A consecutive sample of term and late preterm neonates (>35 weeks’ gestation), aged <6 weeks, who are clinically stable, regardless of ethnicity was recruited. Neonates with skin conditions that interfere with image acquisition (e g., extensive rashes, bruising, or birthmarks over the regions of interest) were be excluded. Ethical approval has been obtained from the SingHealth Centralised Institutional Review Board. The test sites include major healthcare institutions in Singapore: Singapore General Hospital (SGH), KK Women’s and Children’s Hospital (KKH), National University Hospital (NUH), SingHealth Polyclinics (SHP), and National University Polyclinics (NUP).

[00171] Informed consent was obtained from parents or legal guardians, with the following data collected: Demographic information: gestational age, birthweight, sex, and electronic health record-reported ethnicity (Chinese, Malay, Indian, or other, with specification); Clinical information: presence of cephalohematoma, ABO and Rhesus (Rh) blood group, glucose-6- phosphate dehydrogenase (G6PD) deficiency status, and phototherapy status; Skin tone assessment: using both the Fitzpatrick and Monk skin tone scales, performed by trained study personnel; feeding type: breast milk, formula, or mixed, as self-reported; End-user acceptability: assessed through a qualitative questionnaire completed by parents.

[00172] ABO incompatibility in this study was defined as follows: Group O mothers with nongroup O neonates; Group A mothers with group B or AB neonates; Group B mothers with group A or AB neonates.

[00173] Experiments were conducted in two phases:

[00174] - Phase 1 focused on enhancements to the app’s user interface and user experience (UI/UX), enabling image capture from five cephalocaudal regions. Retrospective model refinement and enhancement of the ML model were also conducted (see section: Inclusion of Skin Tone-related Color Metrics as Predictors in the ML Model). Iterative development and prototyping of a photo-opaque occluder with a central window for standardized image acquisition would also take place during this phase.

[00175] - Phase 2 focused on iterative development and cross-validation, followed by prospective temporal validation of the updated ML model. Public and caregiver input were incorporated into the UI/UX refinement process to ensure usability and accessibility.

[00176] Study Population

[00177] Inclusion Criteria: Neonates born at >35 weeks’ gestation; and Aged <6 weeks at the time of enrolment.

[00178] Exclusion Criteria: Neonates with skin conditions that may interfere with accurate image acquisition (e.g., extensive rashes, birthmarks, or bruising over regions of interest)

[00179] Intervention and Procedures [00180] Eligible neonates underwent several bilirubin measurements or estimations within the same hour using smartphone-predicted bilirubin (SpB), TcB, and TSB methods.

[00181] TcB was measured with the Drager JM-105 bilirubinometer (Drager Medical GmbH), 35 which were calibrated on a daily basis. Three measurements from the sternal area were taken, and the mean recorded.

[00182] TSB was measured using capillary heel prick samples. Analysis of the blood samples were performed using the Unistat analyzer through direct spectrophotometry 36 in SGH, NUH and SHP, the Radiometer ABL90 Flex blood gas analyser in KKH, and the DAS Bilirubinometer in NUP. Calibration of the Unistat and DAS analyzers were performed every 6 months, along with standardized maintenance at all study centres. Quality control was conducted twice daily as per manufacturer standards, with trends tracked using Levy-Jennings plots and Westgard rules. All analyzers took part in proficiency testing to ensure measurement accuracy through accuracy -based surveys.

[00183] For neonates who underwent multiple SpB-TSB tests, all results contributed to model development and cross-validation, while only the initial measurement was used for temporal validation.

[00184] SpB estimates were obtained using the smartphone app on an Apple iPhone 12 or 13 model (Apple Inc), as well as a Samsung A55 smartphone by placing a color calibration sticker card with a central aperture on specific areas (forehead, sternum, abdomen, shin, and foot). The app captured and analyzed images of these regions of interest. High-quality images were captured in ambient light and screened for artifacts, with those covering over a significant proportion of the regions of interest excluded. A small group of trained study members performed SpB measurements to minimize variability and blinded to the results. Clinical management was guided by TcB and TSB measurements as per local clinical practice guidelines. The app recorded ambient light intensity (using built-in lux meters or standardized light sensors) at the time of image acquisition. Inter-rater reliability was studied across the 9 study centres.

[00185] Inclusion of Skin Tone-Related Color Metrics as Predictors in the ML Model

[00186] In BiliSG version 1.0, the machine learning (ML) model incorporated the Fitzpatrick skin tone scale; however, SHAP analysis revealed that skin tone was not among the top predictive features. This was unexpected given that darker skin tones can adversely affect the accuracy of non-invasive methods for estimating neonatal jaundice (NNJ), including visual assessment24-26 and transcutaneous bilirubinometry (TcB). It was hypothesized that this discrepancy may be due to the limitations of the Fitzpatrick scale — which consists of only six categories and fails to capture the full spectrum of skin tones — as well as limited sample representation of darker skin tones in the dataset.

[00187] To address this, the more inclusive Monk Skin Tone (MST) scale included with six raters trained to classify skin tones accordingly. The rater team comprised a dermatologist, two primary care physicians, a neonatologist, and two non-medically trained research staff. Training was based on the MST Examples (MST-E) dataset, which included 1,515 exemplar images representing 19 individuals across all 10 MST tones. Raters were trained under standardized conditions, including calibrated monitor display settings, consistent bright and untinted ambient lighting, and an unobstructed physical environment conducive to accurate visual assessment. All raters were screened for color vision deficiency using the Ishihara test.

[00188] To ensure consistency, inter-rater reliability was evaluated. Raters were required to accurately classify a sample of exemplar images from the MST-E dataset. This exercise served both as an assessment and as asynchronous training, allowing raters unlimited attempts to familiarize themselves with the scale During this task, raters answered one question at a time, individually and without communication. Each question presented an image of a body part (forehead, abdomen, or sternum) from the study population, and the rater was asked to select the best corresponding MST tone.

[00189] To reduce visual fatigue (asthenopia), raters were required to take breaks of at least 20 seconds after every 20 images. All rating tasks were performed under standardized conditions to minimize environmental variability and ensure consistent results.

[00190] In addition to previously identified key predictors — such as various color metrics including the yellow channel in the cyan-magenta-yellow-key (CMYK) model, blue chromaticity, the Jaundice Eye Color Index, and the B channel from the CIE LAB color space (which captures yellow and blue components) — new skin tone-related color metrics were also included, such as the Individual Typology Angle (ITA), which is a measure derived from the L* (lightness) and B* (yellow/blue) components of the CIE L*a*b* color space. The selection of these skin tone-related metrics were informed by their correlation and statistical agreement with the Monk Skin Tone (MST) scale, as well as their ability to minimize root mean squared error (RMSE).

[00191] The predictors were determined through automated feature selection during the training of the gradient boosted trees model. This model iteratively constructed decision trees, each learning from the errors of the previous ones. Feature importance was evaluated based on each predictor’s contribution to reducing the model’s overall loss function

[00192] Evaluation of BiliSG performance in babies undergoing phototherapy

[00193] Infants undergoing phototherapy were recruited for evaluation. Regions of interest on the skin were covered with photo-opaque stickers, and an illuminance meter was used to confirm their opacity. Each occluder included a central aperture that may be uncovered for image capture using the BiliSG v2.0 app, allowing for acquisition of "covered" skin bilirubin (SpB-C) and transcutaneous bilirubin (TcB-C) measurements. Readings from "exposed" areas adjacent to the occluded skin were taken using both BiliSG and a TcB device, generating "exposed" values: SpB-E and TcB-E. Each infant had SpB-C, SpB-E, TcB-C, TcB-E, and total serum bilirubin (TSB) measurements recorded at three time points: before phototherapy, during phototherapy (after 6-12 hours of treatment), and 24 hours post-phototherapy.

[00194] Data Collection

[00195] The following data were recorded: demographic and clinical variables, skin tone (using both Fitzpatrick classification and the Monk Skin Tone [MST] scale), device model, ambient lighting conditions, operator identity, and image quality (shadows, focus & sharpness, glare or overexposure). For infants undergoing phototherapy, additional variables included the type of phototherapy device (e.g., overhead, bed, blanket; intense phototherapy; conventional fluorescent lamp, LED, or fiber-optic) and the site of irradiance (front, back, or both). Multiple paired measurements — including SpB, TcB, and total serum bilirubin (TSB) — were collected per participant when feasible. Skin tone classification using the Monk Skin Tone Scale: Six raters were trained following the MST-Google Research methodology (Monk, Ellis. “Monk Skin Tone Scale,” 2019) to ensure consistent skin tone annotations across varying environmental conditions, including both high and low lighting scenarios.

[00196] Sample Size Calculation

[00197] Based on prior data from the previous implementation of BiliSG, SpB readings showed a slight negative bias compared to TSB, with a mean difference of -0.18 mg/dL (95% limits of agreement: -4.2 to 3.84 mg/dL). In Phase 2, an improvement is expected, reducing the mean difference by 10% to -0.16 mg/dL. To detect agreement with a standard deviation of 1.35 mg/dL (approximately a 30% reduction from BiliSG), and assuming a 95% confidence interval for the limits of agreement with 80% power and a maximum allowable difference of 3 mg/dL, a total of 1,113 paired measurements were required.

[00198] To account for an anticipated 15% loss of data due to image quality issues or participant withdrawal, the plan was to recruit 1,350 infants. Of these, approximately 950 paired measurements (70%) were used for model development, and 400 pairs (30%) were allocated for prospective validation. Sample size to be recruited was stratified by age and skin tone to ensure adequate representation

[00199] Sample Size for Evaluation of BiliSG in Infants Undergoing Phototherapy

[00200] For the phototherapy subgroup, assuming an intraclass correlation coefficient (ICC) of 0.8 and a two-way random-effects model, a sample of 250 infants — each providing two measurements (with and without photo-opaque occluder) — were required to achieve a two- sided 95% confidence interval width of 0.10 (±0.05). This accounts for an estimated 20% data loss due to poor image quality or other factors.

[00201] Statistical Analyses

[00202] Descriptive statistics were used to summarize the demographic and clinical characteristics of enrolled neonates. The Pearson correlation coefficient was used to assess the linear relationship between SpB values and both total serum bilirubin (TSB, the gold standard) and transcutaneous bilirubin (TcB). Agreement between SpB and TSB, as well as TcB, were evaluated using Bland-Altman plots, with clinically acceptable limits of ±50 pmol/L (approximately ±3 mg/dL). Subgroup analyses were performed based on skin tone classifications.

[00203] Where appropriate, a Generalised Estimating Equation (GEE) model was applied to account for repeated measures and to evaluate the influence of specific factors on the differences between SpB and TSB/TcB. For each factor, the corresponding regression coefficient ( ) and its 95% confidence interval was reported.

[00204] Secondary analyses included stratification by the method of TSB measurement to confirm that observed discrepancies are attributable to the prediction model rather than inherent variability in TSB measurement techniques. [00205] Diagnostic performance was evaluated using sensitivity, specificity, positive and negative predictive values (PPV, NPV), and likelihood ratios, calculated across multiple predefined TSB thresholds: >100, 150, 175, 200, 225, 250, 275, 300, and 325 pmol/L (equivalent to 5.9, 8.8, 10.2, 11.7, 13.2, 14.6, 16.1, 17.5, and 19 mg/dL). These thresholds corresponded to the age-specific phototherapy cut-offs for neonates in various postnatal age groups (<24 hours, 25-36 hours, 37-48 hours, 49-72 hours, 73-120 hours, 121 hours-7 days, and Day 8-14), for both normal and high-risk infants, as per local clinical guidelines³⁸.

[00206] Within each age group, the optimal SpB threshold was determined using Youden’s Index. Corresponding sensitivity, specificity, PPV, and NPV were reported for each group.

[00207] Inter-observer Reliability

[00208] Given that the skin tone scoring metrics were ordinal in nature, inter-rater agreement was assessed using Gwef s AC2 coefficient, which was well-suited for evaluating reliability in ordinal categorical data and offered robustness against prevalence and marginal distribution issues.

[00209] Model Calibration

[00210] The ML model was re-calibrated as necessary, incorporating new predictors from additional regions of interest (ROIs), as well as relevant device-specific inputs. The utility of all five ROI zones were evaluated, and the optimal combination of ROIs for the ML model was selected based on their contribution to minimizing the root mean squared error (RMSE).

[00211] Outcomes and Impact

[00212] Improved Accuracy and Generalizability: Enhanced performance of the BiliSG app for neonatal jaundice (NNJ) screening across diverse populations, with greater robustness to variability in skin tone, lighting conditions, and device models. [00213] Demonstrated Reliability Across Use Conditions: Validation of consistent performance across different devices, lighting environments, and operators, supporting real- world usability.

[00214] Establishment of Clinically Relevant SpB Thresholds: Development of evidencebased screening thresholds that effectively balance sensitivity and specificity — minimizing false negatives while managing false positives.

[00215] Facilitation of Remote and Decentralized NNJ Screening: Enabling wider access to jaundice screening, particularly in low-resource or community settings, reducing reliance on invasive TSB assays and TcB measurements, and improving early detection and timely intervention.

[00216] The proposed system and method for predicting bilirubin levels may be implemented by a processor system 900 as illustrated in the schematic block diagram of FIG. 17. Components of the processing system 900 may be provided within one or more computing device to carry out the functions of the modules or any other modules. One skilled in the art will recognize that the exact configuration or arrangement illustrated in FIG. 17 is provided by way of example only, e g , each processing system provided may be different and the exact configuration of processing system 900 may vary.

[00217] In embodiments of the present disclosure, the processing system 900 may include a controller 901 and user interface 902. User interface 902 is configured to enable manual interactions between a user and the computing module as required. For this purpose, the processing system 900 includes the input/output components required for the user to enter instructions to provide updates to each of the modules. A person skilled in the art will recognize that components of user interface 902 may vary from embodiment to embodiment but may typically include one or more input devices 935 such as but not limited to a touchscreen, a keyboard, a joystick, a mouse, a microphone, etc. The user interface 902 can also include a media player 940, which can be in the form of one or more playback devices, including but not limited to a display, a speaker, earphones, headsets, etc.

[00218] The controller 901 is configured to be in data communication with the user interface 902 via bus 915. The controller 901 includes memory 920 and processor 905 mounted on a circuit board to process instructions and data, e.g., to perform the method of the present disclosure. The controller 901 includes an operating system 906, an input/output (I/O) interface 930 for communicating with user interface 902, and a communications interface, e.g., a network card 950. The network card 950 may, for example, be configured to send data from the controller 901 via a wired or wireless network to other processing devices or to receive data via the wired or wireless network. Wireless networks that may be utilized by the network card 950 include, but are not limited to, Wireless-Fidelity (Wi-Fi), Bluetooth, Near Field Communication (NFC), cellular networks, satellite networks, telecommunication networks, Wide Area Networks (WAN), and etc.

[00219] Memory 920 and operating system 906 are in data communication with central processing unit (CPU) 905 via bus 910. The memory 920 may include both volatile and nonvolatile memory. The memory 920 may include more than one of each type of memory, e g , Random Access Memory (RAM) 923, Read Only Memory (ROM) 925, and a mass storage device 945. The mass storage device 945 may include one or more solid-state drives (SSDs). One skilled in the art will recognize that the memory described above includes non-transitory computer-readable media and shall be taken to include all computer-readable media except for a transitory, propagating signal. Typically, instructions are stored as program code in the memory but can also be hardwired. Memory 920 may include a kernel and/or programming modules such as a software application that may be stored in either volatile or non-volatile memory. [00220] Herein, the term “processor” is used to refer generically to any device or component that can process computer-readable instructions, including for example, a microprocessor, microcontroller, programmable logic device, or other computational device. That is, processor 905 may be provided by any suitable logic circuitry for receiving inputs, processing them in accordance with instructions stored in memory, and generating outputs (for example to the memory components or media player 940). Tn the present disclosure, processor 905 may be a single core or multi-core processor with memory addressable space. In one example, processor 905 may be multi-core, comprising — for example — an 8 core CPU. In another example, it could be a cluster of CPU cores operating in parallel to accelerate computations.

[00221] Further, one skilled in the art will recognize that certain functional units in this description have been labelled as modules throughout the specification. The person skilled in the art will also recognize that a module may be implemented as circuits, logic chips or any sort of discrete component. Still further, one skilled in the art will also recognize that a module may be implemented in software which may then be executed by a variety of processor architectures. In embodiments of the disclosure, a module may also comprise computer instructions or executable code that may instruct a computer processor to carry out a sequence of events based on instructions received. In further embodiments, the module may comprise a combination of different types of modules or sub-modules. The choice of the implementation of the modules may be determined by a person skilled in the art and does not limit the scope of the claimed subject matter in any way.

[00222] All examples described herein, whether of apparatus, methods, materials, or products, are presented for the purpose of illustration and to aid understanding, and are not intended to be limiting or exhaustive. Modifications may be made by one of ordinary skill in the art without departing from the scope of the invention as claimed.

Claims

1. A system, comprising: memory storing instructions; and a processor coupled to the memory and configured to process the stored instructions to implement: a module configured to perform a method of predicting bilirubin levels based on a plurality of images, the method including: pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject; color calibrating the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; generating color-related features based on the plurality of color calibrated patches; and providing the color-related features as a predictor input to a machine learning model to generate a prediction of bilirubin levels of the subject.

2. The system as recited in claim 1, wherein the plurality of images comprises at least: an image of a skin surface on a first body surface selected from a plurality of unique body surfaces and an image of a skin surface on a second body surface selected from the plurality of unique body surfaces, wherein the first body surface is superior relative to the second body surface.

3. The system as recited in any one of claims land 2, wherein one of the color-related features comprises a yellowness gradient of the plurality of color calibrated patches, wherein the yellowness gradient corresponds to a color difference between each of the plurality of color calibrated patches.

4. The system as recited in claim 3, wherein the yellowness gradient comprises markers selected from one of or a combination of: Cyan-Magenta-Yellow and Key (Black) (CMYK) color space’s Yellow channel, blue chromaticity, Jaundice Eye Color Index (JECI), B channel of LAB color space, and Weighted Average Yellowness.

5. The system as recited in any one of claims 3 to 4, wherein the yellowness gradient corresponds to the color difference between an image of a skin surface on a first body surface and an image of a skin surface on a second body surface, wherein the first body surface is superior relative to the second body surface.

6. The system as recited in claim 5, wherein the first body surface is a skin surface on a sternum of the subject and the second body surface is a skin surface on an abdomen of the subject.

7. The system as recited in any one of claims 1 to 6, wherein one of the color-related features comprises a second yellowness gradient of the plurality of color calibrated patches, wherein the second yellowness gradient corresponds to a color difference between an image of a skin surface on a forehead and an image of a skin surface on a sternum.

8. The system as recited in any one of claims 2 to 7, wherein the plurality of images further comprises: an image of a skin surface on a third body surface, an image of a skin surface on a fourth body surface and an image of a skin surface on a fifth body surface, wherein the second body surface is superior to the third body surface, wherein the third body surface is superior to the fourth body surface, and the further body surface is superior to the fifth body surface.

9. The system as recited in claim 8, wherein the color-related features further comprise a plurality of subsequent yellowness gradient, wherein the plurality of subsequent yellowness gradient corresponds to: a color difference between a skin surface on an abdomen of the subject and a skin surface on a shin of the subject; and a color difference between a skin surface on the shin of the subject and a skin surface on a feet of the subject.

10. The system as recited in any one of claims 1 to 9, wherein the color-related features comprise one of or a combination of: color-space related features, yellowness-related features, clustering features, gradient features and skin tone similarity features.

11. The system as recited in any one of claims 1 to 10, wherein the method further comprises the step of providing user-input features as another predictor input to the machine learning model.

12. The system as recited in any one of claims 10 to 11, wherein the color-space related features are obtained from one of or a combination of color spaces: Red-Green-Blue (RGB),

LAB, Luminance-Chrominance red difference-Chrominance blue difference (YCrCb), Hue Saturation Value (HSV), Cyan-Magenta- Yellow and Key (Black) (CMYK) ,Hue Saturation

Lightness (HSL) and Individual Typology Angle (ITA).

13. The system as recited in any one of claims 10 to 12, wherein the yellowness-related features comprise a magnitude of yellowness selected from one of or a combination of: a colorspace related Y channel of CMYK color space, B channel of LAB color space, blue chromaticity values, and JECI value.

14. The system as recited in any one of claims 10 to 13, wherein the clustering features correspond to an average of respective yellowness value of each of a plurality of cluster centroids, wherein the plurality of cluster centroids are obtained via K-means clustering in a RGB color space, and wherein each of the plurality of cluster centroids comprises yellowness extracted from a corresponding color space, the corresponding color space converted from the RGB color space.

15. The system as recited in any one of claims 10 to 14, wherein the gradient features correspond to a difference between the respective yellowness-related features of respective ones of the plurality of unique body surfaces.

16. The system as recited in any one of claims 10 to 15, wherein the skin tone similarity features are obtained between ones of the plurality of color calibrated patches and respective ones of the plurality of calibration patches.

17. The system as recited in claim 16, wherein the respective ones of the plurality of calibration patches corresponds to at least one of: a Monk Skin Tone (MST) scale and a Fitzpatrick scale.

18. The system as recited in any one of claims 11 to 17, wherein the user-input features comprise one of or a combination of: a gestational age, an hour of life, a mode of delivery, a compatibility of mother-baby blood type (rh/abo), a type of feeding (breastfeeding/formula/mixed), and an indication of Glucose-6-Phosphate Dehydrogenase (G6PD) deficiency.

19. The system as recited in any one of claims 1 to 18, wherein the step of color calibrating the skin patch in each of the plurality of pre-processed images further comprises the step of color calibrating the skin patch in each of the plurality of pre-processed images with the respective calibration patch, based on a polynomial regression.

20. The system as recited in claim 19, wherein the polynomial regression is a second order polynomial regression.

21. The system as recited in any one of claims 1 to 20, wherein the method further comprises the step of: adhering a calibration card to each of the unique body surfaces of the subject, wherein each of the calibration cards corresponds to each of the calibration patches in each of the plurality of images.

22. The system as recited in claim 21, wherein each of the calibration cards comprises a frame portion defining a window, the window corresponding to the skin patch.

23. The system as recited in any one of claims 21 to 22, wherein the frame portion comprises an array of a plurality of colors, the plurality of colors corresponding to the calibration patch.

24. The system as recited in any one of claims 22 to 23, wherein at least a portion of the window is occluded during a prior phototherapy session.

25. The system as recited in any one of claims 1 to 24, wherein the step of pre-processing each of the plurality of images further comprises the step of: segmenting a sclera patch in each of the plurality of images to obtain a plurality of pre-processed images, and wherein the step of color calibrating further comprises the step of: color calibrating the sclera patch in each of the plurality of pre-processed images based on the calibration patch associated with the sclera patch to obtain a plurality of color calibrated patches.

26. The system as recited in any one of claims 1 to 25, wherein the plurality of images of the unique body surfaces of the subject correspond to images taken of selected ones of: a skin surface on a forehead of the subject, a skin surface on a sternum of the subject, a skin surface on an abdomen of the subj ect, a skin surface of an arm of the subj ect, a skin surface of a forearm of the subject, a skin surface of a hand of the subject, a skin surface of a thigh of the subject, a skin surface on a shin of the subject, and a skin surface of a feet of the subject.

27. The system as recited in any one of claims 1 to 26, wherein at least one of the unique body surfaces of the subject is occluded during a prior phototherapy session.

28. The system as recited in any one of claims 26 to 27, wherein at least a portion of the unique body surface of the subject is occluded during a prior phototherapy session.

29. The system as recited in any one of claims 1 to 28, wherein the machine learning model is one selected from: a Support Vector Regression, a Random Forest Regression, and an Extreme Gradient Boosting Regressor, Gradient-boosted Trees Regressor, Extra Trees Regressor, Light Gradient-boosting Machine.

30. The system as recited in any one of claims 1 to 29, further comprising: at least one calibration card comprising a frame portion defining a window, the at least one calibration card being conformable to a body surface of the subject.

31. The system as recited in claim 30, wherein each of the at least one calibration card comprises an adhesive layer for adhering and conforming to the body surface.

32. The system as recited in any one of claims 30 to 31, wherein the frame portion comprises an array of a plurality of colors, the array of the plurality of colors corresponding to the calibration patch.

33. The system as recited in claim 32, wherein selected ones of the plurality of colors correspond to at least one of: a Monk Skin Tone (MST) scale and a Fitzpatrick scale.

34. The system as recited in any one of claims 30 to 33, wherein the frame portion further comprises at least one visual marker (ArUco marker).

35. The system as recited in any one of claims 30 to 34, wherein each of the at least one calibration card further comprises at least one occluding portion movably coupled to the frame portion, wherein the at least one occluding portion is movable to at least partially cover the window, wherein the at least one occluding portion is opaque to a phototherapy wave.

36. The system as recited in any one of claims 30 to 35, further comprising an occluder detachably coupleable to the frame portion, wherein the occluder is attachable to the frame portion to at least partially cover the window, wherein the occluder is opaque to a phototherapy wave.

37. The system as recited in any one of claims 1 to 36, further comprising: a portable device comprising the processor, the portable device being configured to perform the method of predicting bilirubin levels.

38. The system as recited in any one of claims 1 to 37, further comprising: an imaging device in signal communication with the processor, the imaging device being configured to capture the plurality of images.

39. A method of predicting bilirubin levels based on a plurality of images, the method including: pre-processing each of the plurality of images by segmenting a skin patch and a calibration patch in each of the plurality of images to obtain a plurality of pre-processed images, wherein each of the plurality of images corresponds to an image of a unique body surface of a subject; color calibrating the skin patch in each of the plurality of pre-processed images based on the calibration patch associated with the skin patch to obtain a plurality of color calibrated patches; generating color-related features based on the plurality of color calibrated patches; and providing the color-related features as a predictor input to a machine learning model to generate a prediction of bilirubin levels of the subject.