CN118116051A

CN118116051A - Prediction method for evaluating anemia of subject based on facial spectral features

Info

Publication number: CN118116051A
Application number: CN202410237208.9A
Authority: CN
Inventors: 许家佗; 屠立平; 江涛; 张雅涵; 春意; 胡晓娟; 崔骥; 马旭翔; 崔龙涛; 黄景斌
Original assignee: Shanghai University of Traditional Chinese Medicine
Current assignee: Shanghai University of Traditional Chinese Medicine
Priority date: 2024-03-01
Filing date: 2024-03-01
Publication date: 2024-05-31

Abstract

The invention discloses a prediction method for evaluating anemia of a subject based on facial spectral characteristics, and belongs to the field of medical data processing. The prediction method comprises the following steps: obtaining spectral features of a plurality of regions to be evaluated of a subject to be evaluated; and obtaining an anemia evaluation result through the anemia evaluation model by using the spectral characteristics of all the areas to be evaluated. According to the prediction method for evaluating the anemia of the subject based on the facial spectral features, which is provided by the invention, the spectral indexes of the spectral feature wave bands at specific positions are input as input data, so that whether the anemia of the subject to be evaluated is evaluated can be evaluated more quickly and objectively.

Description

Prediction method for evaluating anemia of subject based on facial spectral features

Technical Field

The invention relates to the field of medical data processing, in particular to a prediction method for evaluating anemia of a subject based on facial spectral features.

Background

Anemia refers to a clinically common condition that occurs when the volume of red blood cells in peripheral blood of a human body is lower than a normal value and cannot meet physiological needs. Common symptoms of anemia are fatigue, somnolence, weakness, tachycardia, shortness of breath, increased heart rate, anorexia, hypotension and dizziness.

In recent years, anemia has become an important public health problem in China, and is more common among middle-aged and elderly people in China, particularly the elderly. The study shows that the old people with anemia problem are more than twice as weak as the normal people, and the survival rate of the old anemic people is obviously lower than that of the old non-anemic people. In particular, for the elderly anemic population, low hemoglobin concentration has a great relationship with more disabilities, poor physical performance, reduced cognitive abilities and strength in the elderly, especially for the elderly with a greater age. In addition, preoperative anaemia is an independent factor associated with postoperative complications and increased mortality.

Traditionally, anemia is often assessed by invasive blood sample collection methods, which are a burden for people requiring frequent testing. Frequent blood sampling is uncomfortable for people, especially for the weaknesses of infants, pregnant women, the elderly, patients, and the like.

Currently, non-invasive detection methods such as photoplethysmography, spectroscopy and hyperspectral techniques are beginning to explore. For photoplethysmography waves, acharya S et al developed a multi-model stacked regressor, using non-invasively acquired photoplethysmography waves (PPG) to estimate total hemoglobin (Hb). The noninvasive detection method mainly detects the parts such as pulse wave, nails, fingertips, eyelids, tongue images and the like.

One of the most common techniques used in methods for noninvasive measurement of hemoglobin values is spectroscopy. For example, PLS and BP-ANN models were constructed using fingerend measurements, and a prototype of the portable noninvasive hemoglobin detection system was developed. Mannino RG et al uses the color of the nail bed smartphone photo to estimate hemoglobin levels, uses 11.0g/dL < threshold to define anemia, and in the classification of anemia and healthy individuals, the model used for the study has sensitivities and specificities of 0.92 and 0.76. Bevilacqua V et al, proposed an alternative method of noninvasive hemoglobin estimation based on conjunctival specific region image analysis, studied 77 anaemic and healthy subjects, and modeled using an SVM binary classifier with an accuracy of 0.844, a specificity of 0.824, and a sensitivity of 1.000.Ashwini K et al performed blind and independent comparisons of the presence or absence of pallidum in signs (including conjunctiva, tongue, palm and nail bed) and reference standards (i.e., hemoglobin estimated by electronic cell counters). Diagnostic accuracy was measured by calculating likelihood ratios and 95% confidence intervals for different hemoglobin thresholds and areas under the subject's characteristic curve. While facial spectroscopy for hemoglobin detection is rarely explored in research.

Thus, it is indeed necessary to provide a predictive method for assessing anemia in a subject based on facial spectral features.

Disclosure of Invention

In order to solve at least one aspect of the above problems and disadvantages in the prior art, the present invention provides a prediction method for evaluating anemia in a subject based on facial spectral features, which can at least partially realize that spectral features with relatively high correlation with anemia are obtained by multiple screening of spectral features of a specific part of the subject to be evaluated, thereby improving accuracy of the anemia features and further improving judgment of anemia conditions. The technical scheme is as follows:

According to one aspect of the present invention, there is provided a prediction method for assessing anemia in a subject based on facial spectral features, the prediction method comprising the steps of:

Obtaining spectral features of a plurality of regions to be evaluated in the face of the subject to be evaluated;

and (5) passing the spectral characteristics of all the areas to be evaluated through an anemia evaluation model to obtain anemia evaluation results.

In particular, the spectral features comprise spectral indicators at least one spectral feature band,

The anemia evaluation result is one of anemia or anemia-free anemia probability.

Further, the spectral index includes a spectral reflectance,

The plurality of regions to be evaluated includes the forehead, intereyebrow, nose, mandible, right cheeks, left cheeks, right cheeks and left cheeks of the subject to be evaluated,

At least one spectral characteristic wave band at the forehead, the intereyebrow, the nose, the lower jaw, the right cheeks and the left cheeks is set to 400-700 nm;

At least one spectral feature band at the mandible is set to 400 nm-620 nm; and

At least one spectral feature band at the left cheeks is set to 400nm to 480nm and 490nm to 620nm.

Specifically, the spectral features of all the areas to be evaluated are passed through an anemia evaluation model to obtain anemia evaluation results, including the following steps:

Predicting spectral features of all regions to be assessed by an anemia assessment random forest model to obtain at least one feature region associated with the anemia and selected spectral features at the at least one feature region; and

An anemia evaluation model is constructed based on the at least one feature region and selected spectral features corresponding to the at least one feature region to obtain the anemia evaluation result.

Preferably, the spectral features of all the areas to be assessed are predicted by an anemia assessment random forest model to obtain at least one feature area associated with said anemia and a selected spectral feature at said at least one feature area, comprising the steps of:

Dividing the spectral features of the region to be evaluated into a plurality of spectral sub-features according to a preset interval;

Predicting contribution scores of all spectral sub-features of all the regions to be evaluated through the anemia evaluation random forest model;

And sequencing contribution scores of all the spectral sub-features of all the regions to be evaluated according to a sequence from high to low, selecting the regions to be evaluated positioned in the first N bits and the spectral sub-features corresponding to the regions to be evaluated, wherein each region to be evaluated in the selected regions to be evaluated positioned in the first N bits is used as a characteristic region, and the spectral sub-features corresponding to the regions to be evaluated are selected spectral features of the characteristic region.

Further, obtaining the anemia evaluation result based on the at least one feature region and the selected spectral feature corresponding to the at least one feature region, comprises:

Statistically analyzing the characteristic areas positioned in the front N bits and the selected spectrum characteristics corresponding to the characteristic areas, and selecting and obtaining a plurality of selected areas and the selected spectrum characteristics corresponding to each selected area in the plurality of selected areas;

Constructing an anemia evaluation model based on all selected regions and selected spectral features corresponding to the selected regions;

and obtaining the anemia evaluation result based on the anemia evaluation model.

Specifically, constructing an anemia evaluation model based on all selected regions and selected spectral features corresponding to the selected regions, comprising:

constructing the anemia evaluation model by Logistic regression analysis based on all selected regions and selected spectral features corresponding to the selected regions; or (b)

And obtaining the anemia evaluation model through a preset voting method in the anemia evaluation random forest model based on all characteristic areas in at least one characteristic area and selected spectral characteristics corresponding to the characteristic areas.

Preferably, the anemia evaluation model is a Logistic anemia evaluation model or a random forest prediction model, and the expression of the Logistic anemia evaluation model is:

Logit(p)＝β+α₁X₁-α₂X₂-α₃X₃+α₄X₄-α₅X₅+α₆X₆-α₇X₇

Where Logit (p) represents the anemia probability, β represents the bias value, X ₁ represents the spectral reflectance value of 500nm visible light reflected at the nose, X ₂ represents the spectral reflectance value of 700nm visible light reflected at the nose, X ₃ represents the spectral reflectance value of 420nm visible light reflected at the mandible, X ₄ represents the spectral reflectance value of 570nm visible light reflected at the mandible, X ₅ represents the spectral reflectance value of 310nm visible light reflected at the mandible, X ₆ represents the spectral reflectance value of 400nm visible light reflected at the right cheeks, X ₇ represents the spectral reflectance value of 420nm visible light reflected at the left cheeks, and α _i (i=1, 2, …, 7) represents the coefficient.

Further, the spectral features of the plurality of regions to be evaluated in the face of the subject to be evaluated are spectral feature data screened from all regions to be evaluated and spectral feature data corresponding to each of all regions to be evaluated based on statistical analysis results obtained by statistically analyzing all spectral feature data of all regions to be evaluated in the face of the subject to be evaluated,

The subject to be evaluated is middle-aged and elderly people to be evaluated.

Further, the anemia evaluation random forest model is provided with 4500 decision trees, and judges whether each decision tree of the anemia evaluation random forest model continues to split or not through the non-purity of the keni,

And when the reduction value of the Indonesia is more than or equal to 0, the decision tree corresponding to the Indonesia is continuously split.

The prediction method for evaluating anemia in a subject based on facial spectral features according to an embodiment of the present invention has at least one of the following advantages:

(1) The prediction method for evaluating the anemia of the subject based on the facial spectral features can screen the spectral features of the specific part of the subject to be evaluated for multiple times to obtain the spectral features which have relatively high correlation with the anemia, so that the accuracy of the obtained anemia features is improved, and the judgment of the anemia condition is further improved;

(2) The prediction method for evaluating the anemia of the subject based on the facial spectral features can be used for predicting whether the subject to be evaluated is anemia or not by inputting the spectral index (for example, spectral reflectivity) of the spectral feature wave band of a specific position as input data;

(3) The prediction method for evaluating the anemia of the subject based on the facial spectrum features can evaluate whether the subject is anemic or not more rapidly and objectively;

(4) According to the prediction method for evaluating the anemia of the subject based on the facial spectral features, the random forest screening is adopted to extract the spectral features such as the specific spectral reflectivity of the specific wave band of the specific position of the subject, so that the judgment on whether the anemia is evaluated is more accurate;

(5) The prediction method for evaluating the anemia of the subject based on the facial spectral features can obtain the evaluation results of the anemia and the non-anemia of the middle-aged and the elderly based on machine learning and through facial spectral information, and provide a repeatable spectral reference index for clinic, so that the prediction method can help clinical intervention of the anemia of the middle-aged and the elderly to provide a reference basis.

Drawings

These and/or other aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the preferred embodiments, taken in conjunction with the accompanying drawings, in which:

FIG. 1 is a flow chart of a predictive method for assessing anemia in a subject based on facial spectral features according to one embodiment of the invention;

FIG. 2 is a flow chart of obtaining an evaluation result based on the anemia evaluation model shown in FIG. 1;

FIG. 3 is a schematic illustration of the feature region and corresponding selected spectral features of FIG. 2 prior to the anemia evaluation random forest model contribution score 30;

FIG. 4 is a flow chart of a method of constructing the anemia evaluation random forest model shown in FIG. 2;

FIG. 5A is a schematic diagram of a hybrid matrix of the decision tree model shown in FIG. 4;

FIG. 5B is a schematic diagram of a hybrid matrix of the support vector machine model shown in FIG. 4;

FIG. 5C is a schematic diagram of a hybrid matrix of the random forest model shown in FIG. 4;

FIG. 5D is a hybrid matrix schematic of the K-nearest neighbor model shown in FIG. 4;

FIG. 5E is a schematic diagram of a hybrid matrix of the artificial neural network model shown in FIG. 4;

FIG. 5F is a schematic diagram of a hybrid matrix of the Bayesian classification model shown in FIG. 4;

fig. 6 is a graph of ROC for each model shown in fig. 5A-5F.

Detailed Description

The technical scheme of the invention is further specifically described below through examples and with reference to the accompanying drawings. In the specification, the same or similar reference numerals denote the same or similar components. The following description of embodiments of the present invention with reference to the accompanying drawings is intended to illustrate the general inventive concept and should not be taken as limiting the invention.

For spectroscopy, spectroscopic techniques typically explore the properties of materials through interactions between substances and electromagnetic waves of different frequencies. For hyperspectral imaging techniques, it provides a more detailed segmentation in the spectral dimension, containing more information than red, green, blue (RGB).

The color measurement for traditional Chinese medicine color evaluation based on visible reflectance spectroscopy is similar to clinical color evaluation, and reflects the accuracy, the authenticity and the reliability of the color evaluation result. Raman spectroscopy can analyze the distribution of metabolites, lipids, proteins, water and blood content in tissues by skin detection. Hyperspectral imaging can estimate hemoglobin concentration and blood oxygen saturation by diffuse reflection of tissue.

In traditional Chinese medicine color diagnosis, the facial features of anemic patients are always different from those of healthy people, and are usually yellow or pale, but qualitative and quantitative experimental researches are lacking in the research of the relationship between the anemia and the facial color.

In view of this, the present invention constructs an anemia assessment model by collecting facial images of anemic patients and healthy controls, especially facial images of middle-aged and elderly people, analyzing spectral reflectance and image features of their faces by an anemia assessment random forest model, finding that anemia is highly correlated with facial spectral features (e.g., facial spectral reflectance) in specific locations and bands, and that skin colors are generally more pale especially at the nose, right cheek, right cheekbone, and mandible, while the filtered feature locations and feature spectral bands of the anemia assessment random forest model based on a machine learning method.

Referring to fig. 1, a predictive method for assessing anemia in a subject based on facial spectral features according to one embodiment of the invention is shown, the predictive method comprising the steps of:

and obtaining an anemia evaluation result through the anemia evaluation model based on the spectral characteristics of all the areas to be evaluated.

In one example, the spectral features include visible light spectral features and invisible light spectral features. The wavelength of the visible light spectrum is between 400nm and 760 nm.

In one example, the invisible light may be ultraviolet, infrared, far infrared, or the like.

In one example, the spectral features include a spectral index at least one spectral feature band. The spectral index is spectral reflectance. It will be appreciated by those skilled in the art that the spectral features may also be other spectral indicators at least one spectral feature band, such as spectral absorptance. In one example, the subject to be evaluated includes middle aged and elderly people to be evaluated, young people, etc., preferably the subject to be evaluated is middle aged and elderly people to be evaluated.

In one example, the plurality of regions to be assessed includes the forehead, intereyebrow, nose, mandible, right cheeks, left cheeks, right cheeks, and left cheeks of the subject to be assessed. Of course, it will be apparent to those skilled in the art that other regions of the face or other parts of the body other than the face of the subject to be evaluated may also be selected in order to improve the accuracy of the prediction results.

In one example, at least one spectral signature band at the forehead, the brow, the nose, the mandible, the right cheeks, and the left cheeks is set to 400nm to 700nm. At least one spectral feature band at the mandible is set to 400nm to 620nm. At least one spectral feature band at the left cheeks is set to 400nm to 480nm and 490nm to 620nm.

That is, each region to be evaluated of the subject to be evaluated may correspond to one spectral signature band in which there are multiple spectral indices (e.g., multiple spectral reflectivities). Of course, it will be appreciated by those skilled in the art that each region to be evaluated may correspond to two, three, or more spectral signature bands, each of which may have multiple spectral indices (e.g., multiple spectral reflectivities).

In the same spectral signature band in the same evaluation region, healthy people and anemic subjects have differentiated spectral reflectivities. And, the spectral reflectance of the same spectral characteristic wave band in the same evaluation area among different anemic subjects is also subject to individual differentiation.

The present invention provides for the comprehensive analysis of the anemia condition (e.g., anemia free, anemia) of a subject under evaluation by collecting spectral features of a plurality of different regions under evaluation of the same subject under evaluation. It should be understood by those skilled in the art that the spectrum band of the subject to be evaluated can be other bands than 400nm to 700nm, for example, the ultraviolet spectrum band and the infrared spectrum band can be collected to comprehensively analyze and evaluate the anemia condition of the subject to be evaluated.

In one example, obtaining spectral features of a plurality of regions to be evaluated of a subject to be evaluated is acquiring a facial image of the subject to be evaluated, then extracting facial spectral data of the subject to be evaluated from the facial image (e.g., extracting 400nm to 700nm spectral indicators of the right cheek) and determining respective spectral feature bands of the forehead, the brow, the nose, the chin, the right cheek, the left cheek, the right cheek, and the left cheek based on the facial spectral data.

The method for predicting anemia of the subject based on the facial spectral features is exemplified by taking a plurality of regions to be evaluated as regions of the face and taking the spectral indexes as spectral reflectances, and the method and principle for predicting anemia of the subject to be evaluated by collecting other body parts and/or spectral indexes are substantially the same as the method and principle in which the regions of the face are taken as regions to be evaluated and the spectral indexes are spectral reflectances, and will not be described in detail here.

In one example, a method of determining spectral features of a plurality of regions to be evaluated of a subject to be evaluated includes the steps of:

Step S11, clinical data acquisition: eight points (i.e., the region to be evaluated) were acquired with a CS-600CG spectrometer for the subject to be evaluated: spectral features of forehead, intereyebrow, nose, mandible, right cheek, left cheek, for example, spectral reflectivities of eight points to visible light in 400 nm-700 nm wave bands (i.e. spectral feature data of face) can be obtained;

Step S12, selecting spectral reflectivities of forehead, eyebrow, nose, mandible, right cheek and left cheek in 400-700 nm wave band, mandible in 400-620 nm wave band, left cheek in 400-480 nm and 490-620 nm wave band from facial spectral characteristic data as spectral characteristics.

Of course, the person skilled in the art can also obtain spectral features of the individual by statistical analysis of the spectral feature data of the face of the subject to be evaluated. Specifically, the spectral reflectances of the 8 points in the wave bands of 400nm to 700nm are respectively statistically analyzed by adopting SPSS25.0, wherein the data conforming to the normalization are represented by average numbers and Standard Deviations (SD), and the data not conforming to the normalization are represented by median numbers and quartiles. At the time of statistical analysis, the test level was set to α=0.05. When two groups of data, namely normal data and non-normal data, are compared, the normal and variance uniformity is detected by adopting independent samples, and the non-normal and variance uniformity is detected by adopting Mann-Whitney U. All tests used a two-tailed method, with P <0.05 being statistically significant for differences. The statistical analysis shows that except the spectral reflectance of the forehead in the wave band of 610nm-700nm, the spectral reflectance of the mandible in the wave band of 620-700nm, the spectral reflectance of the left cheek in the wave band of 480nm-490nm and the spectral reflectance in the wave band of 620-700nm are not statistically significant, and the spectral reflectance in the rest wave bands are statistically significant, namely, the spectral reflectance of the forehead, the brow, the nose, the mandible, the right cheek and the left cheek in the wave band of 400 nm-700nm and the spectral reflectance of the mandible in the wave band of 400 nm-620 nm; the spectral reflectivities of the left cheeks at the wave bands of 400-480 nm and 490-620 nm have statistical significance, so that the preliminary screening of the spectral characteristics is realized, and the spectral characteristics (namely the spectral characteristics after screening) which are relatively related to anemia are reserved.

In one example, as shown in fig. 2, the spectral features of all the regions to be evaluated are passed through an anemia evaluation model to obtain anemia evaluation results, comprising the steps of:

For example, the spectral reflectances of the forehead, the eyebrow space, the nose, the mandible, the right cheek, and the left cheek, each for 400nm to 700nm visible light, the spectral reflectances of the mandible for 400nm to 620nm visible light, and the spectral reflectances of the left cheek for 400nm to 480nm and 490nm to 620nm visible light are all input into the anemia evaluation random forest model to score the contribution of the spectral reflectances of a specific portion at a specific band.

In one example, the quantitative relationship between the at least one feature region and the selected spectral feature corresponding to the feature region is one-to-many or many-to-many. For example, when the feature area is 1, its corresponding selected spectral feature may be 2,3, or more. For another example, when the number of feature regions is 8, each feature region may be 1 selected spectral feature, or may be a plurality of selected spectral features, or may be a mixture of both.

In one example, predicting spectral features of all regions to be evaluated by an anemia evaluation random forest model to obtain at least one feature region associated with the anemia and a selected spectral feature at the at least one feature region, comprising the steps of:

The contribution degree of all the spectral sub-features of all the regions to be evaluated is sequenced from high to low, the regions to be evaluated positioned in the first N bits and the spectral sub-features corresponding to the regions to be evaluated are selected, each region to be evaluated in the selected regions to be evaluated positioned in the first N bits is used as a characteristic region, and the spectral sub-features corresponding to the regions to be evaluated are selected spectral features of the characteristic region.

In one example, the screened spectral feature data, such as the spectral reflectance at 400nm to 700nm for each of the forehead, the glabella, the nose, the mandible, the right cheek, and the left cheek, the spectral reflectance at 400 to 620nm for the mandible, and the spectral reflectance at 400nm to 480nm and 490nm to 620nm for the left cheek, may be statistically extracted as spectral features by Pycharm invoking various functions of sklearn codebase.

As shown in fig. 2, the person skilled in the art can divide the spectral features having statistical significance in eight points (i.e. the region to be evaluated) according to needs, namely, the spectral reflectivities of the forehead, the brow, the nose, the mandible, the right cheek and the left cheek in the wave bands of 400nm to 700nm, the spectral reflectivities of the mandible in the wave bands of 400nm to 620nm, the spectral reflectivities of the left cheek in the wave bands of 400nm to 480nm and the spectral reflectivities of the left cheek in the wave bands of 490nm to 620nm, respectively, by 10nm each interval, for example, the spectral reflectivities of the forehead in the wave bands of 400nm to 700nm can be divided into the spectral reflectivities at 400nm, the spectral reflectivities at 410nm, … …, the spectral reflectivities at 680nm, the spectral reflectivities at 690nm and the spectral reflectivities at 700nm, thereby obtaining the spectral sub-features of the forehead. It will be understood by those skilled in the art that when dividing and extracting the spectral sub-features, the interval wavelength may be set to 5nm, 10nm, 15nm, etc., and of course, the art may also divide according to the spectral reflectivities, for example, the spectral reflectivities of different spectral bands are randomly selected according to a preset number in the same region to be evaluated, and those skilled in the art may choose the spectral reflectivities as required, so long as the accuracy of the result of the anemia evaluation random forest model and the generalization capability of the model can be ensured.

For another example, the spectral reflectance at 400nm to 480nm of the left cheeks can be divided into the spectral reflectance at 400nm, the spectral reflectance at 410nm, … …, the spectral reflectance at 470nm and the spectral reflectance at 480 nm. And the spectral reflectance at 490nm to 620nm at the left cheeks is divided into the spectral reflectance at 490nm, the spectral reflectance at 500nm, … …, the spectral reflectance at 610nm, the spectral reflectance at 620nm, and so on.

Those skilled in the art will appreciate that the spectral signature band may be specifically one wavelength (e.g., 400nm, 610nm, etc.) or may be a range of wavelengths (400 nm to 480nm, 400nm to 700 nm).

In one example, the extracted or partitioned spectral sub-features of the forehead, brow, nose, mandible, right cheeks, and left cheeks are all input into an anemia evaluation random forest model, and then the contribution score of each spectral sub-feature of each region to be evaluated is predicted by the model.

For example, as can be seen from fig. 3, the extracted or divided spectral sub-features of the forehead, the brow, the nose, the mandible, the right cheek and the left cheek, each in the 400nm to 700nm band, i.e. each of the spectral reflectance at 400nm, … …, the spectral reflectance at 680nm, the spectral reflectance at 690nm and the spectral reflectance at 700nm, are all input into the anemia evaluation random forest model, and the contribution score of the spectral reflectance at 400nm, 410nm, 430nm, 440nm, … … and 700nm can be obtained for each region to be evaluated (i.e. 8 points).

And sequentially arranging contribution degree scores of all the spectral sub-features of the eight points in a sequence from high to low, and selecting the region to be evaluated (i.e. the point) positioned in the first 30 bits and the spectral sub-features corresponding to the point, wherein each region to be evaluated in the selected region to be evaluated positioned in the first 30 bits is used as a characteristic region, and the spectral sub-features corresponding to the region to be evaluated are selected spectral features of the characteristic region.

For example, as shown in fig. 3, the spectral reflectance of visible light of the nose to 400nm is taken as a characteristic region (i.e., nose) and the spectral reflectance at 400nm is taken as a selected spectral feature, the spectral reflectance of visible light of the left cheekbone to 410nm is taken as a characteristic region (i.e., left cheekbone) and the selected spectral feature (spectral reflectance at 410 nm), and so on, and are not exemplified here.

It will be appreciated by those skilled in the art that after all the spectral feature sub-bands of the eight points are ranked according to the contribution degree, the first 20 bits of the region to be evaluated and the 50 bits of the region to be evaluated and the spectral sub-features corresponding to the regions to be evaluated can be selected as the feature region and the selected spectral features. Of course, the selected spectral features of the feature region corresponding to the feature region may also be determined based on a preset percentile. For example, a spectral sub-feature before the 80 th bit may be selected as the selected spectral feature, and a point corresponding thereto may be a feature region. This example is merely an illustrative example and those skilled in the art should not be construed as limiting the invention.

In one example, as shown in fig. 3, the feature area at the first 30 positions includes nose, right cheek, lower jaw, intereyebrow, left cheek. The characteristic region and the spectrum characteristic which are more relevant to anemia are further screened out through the anemia evaluation random forest model, so that not only can the pre-exploration be carried out for the non-contact hyperspectral detection technology, but also the accuracy of the follow-up anemia evaluation prediction can be improved.

In one example, the anemia assessment random forest model is provided with 4500 decision trees. Since the non-purity of the keni has the following advantages: (1) High computational efficiency, generally faster than information gain computation, can provide efficient feature selection; (2) For classification problems, such as processing multi-class classification problems and classification problems, the performance is very good and effective, and the classification problems can also work well under the condition of unbalanced classes; (3) The processing of the numerical features (e.g., spectral feature bands, spectral feature sub-bands) is more convenient than the information gain, i.e., the base-no-purity can be directly used to compare the numerical features, which the information gain needs to discretize; (4) robustness: the non-purity of the keni has a certain robustness to the presence of outliers. In the presence of noise or outliers in the data, the base purity performance will be better because it has less impact on individual outliers. Thus, the split criteria for the anemia assessment random forest model are determined by the keni unreliability. And when the reduction value of the Indonesia is more than or equal to 0, the decision tree corresponding to the Indonesia is continuously split. In one example, the stopping condition of the anemia evaluation random forest model includes: (1) Maximum depth each decision tree will grow until each leaf node contains fewer samples than the minimum number of samples split, stopping splitting. The minimum sample number split is set to, for example, 2, indicating that a node needs to contain at least two samples to continue splitting. (2) The minimum number of samples of the leaf node is set to, for example, 1, indicating that the leaf node contains at least one sample, and when there is less than one sample, splitting is stopped. (3) minimal reduction in non-purity: cleavage is stopped when the reduction in post-cleavage purity is less than 0. When the decision tree satisfies the at least one condition, the decision tree stops splitting.

In one example, as shown in fig. 4, the anemia evaluation random forest model may be constructed by:

Step S21 inputs spectral features (e.g., positions with statistical differences and spectral sub-feature bands) of all regions to be assessed of the subject population to be assessed (including anemic subjects and healthy people) into decision trees, random forests, decision trees, support Vector Machines (SVMs), K-nearest neighbors (KNNs), artificial Neural Networks (ANNs), and bayes classifiers, respectively, to construct machine-learning-based anemia assessment and prediction models.

Step S22, evaluating each anemia evaluation and prediction model according to the performance evaluation index to determine an optimal model, and taking the optimal model as a anemia evaluation random forest model.

In one example, the spectral features of the population of subjects to be evaluated that are input into each model are labeled with their own corresponding blood routine index (e.g., hemoglobin detection result), i.e., when the hemoglobin detection result of the subject to be evaluated shows that it is a healthy person, the spectral features of the subject to be evaluated are labeled as healthy person. When the hemoglobin detection result of the subject to be evaluated shows that the subject has anemia, the spectral characteristics of the subject to be evaluated are marked as an anemic subject.

In one example, hemoglobin detection results for each subject to be evaluated in a population of subjects to be evaluated are obtained using invasive blood sample testing.

In one example, a method of obtaining spectral features of all regions to be assessed of a population of subjects to be assessed comprises the steps of:

Step S11, clinical data acquisition: anemic subjects and healthy persons meeting the criteria are selected as a population of subjects to be evaluated, and eight points (i.e., areas to be evaluated) of each subject to be evaluated are acquired by a CS-600CG spectrometer: spectral features of the forehead, intereyebrow, nose, mandible, right cheek, left cheek, for example, spectral reflectivities of eight points each in the 400 nm-700 nm band can be obtained;

Step S12, statistical analysis: the spectral reflectivities of 8 points of the subject to be evaluated in the wave bands of 400 nm-700 nm are respectively and statistically analyzed by adopting SPSS25.0, wherein the data conforming to the normalization are represented by mean and Standard Deviation (SD), and the data not conforming to the normalization are represented by median and quartile. At the time of statistical analysis, the test level was set to α=0.05. When two groups of data, namely normal data and non-normal data, are compared, the normal and variance uniformity is detected by adopting independent samples, and the non-normal and variance uniformity is detected by adopting Mann-Whitney U. All tests used a two-tailed method, with P <0.05 being statistically significant for differences.

The statistical analysis shows that the spectral reflectance of the forehead in the wave band of 610-700nm, the spectral reflectance of the mandible in the wave band of 620-700nm, the spectral reflectance of the cheeks in the wave band of 480-490nm and the spectral reflectance in the wave band of 620-700nm are not statistically significant, and the spectral reflectance in the rest wave bands are statistically significant, thereby realizing the primary screening of the spectral characteristics and retaining the spectral characteristics (namely the spectral characteristics after screening) which are relatively related to anemia.

As shown in Table one, baseline information comparisons were also made for age and gender, as age was not normally distributed, and thus gender was tested using chi-square with a nonparametric test. P > 0.05 indicates comparability. As shown in table two, the spectral reflectance of the portion 30 before the random forest contribution was selected for statistical result display because the spectral dimension was high and all data could not be displayed.

Table-anemic subjects and healthy person age sex baseline comparison table

Statistical analysis of spectral reflectance of the sites with random forest contribution scores of the Subjects with Table II anemia and healthy people at the top 30

In one example, to avoid occasional results from modeling results, light, medium, and severe anemic subjects were classified using an average random sampling method prior to model training. One is to divide the data into training sets (108 cases) and test sets (48 cases) in a 7:3 ratio. The other is to remove mild anemic subjects from the test set based on the training set and the test set, and randomly remove an equivalent number of healthy controls. The performance of the model is calibrated through two independent test sets, and the anemia group and the health group are used as the test sets for calibration.

In one example, in each decision tree in the decision number and random forest, one node of the decision tree contains at least 24 samples at each sub-node after the branch. In the support vector machine, it tolerates error tol set to 1e-6 and penalty term set to 35. In K-neighbors, the number of neighbors is set to 6. In the artificial neural network, the iteration number is set to 1200000, the hidden layer is set to 120 layers, and the maximum variance of all features is set to 1e-11 in the naive Bayes classification. This example is merely an illustrative example and those skilled in the art should not be construed as limiting the invention.

In one example, as shown in fig. 5A-5F, fig. 6, and table three, the performance evaluation index includes Sensitivity (Sensitivity), specificity (SPECIFICITY), F1 score (f1_score), accuracy (Accuracy), AUC, positive Predictive Value (PPV), negative Predictive Value (NPV) to evaluate the performance of the model.

Table three evaluation table of classification results of each model

From fig. 5A-5F, fig. 6 and table iii, it can be seen that the random forest prediction model performs best in the classification model based on the facial spectral feature data, the accuracy is 0.875, the auc is 0.925, and the true negative rate and the true positive rate are the highest. As can be seen from fig. 6, the AUC of the decision tree model is 0.806, the AUC of the svm model is 0.903, the AUC of the knn model is 0.880, the AUC of the ann model is 0.911, and the AUC of the bayesian classification model is 0.802.

In one example, constructing an anemia assessment model based on the at least one feature region and selected spectral features corresponding to the at least one feature region to obtain the anemia assessment result, includes:

statistically analyzing the characteristic areas positioned in the front N bits and the selected spectral characteristics corresponding to the characteristic areas, and selecting and obtaining a plurality of selected areas and selected characteristic wave bands corresponding to each selected area in the plurality of selected areas;

In one example, the anemia evaluation result is one of anemia or non-anemia, anemia probability. For example, anemia results in anemia. Also for example, the anemia results in anemia. As another example, anemia results in a anemia of 68% probability. This example is merely an illustrative example and those skilled in the art should not be construed as limiting the invention.

In one example, selected regions with statistical significance (p < 0.05) and selected spectral features corresponding to and having statistical significance respectively to the selected regions are obtained by statistical analysis (e.g., by SPSS 25.0) using the selected regions (i.e., selected) located at the random forest contribution front 30 and the selected spectral features corresponding to and selected from the selected regions.

For example, after statistical analysis, the 7 index noses were determined to have a spectral reflectance at 500nm, a spectral reflectance at 700nm, a spectral reflectance at 420nm, a spectral reflectance at 570nm, a spectral reflectance at 610nm, a spectral reflectance at 400nm for the right cheeks, and a spectral reflectance at 420nm for the right cheeks were statistically significant selected regions and corresponding selected spectral features.

And then, analyzing each index parameter (shown in a table four) based on the Logistic regression of the multi-position multi-band spectrum characteristic data to form a Logistic anemia evaluation model.

In one example, the Logistic anemia assessment model was tested by R ² for goodness of fit, which was 0.808, and Hosmer-Lemeshow, with p=0.562, indicating that the Logistic anemia assessment model fits well.

Table four spectral feature logistic regression analysis results of first 30 bits of random forest contribution degree

In one example, the expression of the Logistic anemia assessment model is:

Where Logit (p) represents the anemia probability of the subject to be evaluated, β represents the bias value, X ₁ represents the spectral reflectance of 500nm visible light reflected at the nose of the subject to be evaluated, X ₂ represents the spectral reflectance of 700nm visible light reflected at the nose, X ₃ represents the spectral reflectance of 420nm visible light reflected at the mandible, X ₄ represents the spectral reflectance of 570nm visible light reflected at the mandible, X ₅ represents the spectral reflectance of 610nm visible light reflected at the mandible, X ₆ represents the spectral reflectance of 400nm visible light reflected at the right cheeks, X ₇ represents the spectral reflectance of 420nm visible light reflected at the left cheeks, and α _i (i=1, 2, …, 7) represents the coefficient.

In one example, the anemia evaluation model may also be obtained by a voting method preset in the anemia evaluation random forest model based on all feature regions and selected spectral features (i.e. spectral sub-features) corresponding to the feature regions. For example, the results of all decision trees in a random forest may be voted based on a majority voting rule, for example, to obtain a feature region and a corresponding feature band with the highest voting proportion, and whether to suffer from anemia, for example, anemia or anemia, is determined based on the highest feature region and the corresponding spectral sub-features. In one example, the anemia evaluation random forest model may be constructed in combination with the Bagging method to obtain the anemia evaluation model.

(1) The prediction method for evaluating the anemia of the subject based on the facial spectral features can screen the spectral features of the specific part of the subject to be evaluated for multiple times to obtain the spectral features with relatively high correlation with the anemia, thereby improving the accuracy of the obtained anemia features and further improving the judgment of the anemia condition

Although a few embodiments of the present general inventive concept have been shown and described, it would be appreciated by those skilled in the art that changes may be made in these embodiments without departing from the principles and spirit of the general inventive concept, the scope of which is defined in the claims and their equivalents.

Claims

1. A predictive method for assessing anemia in a subject based on facial spectral features, the predictive method comprising the steps of:

2. The prediction method according to claim 1, wherein,

The spectral features include spectral indicators at least one spectral feature band,

3. The prediction method according to claim 2, wherein,

The spectral index comprises a spectral reflectance and,

At least one spectral feature band at the mandible is set to 400 nm-620 nm; and

4. A prediction method according to any one of claim 1 to 3, wherein,

Obtaining an anemia evaluation result through an anemia evaluation model based on the spectral characteristics of all the areas to be evaluated, wherein the method comprises the following steps:

5. The prediction method according to claim 4, wherein,

Predicting by an anemia evaluation random forest model based on spectral features of all regions to be evaluated to obtain at least one feature region associated with the anemia and a selected spectral feature at the at least one feature region, comprising the steps of:

predicting contribution scores of all spectral sub-features of all the regions to be evaluated through the anemia evaluation random forest model; and

6. The prediction method according to claim 5, wherein,

Constructing an anemia assessment model based on the at least one feature region and selected spectral features corresponding to the at least one feature region to obtain the anemia assessment result, comprising:

Constructing an anemia evaluation model based on all selected regions and selected spectral features corresponding to the selected regions; and

7. The prediction method according to claim 6, wherein,

Constructing an anemia evaluation model based on all selected regions and selected spectral features corresponding to the selected regions, comprising:

8. The prediction method according to any one of claims 1 to 3 and 7,

The anemia evaluation model is a Logistic anemia evaluation model or a random forest prediction model,

The expression of the Logistic anemia evaluation model is:

9. The prediction method according to any one of claims 1 to 4, wherein,

The spectral features of a plurality of regions to be evaluated in the face of the subject to be evaluated are spectral feature data screened from all regions to be evaluated and spectral feature data corresponding to all regions to be evaluated respectively based on statistical analysis results obtained by statistically analyzing all spectral feature data of all regions to be evaluated in the face of the subject to be evaluated,

The subject to be evaluated is middle-aged and elderly people to be evaluated.

10. The prediction method according to claim 4, wherein,

The anemia evaluation random forest model is provided with 4500 decision trees, and judges whether each decision tree of the anemia evaluation random forest model continues to split or not through the genie unrepeace,