IT202000016252A1 - ASSESSMENT SYSTEM OF THE TERRITORIAL INCIDENCE OF SYMPTOMS - Google Patents

Description

Evaluation system of the territorial incidence of symptoms

Description

The object of the present invention is a system for evaluating the territorial incidence of certain symptoms, capable of providing an indication of their density, in particular for preparing containment measures.

As ? known, the onset of epidemics can? lead a territorial health system to structural suffering, with even heavy repercussions on the ability to of a health system to prepare both therapeutic treatments and containment measures, which require the preparation of resources for this purpose.

Among the containment measures, can be included the supply of medicines, health aids, personal protective equipment, the availability of of beds, medical and nursing staff, the adoption of social distancing rules.

To obtain an evaluation of what could be the density? of predetermined symptoms linked to epidemics and/or endemics, the tracing of contacts with subjects with symptoms and tests on potentially infected people cannot generally respond to statistical needs, but must be aimed at identifying subjects to isolate and/or to treat.

Furthermore, normal tracing techniques risk providing the necessary information too late with respect to planning needs in the area.

To overcome this inconvenience, tracking systems based on the premature identification of symptoms have been proposed: this route can be considered promising, but ? however based on information provided by subjects who may not be able to make correct assessments of their symptoms.

Furthermore, individuals who should autonomously provide their own assessments may have concerns about the disclosure of sensitive data concerning their health.

These fears would be unjustified if the privacy laws in force were respected, but even more so would they have no reason to exist if the purpose of the evaluation? that of understanding what the needs of a territory could be in order to respond to a health emergency.

The technical problem that ? at the basis of the present invention ? to provide an evaluation system which allows the drawback mentioned with reference to the prior art to be overcome.

This problem is solved by an evaluation system as specified above which is characterized by the fact that it includes:

? a wearable device that takes measurements of body data of a plurality? of users associated with the system;

? a digital video and image detector, which allows you to detect biometric data of said plurality? of users;

? a geo-positioning device to determine the positioning data of said plurality? of users, and associates them with geographic coordinates of a reference area or region;

? a device for transmitting body data, video and digital images and geographical coordinates to a server, on which installed artificial intelligence software and a classifier, which processes this evaluation.

The main advantage of the evaluation system according to the present invention lies in allowing an association of a set of information to geographical coordinates which are not linked to each of the participating users.

In particular, an ecosystem is created to care for the health condition of a community, which is based on the multi-source collection of symptoms, correlating the measurement of the medical device to a model managed by artificial intelligence.

The user has the possibility? to activate the optimization of the model on the recent history of your biometric data through an active calibration. This possibility? of customization? based on local information processing through an accumulation of data that takes place only in the physical memory of the mobile device (these are text files of a few bytes, for each measurement set) and on data processing carried out through the software who works locally, in the application installed on the user's mobile device. The mode? described above, developed following the privacy-by-design concept, ? to be understood from the point of view of tele-prevention tools.

The result of the user-ecosystem interaction? what? a feedback on the health condition and a possible guideline as support in the daily routine of the user.

This invention will come described hereinafter according to a preferred embodiment thereof, provided by way of non-limiting example with reference to the attached drawings in which:

? Figure 1 shows a block diagram illustrating the evaluation system according to the present invention;

? Figure 2 shows a block diagram illustrating a wearable device used in the evaluation system according to the invention; And

? Figure 3 illustrates the data processing process in the evaluation system according to the invention.

With reference to the figures, a system of evaluation of the territorial incidence of certain symptoms is described.

As explained above, the purpose of the system ? to provide an? indication on their density? in relation to a predetermined territory.

Such a system includes a wearable device that takes measurements of a plurality of body data. of users associated with the system

This wearable device can be made by assembling components designed to collect biometric data such as heart rate, respiratory rate, blood pressure, body temperature and blood oxygen saturation.

In addition to the aforementioned sensors, the device ? equipped with a microprocessor, a transmission module, such as a Bluetooth driver?, and an interface with the user, such as a touchscreen and an acoustic signal. It can include a built-in battery in the device so that it can operate wirelessly and on the go.

The wearable device collects data from its dedicated sensors, and they are sent to a server through the intermediary of a communication device, for example a smartphone that has the necessary functionalities. Bluetooth ? and a connection to the network (figure 2).

It is understood that this data is sent to a server which constitutes a computer center which uses procedures adequate to the legislation on privacy and data security.

The evaluation system therefore comprises a detector which allows images and biometric data of said plurality to be detected. of users.

This detector can be made up of the same smartphone already? previously mentioned, it pu? be used both to video-photograph the user's face and to capture the fundus of the retina.

Once obtained, the images can be sent to the aforementioned server, possibly using an application installed on the smartphone.

The software installed in it ? capable of obtaining significant data from them automatically, i.e. without a user performing any type of intervention.

The collection of images with the aforementioned device requires the user to make a video of the face with the front camera of the smartphone, while reading subtitles formulated as short and simple sentences, associated with a slideshow shown on the screen.

The software operates on the basis of a predictive/diagnostic support logic that is based on two macro-levels of a convolutional neural network (CNN). A first level identifies the presence or absence of the eye in the video frames and chooses for the second level CNN which development to undertake between two cases.

In the first case, the eye? present in the data, and the software selects some frames and crops the images in the eye area. The algorithm searches for significant clusters, by searching for patterns of regularity, within the ocular fundus for the study of the retina, related to certain pathologies of the eye itself and of the individual in a more general sense. extended.

In the second case, the eye is not present in the data, and therefore the images or the video are processed to study particular asymmetries and micro-expressions such as reduced fluidity? expression, tensions in the area of the mouth, browbone and forehead. Visual scanning behaviors (VSBs) techniques can be used, adopted for the study of the response to emotional or non-emotional stimuli, which allow the detection of information from the psychological sphere. An example of this application? the research of manifestation of apathy in Alzheimer's patients.

The presence of the two levels makes it possible to optimize resources in terms of use of computational power, calculation time and efficiency in the convergence of the algorithm. The neural network? structured on pi? levels in such a way as to follow a decision-making path which provides in-depth sections whose access (and therefore activation) takes place according to the flow chart in figure 4.

Therefore, in the first level, the presence or absence of the eye in the video frames is identified by studying the point-of-gaze (PoG: point of view), a point within the visual field which is displayed on the fovea centralis, region of the retina in the center of the macula lutea. This part of the eye is identified as the zone of maximum visual acuity and the central focus of visual attention. The study of PoG is used in various fields such as the optimization of assistive devices for disabilities? motor skills, feedback on the quality? of user interface with digital and studies on mood disorders. Typical features considered are the center of the pupil and reflections from the cornea.

Although video information allows you to collect a huge amount of additional information compared to a photographic image, the search for the aforementioned characteristics is complicated by the movement of the target, hence the need? of AI support based on deep learning methods.

A first phase of definition of a reference level against which to evaluate more? accurate the state of health of the system user provides a guided calibration procedure to determine some geometric references such as the radius of curvature of the cornea, the distance between the center of the curvature of the cornea and the center of the pupil, and the vertical offset angles and horizontal between the optical and visual axes.

The procedure ? based on focusing the PoG of sight of the individual towards targets with known spatial reference and on the acquisition of image and video samples.

Starting from this first interaction, differential comparison techniques are employed to support the software and the method optimization routine.

To identify regions of interest and reference patterns in the ocular image yes ? chosen to use algorithms based on a specific configuration of feed-forward and multilevel neural networks, i.e. a convolutional neural network optimized for the adaptive processing of images of the eye that are able to perform of pre-processing and classification in a single flow with multiple maps, each intended for the detection and classification of a given characteristic. Each map ? equipped with a set of weights and bias.

The outputs of the units? of convolution are modeled with a non-linear sigmoid activation function, to then be sent as input for a sampling level more? inside the network. For every map in the convolution layer, is there a unit plane? in the internal sampling level. Each unit? of this internal sampling level calculates the average of the input values from a region, in the corresponding feature map. To reveal the presence or absence of the eye, a 2x2 region is enough. The calculated mean is then multiplied by a weight and a bias value is then added (figure 5).

For the first level of the convolutional neural network, the one relating to the detection of the presence of the eye, the neural network has two convolutional stages Cs1, Cs2 and two internal sampling levels S1, S2.

A first stage of convolution Cs1 ? aimed at the extraction of basic information from the input image, using a spatial resolution at maximum density? of pixels and differential analysis techniques with respect to the reference.

A second stage of Cs2 convolution learns pattern recognition as recurring spatial references, generating feature sets.

To extract different types of characteristics from the second convolution stage Cs2, a first internal sampling level S1 sends different input sets. In this way, you can? wait for the second internal sampling level S2 to learn to extract characteristics unrelated to what was done by the first level S1, and related to specific points of the image.

These latter characteristics can be used for classification. a third stage of convolution Cs3 ? what? a stage completely connected (fully-connected), where each unit? ? connected to all units? of a single corresponding map in the second internal sampling level S2.

Finally, all the units? of are totally connected cos? to generate two outputs with a softmax type activation function, with normalized exponential function:

k = 0 for the images with the eye present and k = 1 for the images with the absent eye. ?k ? the output activation parameter k

With wki and wk0 as network tuning parameters for output k and zi ? the?output for the?unit? hidden i in Cs3 and N the number of units? in level Cs3,

In the example of a convolutional neural network described here, used to identify the eye in the image, the architecture uses 2162 parameters distributed as follows.

Cs1 ? a convolution layer with 4 feature maps. Each unit? in each feature map ? connected to a 5x5 field around the input image (with dimensions of 38x38 pixels). The size of each feature map ? of 34x34 pixels, result of the convolution of the 5x5 kernel with the input image. Each map has 40 adaptive parameters (160 in total, in Cs1).

The internal sampling level S1 ? composed of 4 maps, one for each map in Cs1 (16x16 pixels). Each map has 2 adaptive parameters, 8 in total for S1. Cs2 ? a 15-map convolution layer that receives input from subsets of maps from S1. In the maps in S1, each unit? in each map ? connected to a 4x4 field at that location. In Cs2 every map ? from 18x18 pixels and the total parameters in Cs2 are 400.

S2 ? consisting of 15 8x8 pixel maps and 30 adaptive parameters are employed in this layer. Cs3 ? composed of 15 units? each fully-connected to all units? of a map in S2, with 1500 total adaptive parameters. All units? in Cs3 they are finally connected to generate 2 outputs, with 64 adaptive parameters in this final level.

For what? which concerns the training of the convolutional neural network for the identification of the eye, ? an experiment described below was carried out.

Fifteen people participated by collecting 100 photographs of each face (50 with eyes open and 50 with eyes closed). The 1500 original photographs have been converted to grayscale to minimize the variance related to chromatic effects and have been cropped to 36x36 pixels around the right and left eyes. For each image, 10 were produced by artificially varying parameters such as brightness, contrast and saturation. The 30,000 images labeled ?open eye? or ?closed eye? were provided as serial input to the dedicated convolutional neural network.

From a technical point of view, the training of the convolutional neural network consists in modifying the parameters w of the neural network through a fit of the input data with the activation function YK. By minimizing the cross-entropy loss function which measures the discrepancy between the function returned by the neural network for each value of w and the input training data.

With Yk network output, tk presence or absence of the eye (1,0), n = 1..N the input images.

The 1227 parameters employed in the convolutional neural network have been trained using the Stochastic Diagonal Levenberg Marquardt method, engaged in the estimation of the learning rate of each parameter, to optimize the speed? of convergence of the neural network. To predict abilities? of network generalization ? the early shutdown procedure was employed (W. S. Sarle, ?Stopped training and other remedies for overfitting,? in Proceedings of the 27th Symposium on the Interface of Computing, Science and Statistics, 1995, pp. 352?360).

The learning is finally interrupted when a minimum error is recognized in the validation of the dataset.

As regards the second macro-level of convolutional neural network, the general idea of the application ? the following: given the input images, obtain significant data from them automatically, i.e. without a user performing any type of intervention.

Therefore, the application collects sensor data and sends them to an online repository where they are processed by a convolutional neural network (CNN) which constitutes an artificial intelligence, with an architecture based on the U-Net type, for semantic segmentation of images of the ocular fundus or video content for the extrapolation of indications on the psyche and/or manifestations of physical pathologies (Face study).

Artificial intelligence (AI) assigns a definition and localization label to each pixel of each image.

Examples of pathologies identifiable by an eye scan in the visible spectrum are: uveitis, hypertension, multiple sclerosis, diabetic retinopathy, early detection of pathological leukocoria and distinction from physiological.

The workflow of this software basically consists of 4 steps:

1. Creation of pixel-label images for network training: each pixel is artificially labeled with the category it belongs to on the basis of correlation metrics obtained during the training phase with the acquired data;

2. Creation of the U-net neural network architecture, in which ? the coding/decoding structure necessary for the classification at the pixel level has been defined;

3. Setting of hyper-parameters, training and validation of the network through cross-validation;

4. Post-processing phase.

The adopted U-Net structure foresees a contraction path, useful for capturing the context in the image and characterized by the typical structure of a CNN, followed by a symmetrical expansion path to enable a precise localization (figure 3).

The activation maps obtained in the coding blocks before the under-sampling are stored and transferred to the corresponding decoding blocks in order to be able to concatenate them with the respective over-sampled maps through an up-convolution operation which, in this case, can? be obtained via a transposed convolution.

Does the evaluation system include a geo-positioning device for determining the positioning data of said plurality? of users.

This geo-positioning, made up of geographic coordinates, is associated with geographic coordinates which, instead of being relative to the user, are connected to a reference area or region.

In a parallel data flow, the information relating to these geographical coordinates ? made anonymous and is therefore associated with a set of spatial coordinates which does not directly refer to the user's position.

The georeferencing of the data is saved only after an elaboration of the coordinates aimed at guaranteeing anonymity considering problems such as the different density of population in the area. Once associated, the hardware device with the smartphone, the user starts the activity? of data disclosure. Here begins the phase of information spatial i, with the following expedients:

? the reference not ? the GPS coordinate well? that of the network (Wi-Fi or 3G cell); And

? a pair of pseudo-random values generated with Monte Carlo logic is superimposed on the X and Y coordinates, superimposed on the part of the data relating to the spatial resolution pi? high. The datum is then truncated to a certain decimal figure (e.g. the third decimal figure, corresponding to about 100 m).

Finally, does the evaluation system include a device for transmitting body data, digital images and geographical coordinates to said server, on which ? installed artificial intelligence software and a classifier, which processes this evaluation.

It is understood that this transmission device can also? be made by the user's smartphone.

Data processing will carried out by exploiting the hardware of a computer center optimized for processing image format data in parallel, thanks to the Graphics Processing Unit (GPU). Techniques of this type have been used to produce diagnostic algorithms for pathologies such as melanoma and diabetic retinopathy starting from the analysis of images collected in a medical context, generating results of accuracy comparable to those performed by qualified personnel.

Two types of user can be expected: the normal user and the superuser.

The normal user has display constraints aimed at showing only a neighborhood of his position. Access to the map? limited by fixed boundaries (maximum zoom-out for example to 1 Km<2>) and a certain level of detail (the data marker is a colored circular spot with a minimum diameter of 100 m).

The super user, or health administrations, civil protection, local health company, has the possibility? to move around the map without restrictions.

This evaluation system is proposed as an infrastructure that finds greater expression in a context extended to the community, promoting the individual as an active, responsible and interested component in the collective well-being, an ideal condition for improving personal experience as well. The single individual monitors some biometric parameters with the hardware device and at the same time uses the self-learning artificial intelligence system, to get more accurate feedback. advanced.

The outcome for the benefit of the community? comes from two types of activity:

? data collection by individuals to refine the AI response; And

? provide metrics to the healthcare system, based on interactive functional mapping.

To the evaluation system described above, a person skilled in the art, in order to satisfy further and contingent needs, will be able to make numerous further modifications and variations, all however included in the scope of protection of the present invention, as defined by the attached claims.

Claims (9)

1. Evaluation system of the territorial incidence of certain symptoms, including: ? a wearable device that takes measurements of body data of a plurality? of users associated with the system; ? a digital image detector, which allows you to detect biometric images of said plurality? of users; ? a geo-positioning device to determine the positioning data of said plurality? of users, and associates them with geographic coordinates of a reference area or region; ? a device for transmitting body data, digital images and geographic coordinates to a server, on which installed artificial intelligence software and a classifier, which processes this evaluation. The evaluation system according to claim 1, wherein the wearable device is designed to collect biometric data selected from heart rate, respiratory rate, blood pressure, body temperature and blood oxygen saturation. The evaluation system according to claim 1, wherein the wearable device is equipped with a microprocessor, a transmission module, a user interface, and a battery integrated into the wearable device. 4. Evaluation system according to claim 1, wherein said artificial intelligence software and said server form a convolutional neural network (CNN) with an architecture based on the U-Net type, arranged to carry out a semantic segmentation of images of the ocular fundus , assigning a definition and localization label to each pixel of each image. The evaluation system according to claim 4, wherein said artificial intelligence software performs the following steps: ? creation of pixel-label images for network training: each pixel is artificially labeled with the category it belongs to on the basis of correlation metrics obtained during the training phase with the acquired data; ? creation of? architecture of neural network U-net, in which? the coding/decoding structure necessary for the classification at the pixel level has been defined; ? setting of hyper-parameters, training and validation of the network through cross-validation; And ? post-processing. The evaluation system according to claim 5, wherein said post-processing provides a contraction path, which captures the context in the image, followed by a symmetrical expansion path to enable a precise localization. Evaluation system according to claim 6, wherein activation maps are obtained from coding blocks before an undersampling, which are stored and transferred to the corresponding decoding blocks in order to be able to concatenate them with the respective oversampled maps through a ?upconvolution operation using a transpose convolution. 8. Evaluation system according to claim 1, wherein said positioning data is processed in a parallel data stream, and is associated with a set of spatial coordinates which does not directly refer to the user's position. 9. Evaluation system according to claim 8, wherein said spatial coordinates are obtained using as reference the coordinate of the cell of the network used for data transmission, to which a pair of pseudo-random values generated with Monte Carlo logic is superimposed, superimposed to the part of the data relating to the spatial resolution pi? high, and the data is then truncated to a predetermined decimal place.