Personalization of deep neural network based models for processing health information

Abstract

A hybrid approach to providing a personalized health information prediction model is described. The personalized health information prediction model is a machine learning model having a hybrid architecture that advantageously incorporates both a personalized component and a universal component. The personalized component and the universal component are separable in a clear manner, which enables the components to be trained separately using different training data and using different computing devices. Particularly, the personalized component is trained solely using personal health data collected with respect to the particular person to which the model is personalized. Conversely, the universal component is trained using a corpus of health data collected with respect to a plurality of persons. Thus, the hybrid architecture has the advantages of both a population-based model and a fully personalized model, while mitigating the disadvantages of each.

Description

P39429-WO (2668-0006PCT) PERSONALIZATION OF DEEP NEURAL NETWORK BASED MODELS FOR PROCESSING HEALTH INFORMATIONFIELD

[0001] The devices and methods disclosed in this document relate to machine learning models and, more particularly, to personalization of deep neural network-based models for processing health information.BACKGROUND

[0002] Unless otherwise indicated herein, the materials described in this section are not admitted to be the prior art by inclusion in this section.

[0003] Deep neural network-based models are often used for classification or regression in health care use cases. Such models are typically population-based models that are trained with data collected from a representative population of patients in a central server. This population-based approach has the benefit of a large amount of available training data. However, this representative population-based approach fails to perform well when there are significant differences between individuals due to ethnicity, age, gender, etc. In such cases, it is possible for a personalized model to perform significantly better than a population-based model.

[0004] However, developing personalized models presents significant challenges. Particularly, it may be challenging to collect an adequate amount of personal health data to train a fully personalized model that achieves adequate performance. Additionally, it may not be practical to train a fully personalized model on an edge device or in a manner that otherwise complies with data privacy laws or regulations.

[0005] What is needed is an approach for training and deploying a model for processing health information that has the benefits of both personalized and population-based approaches.P39429-WO (2668-0006PCT) SUMMARY

[0006] A method for providing health information is disclosed. The method comprises receiving, with an electronic device, third health information of a first person. The method further comprises determining, with the electronic device, fourth health information of the person based on the third health information using a first neural network model. The first neural network model has a first subnetwork and a second subnetwork. The first subnetwork has been trained solely using first health information of the first person. The second subnetwork has been trained using second health information of a plurality of second persons. The method further comprises providing at least one of (i) a visualization of the fourth health information on a display of the electronic device, (ii) a perceptible warning of a possible future health status via the electronic device, the possible future health status being indicated by the fourth health information, or (iii) a message including the fourth health information that is transmitted to a remote computing device.

[0007] In one embodiment, the method further comprises, prior to determining the fourth health information, receiving pretrained parameters of the second subnetwork of the first neural network model. In one embodiment, the method further comprises, prior to determining the fourth health information, receiving the first health information of the first person. The first health information includes a plurality of first training samples. In one embodiment, the method further comprises, prior to determining the fourth health information, training the first neural network model using the first health information of the first person, the pretrained parameters of the second subnetwork being frozen as parameters of the first subnetwork are learned during the training of the first neural network model.

[0008] In one embodiment, the receiving the first health information further comprises measuring, with a sensor of the electronic device, at least some of the first health information of the first person.

[0009] In one embodiment, the training the first neural network model further comprises training the first neural network model with the electronic device.

[0010] In one embodiment, the method further comprises receiving the second health information of the plurality of second persons, the second health information including a plurality of second training samples. In one embodiment, the method further comprisesP39429-WO (2668-0006PCT) training a second neural network model using the second health information of the plurality of second persons. The second neural network model includes a third subnetwork and the second subnetwork. The pretrained parameters of the second subnetwork are learned during the training of the second neural network model.

[0011] In one embodiment, the training the second neural network model further comprises training the second neural network model with a server that is remote from the electronic device.

[0012] In one embodiment, the training the second neural network model further comprises training the second neural network model with a plurality of computing devices that are remote from the electronic device, using federated learning.

[0013] In one embodiment, the fourth health information is one-hot encoded to associate each second training sample in the plurality of second training samples with a respective second person of the plurality of second persons.

[0014] In one embodiment, the first subnetwork of the first neural network model has a first architecture. In one embodiment, the third subnetwork of the second neural network model includes a plurality of instances of at least a portion of the first architecture. The third subnetwork is configured such that each respective instance in the third subnetwork is activated only in response to second training samples in the plurality of second training samples that are one-hot encoded in association with a respective second person of the plurality of second persons.

[0015] In one embodiment, the plurality of instances of at least the portion of the first architecture include a respective instance associated with each respective second person of the plurality of second persons.

[0016] In one embodiment, the training the first neural network model further comprises initializing the parameters of the first subnetwork in the first neural network model depending on parameters of the third subnetwork learned during the training of the second neural network model.

[0017] In one embodiment, the training the first neural network model further comprises randomly initializing the parameters of the first subnetwork.

[0018] In one embodiment, the first subnetwork receives the third health information and is configured to transform the third health information into an input space of the secondP39429-WO (2668-0006PCT) subnetwork. In one embodiment, the second subnetwork in the first neural network model is configured to receive the transformed third health information from the first subnetwork and to determine the fourth health information.

[0019] In one embodiment, the second subnetwork in the first neural network model receives the third health information and is configured to determine the fourth health information in an output space of the second subnetwork. In one embodiment, the first subnetwork receives the fourth health information in the output space of the second subnetwork and is configured to transform the fourth health information into an output space of the first neural network model.

[0020] In one embodiment, the fourth health information is a regression of at least one health metric based on the third health information.

[0021] In one embodiment, the fourth health information is a classification of the third health information.

[0022] In one embodiment, the third health information includes historical blood glucose measurements of the first person. In one embodiment, the fourth health information includes at least one of (i) predicted future blood glucose measurements of the first person, (ii) a prediction of a future hypoglycemic status, (iii) a prediction of a future hyperglycemic status, or (iv) a predicted risk for diabetic neuropathy.

[0023] In one embodiment, the third health information further includes at least one of (i) insulin intake data indicating insulin doses taken by the first person or (ii) carbohydrate consumption data indicating carbohydrates consumed by the first person.

[0024] A method for training a machine learning model configured to provide health information is disclosed. The method comprises receiving first health information of a first person. The first health information includes a plurality of first training samples. The method further comprises receiving second health information of a plurality of second persons. The second health information includes a plurality of second training samples. The method further comprises training a second neural network model using the second health information of the plurality of second persons. The second neural network model includes a third subnetwork and a second subnetwork. The method further comprises training a first neural network model using the first health information of the first person. The first neural network model has a first subnetwork and the second subnetwork. TheP39429-WO (2668-0006PCT) parameters of the second subnetwork that were learned during the training of the second neural network model arc frozen during the training of the first neural network model. The first neural network model is configured to receive third health information of a first person and determine fourth health information of the person based on the third health information.

[0025] A non-transitory computer-readable medium that stores program instructions for providing health information is disclosed. The program instructions are configured to, when executed by a processor, cause the processor to receive third health information of a first person. The program instructions are configured to, when executed by the processor, cause the processor to determine fourth health information of the person based on the third health information using a first neural network model. The first neural network model has a first subnetwork and a second subnetwork. The first subnetwork has been trained solely using first health information of the first person. The second subnetwork has been trained using second health information of a plurality of second persons. The program instructions are configured to, when executed by the processor, cause the processor to provide at least one of (i) a visualization of the fourth health information on a display operably connected to the processor, (ii) a perceptible warning of a possible future health status via an output device operably connected to the processor, the possible future health status being indicated by the fourth health information, or (iii) a message including the fourth health information that is transmitted to a remote computing device via a transceiver operably connected to the processor.BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The foregoing aspects and other features of the system and methods are explained in the following description, taken in connection with the accompanying drawings.

[0027] FIG. 1 summarizes a hybrid approach to providing a personalized health information prediction model.

[0028] FIG. 2 shows an exemplary health management system, which is used by a person and their healthcare provider to manage their health.P39429-WO (2668-0006PCT)

[0029] FIG. 3 shows possible configurations of the personalized health information prediction model.

[0030] FIG. 4 shows simplified exemplary neural networks for implementing the personalized health information prediction model in deployment and training.

[0031] FIG. 5 shows a flow diagram for a method for training a personalized health information prediction model.

[0032] FIG. 6 shows a flow diagram for a process for training the universal model component in a distributed manner using federated learning.

[0033] FIG. 7 shows a flow diagram for a method for providing health information.

[0034] FIG. 8 shows a portion of an exemplary graphical user interface including a data plot that visualizes both historical and predicted blood glucose measurements.

[0035] FIG. 9 shows a portion of a graphical user interface including a visual alert.DETAILED DESCRIPTION

[0036] For the purposes of promoting an understanding of the principles of the disclosure, reference will now be made to the embodiments illustrated in the drawings and described in the following written specification. It is understood that no limitation to the scope of the disclosure is thereby intended. It is further understood that the present disclosure includes any alterations and modifications to the illustrated embodiments and includes further applications of the principles of the disclosure as would normally occur to one skilled in the art to which this disclosure pertains.Overview

[0037] FIG. 1 summarizes a hybrid approach to providing a personalized health information prediction model 110. The personalized health information prediction model 110 is a machine learning model having a hybrid architecture that advantageously incorporates both a personalized component 120 and a universal component 130.

[0038] As discussed above, a population-based model 140 would typically be trained using data collected from a representative population of patients or users (e.g., all of the persons #1-6). The population-based model 140 has the advantage of being trained with a larger dataset but may have poor performance when there are significant differencesP39429-WO (2668-0006PCT) between individuals due to ethnicity, age, gender, etc. However, the population-based model 140 also has the lowest data needs from each individual. Instead, the necessary training data volume is achieved by assembling data from a plurality of individuals. Thus, the population-based model 140 requires the least effort to develop and, similarly, has the lowest burden for regulatory approval. Training and validation are performed in-house and the model is submitted with the regulatory filing.

[0039] In contrast, a fully personalized model 150 would typically be trained using a dataset that is specific to the individual person (e.g., a particular one of the persons #1-6). The fully personalized model 150 can enable superior performance when applied to problems in which there are significant differences between individuals due to ethnicity, age, gender, etc., but has the disadvantage of being trained on a smaller dataset. In this way, to achieve good performance, the fully personalized models 150 have the highest need for data from each individual. Thus, fully personalized models 150 require the highest amount of research and development effort, since each individual's model has to be developed separately. Moreover, the fully personalized modelsl50 have the highest burden for regulatory approval. It is necessary to provide assurance that each individual fully personalized model 150 is safe to the consumer.

[0040] The hybrid architecture of the personalized health information prediction model 110 has the advantages of both the population-based model 140 and the fully personalized model 150, while mitigating the disadvantages of each. Using the hybrid architectures described herein, it should be appreciated that a spectrum of models can be constructed, ranging from population models to fully personalized models. Hybrid model architectures advantageously provide optimal performance in special cases where the personalized component 120 and the universal component 130 can be decoupled. Hybrid model architectures have intermediate data needs and can even have low data needs if the universal component 130 makes up the bulk of the model. Thus, hybrid architectures have a simplified development effort in which the bulk of the model development is performed in-house, and the balance is performed in the field. Moreover, hybrid architectures have only an intermediate burden for regulatory approval and can even have a low burden for regulatory approval if the universal component 130 makes up the bulk of the model.P39429-WO (2668-0006PCT)

[0041] Due to the decoupled architecture, the personalized component 120 and the universal component 130 can advantageously be trained separately using different training data and using different computing devices. Particularly, the hybrid approach of the disclosure advantageously adopts a two-stage training process.

[0042] In a first training phase, the universal component 130 is trained using a corpus of health data collected with respect to a plurality of persons (e.g., all of the persons #1-6), and thereby has the advantages of a large training dataset. In some embodiments, the universal component 130 is advantageously trained by a central server or in a distributed manner using federated learning. In some embodiments, the universal component 130 is trained jointly with a placeholder component (not shown) which employs one-hot encoding of the training samples associated with each particular person in the population-based corpus of health data. As a result of the separate training of the universal component 130, the universal component 130 can be deployed in a shared and portable manner for the purpose of building personalized health information prediction models 110 to any number of end-users.

[0043] In a second training phase, the parameters of the universal component 130 are frozen and the personalized component 120 is trained solely using personal health data collected with respect to the particular person to which the model 110 is personalized. In some embodiments, the personalized component 120 is advantageously trained locally by an edge device in the possession of the person for which the model 110 is being personalized. In this way, the deployed model 110 leverages the benefits of a populationbased model, but is also personalized to a particular person. Moreover, the personalization of the deployed model 110 is achievable without exposing or sharing personal health information of the person to any devices other than their own devices.Exemplary Health Management System

[0044] FIG. 2 shows an exemplary health management system 200, which is used by a person and their healthcare provider to manage their health. The health management system 200 includes at least one measurement device 210, a computing device 250, and a remote server 270. In some embodiments, the health management system 200 also includes an administration device 230. It should be appreciated, however, that the components of theP39429-WO (2668-0006PCT) health management system 200 shown and described are merely exemplary and that the health management system 200 may comprise any alternative configuration.

[0045] The measurement device(s) 210 include at least one sensor 212, a memory device 214, and a transceiver 216, each operably connected to a processor 218. The measurement device(s) 210 are configured to measure a variety of health data 228 of a person 220. Health data may be in a raw measured form or in a processed form. Health data may be automatically measured, sensed, or collected by the one or more sensors 212 of the measurement device(s) 210, but may also be entered manually by the user via the measurement device 210 and / or the computing device 250. Examples of such health data may include physical activity data, acceleration data, gyroscopic data, heart rate data, blood pressure data, blood oxygenation data, sleep data, or body temperature data.

[0046] In at least some embodiments, one or more of the measurement device(s) 210 are body-wom devices. For example, in one embodiment, a measurement device 210 is in the form of a so-called “smart” watch which is worn on the user’s wrist. However, the measurement device(s) 210 may also be designed and dimensioned to be worn elsewhere by the user, for example, on his or her waist, arm, belly, or ankle. In some embodiments, the measurement device(s) 210 may include a 3-axis accelerometer, altimeter, and / or GPS receiver configured to detect the motions of the wearer during running or walking, or while sleeping for the purpose of sleep tracking. Additionally, the measurement device(s) 210 may include an optical sensor configured to detect heart rate, blood pressure, and / or blood oxygenation.

[0047] However, the illustrated measurement device 210 of FIG. 2 is in the form of a body-worn continuous glucose monitor (“CGM”) that is used to generate blood glucose measurements that correspond to a person’s blood glucose concentration. In the illustrated embodiment, a sensor 212 is mounted on the skin 222 of the person 220 with an adhesive and includes a probe 224 that is positioned just under the skin 222. The probe 224 is in contact with interstitial fluid 226 of the person 220. In one embodiment, the probe 224 is an enzyme-based amperometric biosensor that is configured to measure blood glucose measurements in the interstitial fluid 226. In other embodiments, the sensor 212 measures glucose concentrations according to other suitable structural configurations andP39429-WO (2668-0006PCT) methodologies. The measurement device 210 may operate with or without a corresponding insulin pump (shown as an embodiment of the administration device 230, for example).

[0048] The processor 218 of the measurement device 210 is configured to execute instructions to operate the measurement device 210 to enable the features, functionality, characteristics, and / or the like as described herein. The processor 218 generally comprises one or more processors which may operate in parallel or otherwise in concert with one another. It will be recognized by those of ordinary skill in the art that the term “processor” as used herein includes any hardware system, hardware mechanism, or hardware component that processes data, signals, or other information. Accordingly, the processor 218 may include a system with a central processing unit, graphics processing units, multiple processing units, dedicated circuitry for achieving functionality, programmable logic, or other processing systems.

[0049] The memory device 214 of the measurement device 210 is configured to store data and program instructions that, when executed by the processor 218, enable the measurement device 210 to perform various operations described herein. The memory device 214 may be any type of electronic device capable of storing information accessible by the processor 218, such as a memory card, read only memory (“ROM”), random access memory (“RAM”), a hard drive, a solid-state drive, a disc, flash memory, or any of various other computer-readable media serving as data storage devices, as will be recognized by those of ordinary skill in the art. The memory device 214 is also referred to herein as a non-transitory computer readable medium, a non-transitory memory device, and a non-transitory memory. The memory device 214 stores the health data 228 generated by the processor 218 and as measured by the sensor 212. Additionally, in some embodiments (not shown), the memory device 252 is configured to store the personalized health information prediction model 110, which incorporates both the universal component 130 and the personalized component 120.

[0050] The transceiver 216 of the measurement device 210, in one embodiment, is configured for the wired and / or wireless exchange of data with the computing device 250. The transceiver 216 includes one or more modems, processors, memories, oscillators, antennas, or other hardware conventionally included in a communications module to enable electronic communications with various other devices. For example, the transceiverP39429-WO (2668-0006PCT) 216 may exchange electronic data using a wireless local area network (“Wi-Fi”), a personal area network, Bluetooth®, ncar-ficld communication (“NFC”), ultra-wide band (“UWB”), a cellular network, and / or any other wireless network protocol. Accordingly, the transceiver 216 is compatible with any desired wireless communication standard or protocol including, but not limited to, IEEE 802.11, IEEE 802.15.1 (“Bluetooth®”), Global System for Mobiles (“GSM”), and Code Division Multiple Access (“CDMA”). In one embodiment, the transceiver 216 operably connects the measurement device 210 to the Internet for data exchange with any other Internet-connected device. In another embodiment, the transceiver 216 transmits and receives data directly to the computing device 250 without being connected to the Internet 290. The transceiver 216 is also referred to herein as a network adapter, a network device, and / or a network communication module.

[0051] With continued reference to FIG. 2, the administration device 230 of the health management system 200 is a medicament delivery device. In the illustrated embodiment, the administration device 230 is in particular an insulin bolus delivery device. However, other administration devices 230 may be incorporated into other embodiments. In the illustrated embodiment, the administration device 230 is operably connected to the measurement device 210 and the computing device 250 to receive insulin bolus administration data, for example. In some embodiments, the administration device 230 is also operably connected to the Internet 290 via a corresponding transceiver (not shown). Thus, the administration device 230 is operably connected to other Internet-connected devices, such as the remote server 270 and / or a computer (not shown) used by a doctor or a pharmacist. The administration device 230 is connected or is connectable to the person 220 to deliver the medicament 232.

[0052] In an exemplary embodiment, the administration device 230 is a smart insulin pump, and the medicament 232 is insulin, a rapid-acting insulin analog, or a regular insulin analog, each of which is referred to interchangeably herein as “insulin.” The insulin pump delivers the medicament 232 to the person 220 through a thin tube (cannula, not shown) that goes under the person’s skin. In another embodiment, the administration device 230 is a smart insulin “pen.” The smart insulin pen delivers the medicament 232 to the person 220 through a reusable injection device (i.e., a needle).P39429-WO (2668-0006PCT)

[0053] The administration device 230 is electronically configurable to deliver a predetermined dosage (i.c., a bolus dose) of the medicament 232 to the person 220 based on insulin bolus administration data received from the computing device 250, for example. The medicament 232 is capable of reducing the blood glucose concentration level of the person 220, and, therefore, is provided to the person 220 when a hyperglycemic condition is present or is predicted in order to assist the person 220 in returning their glucose concentration level to a lower desired level.

[0054] With continued reference to FIG. 2, the computing device 250 of the health management system 200 includes a memory device 252, a transceiver 254, an input device 256, and a display screen 258 each operably connected to a processor 260. The computing device 250 is described and illustrated herein as a smartphone. It will be appreciated that the illustrated embodiment of the computing device 250 is only one exemplary embodiment and is merely representative of any of various manners, configurations, or combinations of a personal computer, a desktop computer, a laptop computer, a smartwatch, a mobile phone, a tablet computer, or any other computing device that is operative in the manner set forth herein.

[0055] The processor 260 of the computing device 250 is configured to execute instructions to operate the computing device 250 to enable the features, functionality, characteristics, and / or the like as described herein. The processor 260 generally comprises one or more processors which may operate in parallel or otherwise in concert with one another. The processor 260 may include a system with a central processing unit, graphics processing units, multiple processing units, dedicated circuitry for achieving functionality, programmable logic, or other processing systems. The processor 260 is configured to run applications (i.e., “apps”), at least including a health management application 262.

[0056] The memory device 252 of the computing device 250 is configured to store data and program instructions that, when executed by the processor 260, enable the computing device 250 to perform various operations and methods described herein. The memory device 252 may be any type of electronic device capable of storing information accessible by the processor 260, such as a memory card, ROM, RAM, a hard drive, a solid-state drive, a disc, flash memory, or any of various other computer-readable media serving as data storage devices, as will be recognized by those of ordinary skill in the art. The memoryP39429-WO (2668-0006PCT) device 252 is also referred to herein as a non-transitory computer readable medium, a non-transitory memory device, and a non-transitory memory. The memory device 252 is configured to store health data 228 measured or otherwise collected by the health management system 100. Additionally, the memory device 252 is configured to store the personalized health information prediction model 110, which incorporates both the personalized component 120 and the universal component 130.

[0057] The transceiver 254 of the computing device 250 is configured for the wired and / or wireless exchange of data with the measurement device 210, the administration device 230, the remote server 270, and the Internet 290. The transceiver 254 includes one or more modems, processors, memories, oscillators, antennas, or other hardware conventionally included in a communications module to enable electronic communications with various other devices. For example, the transceiver 254 may exchange data using WiFi, a personal area network, Bluetooth®, NFC, UWB, a cellular network, and / or any other wireless network protocol. Accordingly, the transceiver 254 is compatible with any desired wireless communication standard or protocol including, but not limited to, IEEE 802.11, Bluetooth®”, GSM, and CDMA. The transceiver 254 operably connects the computing device 250 to the Internet 290 for data exchange with any other Internet-connected device. Additionally, the transceiver 254 transmits and receives data from the measurement device 210 either directly or indirectly. The transceiver 254 is also referred to herein as a network adapter and / or a network device.

[0058] The display screen 258 of the computing device 250 is configured to render and to display a graphical user interface including text, images, and other user sensible outputs and visually comprehensible data including, but not limited to, a visualization of the health data 228 and a visualization of predicted future health data. Additionally, graphical user interfaces of the display screen 258 may include one or more warnings, alerts, notifications, or messages relating to a current or predicted health status, such as a hypoglycemic condition or hyperglycemic condition, of the person 220. The display screen 258 may comprise any of various known types of displays, such as liquid crystal displays (“LCD”) or organic light emitting diode (“OLED”) screens.

[0059] In at least some embodiments, the input device 256 of the computing device 250 includes a touchscreen applied over the display screen 258 that is configured toP39429-WO (2668-0006PCT) generate user input data in response to the touch of a finger or a stylus. The input device 256 may also include at least one button, switch, keyboard, and / or keypad that is configured to generate user input data when touched or moved by a user. Additionally, or alternatively, the input device 256 includes a microphone configured to generate user input data in response to sounds, such as the voice of a user of the computing device 250. In yet another embodiment, the input device 256 is any device configured to generate user input data, as recognized by those of ordinary skill in the art.

[0060] With continued reference to FIG. 2, the remote server 270 of the health management system 200 includes a transceiver 272 and a memory device 274 operably connected to a processor 276. The processor 276 is configured to execute instructions to operate the remote server 270 to enable the features, functionality, characteristics and / or the like as described herein. The processor 276 generally comprises one or more processors which may operate in parallel or otherwise in concert with one another. The processor 276 may include a system with a central processing unit, graphics processing units, multiple processing units, dedicated circuitry for achieving functionality, programmable logic, or other processing systems.

[0061] The transceiver 272, in one embodiment, is configured for the wired and / or wireless exchange of data with the computing device 250 and the Internet 290. The transceiver 272 includes one or more modems, processors, memories, oscillators, antennas, or other hardware conventionally included in a communications module to enable electronic communications with various other devices. For example, the transceiver 272 may exchange data using Wi-Fi, a personal area network, Bluetooth®, NFC, UWB, a cellular network, and / or any other wireless network protocol. Accordingly, the transceiver 272 is compatible with any desired wireless communication standard or protocol including, but not limited to IEEE 802.11, Bluetooth®, GSM, and CDMA. The transceiver 272 is also referred to herein as a network adapter and / or a network device.

[0062] The memory device 274 is configured to store data and program instructions that, when executed by the processor 276, enable the remote server 270 to perform various operations and methods described herein. The memory device 274 may be of any type of electronic device capable of storing information accessible by the processor 276, such as a memory card, ROM, RAM, hard drives, solid-state drives, discs, flash memory, or any ofP39429-WO (2668-0006PCT) various other computer- readable medium serving as data storage devices, as will be recognized by those of ordinary skill in the art. The memory device 274 is also referred to herein as a non-transitory computer readable medium, a non-transitory memory device, and a non-transitory memory. The memory device 274 at least stores the universal component 130 that is provided to the computing device 250 and / or the measurement device 210 and used to build the personalized health information prediction model 110. Additionally, in at least some embodiments, the memory device 274 stores the personal health data of a plurality of persons (not shown), which is used to train the universal component 130. The plurality of persons may, for example, be persons who have provided the necessary authorizations to use their health data to train the universal component 130 of the model 110.Model Architecture

[0063] As mentioned above, the personalized health information prediction model 110 is a machine learning model having both a personalized component 120 and a universal component 130. In particular, in at least some embodiments, the personalized health information prediction model 110 is a neural network having an architecture that includes a personalized subnetwork 310 corresponding to the personalized component 120 and a universal subnetwork 320 corresponding to the universal component 130.

[0064] In at least some embodiments, the universal subnetwork 320 is trained using a corpus of health data collected with respect to a plurality of persons. In some embodiments, the plurality of persons may be persons who have provided the necessary authorizations to use their health data to train the model 110. Accordingly, in some embodiments, the universal subnetwork 320 can be trained by a centralized data processing system, such as the remote server 270. Alternatively, in some embodiments, the universal subnetwork 320 can be trained in a distributed manner by a plurality of computing devices using federated learning.

[0065] In contrast, in at least some embodiments, the personalized subnetwork 310 is trained solely using personal health data collected with respect to the particular person 220 to which the model 110 is personalized. Additionally, in at least some embodiments, the personalized subnetwork 310 is advantageously trained locally by an edge device, such asP39429-WO (2668-0006PCT) the computing device 250 or the measurement device 210, rather than by a centralized data processing system, such as the remote server 270. In this way, the model 110 is personalized without exposing or sharing personal health information of the person 220 to which the model 110 is personalized.

[0066] The personalized health information prediction model 110 is configured to receive personal health data collected with respect to the particular person 220 and output further health data regarding the particular person 220. The personal health data input to the model 110 includes any measurable or otherwise determined health metric relating to a health or wellbeing of the person 220. In some embodiments, the personal health data is in the form of time series data in which each measurement is associated with a respective timestamp indicating a time at which the value was measured. In some embodiments, the personal health data input to the model 110 is limited to a predetermined time window (e.g., the last 30 minutes). In some embodiments, the further health data output by the model 110 include a classification of the personal health data collected with respect to the person 220 (e.g., a prediction of a future health status or condition). In some embodiments, the further health data output by the model 110 include a regression of values for at least one health metric that is determined based on the input personal health data. In one embodiment, the further health data output by the model 110 includes a regression of future values for at least one health metric during a predetermined prediction time window (e.g., 30 minutes into the future).

[0067] It should be appreciated that one of the many possible applications of the health management system 200 is, in particular, to diabetes or blood glucose management for a person with diabetes. In such an application, the personal health data received by the model 110 may include, for example, historical blood glucose measurements, insulin intake data, and carbohydrate consumption data.

[0068] In some embodiments, the historical blood glucose measurements may take the form of a time series of blood glucose concentration measurements, for example having the units ‘mg / dL’ (milligrams per deciliter). Each blood glucose concentration measurement is associated with a respective timestamp indicating a time at which the value was measured.P39429-WO (2668-0006PCT)

[0069] In some embodiments, the insulin intake data may take the form of a time series of insulin bolus dosage amounts (c.g., of the medicament 232), as previously received by the person 220, with associated timestamps. The insulin bolus dosage data tracks both correction bolus and meal bolus received by the person 220. Additionally, the insulin intake data may be smoothed to provide an insulin-on-board time series.

[0070] In some embodiments, the carbohydrate consumption data may take the form of a time series of amounts of carbohydrates (“carbs”) consumed by the person 220 in each meal and snack, with associated timestamps. In at least some embodiments, each amount of carbohydrates in the carbohydrate intake data is classified or otherwise labeled with a particular carbohydrate type. Exemplary carbohydrate types may include simple carbohydrates, complex carbohydrates, and other carbohydrates. It will be appreciated that such carbohydrate type classifications generally reflect different rates at which the consumed carbohydrates are absorbed by the human body. Additionally, the carbohydrate consumption data may be smoothed to provide a carbohydrates-on-board time series.

[0071] In diabetes or blood glucose management applications, the further health data that is output by the model 110 may include, for example, predicted future blood glucose measurements for a prediction time window (PTW), a prediction of a future hypoglycemic status, a prediction of a future hyperglycemic status, or a predicted risk for diabetic neuropathy.

[0072] It should be appreciated, however, that the health management system 200 may be used for a wide variety of other health management scenarios. Accordingly, the personal health data input to the model 110 may include a variety of other health data such as physical activity data, acceleration data, gyroscopic data, heart rate data, blood pressure data, blood oxygenation data, sleep data, body temperature data, or electronic medical records. Likewise, the further health data that is output by the model 110 may include a variety of other health data including such as predicted risk for heart disease, chronic kidney disease, and the like, categorization of the person 220 into ‘buckets’ (for example that categorize patients depending on overall health risk), or predicted future values of any health metric.

[0073] FIG. 3 shows possible configurations of the personalized health information prediction model 110. As mentioned above, the neural network architecture of theP39429-WO (2668-0006PCT) personalized health information prediction model 110 includes a personalized subnetwork 310 and a universal subnetwork 320. The arrangement of the personalized subnetwork 310 and the universal subnetwork 320 can have at least two configurations referred to herein as “Type-1” and “Type-2” configurations.

[0074] A “Type 1” configuration 300A of the model 110 is one in which one or more neural network layers at or near an input layer 330 are personalized for the particular person 220. In other words, the personalized subnetwork 310 precedes the universal subnetwork 320 in the architecture. In a “Type 1” configuration, the personalized subnetwork 310 receives the personal health data collected with respect to the person 220 in a ‘personalized’ or ‘strata- specific’ input data space of the model 110 and is configured to transform the personal health data into a ‘common’ or ‘universal’ input data space of the universal subnetwork 320. The universal subnetwork 320 is configured to receive the response from the personalized subnetwork 310 (i.e., the transformed personal health data) and to determine the further health data regarding the person 220 in the ‘personalized’ or ‘strata-specific’ output data space of the model 110 by applying a universal transformation.

[0075] A “Type 2” configuration 300B of the model 110 is one in which one or more neural network layers at or near an output layer 340 are personalized for the particular person 220. In other words, the personalized subnetwork 310 follows the universal subnetwork 320 in the architecture. In a “Type 2” configuration, the universal subnetwork 320 receives the personal health data collected with respect to the person 220 in the ‘personalized’ or ‘strata- specific’ input data space of the model 110 and is configured to determine further health data regarding the person 220 in the ‘common’ or ‘universal’ output space of the universal subnetwork 320 by applying a universal transformation. The personalized subnetwork 310 receives the response from the universal subnetwork 320 and is configured to transform the response into the ‘personalized’ or ‘strata- specific’ output data space of the model 110 so as to provide the further health data that is finally output by the model 110.

[0076] It should be appreciated that further configurations are possible. Particularly, in some embodiments, the architecture of the personalized health information prediction model 110 may include multiple personalized subnetworks 310 and / or multiple universal subnetworks 320. As such, the subnetworks can be organized into any arbitraryP39429-WO (2668-0006PCT) configuration. For example, in one embodiment, a personalized subnetwork 310 may be sandwiched between two distinct universal subnetworks 320. Likewise, in another embodiment, a universal subnetwork 320 may be sandwiched between two distinct personalized subnetworks 310.

[0077] FIG. 4 shows simplified exemplary neural networks for implementing the personalized health information prediction model 110 in deployment and training. On the right-hand side, a deployed neural network 400 is illustrated. The deployed neural network 400 represents one of many possible implementations of the personalized health information prediction model 110, which may have any of a variety of alternative architectures. In contrast, on the left-hand side, a training neural network 450 is illustrated. The training neural network 450 is similar to the deployed neural network 400 and directly incorporates the same universal subnetwork 320. However, rather than the personalized subnetwork 310, the training neural network 450 incorporates a placeholder subnetwork 460 that enables the universal subnetwork 320 to be pretrained prior to deployment and prior to training of the personalized subnetwork 310.

[0078] The deployed neural network 400 is shown with a simplified architecture in a “Type-1” configuration. In the simplified architecture of the deployed neural network 400, the personalized subnetwork 310 has an input layer (Input) that receives a respective input feature. The input layer (Input) is connected to a fully connected layer (FC1). Finally, rectified linear unit activation (relu_l) is applied to the output of the fully connected layer (FC1).

[0079] Additionally, in the simplified architecture of the deployed neural network 400, the universal subnetwork 320 has a fully connected layer (FC2) that receives the output from the personalized subnetwork 310. Rectified linear unit activation (relu_2) is applied to the output of the fully connected layer (FC2). A fully connected layer (FC3) is connected to the rectified linear unit activation (relu_2). Finally, the fully connected layer (FC3) is connected to a SoftMax output layer (SMI).

[0080] As mentioned above, the universal subnetwork 320 is trained using a corpus of health data collected with respect to a plurality of persons. Before describing the simplified architecture of the training neural network 450 that is used to perform this training, it should be understood that, in at least some embodiments, the corpus of health data that is used toP39429-WO (2668-0006PCT) train the universal subnetwork 320 is one-hot encoded to associate each training sample in the corpus of health data with a respective person of the plurality of persons.

[0081] The corpus of health data includes a plurality of training samples that include personal health data of the plurality of persons. The training samples take the form of tabular data organized in rows and columns. Each row corresponds to a respective training sample. The columns include columns corresponding to a plurality of respective input features and further columns corresponding to a plurality of respective ‘ground truth’ output features. Additionally, the columns include columns that one-hot encode the identity of the respective person whose personal health data is included in each training sample. Particularly, for each respective person of the plurality of persons whose personal health data is included in the corpus of health data, the training samples include a column having a binary value ‘1’ or ‘TRUE’ if the training sample is associated with the respective person or a binary value ‘0’ or ‘FALSE if the training sample is not associated with the respective person. Thus, the corpus of health data includes a number of one-hot encoding columns equal to a number of persons whose personal health data is included in the corpus of health data.

[0082] In contrast with the deployed neural network 400, the training neural network 450 incorporates the placeholder subnetwork 460 rather than the personalized subnetwork 310. In some embodiments, the placeholder subnetwork 460 has an architecture that is more complex than, but directly related to that of the personalized subnetwork 310. Particularly, in at least one embodiment, the architecture of the placeholder subnetwork 460 has a plurality of instances of at least a portion of the architecture of the personalized subnetwork 310. Moreover, the placeholder subnetwork 460 is configured such that each respective instance of the portion of the architecture is activated only responsive to training samples that are one-hot encoded in association with a respective person of the plurality of persons. In other words, for each respective person whose personal health data is included in the corpus of health data, the placeholder subnetwork 460 includes a respective instance of at least a portion of the architecture of the personalized subnetwork 310 that is activated only responsive to training samples including the respective person’s personal health data.

[0083] With continued reference to FIG. 4, the training neural network 450 is shown with a simplified architecture that is analogous to that of deployed neural network 400. InP39429-WO (2668-0006PCT) the simplified architecture of the training neural network 450, the placeholder subnetwork 460 has an input layer (Input), which receives a respective input feature, which is equivalent to that of the personalized subnetwork 310. Additionally, the placeholder subnetwork 460 has four one-hot encoding inputs (S_l, S_2, S_3, and S_4), that receive hot-encoding values from four columns in the corpus of health data corresponding respectively to four different persons.

[0084] The placeholder subnetwork 460 has four instances of an architecture that corresponds to a portion of the personalized subnetwork 310. Particularly, for each of the four different persons, the placeholder subnetwork 460 has a respective fully connected layer that is similar to the fully connected layer (FC1) in the personalized subnetwork 310 and a rectified linear unit activation that is similar to the rectified linear unit activation (relu_l) in the personalized subnetwork 310. A first instance associated with a first person includes a fully connected layer (FC1), with rectified linear unit activation (relu_l) applied to the output of the fully connected layer (FC 1). A second instance associated with a second person includes a fully connected layer (FC2), with rectified linear unit activation (relu_2) applied to the output of the fully connected layer (FC2). A third instance associated with a third person includes a fully connected layer (FC3), with rectified linear unit activation (relu_3) applied to the output of the fully connected layer (FC3). Finally, a fourth instance associated with a fourth person includes a fully connected layer (FC4) and rectified linear unit activation (relu_4) applied to the output of the fully connected layer (FC4). The input layer (Input) is connected to each of these connected layers (FC1, FC2, FC3, and FC4).

[0085] In the placeholder subnetwork 460, multiplication layers multiply the outputs of the rectified linear unit activation in each instance by a corresponding respective one-hot encoding input, such that each instance is activated only responsive to training samples including the respective person’s personal health data. Particularly, a multiplication layer (mult_l) multiplies the output of the rectified linear unit activation (relu_l) with the one-hot encoding input (S_l), such that the fully connected layer (FC1) is activated in response to the one-hot encoding of training samples associated with the first person. Similarly, a multiplication layer (mult_2) multiplies the output of the rectified linear unit activation (relu_2) with the one-hot encoding input (S_2), such that the fully connected layer (FC2) is activated in response to the one-hot encoding of training samples associated with theP39429-WO (2668-0006PCT) second person. Similarly, a multiplication layer (mult_3) multiplies the output of the rectified linear unit activation (rclu_3) with the one-hot encoding input (S_3), such that the fully connected layer (FC3) is activated in response to the one-hot encoding of training samples associated with the third person. Finally, a multiplication layer (mult_4) multiplies the output of the rectified linear unit activation (relu_4) with the one-hot encoding input (S_4), such that the fully connected layer (FC4) is activated in response to the one-hot encoding of training samples associated with the fourth person.

[0086] In the placeholder subnetwork 460, the multiplication layers (mult_l, mult_2, mult_3, and mult_4) are connected to an addition layer (add). In this way, the multiplication layers and the addition layer act, in effect, to multiplex between the fully connected layer activations corresponding to each of the four different persons whose personal health data is included in the corpus of health data.

[0087] Lastly, in the training neural network 450, the universal subnetwork 320 has a same structure as discussed above with respect to the deployed neural network 400. As will be discussed in greater detail below, parameters of the universal subnetwork 320 are learned during the training of the training neural network 450 using the corpus of health data having the personal health data of the plurality of persons. Once this training is performed, the parameters of the universal subnetwork 320 are frozen and the parameters of the personalized subnetwork 310 are learned during the training of the deployed neural network 400 solely using personal health data collected with respect to the particular person 220 to which the model 110 is personalized.Methods for Training a Personalized Health Information Prediction Model

[0088] A variety of operations and processes are described below for operating the health management system 200 to train a personalized health information prediction model 110 for a particular person 220. In these descriptions, statements that a method, processor, and / or system is performing some task or function refers to a controller or processor (e.g., the processor 218 of the measurement device 210, the processor 260 of the computing device 250, and the processor 276 of the remote server 270) executing programmed instructions stored in non-transitory computer readable storage media (e.g., the memory device 214 of the measurement device 210, the memory device 252 of the computingP39429-WO (2668-0006PCT) device 250, and the memory device 274 of the remote server 270) operatively connected to the controller or processor to manipulate data or to operate one or more components in the health management system 200 to perform the task or function. Additionally, the steps of the methods may be performed in any feasible chronological order, regardless of the order shown in the figures or the order in which the steps are described.

[0089] FIG. 5 shows a flow diagram for a method 500 for training a personalized health information prediction model. The method 500 advantageously adopts a two- stage training process in which a universal component is first trained using health data collected from a representative population of patients or users and, subsequently, a personalized component is trained using a data collected only from the individual person. The universal component is advantageously trained jointly with a placeholder component which employs one-hot encoding of the training samples associated with each particular person in the populationlevel health data.

[0090] The method 500 begins with universal component training (block 510). Particularly, a corpus of health data is provided including a plurality of training samples that include personal health data of the plurality of persons. In at least some embodiments, the plurality of persons may be persons who have provided the necessary authorizations to use their health data to train the model 110. Next, a training model is trained using the plurality of training samples of the corpus of health data. The training model used at this stage of training includes the universal component 130 but does not include the personalized component 120. The training neural network 450 including the universal subnetwork 320 and the placeholder subnetwork 460 is one example of such a training model. During this training process, the parameters of the universal component 130 (e.g., the universal subnetwork 320) are learned based on the relatively larger population of persons whose personal health data are included in the corpus of health data.

[0091] In some embodiments, the training of the universal component 130 in this manner may be performed by a centralized data processing system, such as the processor 276 of the remote server 270. Alternatively, in some embodiments, the training of the universal component 130 is performed in a distributed manner using federated learning by a plurality of computing devices that are different and remote from the computing deviceP39429-WO (2668-0006PCT) 250 and the measurement device 210 of the health management system 100 for the particular person 220.

[0092] FIG. 6 shows a flow diagram for a process for training the universal model component in a distributed manner using federated learning. Federated learning is a method of training machine learning models across multiple decentralized devices while keeping data localized. The process begins with a central server 610 (e.g., the remote server 270), which may also be referred to as the aggregation server, distributing the initial model parameters to a network of distributed computing devices 620. In one embodiment, each of the distributed computing devices 620 initializes a placeholder personalized subnetwork 640 to (e.g., sets to {1,0}) and freezes the weights. Additionally, the distributed computing devices 620 shuffle and batch their local datasets 630. For each batch, the weights in placeholder personalized subnetworks 640 are frozen, and the weights of the universal subnetwork 320 is unfrozen. These distributed computing devices train the model on their own local datasets 630. Importantly, each distributed computing device 620 works with its own data independently, ensuring that sensitive information remains on the distributed computing device and is not transmitted over the network.

[0093] After completing the local training process, each of the distributed computing devices 620 calculates an update to the model parameters based on its training. These updates, rather than the raw data, are sent back to the central server 610. The central server 610 aggregates these updates, for example using methods such as averaging, to create a consolidated model update. This aggregation step combines the knowledge gained from each distributed computing device 620, allowing the central server 610 to produce a new, updated model that incorporates insights from the diverse local datasets.

[0094] The updated model is then sent back to the distributed computing devices 620, where the process repeats iteratively. Each distributed computing device 620 receives the latest model parameters, applies them to their local training, and sends back further updates. This cycle continues until the model reaches a desired level of performance or convergence (e.g., until a termination criterion is reached).

[0095] It should be appreciated that, in the case that the local datasets 630 only include personal health data of a single person, the placeholder personalized subnetwork 640 used by the distributed computing devices 620 during each training cycle generally takes aP39429-WO (2668-0006PCT) simpler form compared to the previously described placeholder personalized subnetwork 460 used for centralized training. Instead, in such cases, the placeholder subnetwork 640 used by the distributed computing device 620 is essentially similar in form to the personalized subnetwork 310 that is used in deployment. Additionally, in such cases, for similar reasons, the local datasets 630 need not be one-hot encoded in the manner discussed above.

[0096] Returning to FIG. 5, the method 500 continues with personalized component training (block 520). Particularly, when the personalized health information prediction model 110 is deployed and personalized for the particular person 220, the processor 260 of the computing device 250 and / or the processor 218 of the measurement device 210 receives the pretrained parameters of the universal component 130 (e.g., the universal subnetwork 320). In one embodiment the processor 260 and / or the processor 218 operates the transceiver 254 and / or the transceiver 216, respectively, to download the pretrained parameters of the universal component 130 from the remote server 270.

[0097] Next, the processor 260 and / or the processor 218 collects a plurality of personal health data with respect to the particular person 220, which may for example be measured by the sensor(s) 212 or collected from the person 220 in the form of user inputs via the input device 256. The processor 260 and / or the processor 218 forms a plurality of training samples based on and including the personal health data collected with respect to the particular person 220. In general, the personal health data that is used to train the personalized health information prediction model 110 is collected in a similar manner as discussed below with respect to block 710 in the method 700 of FIG. 7.

[0098] Next, the processor 260 and / or the processor 218 trains the personalized health information prediction model 110 (e.g., the deployed neural network 400), which includes both the universal component 130 (e.g., the universal subnetwork 320) and the personalized component 120 (e.g., the personalized subnetwork 310). During the training, the parameters of the universal component 130 (e.g., universal subnetwork 320) are frozen and the parameters of the personalized component 120 (e.g., the personalized subnetwork 310) are learned solely using personal health data collected with respect to the particular person 220.P39429-WO (2668-0006PCT)

[0099] In one embodiment, during the training of the personalized health information prediction model 110, the processor 260 and / or the processor 218 randomly initializes the parameters of the personalized component 120 (e.g., the personalized subnetwork 310).

[0100] However, in another embodiment, the processor 260 and / or the processor 218 initializes the parameters of the personalized component 120 (e.g., the personalized subnetwork 310) depending on previously learned parameters of a placeholder component (e.g., the placeholder subnetwork 460) of the training model previously used to train the universal component 130 (e.g., universal subnetwork 320).

[0101] Particularly, in one embodiment, the processor 260 and / or the processor 218 initializes the parameters of the personalized component 120 (e.g., the personalized subnetwork 310) according to average values for persons with similar demographics to the particular person 220 for which the personalized health information prediction model 110 is being deployed and personalized. For example, if the person 220 is male, then average parameter values for layers in the placeholder component (e.g., the placeholder subnetwork 460) associated with other men are used to initialize the parameters of the personalized component 120 (e.g., the personalized subnetwork 310).Method for Providing Health Information

[0102] A variety of operations and processes are described below for operating the health management system 200 to provide health information to a person 220. In these descriptions, statements that a method, processor, and / or system is performing some task or function refers to a controller or processor (e.g., the processor 218 of the measurement device 210, the processor 260 of the computing device 250, and the processor 276 of the remote server 270) executing programmed instructions stored in non-transitory computer readable storage media (e.g., the memory device 214 of the measurement device 210, the memory device 252 of the computing device 250, and the memory device 274 of the remote server 270) operatively connected to the controller or processor to manipulate data or to operate one or more components in the health management system 200 to perform the task or function. Additionally, the steps of the methods may be performed in any feasible chronological order, regardless of the order shown in the figures or the order in which the steps are described.P39429-WO (2668-0006PCT)

[0103] FIG. 7 shows a flow diagram for a method 700 for providing health information. The method 700 advantageously leverages the personalized health information prediction model 110, trained for example using the method 500 discussed above, to provide further health data based on health data collected with respect to the person 220. As a result of using the personalized health information prediction model 110, the method 700 advantageously provides further health data in a more accurate and robust manner.

[0104] The method 700 begins with data collection (block 710). Particularly, the health management system 200 collects and records a plurality of personal health data with respect to the person 220, which will be input into the personalized health information prediction model 110 to provide further health information. As discussed above, in the case that the health management system 200 is used for diabetes or blood glucose management for a person with diabetes, the personal health data input into the model 110 may include, for example, historical blood glucose measurements, insulin intake data, and carbohydrate consumption data. However, the personal health data input to the model 110 may include a variety of other health data such as physical activity data, acceleration data, gyroscopic data, heart rate data, blood pressure data, blood oxygenation data, sleep data, body temperature data, or electronic medical records. In some embodiments, the personal health data is in the form of time series data in which each measurement is associated with a respective timestamp indicating a time at which the value was measured. In some embodiments, the personal health data input to the model 110 is limited to a predetermined time window (e.g., the last 30 minutes).

[0105] The processor 218 of the measurement device 210 operates the sensor 212 to measure health data values (e.g., blood glucose measurements) with respect to the person 220. As the health data values are measured over time, they are stored in the memory device 214 of the measurement device 210. Upon request or at some periodic interval, the processor 218 operates the transceiver 216 to transmit these health data values that are stored in the memory device 214 to one or both of the computing device 250 and the remote server 270. In addition to the health data values measured by the measurement device 210, the person 220 may manually input additional health data values (e.g., insulin bolus dosage data or the carbohydrate intake data) via the input device 256, for example by interacting with a graphical user interface of the health management application 262 on the displayP39429-WO (2668-0006PCT) screen 258 via the input device 256. However, in other embodiments, such health data may be determined or measured automatically.

[0106] The method 700 continues with prediction (block 720). Particularly, the processor 260 of the computing device 250 and / or the processor 218 of the measurement device 210 determines further health data of the person based on the collected personal health data. The processor 260 and / or the processor 218 executes the personalized health information prediction model 110 to determine the further health data of the person.

[0107] In some embodiments, the further health data output by the model 110 include a regression of values for at least one health metric that is determined based on the input personal health data. In one embodiment, the further health data output by the model 110 includes a regression of future values for at least one health metric during a predetermined prediction time window (e.g., 30 minutes into the future).

[0108] In some embodiments, the further health data output by the model 110 include a classification of the personal health data collected with respect to the person 220. Such classifications may include, for example, a predicted health status of the person 220, such as a prediction of a future hypoglycemic status, a prediction of a future hyperglycemic status, a predicted risk for diabetic neuropathy, a predicted risk for heart disease, a predicted risk for chronic kidney disease, or a categorization of the person 220 into ‘buckets’ (for example that categorize patients depending on overall health risk).

[0109] The method 700 continues with output (block 730). Particularly, the computing device 250 may be configured to provide a wide variety of useful and advantageous features using the further health data determined using the model 110. In some embodiments, the computing device 250 displays to a user (e.g., the person 220) a visualization, such as a data plot, of the further health data, e.g., a regression of predicted future values for at least one health metric. In another embodiment, the computing device 250 provides a perceptible warning, alert, notification, or message of a predicted health status of the person 220 (or regarding any other classification or prediction based on the further health data). In another embodiment, the computing device 250, additionally or alternatively, transmits the same types of warnings, alerts, notifications, or messages to a remote computing device, for example, a computer or online system used by a physician or a caretaker of the person.P39429-WO (2668-0006PCT)

[0110] In one example relating to diabetes or blood glucose management applications, the processor 260 operates the display screen 258 to display a data plot showing the historical blood glucose measurements and / or the predicted future blood glucose measurements. In some embodiments, a data plot may further show the insulin bolus dosage data and the carbohydrate intake data. Such a data plot advantageously enables the person 220 to more easily and better understand their health and how to better manage their blood glucose concentration throughout the day. The predicted future blood glucose measurements can help the person 220 to better appreciate the relationships that time of day, insulin bolus dosing, and the carbohydrate intake can have on his or her blood glucose concentration.

[0111] It should be appreciated that visualizations can be provided with respect to any regression or classification outputs from the personalized health information prediction model 110 or otherwise derived from the further health data generated thereby. Such visualizations are not limited to diabetes or blood glucose management applications.

[0112] FIG. 8 shows a portion of an exemplary graphical user interface including a data plot that visualizes both historical and predicted blood glucose measurements. Particularly, the data plot 800 includes a solid-line curve 810 that represents a subset of historical blood glucose measurements that were measured immediately prior to the current time. Similarly, the data plot 800 includes a dotted-line curve 820 that represents predicted future blood glucose measurements and extends two hours into the future. The solid-line curve 810 and the dotted-line curve 820 are superimposed upon a vertical band 830, which is shaded to indicate a range of ideal blood glucose measurement values (e.g., between 70 mg / dL and 180 mg / dL) in which the person 220 is not considered to be in a hypoglycemic or hyperglycemic condition. Additionally, each prediction of the dotted-line curve 820 (indicated by an individual dot) is paired with an error band 840, which is shaded to represent upper and lower error bounds for the particular prediction. Finally, the data plot 800 includes data annotations 850 and 860, which indicate amounts of carbohydrates consumed (e.g., “50g”) and amounts of insulin intake (e.g., “2U”), respectively.

[0113] In at least some embodiments, the processor 260 of the computing device 250 is configured to determine, based on predicted future blood glucose measurements, whether a possible hypoglycemic condition of the person 220 or a possible hyperglycemic conditionP39429-WO (2668-0006PCT) of the person 220 is likely within a prediction time window. In one embodiment, the processor 260 determines the possible hypoglycemic condition of the person 220 may occur within the prediction time window in response to any of the predicted future blood glucose measurements being less than a predetermined low glucose alarm limit. In one embodiment, the processor 260 determines the possible hyperglycemic condition of the person 220 may occur within the prediction time window in response to any of the predicted future blood glucose measurements being greater than a predetermined high glucose alarm limit. Alternatively, in some embodiments, the processor 260 determines, as a direct classification output of the personalized health information prediction model 110, whether the possible hypoglycemic condition of the person 220 or the possible hyperglycemic condition of the person 220 is likely within the prediction time window.

[0114] In response to determining that the possible hypoglycemic condition of the person 220 or the possible hyperglycemic condition of the person 220 is likely within the prediction time window, the processor 260 operates the display screen 258 or other output devices to output a perceptible warning, alert, notification, or message. The visual alert is a typical smartphone notification or any other image and / or text that informs the person 220 of the possible hyperglycemic or hypoglycemic condition. Additionally, or alternatively, the computing device 250 generates an auditory alert using a speaker (not shown). The auditory alert is a sound that the person 220 associates with the possible hyperglycemic or hypoglycemic condition, such as a typical smartphone notification sound or any other brief sound. Additionally, or alternatively, the computing device 250 generates a tactile alert using a haptic actuator (not shown). The tactile alert causes the computing device 250 to vibrate thereby informing the person 220 of the possible hyperglycemic or hypoglycemic condition. The computing device 250 may be additionally or alternatively configured to inform the person 220 of the possible hyperglycemic or hypoglycemic condition using any other notification method, such as by sending the person 220 a text message, sending the person 220 an email, and / or by making an automated phone call to the person 220. In another embodiment, the computing device 250, additionally or alternatively, transmits the same types of warnings, alerts, notifications, or messages to a remote computing device, for example, a computer or online system used by a physician or a caretaker of the person.P39429-WO (2668-0006PCT)

[0115] It should be appreciated that equivalent warnings, alerts, notifications, or messages can be provided with respect to any regression or classification outputs from the personalized health information prediction model 110 or otherwise derived from the further health data generated thereby. Such warnings, alerts, notifications, or messages are not limited to diabetes or blood glucose management applications.

[0116] FIG. 9 shows a portion of a graphical user interface including a visual alert 900 that lets a user know that his or her blood glucose has a high probability of falling below a threshold value. Particularly, the visual alert 900 is a typical smartphone notification including text 910 (e.g., “Low glucose soon!” “High probability” and “Between 12:30 -13:00, you might go below 70 mg mg / dL”) that informs the person 220 of the possible hypoglycemic condition. In addition, the visual alert 900 includes additional text 920 (e.g., “What can I do” and “Have something sugary (approx. 15 g of carbs).”) that informs the person 220 of a recommended action that can be taken to avoid the possible hypoglycemic condition.

[0117] In response to receiving the alert and / or the notification regarding a possible hyperglycemic condition, the person 220 may want to provide themselves with a bolus dose of the medicament 232 from the administration device 230. In some embodiments, the processor 260 may determine, and display on the display screen 258, information concerning a recommended dosage amount of the medicament 232 that should be dosed to the person 220 to attempt to prevent the hyperglycemic condition that was predicted in the predicted future blood glucose measurements.

[0118] For certain types of administration devices 230, the computing device 250 is configured to transmit bolus advisor data to the administration device 230 using the transceiver 254, so that the administration device 230 can deliver a desired dosage of the medicament 232 to the person 220. The amount of the medicament 232 that is delivered is based on the historical blood glucose measurements and / or the predicted future blood glucose measurements, with the objective being to reduce the glucose concentration level of the person 220 to a normal or desired level. In an example, the administration device 230 is an insulin pump that is configured to automatically deliver a bolus dose of insulin to the person 220. In another example, the administration device 230 is a small insulin pen that is configured to deliver a bolus dose for injection. In each example, the amount of theP39429-WO (2668-0006PCT) insulin that is delivered in the bolus dose is based on the historical blood glucose measurements and / or the predicted future blood glucose measurements.

[0119] Embodiments within the scope of the disclosure may also include non-transitory computer-readable storage media or machine-readable medium for carrying or having computer-executable instructions (also referred to as program instructions) or data structures stored thereon. Such non-transitory computer-readable storage media or machine-readable medium may be any available media that can be accessed by a general purpose or special purpose computer. By way of example, and not limitation, such non-transitory computer-readable storage media or machine-readable medium can comprise RAM, ROM, EEPROM, CD-ROM or other optical disk storage, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to carry or store desired program code means in the form of computer-executable instructions or data structures. Combinations of the above should also be included within the scope of the non-transitory computer-readable storage media or machine-readable medium.

[0120] Computer-executable instructions include, for example, instructions and data which cause a general-purpose computer, special purpose computer, or special purpose processing device to perform a certain function or group of functions. Computerexecutable instructions also include program modules that are executed by computers in stand-alone or network environments. Generally, program modules include routines, programs, objects, components, and data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of the program code means for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps.

[0121] While the disclosure has been illustrated and described in detail in the drawings and foregoing description, the same should be considered as illustrative and not restrictive in character. It is understood that only the preferred embodiments have been presented and that all changes, modifications and further applications that come within the spirit of the disclosure are desired to be protected.

Claims

P39429-WO (2668-0006PCT) What is claimed is:

1. A method for providing health information, the method comprising:receiving, with an electronic device, third health information of a first person; determining, with the electronic device, fourth health information of the first person based on the third health information using a first neural network model, the first neural network model having a first subnetwork and a second subnetwork, the first subnetwork having been trained solely using first health information of the first person, the second subnetwork having been trained using second health information of a plurality of second persons; andproviding at least one of (i) a visualization of the fourth health information on a display of the electronic device, (ii) a perceptible warning of a possible future health status via the electronic device, the possible future health status being indicated by the fourth health information, or (iii) a message including the fourth health information that is transmitted to a remote computing device.

2. The method according to claim 1 further comprising, prior to determining the fourth health information:receiving pretrained parameters of the second subnetwork of the first neural network model;receiving the first health information of the first person, the first health information including a plurality of first training samples; andtraining the first neural network model using the first health information of the first person, the pretrained parameters of the second subnetwork being frozen as parameters of the first subnetwork are learned during the training of the first neural network model.

3. The method according to claim 2, the receiving the first health information further comprising:measuring, with a sensor of the electronic device, at least some of the first health information of the first person.P39429-WO (2668-0006PCT) 4. The method according to claim 2, the training the first neural network model further comprising:training the first neural network model with the electronic device.

5. The method according to claim 2 further comprising:receiving the second health information of the plurality of second persons, the second health information including a plurality of second training samples; and training a second neural network model using the second health information of the plurality of second persons, the second neural network model including a third subnetwork and the second subnetwork, the pretrained parameters of the second subnetwork being learned during the training of the second neural network model.

6. The method according to claim 5, the training the second neural network model further comprising:training the second neural network model with a server that is remote from the electronic device.

7. The method according to claim 6, the training the second neural network model further comprising:training the second neural network model with a plurality of computing devices that are remote from the electronic device, using federated learning.

8. The method according to claim 5, wherein the fourth health information is one-hot encoded to associate each second training sample in the plurality of second training samples with a respective second person of the plurality of second persons.

9. The method according to claim 8, wherein:the first subnetwork of the first neural network model has a first architecture; and the third subnetwork of the second neural network model includes a plurality of instances of at least a portion of the first architecture, the third subnetwork being configured such that each respective instance in the third subnetwork is activated only inP39429-WO (2668-0006PCT) response to second training samples in the plurality of second training samples that are one-hot encoded in association with the respective second person of the plurality of second persons.

10. The method according to claim 9, wherein the plurality of instances of at least the portion of the first architecture include a respective instance associated with each respective second person of the plurality of second persons.

11. The method according to claim 5, the training the first neural network model further comprising:initializing the parameters of the first subnetwork in the first neural network model depending on parameters of the third subnetwork learned during the training of the second neural network model.

12. The method according to claim 2, the training the first neural network model further comprising:randomly initializing the parameters of the first subnetwork.

13. The method according to claim 1, wherein:the first subnetwork receives the third health information and is configured to transform the third health information into an input space of the second subnetwork; and the second subnetwork in the first neural network model is configured to receive the transformed third health information from the first subnetwork and to determine the fourth health information.

14. The method according to claim 1, wherein:the second subnetwork in the first neural network model receives the third health information and is configured to determine the fourth health information in an output space of the second subnetwork; andP39429-WO (2668-0006PCT) the first subnetwork receives the fourth health information in the output space of the second subnetwork and is configured to transform the fourth health information into an output space of the first neural network model.

15. The method according to claim 1, wherein the fourth health information is a regression of at least one health metric based on the third health information.

16. The method according to claim 1, wherein the fourth health information is a classification of the third health information.

17. The method according to claim 1, wherein:the third health information includes historical blood glucose measurements of the first person; andthe fourth health information includes at least one of (i) predicted future blood glucose measurements of the first person, (ii) a prediction of a future hypoglycemic status, (iii) a prediction of a future hyperglycemic status, or (iv) a predicted risk for diabetic neuropathy.

18. The method according to claim 17, wherein the third health information further includes at least one of (i) insulin intake data indicating insulin doses taken by the first person or (ii) carbohydrate consumption data indicating carbohydrates consumed by the first person.

19. A method for training a machine learning model configured to provide health information, the method comprising:receiving first health information of a first person, the first health information including a plurality of first training samples;receiving second health information of a plurality of second persons, the second health information including a plurality of second training samples;P39429-WO (2668-0006PCT) training a second neural network model using the second health information of the plurality of second persons, the second neural network model including a third subnetwork and a second subnetwork; andtraining a first neural network model using the first health information of the first person, the first neural network model having a first subnetwork and the second subnetwork, parameters of the second subnetwork that were learned during the training of the second neural network model being frozen during the training of the first neural network model,wherein the first neural network model is configured to receive third health information of the first person and determine fourth health information of the person based on the third health information.

20. A non-transitory computer-readable medium that stores program instructions for providing health information, the program instructions being configured to, when executed by a processor, cause the processor to:receive third health information of a first person;determine fourth health information of the person based on the third health information using a first neural network model, the first neural network model having a first subnetwork and a second subnetwork, the first subnetwork having been trained solely using first health information of the first person, the second subnetwork having been trained using second health information of a plurality of second persons; and provide at least one of (i) a visualization of the fourth health information on a display operably connected to the processor, (ii) a perceptible warning of a possible future health status via an output device operably connected to the processor, the possible future health status being indicated by the fourth health information, or (iii) a message including the fourth health information that is transmitted to a remote computing device via a transceiver operably connected to the processor.

Images