Systems and methods for improving cardiovascular health in a person

Abstract

An estimate of functional capacity, such as VO2Max, is made by applying the monitored person's vital signs to a trained encoding neural network to produce a cardiac profile vector. This vector is applied to a trained functional capacity (VO2Max) neural network to estimate the functional capacity. Once estimated, action is taken.

Description

Technical Field

[0001] The present disclosure relates to performing actions to treat medical conditions and improve a person's health, and more particularly, to determining these actions based on analysis of data obtained from a person. Background Art

[0002] There are many situations in the medical field where it is of great value to determine the comparative health of an individual at a point in time compared to other people (or compared to a known category) or compared to the individual's history. In particular, such health determinations often seek to measure the cardiopulmonary function of an individual, since blood circulation and gas exchange are the first line of survival compared to dietary metabolism or cognitive function. In some aspects and as used herein, "health" refers to a person's cardiopulmonary fitness, but it will be understood that other types of fitness are possible.

[0003] Determining the relative health of a person is useful in various aspects. For example, these determinations can be used to determine the efficacy / side effects of drugs in drug trials (e.g., whether drug use is improving the health of the subject, maintaining the health of the subject in view of expected degradation, or having negative side effects on the health of the subject). Comparative health determinations can also be used to obtain the efficacy of device / procedure interventions in clinical trials, or determine therapeutic capabilities (e.g., in oncology treatment / chemotherapy, determine whether the patient has fully recovered from the last treatment to endure more, whether the patient is being overtreated, or whether the treatment otherwise harms the patient). These determinations can be further used for health care monitoring (e.g., tracking the day-to-day trajectory of patients with chronic diseases (e.g., heart failure or COPD) or discharged after surgery). These determinations can also be used to identify individuals with undiagnosed diseases in a group, or to achieve early intervention (e.g., NYHA Class I heart failure, which is usually underdiagnosed and asymptomatic). In addition, these determinations can be used to evaluate the effectiveness of rehabilitation programs (e.g., whether the program is helping the patient or whether the patient is executing the program). Other examples are possible.

[0004] Indicators of health status are often measured under "effort-challenging" activities (activities that individuals may be reluctant to perform), such as exercise tests. In general, exercise tolerance is determined by three factors: pulmonary gas exchange; cardiovascular performance, including the peripheral vascular tree; and skeletal muscle metabolism. Methods for directly and objectively measuring an individual's cardiopulmonary capacity are often invasive or difficult to establish and perform. They are not easily deployed for long-term monitoring or repeated measurements, either in clinical trials or for long-term monitoring of patients, due to expense or inconvenience.

[0005] For example, objective measures of cardiopulmonary function include the ventilatory anaerobic threshold (VAT) test. VAT measurement involves drawing continuous blood samples from a subject during an exercise challenge to determine the inflection point where metabolism moves from aerobic to anaerobic. This method involves continuous blood gas testing and a challenge set up to obtain blood samples in an exercise equipment setting.

[0006] Another objective measure of cardiopulmonary function includes cardiopulmonary exercise testing (CPET), in which the subject is fitted with a breathing mask for measuring oxygen uptake and carbon dioxide production. The mask and attached equipment are able to measure the O2 (and / or CO2) content from the breath, allowing an objective estimate of oxygen uptake, particularly when the subject reaches a steady level of maximal exercise capacity. This method requires special equipment and a laboratory-type exercise equipment setup.

[0007] Yet another objective measure of cardiopulmonary function includes pulmonary function testing (PFT), which uses a spirometer or plethysmograph chamber to measure the amount and rate of air during breathing. This method measures only breathing and therefore does not capture the "cardiac" aspects of health assessment. Nevertheless, it provides a way to objectively measure structural gas exchange volume and flow.

[0008] Another objective measure of cardiopulmonary function includes a cardiac echo Doppler test, which is a measurement of a person's cardiac output, body expansion, and valve health. These measurements are not expected to change much over a short period of time (a few weeks). This measurement does not address aspects of lung performance or skeletal muscle metabolic health.

[0009] As an alternative to these objective measurements, more easily deployable tests such as the "6-minute walk test" (6MWT) are used in health monitoring and clinical research. In the 6MWT, the clinician marks out a corridor in which the individual being tested walks back and forth as far as possible in 6 minutes. The walking distance is a proxy for cardiopulmonary exercise capacity. One problem with the 6MWT is its simplicity: the setup may vary due to obstacles or distractions; the individual being tested may walk less distance for reasons other than cardiopulmonary fitness (such as arthritis, headaches, balance problems, or joint pain). Therefore, the results are interfered with and only closely correlated with true health. In addition, it is inconvenient for the individual being tested to travel to the clinician's location to take the test.

[0010] Estimators of VO2Max are derived based on physical attribute values ​​such as age or gender. However, these are also only roughly accurate and certainly do not provide information about day-to-day variations in individual performance. For example, such estimates are useless for monitoring the status of a patient with a chronic disease who may develop disease exacerbations over days or weeks. Recently, VO2Max estimators have begun to be developed using features from fitness trackers (primarily heart rate) in combination with physical attribute values, but these lack accuracy and typically still require the subject to perform a specific exercise routine during which the assessment is performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] For a more complete understanding of the present disclosure, reference should be made to the following detailed description and accompanying drawings, wherein:

[0012] Figure 1 includes illustrations of systems configured according to various embodiments of these teachings;

[0013] Figure 2 Includes flow charts configured according to various embodiments of these teachings;

[0014] Figure 3 includes illustrations of systems configured according to various embodiments of these teachings;

[0015] Figure 4 includes illustrations of systems configured according to various embodiments of these teachings;

[0016] Figure 5A and Figure 5B includes illustrations of systems configured according to various embodiments of these teachings;

[0017] Figure 6 includes illustrations of systems configured according to various embodiments of these teachings;

[0018] Figure 7 including illustrations of cardiac profile vectors and creation vectors configured according to various embodiments of these teachings;

[0019] Figure 8 includes illustrations of systems configured according to various embodiments of these teachings;

[0020] Fig. 9 includes illustrations of systems configured according to various embodiments of these teachings;

[0021] Fig.10 includes illustrations of systems configured according to various embodiments of these teachings;

[0022] Fig.11 includes illustrations of systems configured according to various embodiments of these teachings;

[0023] Fig.12 includes illustrations of systems configured according to various embodiments of these teachings;

[0024] Fig.13 includes illustrations of systems configured according to various embodiments of these teachings;

[0025] Fig.14 includes illustrations of systems configured according to various embodiments of these teachings;

[0026] Fig.15 Included are illustrations of systems configured according to various embodiments of these teachings. DETAILED DESCRIPTION

[0027] In general, the present invention relates to machine learning-based computing systems and methods that obtain wearable sensor data from individuals performing walking activities (e.g., activities involving movement) or other activities of daily life and subsequently utilize the wearable sensor data. After obtaining the data, the methods described herein generate one or more of: an estimated VO2Max (maximum oxygen uptake); a category or diagnosis of a health or medical condition (e.g., selected from a range such as normal, pre-heart failure, heart failure (HF) Class I, HF Class II, HF Class III, HF Class IV), an estimated performance capacity for an ideal activity (such as a 6-minute walk test (6MWT)); or a day-by-day index of deterioration or improvement in cardiorespiratory fitness. Other examples are possible.

[0028] In one example, a chest adhesive wearable patch is secured to a human user and electrocardiogram ("ECG") and 3-axis accelerometer ("XYZ") waveform data are obtained, and vital sign data is obtained from the sensor data. In an example, the vital sign data may include one-minute average activity (e.g., total accelerometer RMS within the last moving window of two minutes), one-minute average heart rate (HR), and (optionally) one-minute average respiratory rate (RR). Other examples are possible. As an example, the input to the method described herein is 1,440 minutes of data (or a substantial subset thereof), including one day. However, in other examples, this may be two or more days. Other time windows may also be used.

[0029] The present method creates, trains, and utilizes a mathematical model in the form of a neural network (e.g., a convolutional neural network (CNN)). Once created and trained, data from an individual is selectively applied to the model to obtain an output. In one example, the output is an estimated VO2Max for the individual.

[0030] The output can be used in various ways. For example, a medical clinician can use the output of estimated VO2Max, category, idealized exercise capacity metric (6MWT), etc. to assess the trend of an individual's health over time (e.g., a heart failure patient who is being monitored at home after a recent discharge) or diagnose the stage of an individual's progressive disease. The output can be used to control medical equipment, for example, to adjust the drip rate of an intravenous (IV) machine. Alarm messages can also be sent to the patient. In addition, medical personnel can be directed and navigated to the patient to provide assistance.

[0031] In other examples, the output of the methods provided herein is used in clinical trials as a measure of the effectiveness of a drug, device or procedure on cardiopulmonary fitness between a test group of humans and a control group of humans. The output of the methods of the present invention may also be tracked to understand its negative side effects.

[0032] In various aspects, the method uses the dynamics of vital signs over a certain period of time (e.g., a day, at least one diurnal cycle) to determine a measure of the health of a subject. Information is obtained indirectly from the pattern of activity and vital sign response, such as the lag between the start or stop of activity and the rise or return to baseline of heart rate. These dynamics potentially encode the functional ability of the subject without requiring the subject to perform any specific activity or activity level or other prescribed exercise test. The method obtains this potentially encoded information to produce an objective measure of health status.

[0033] In other aspects, the methods described herein employ artificial intelligence techniques to train deep neural networks. The network architecture described herein provides benefits by not overfitting to everyday training data, and at the same time, it does not require a large amount of labeled training data to train effectively.

[0034] Advantageously, the methods provided herein accurately characterize cardiopulmonary exercise capacity on an ongoing basis and are customized for individuals, i.e., based on real-time measurements from individuals, without the inconvenience of existing or invasive testing. These characterizations are accurately determined using sensor data from wearable devices. Hints, markers, or overt signs of the potential functional capacity of individuals on the day are encoded in all aspects and based on various efforts, activities, and vital sign responses (e.g., heart rate recovery after climbing stairs) throughout the day.

[0035] Ideally, deterministic metrics are generated that (a) are independent of reliance on the individual being measured to perform a prescribed activity and (b) can be collected from the normal context of daily activities.

[0036] The present method trains a neural network from wearable sensor data, but does not require such a large number of VO2Max results. These methods provide health status indicators such as VO2Max and 6MWT that can be recognized and acted upon by both humans (e.g., clinicians) and machines.

[0037] Advantageously, the present method provides a cardiac profile vector similar to a "fingerprint" for an individual as expressed in a day's vital signs data. This fingerprint encapsulates the individual's basic cardiorespiratory behavior as expressed in their activities of daily living (ADL).

[0038] With this approach, individuals do not need to be forced to perform any specific activity every day. The cardiac profile is captured based on their ADL, without constraints. The benefit of this is to achieve higher compliance and avoid the problem of human subjects being unable to reliably perform a set of prescribed activities.

[0039] Furthermore, the present approach yields a significant savings in required training data by using readily collected vital sign data to train the network to generate a cardiac profile, and then regressing it solely against a substantially smaller set of VO2Max (or other final assessment) results that may be too expensive or difficult to obtain in sufficient quantities for normal direct neural network training.

[0040] By the present method, the regression of the cardiac profile against a known, objective cardiorespiratory fitness metric over time to provide an estimate of that metric continuously greatly enhances visibility into the subject's ongoing condition. By placing the output of the assessment in terms of a known metric, the subject's health status can be made very easy to interpret to the clinician.

[0041] Advantageously, when tracked over time, a regressed metric such as VO2Max can show the directionality of a subject's health simply by observing the subject's vital sign responses to ADLs.

[0042] Furthermore, regression measures similar to estimated VO2Max can replace noisier and less reliable tests like the 6MWT used in clinical trials and clinical fitness assessments. Because the accuracy of assessments according to the present invention is superior to crude tests like the 6MWT, it is possible that clinical studies can be reduced in size or reach conclusions more quickly because smaller statistics are required to achieve similar abilities when the accuracy of the measure is more stringent.

[0043] As explained elsewhere herein, the generation of cardiac profile vectors allows direct comparisons between groups of tested subjects with high variability in their health status and activities of daily living. With a 100 element vector, multiple classifications can be achieved, and individuals can even be grouped through self-organizing clusters to discover unforeseen commonalities between subgroups.

[0044] In many of these embodiments, methods for treating and improving a person's cardiopulmonary health are provided. In embodiments of the invention, an encoding neural network is configured to accept as input the values ​​of a selected plurality of cardiopulmonary variables for a time series duration (e.g., a day) from a subject, and to generate a multi-element cardiac profile vector for the subject. When supplied with the cardiac profile vector as input, the function generator neural network is configured to generate values ​​of neural network connection weights for use in another network. An estimation neural network is configured to accept as input the values ​​of a subset of a selected plurality of cardiopulmonary variables for another time series duration from the subject. The subset omits at least one excluded variable, and the estimation neural network is configured to generate an estimated value of at least one excluded variable. The estimation neural network is configured using the weights output by the function generator network as the weights of the connections between nodes in the estimation neural network.

[0045] The encoding neural network is concatenated and iteratively trained through a plurality of iterations to produce a trained encoding neural network, and the function generator neural network is concatenated and iteratively trained through a plurality of iterations to produce a trained function generator neural network. The training uses an error metric that includes a difference between at least one excluded variable and an estimated value of the at least one excluded variable generated by the estimator neural network.

[0046] For each of the iterations, training data of a first duration comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from a training human subject is input to an encoding neural network. For each of the iterations, training data of a second duration (substantially non-overlapping in time with the first duration) comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from a training human subject is input to an estimation neural network. The estimation neural network provides an estimate of at least one excluded variable. Data from the time series duration from one or more training human subjects may be used for training. Once trained, the encoding neural network may be used to generate a cardiac profile vector for any new human subject (e.g., a subject that may not be any of those human subjects used for training).

[0047] The functional capacity estimator is configured to accept as input a cardiac profile vector of a human subject generated by a trained encoding neural network. The cardiac profile vector may be augmented by a set of preselected physical property values ​​of the human subject, and the functional capacity estimator outputs an estimate of a measure of the functional capacity of the human subject.

[0048] A functional capacity estimator is trained using a plurality of matched examples to produce a trained functional capacity estimator. Each of the matched examples includes a cardiac profile vector that, in some embodiments, is augmented with physical attribute values ​​from a person, matched to a corresponding measured functional capacity metric of the person, to enable regression of the input of the physical attribute augmented cardiac profile vector to an estimate of the functional capacity metric.

[0049] A trained encoding neural network and a trained functional capacity estimator are deployed to accept as input monitored data for a duration of time from a monitored human subject. The monitored data includes a substantially continuous time series of a selected plurality of cardiopulmonary variables. The trained encoding neural network generates a current cardiac profile vector for the monitored human subject. The current cardiac profile vector may be augmented by physical attributes of the monitored human subject. The trained functional capacity estimator generates a current estimate of a measure of functional capacity from application of the current cardiac profile vector augmented by the physical attributes to the trained functional capacity estimator.

[0050] One or more actions are performed based on the current estimate of functional capacity. In one example, the current estimate in a time series with previous estimates is displayed to a clinician for review of possible health changes in the monitored human subject. In other examples, the current estimate is compared to previous estimates and tests are performed for changes. If a change is detected, an alert is triggered to a clinician to investigate the health of the monitored person.

[0051] In other examples, the current estimate is compared to a previous estimate and a test is performed for changes, whereupon a questionnaire is sent to the monitored human subject. In other examples, the current estimate is compared to a previous estimate and a test is performed to detect changes, whereupon an entry is made in the monitored human subject's medical record indicating a change in the monitored human subject's health.

[0052] In other examples, the current estimate is compared to estimates from other monitored human subjects to quantify the health impact between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial. In other examples, the operation of a medical device is controlled or a parameter of a medical device is set, wherein the medical device is associated with treating or monitoring a monitored human subject.

[0053] In various aspects, the functional capacity is VO2Max. Other examples are possible.

[0054] In other aspects, the time series of the selected plurality of cardiopulmonary variables is obtained by a sensor non-invasively affixed to the skin of the person. In other examples, the time series of the selected plurality of cardiopulmonary variables is obtained by a sensor such as a camera, an electrocardiogram sensor, a light sensor, an accelerometer, an RF energy sensor, a temperature sensor, and a gyroscope sensor. Other examples are possible.

[0055] In other examples, the functional capacity estimator generates a current estimate of the measure of functional capacity using one of: a linear regression technique, a decision tree, a random forest estimate, or a gradient boosting model. In other examples, the functional capacity estimator includes estimating a neural network. Other examples are possible.

[0056] In aspects, the cardiopulmonary variables include one or more of: heart rate, a measure of physical activity or exercise, temporal heart rate variability, respiratory rate, tilt angle, probability of atrial fibrillation, and sample time difference.Other examples are possible.

[0057] In other aspects, the current cardiac profile vector comprises a plurality of real numbers, each of the real numbers being associated with one or more of the cardiopulmonary variables.

[0058] In other of these embodiments, a system configured to treat and improve cardiopulmonary fitness of a person includes a sensor and control circuitry. The sensor is configured to obtain monitored data from a monitored human subject for a duration of time, the monitored data including a substantially continuous time series of a selected plurality of cardiopulmonary variables, the sensor being deployed with the monitored human subject.

[0059] The control circuit is coupled to the sensor and is configured to: receive the monitored data for the duration; apply the monitored data for the duration to a trained encoding neural network and responsively create a current cardiac profile vector for the monitored human subject; enhance the current cardiac profile vector with physical attributes of the monitored human subject as needed; and generate a current estimate of a measure of functional capacity from the current cardiac profile vector enhanced with physical attributes by applying the current cardiac profile vector enhanced with physical attributes to a trained functional capacity estimation neural network.

[0060] A trained functional capacity estimation neural network is created using a plurality of matched examples to produce a trained functional capacity estimation neural network. Each of the matched examples includes a cardiac profile vector augmented by physical attribute values ​​from a person matched to a corresponding measured functional capacity metric of the person such that an input of the physical attribute augmented cardiac profile vector is regressed to an estimated value of the functional capacity metric.

[0061] One or more actions are performed based on the current estimate of functional capacity. The actions include the control circuit presenting a display of the current estimate in a time series with previous estimates on a display screen to a clinician for review of possible health changes in the monitored human subject; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit triggers and transmits an electronic alert to the clinician to investigate the health of the monitored person; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit transmits an electronic questionnaire to the monitored human subject; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit records an entry in the electronic medical record of the monitored human subject indicating that the health of the monitored human subject has changed; the control circuit compares the current estimate with estimates from other monitored human subjects to quantify the health impact between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial; and the control circuit transmitting an electronic control signal that controls the operation or setting of a parameter of a medical device associated with treating or monitoring the monitored human subject.

[0062] With respect to training, the encoding neural network is configured to accept as input a time series of a selected plurality of cardiopulmonary variables and generate a multi-element cardiac profile vector. The function generator neural network is configured to generate weights when supplied with the cardiac profile vector as input. The estimation neural network is configured to accept as input a time series of a subset of the selected plurality of cardiopulmonary variables, the subset omitting at least one excluded variable, and generate an estimate of at least one excluded variable. The estimation neural network is configured using weights output by the function generator network as weights for connections between nodes in the estimation neural network.

[0063] The encoding neural network and the function generator neural network are connected in series and trained through multiple iterations. The training of the encoding neural network produces a trained encoding neural network, and the training of the function generator neural network produces a trained function generator neural network. The training of the encoding neural network and the function generator neural network uses an error metric that includes the difference between at least one excluded variable and the estimated value of at least one excluded variable generated by the estimation neural network. The estimation neural network itself is not trained, and its weights are dynamically supplied by the function generator neural network through each set of inputs at the encoding neural network; however, the estimation network is used to create the error metric. To further clarify, the weights of the estimation neural network are never truly "fixed", but change with each set of inputs to the encoding neural network and are dynamically determined by the function generator neural network and are therefore not trained. In contrast, the function generator neural network is trained to be good at providing effective weights for use by the estimation neural network. In contrast, the weights of the encoding neural network and the function generator neural network remain "fixed" regardless of the input data unless they are changed through the training process.

[0064] As known to those skilled in the art, training a neural network involves two distinct stages. The first stage is a forward pass through the neural network, where the network parameters are frozen, inputs are provided to the network, and the network generates an estimate of the target output value. The second stage (referred to as the backward pass through the neural network) involves calculating the amount of error between the estimate of the target value and the target value itself. The partial derivatives of each network parameter with respect to the error are then calculated for each layer in the network (this process is referred to in the art as back propagation of the error). The neural network parameters are then updated by subtracting small multiples of the partial derivatives of each corresponding parameter. This two-stage process is then repeated until the network has learned to accurately generate an estimate of the target output.

[0065] For each of the iterations: inputting training data of a first duration comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects to an encoding neural network; and providing input to an estimation neural network of training data of a second duration comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects that is substantially non-overlapping in time with the first duration (excluding at least one excluded variable). The estimation neural network provides an estimate of the at least one excluded variable.

[0066] In other of these embodiments, a system configured to treat and improve a person's cardiopulmonary fitness comprises: an encoding neural network configured to accept as input a time series of a selected plurality of cardiopulmonary variables and generate a multi-element cardiac profile vector; a function generator neural network configured to generate weights when supplied with the cardiac profile vector as input; an estimation neural network configured to accept as input a time series of a subset of a selected plurality of cardiopulmonary variables that omits at least one excluded variable and generate an estimate of the at least one excluded variable, wherein the estimation neural network is configured using weights output by the function generator network as weights for connections between nodes in the estimation neural network; and a control circuit.

[0067] The control circuit is configured to iteratively train (1) the encoding neural network in series and over a plurality of iterations to produce a trained encoding neural network, and iteratively train (2) the function generator neural network to produce a trained function generator neural network. The training uses an error metric that includes a difference between at least one excluded variable and an estimated value of the at least one excluded variable generated by the estimation neural network.

[0068] For each of the iterations, a first duration of training data comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects is input to the encoding neural network; and a second duration of training data comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects that substantially does not overlap in time with the first duration (excluding at least one excluded variable) is provided as input to the estimation neural network. The estimation neural network provides an estimate of the at least one excluded variable.

[0069] The functional capacity estimator is configured to accept as input a cardiac profile vector generated by a trained encoding neural network. The cardiac profile vector may be augmented by a set of preselected physical property values ​​of a human subject and output an estimate of a measure of functional capacity.

[0070] The control circuit is further configured to train the functional capacity estimator using multiple matched examples to produce a trained functional capacity estimator, each of the matched examples comprising a cardiac profile vector augmented by physical attribute values ​​from a person (if any), matched to a corresponding measured functional capacity metric of the person, so as to enable regression of the input of the physical attribute augmented cardiac profile vector to an estimate of the metric of functional capacity.

[0071] A trained encoding neural network and a trained functional capacity estimator are deployed to accept as input monitored data of duration from a monitored human subject, the monitored data comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables, the encoding neural network and the functional capacity estimator to generate a current cardiac profile vector for the monitored human subject, to augment the current cardiac profile vector with physical attributes of the monitored human subject, and to generate a current estimate of a measure of functional capacity from the current cardiac profile vector augmented with the physical attributes.

[0072] One or more actions are performed based on the current estimate of functional capacity. The one or more actions include: the control circuit presenting a display of the current estimate in a time series with previous estimates on a display screen to a clinician for review of possible health changes in the monitored human subject; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit triggers and transmits an electronic alert to the clinician to investigate the health of the monitored person; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit transmits an electronic questionnaire to the monitored human subject; the control circuit comparing the current estimate with the previous estimate and testing for changes, whereupon the control circuit records an entry in the electronic medical record of the monitored human subject indicating that the health of the monitored human subject has changed; the control circuit compares the current estimate with estimates from other monitored human subjects to quantify the health impact between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial; and the control circuit transmits an electronic control signal that controls the operation or setting of a parameter of a medical device associated with treating or monitoring the monitored human subject.

[0073] Reference now Figure 1 , depicts an example of a system 100 for determining an estimate of a patient's relevant cardiopulmonary value or functional capacity, such as VO2Max. The system 100 includes a training population 102 of people, including a first person 104, a second person 106, and a third person 108. A first sensor 105 is associated with the first person 104, a second sensor 107 is associated with the second person 106, and a third sensor 109 is associated with the third person 108. The sensors 105, 107, and 109 are coupled to an electronic network 110. The electronic network 110 is coupled to a central processing center 112. The central processing center includes a control circuit 114, a database 116 (including a mathematical model 118).

[0074] The network 110 is coupled to equipment 120 , a fourth sensor 122 (associated with a person being tested or monitored 124 ), a mobile electronic device 126 , a treating person or equipment 128 , a medical person or equipment 130 , and a display 132 .

[0075] The training population 102 includes a plurality of individuals. In this example, only three individuals 104, 106, 108 are shown, but it should be understood that the training population is typically hundreds, thousands, tens of thousands or even millions of individuals.

[0076] Each of the people is associated with a sensor (in this case, sensors 105, 107, and 109). As provided, sensors 105, 107, and 109 are configured to sense one or more vital signs (or other sensed values) from people 104, 106, and 108. In various aspects, the cardiopulmonary state of an individual is manifested as various dynamics of its vital signs during a day's daily living activities (ADL). Examples of vital signs include heart rate (HR), heart rate variability (variance of HR per unit time) (HRV), respiratory rate (RR), gross movement / effort (activity from a 3-axis accelerometer on the body) (ACT or XYZ), body trunk tilt (e.g., angle / degree relative to upright) (TILT), and the presence of atrial fibrillation (a sign when HR and HRV may not be typical) (AF or AFIB). Other examples are possible. In an example, the sensor obtains various types of physical signals (e.g., electrical signals, movement, or ECG), and these signals are converted into vital signs at the sensor or at some other location.

[0077] All of the above vital signs may be obtained continuously (or periodically) from sensors 105, 107, 109. In one example, sensors 105, 107, and 109 are adhesive torso patches with one lead ECG (two electrodes, typically on the upper left chest) plus a 3-axis accelerometer (capturing acceleration measurements reflecting breathing movement and effort used to move the torso (the main part of the body)). In other aspects, vital signs may be calculated at a high sampling rate (e.g., HR once per heartbeat; RR once every 5 or 10 seconds, or TILT once per second) and averaged to some standard rate, such as once per minute. This may be performed at sensors 105, 107, and 109 or at control circuit 114.

[0078] Sensors 105, 107, and 109 may be placed at other locations other than the torso of the person. For example, sensors 105, 107, and 109 may be worn on the wrist, from which a pulse signal indicative of the pumping action of the blood may be obtained optically, typically. Wrist devices also typically have an accelerometer, although this reflects both major physical efforts (standing up, walking, running) as well as minor peripheral activities (raising arms, hand movements). Respiration may be obtained from the pulse signal as an envelope and respiratory sinus arrhythmia may also be seen in the timing of the pulse peaks. Although the wrist location is subject to more motion and the pulse signal is more easily lost than a chest-based ECG, a large amount of good data may still be collected, given the sedentary lifestyle of many people.

[0079] If a wrist sensor is used, the neural network used in the model described herein will use inputs including pulse rate, respiration rate, activity, sample time intervals, pulse rate variability, and optionally filtered versions thereof. Other examples are possible.

[0080] In one specific example, the data obtained from the training population 102 includes "Day 1" data and "Day 2" data. In one example, the data for Day 1 is from a first calendar date and the data for Day 2 is from a second calendar date. However, it should be understood that the start time of such a data requirement day may occur at a time other than midnight, and the data for Day 1 and Day 2 may overlap on the same calendar date, such as when Day 1 overflows into the first part of the calendar date and Day 2 begins on that calendar date.

[0081] In another example, vital signs for day 1 and day 2 are each obtained from eight channels at a sampling rate of 1 minute (1440 samples per day) over a 24 hour period, each normalized by z-scores on the training data. The eight channels are: low pass heart rate (e.g., heart rate after a low pass filter); high pass heart rate (e.g., heart rate after a high pass filter); time domain heart rate variability (e.g., HRV within a moving time window); overall activity; respiratory rate; tilt angle (e.g., from a torso patch); atrial fibrillation probability (e.g., 0-1 probability of the presence of atrial fibrillation in the ECG); and sample time difference (e.g., which informs gaps in the data).

[0082] In yet another example, vital signs for day 1 and day 2 include low heart rate; high heart rate; time domain heart rate variability; overall activity; respiratory rate; and tilt angle (eg, from a torso patch).

[0083] In another example, the vital sign signals for day 1 and day 2 include heart rate; time-domain heart rate variability; overall activity; respiratory rate; and blood pressure (eg, from a wrist-worn device).

[0084] While the above examples utilize one minute averages of vital signs, other examples are possible. For example, a sampling rate of 15 second average vital sign values ​​to 5 minute average vital sign values ​​may be used. In one embodiment, all vital signs used are sampled at the same rate, regardless of whether the rate is within the range of 15 second to 5 minute average sampling. It is also contemplated that different statistical values ​​(such as median or quartile values) may be used in place of the average of the vital signs in such sampling windows.

[0085] In all aspects, the vital signs are sampled by sensors 105, 107 and 109 at a rate that is related in some way to the cardiopulmonary control system response rate of the human physiology. In other words, in order to capture the relevant behavior of human physiology related to the analysis of cardiopulmonary fitness, the vital signs are sampled at a rate that reveals the way the human system responds to a cardiopulmonary challenge or effort challenge. This is typically measured in the range of seconds to low minutes.

[0086] In other respects, and although a full day has been described as a preferred time window upon which a cardiac profile is based, other time windows may be used in the present method. For example, the time window may be longer. An individual's health is unlikely to change dramatically over a two-day period compared to a one-day period. Therefore, more than 1440 minutes (a day) of one-minute data may be used, and the time window may be expanded to a (small) multiple of that number. Expanding the window too much begins to risk having data that spans actual health changes, particularly for chronic patients who may significantly deteriorate over a few days. In an alternative, for example, a two-day (2880-minute) window may be used to generate a cardiac profile. This window may be moved in overlapping increments of one day or with some other overlap to provide an output each day.

[0087] The risk of a window shorter than one day is that all relevant physiological behaviors are not included, so that different time windows represent substantially different types of behaviors. For example, if the time window is 8 hours, the first time window may mainly represent sleep, while the next 8-hour period may exclusively represent waking activities. This can lead to high output variation between assessments. In view of all this, an alternative approach selects a time window to focus specifically on one behavior pattern, such as only waking or only sleeping. Such windows can be selected by fixed time periods, or can be selected by different analysis algorithms known in the art, such as sleep detection using actigraphy.

[0088] In other examples, other sets of vital signs may be used. Heart rate (or pulse rate) and activity are the main indicators of human cardiorespiratory dynamic behavior. Respiration rate is also important, but it is a pseudo-voluntary vital sign that can introduce modeling noise because, for example, an individual may choose to talk (interrupting normal breathing) even while climbing stairs. HRV, tilt, arrhythmias (such as atrial fibrillation) are useful contextual vital signs.

[0089] As mentioned, filtered versions of the vital signs may be used as different inputs (such as low-pass and high-pass HR). For example, this may be done for RR, ACT, or HRV. This may be desirable in order to separate out different underlying aspects of behavior (e.g., overall movement vs. fidgeting).

[0090] Measurements of respiratory tidal volume, cardiac output, blood oxygenation (SpO2), and blood pressure are prime candidates for expansion or modification of the set of vital signs used, as they reflect the same cardiopulmonary performance being captured. For example, the set of vital signs may be expanded by wearing a patch and watch worn in tandem to provide at least HR, HRV, tilt, RR, pulse, activity, and continuous BP. In another embodiment, only the watch may be worn to provide pulse, HRV, activity, RR, and continuous BP.

[0091] Now return to Figure 1 , electronic network 110 is any type of electronic communication network or combination of networks, such as the Internet, a cloud network, a wide area network, a local area network, a wireless network, or a cellular network. Other examples are possible. Central processing center 112 is any central location (e.g., an office, a corporate headquarters, a data center, a hosted virtual private cloud, or a central laboratory) where data is collected and analyzed. In this example, data from people 105, 107, and 109 is used to train model 118 (or a selected portion of model 118 as explained elsewhere herein). Once trained, data obtained from monitored person 124 is applied to the model to produce a current estimate of functional capacity (e.g., VO2Max), and this is used to perform various actions.

[0092] It should be understood that, as used herein, the term "control circuitry" broadly refers to any microcontroller, computer, or processor-based device having a processor, memory, and programmable input / output peripherals that is generally designed to govern the operation of other components and devices. It is further understood that common accompanying accessory devices are included, which include memory, transceivers for communicating with other components and devices, etc. These architectural options are well known and understood in the art and do not require further description here. The control circuit 114 can be configured (e.g., by using corresponding programming stored in memory, as will be well understood by those skilled in the art) to perform one or more of the steps, actions, and / or functions described herein.

[0093] Database 116 is any type of electronic memory device that electronically stores information.

[0094] The mathematical model 118 may be one or more mathematical models stored electronically. In an example, the mathematical model may include an encoding neural network, a function generator neural network, an estimation neural network, and a functional capacity estimator, the function generator neural network being used to generate weights when supplied with a cardiac profile vector as input. These elements are described in more detail elsewhere herein.

[0095] Equipment 120 may be any type of equipment that may be used to treat person 124. In an example, equipment 120 may be a drone or an autonomous vehicle that sends supplies or transportation to person 124. For example, a drone may navigate through the physical environment to the home of person 124 to deliver supplies.

[0096] A fourth sensor 122 (associated with a test or monitored person 124) senses one or more vital signs of the monitored person 124. The fourth sensor 122 may take the form of sensors 105, 107, or 109, or it may have some other form or configuration that delivers the same vital signs used in training.

[0097] The mobile electronic device 126 is any type of mobile electronic device, such as a smart phone, a cellular phone, a notebook computer, or a tablet computer. Other examples are possible. As explained elsewhere herein, various messages, alerts, or other types of information may be presented at the mobile electronic device 126.

[0098] Therapeutic personnel or equipment 128 is medical personnel (eg, doctors, nurses, etc.) and / or medical equipment (eg, cardiac monitoring equipment, surgical equipment, IV equipment) used by medical personnel to treat human patients.

[0099] Medical personnel or equipment 130 is a person or equipment used to help a person. For example, medical personnel or equipment 130 can be emergency personnel and vehicles sent to help person 124. Display 132 is any type of display device that presents information to a user in any visual form (e.g., graphics, numerical values, alphanumeric characters, or charts). For example, the display device can be of the same type as mobile electronic device 126, or can be a personal computer that is not mobile.

[0100] exist Figure 1 In one example of the operation of the system, the model 118 includes an encoding neural network, a function generator neural network, an estimation neural network, and a functional capacity estimator. The encoding neural network is configured to accept as input a time series of a selected plurality of cardiopulmonary variables obtained from sensors 105, 107, and 109, and generate a multi-element cardiac profile vector. The function generator neural network of the model 118 is configured to generate weights when supplied with the cardiac profile vector as input.

[0101] The estimation neural network of model 118 is configured to accept as input a time series of a selected subset of the plurality of cardiopulmonary variables obtained from sensors 105, 107, and 109. The subset omits at least one excluded variable (e.g., heart rate), and the estimation neural network is configured to generate an estimate of the at least one excluded variable. The estimation neural network is configured by control circuit 114 using weights output by the function generator network as weights for connections between nodes in the estimation neural network.

[0102] The concatenation and iterative training through a plurality of iterations are performed by control circuit 114 on the encoder neural network of model 118 to produce a trained encoder neural network (which is also stored in database 116), and the concatenation and iterative training through a plurality of iterations are performed on the function generator neural network to produce a trained function generator neural network (which is also stored in database 116). The training uses an error metric that includes a difference between at least one excluded variable and an estimated value of the at least one excluded variable generated by the estimation neural network.

[0103] For each of the iterations, a first duration of training data 150 comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from one of the training human subjects 104, 106, and 108 is input to the encoding neural network of the model 118. For each of the iterations, a second duration of training data 150 (substantially non-overlapping in time with the first duration) comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from the training human subjects is input to the estimation neural network of the model 118. The estimation neural network provides an estimate of at least one excluded variable that can be used in an error metric for training the encoding neural network and the function generator neural network.

[0104] The functional capacity estimator of the model 118 is configured to accept as input a cardiac profile vector generated by a trained encoding neural network. The cardiac profile vector may be augmented by a set of preselected physical property values ​​of the monitored human subject 124, and the functional capacity estimator outputs an estimate of a measure of functional capacity.

[0105] The functional capacity estimator of the model 118 is trained using a plurality of matched examples to produce a trained functional capacity estimator. Each of the matched examples includes a cardiac profile vector augmented by physical attribute values ​​from a person (if used in an embodiment), matched to a corresponding measured functional capacity metric of the person, to enable regression of the input of the physical attribute augmented cardiac profile vector to an estimated value of the functional capacity metric.

[0106] The trained encoding neural network and the trained functional capacity estimator are arranged to accept as input (via control circuit 114) monitored data 152 from a monitored human subject 124 for a duration of time obtained from sensor 122. The monitored data 152 includes a substantially continuous time series of a selected plurality of cardiopulmonary variables. The trained encoding neural network of model 118 generates a current cardiac profile vector for the monitored human subject 124. The current cardiac profile vector may be augmented by physical attributes of the monitored human subject 124, such as weight, gender, body mass index, body fat percentage, height, age, etc. The cardiac profile vector may additionally be augmented by an estimate of VO2Max calculated from the above physical attributes. Other physical attributes of the monitored human subject are also possible, such as measurements of lung volume, skeletal muscle mass, etc. The trained functional capacity estimator generates a current estimate of a measure of functional capacity (e.g., VO2Max) from applying the current cardiac profile vector augmented with physical attributes to the trained functional capacity estimator.

[0107] One or more actions are performed based on the current estimate of functional capacity. In one example, the current estimate in a time series with the previous estimate is displayed to a clinician (e.g., on a display 132 or mobile device 126) for checking possible health changes in the monitored human subject. In other examples, the current estimate is compared with the previous estimate and a test is performed for the change. If a change is detected, an alarm 154 is triggered to the clinician to investigate the health of the monitored person. The alarm 154 can be presented on the mobile device 126 or the display 132.

[0108] In other examples, the current estimate is compared to the previous estimate, and tests are performed for changes, whereupon a questionnaire 156 (electronic or paper) is sent to the monitored human subject 124. For example, the questionnaire 156 may be electronically displayed on the display 132 or device 126 while these elements are with the monitored person 124. In other examples, the current estimate is compared to the previous estimate and tests are performed to detect changes, whereupon an entry is noted in the medical record of the monitored human subject indicating that a change has occurred in the health of the monitored human subject. The record may be at the database 116 or at some other database or location.

[0109] In another example, the current estimate is compared with estimates from other monitored human subjects to quantify the health impact between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial. This information can be displayed to a user at devices 126 and 132, or annotated in database 116 or at some other database or location for collecting clinical trial results. In other examples, the operation of a medical device (e.g., equipment 120) is controlled by a control signal 162 sent by control circuit 114 or a parameter of the medical device is set by the control signal, wherein the medical device is associated with treating or monitoring a monitored human subject. In another example, a dispatch signal 158 can be sent to medical personnel or equipment 130.

[0110] In another example of an action, a treatment plan 160 may be created by the control circuit 114 and sent to a treatment provider or equipment 128. The treatment plan 160 may result in a medication being dispensed or medical equipment being adjusted, directed, or controlled.

[0111] In other aspects, data augmentation can be applied to data obtained from the training population 102. Data augmentation is the manipulation of existing data sets available for training to create a much larger set of examples for training. In the methods described herein, this can be achieved by selecting multiple start times for a 24-hour period including Day 1; selecting multiple offsets from the end of Day 1 to the beginning of Day 2 (e.g., immediately, delayed by 2 hours, delayed by 36 hours, delayed by 12 hours and 34 minutes, etc.); reversing the data of Day 1 and Day 2; and generating a day including N hours from another 24-hour period (e.g., from the morning of Day Z, from the afternoon of Day Z+1), which is from a 24-hour period with 24-N hours (which may be similar times of day). Other examples are possible.

[0112] Thus, a large number of Day 1 / Day 2 combinations can be generated from a modest number of subjects for whom one to many weeks of continuous data are available. Likewise, the number of VO2Max samples (used to train a VO2Max network as described elsewhere herein) can be similarly scaled up; a large number of Day 1 / Day 2 pairs can be created for each VO2Max value from a modest set of individuals for whom VO2Max results are available. All of these manipulations faithfully maintain data authenticity (i.e., still represent true pairings and are physiologically related) while presenting the network(s) with data variations that allow them to generalize their feature extraction.

[0113] The data for each of the channels (vital signs) may be normalized. In various aspects, this involves establishing a z-score parameter for each input variable (e.g., HR or RR) based on statistical values ​​across all samples of all individuals including the training data set. These statistical values ​​are then also used to convert all inputs (included in the estimation mode) to z-scores. The z-score or "standard score" for a given vital sign value is equal to the number of vital sign values ​​minus the mean of the vital sign across all samples, divided by the standard deviation of the vital sign across all samples.

[0114] Reference now Figure 2 , describes one example of a method for determining an estimate of VO2Max (or some other functional capacity measure) for a monitored human subject. Figure 2 The method utilizes an encoder neural network, a function generator neural network, an estimation neural network and a functional capacity estimator that together constitute the model.

[0115] At step 202, an encoding neural network is configured to accept as input a time series of a selected plurality of cardiopulmonary variables and generate a multi-element cardiac profile vector.

[0116] At step 204, a function generator neural network is configured to generate weights when supplied with a cardiac profile vector as input.

[0117] At step 206, the estimation neural network is configured to accept as input a time series of a subset of the selected plurality of cardiopulmonary variables. The subset omits at least one excluded variable. The estimation neural network generates an estimate of the at least one excluded variable, and the estimation neural network is configured using weights output by the function generator network as weights for connections between nodes in the estimation neural network.

[0118] In step 207, multivariate time series vital sign data of first duration and second duration are obtained for training. Each pair of durations represents temporally close data that represent the same physiological behavior of the training human subject in the vital sign data, typically unrestricted and across typical activities of daily life.

[0119] At step 208, iterative training of the encoding neural network and the function generator neural network is performed. Iterative training occurs in series and is performed by multiple iterations of the encoding neural network to produce a trained encoding neural network, and by multiple iterations of the function generator neural network to produce a trained function generator neural network. The training uses an error metric that includes a difference between at least one excluded variable and an estimated value of at least one excluded variable generated by the estimation neural network.

[0120] For each of the iterations, a first duration of training data comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects is input to the encoding neural network. A second duration of training data (substantially non-overlapping in time with the first duration, but proximate in time so as to represent the same physiological behavior) comprising a substantially continuous time series of a selected plurality of cardiopulmonary variables from training human subjects is provided as input to the estimation neural network, and wherein the estimation neural network provides an estimate of at least one excluded variable.

[0121] Within a given iteration, in step 210, the time series data of the first duration is input to the encoding neural network, which generates a cardiac profile vector. In step 212, the cardiac profile vector is input to the function generator neural network, and expected network parameters (such as weights) are generated for use by the estimation network as its weights. The network parameters generated by the function generator network can be any network parameters (including weights and biases) that define the neural network, but for the purpose of this description should be collectively referred to as weights. In step 214, the time series data of the second duration is input to the estimation neural network, which enables it to generate estimates of variables missing from the second duration input, as described elsewhere herein. In step 216, an error metric is determined based on the difference between the estimated variables and the retained (but known) variables from the time series data of the second duration. In step 218, the error metric is used to adjust the trainable weights and biases of the function generator neural network and the encoding neural network to move them toward producing outputs that enable the estimation network to produce better estimates of the retained variables. The iterative training continues at step 220 until the network in question is considered to be fully trained.

[0122] Then, at step 222, the functional capacity estimator is configured to accept as input the cardiac profile vector generated by the trained encoding neural network. At step 224, the following matching examples are obtained from multiple human subjects: (1) data for a duration (variables from the same sensor types and sample counts as substantially the same as those of step 207) and (2) measured functional capacity metrics collected from the same people using a reference "gold standard" measurement technique (such as cardiopulmonary exercise testing via a metabolic cart for metrics such as VO2Max). At step 226, the functional capacity estimator uses the multiple matching examples to produce a trained functional capacity estimator. For each pair, at step 228, the time series data for that duration is input to the encoding neural network to generate a cardiac profile. In step 230, the cardiac profile is input to the functional capacity estimator to generate an estimate of functional capacity that can be compared to the measured values ​​for the pair for error-driven training. If a neural network is used as a functional capacity estimator, the training may be iterated between steps 226 and 232 until it is considered trained, i.e., where the error is sufficiently low and network improvements have tapered off, as is known in the art. If the functional capacity estimator uses a different regression technique, the duration of functional capacity (and resulting cardiac profile) and a "gold standard" reference measurement are fitted as is known in the art to define the parameters that make up the functional capacity estimator.

[0123] In some embodiments, the cardiac profile vector of each of the matching examples may be augmented with physical attribute values ​​from the person, such as weight, gender, body mass index, height, age, etc. The augmented cardiac profile vector then becomes the input to the functional capacity estimator.

[0124] At step 233, the trained encoding neural network and the trained functional capacity estimator are deployed to accept as input, at step 234, monitored data for a duration of time from a monitored human subject. The data includes a substantially continuous time series of a selected plurality of cardiopulmonary variables. At step 236, a current cardiac profile vector is generated for the monitored human subject. The current cardiac profile vector may be augmented by physical attributes of the monitored human subject. From the current cardiac profile vector augmented with physical attributes applied to the trained functional capacity estimator at step 238, a current estimate of a measure of functional capacity is generated at step 240.

[0125] At step 242, one or more actions are performed based on the current estimate of functional capacity. In one example, the action is to display the current estimate in a time series with previous estimates to a clinician for reviewing possible health changes in the monitored human subject. In another example, the action is to compare the current estimate with the previous estimate and test the change, so as to trigger an alarm to the clinician to investigate the health of the monitored person. In yet another example, the action is to compare the current estimate with the previous estimate and test the change, so as to send a questionnaire to the monitored human subject. In another example, the action is to compare the current estimate with the previous estimate and test the change, so as to record an entry indicating that the health of the monitored human subject has changed in the medical record of the monitored human subject. In yet another example, the action is to compare the current estimate with the estimate from other monitored human subjects in order to quantify the health impact between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial. In another example, the action is to control the operation or set the parameters of a medical device associated with treating or monitoring the monitored human subject.

[0126] Reference now Figure 3 , describes basic aspects of a method for training a system of the present invention for determining an estimate of VO2Max (or some other functional capacity estimate) for a monitored human subject based on vital signs reflecting activities of daily living. The method utilizes training a neural network to perform a first task (i.e., inferred estimates of maintained vital signs such as heart rate) to produce a trained network, or subcomponent thereof, that is well suited to provide input to a second task (i.e., estimation of VO2Max or other functional capacity measure). Figure 3 A different two-step training process is depicted. Known input vital signs 304 (including the input vital sign heart rate (HR)) are obtained from samples 302. Samples 302 include a large group of samples from, for example, hundreds of different human subjects over tens of thousands of hours. The input vital signs 304 are applied to a neural network 306 to produce an estimate of HR 307. The estimate 307 is compared to the actual value 305 and the difference (cost function) 309 is obtained. The cost function 309 represents the error that is back-propagated to the neural network 306 to train at least a portion of the neural network. In this way, tens of thousands or hundreds of thousands of samples from at least hundreds of human subjects are used to train the neural network 306.

[0127] The neural network 306 may be understood as comprising a composed neural architecture. Once trained, a subset 310 of the neural network 306 (e.g., the aforementioned encoding neural network) has applied vital signs 312 from a human subject for which a known reference VO2Max has been measured. Applying the vital signs 312 to the subset 310 produces a cardiac profile vector 314. Using tens or hundreds of matched cardiac profile vectors 314 and known VO2max samples 316, a regression to an estimated VO2Max 318 may be trained or calculated. Thus, once the above-described set of networks is trained to generate a cardiac profile vector, a regression of the cardiac profile may be performed using a much smaller set of VO2Max results to obtain an estimated VO2Max 318. This may be accomplished by any of a number of known regression techniques.

[0128] For example, linear regression may be used. Cardiac profile 314 is a vector of, for example, 100 elements of daily vital sign data from human subjects 312. At 316, the data may be fit to the corresponding measured VO2Max results for each human subject, resulting in a linear model at 318 that maps the person's cardiac profile vector to an estimated VO2Max.

[0129] Decision trees, random forests, or gradient boosting models may also be used. Decision trees may be configured to output a continuous range of parameter estimates (VO2Max) for each cardiac profile as input. Random forests and gradient boosting are extensions of the basic decision tree approach.

[0130] A neural network estimator may be used at regression 318, as described below with respect to FIGS. 5 and Figure 6 and Fig.15 In this case, the error in estimating VO2Max compared to the known samples from 316 can be back-propagated to tune the weights of a neural network that takes the cardiac profile 314 as input. A fully connected network can play this role.

[0131] Reference now Figure 4 , describes an example of a method for determining an estimate of VO2Max (or some other functional capacity value) for a monitored human subject. In various aspects, this represents a method for training Figure 3 A more detailed version of the method of neural network 306.

[0132] A known sample 402 of vital signs from several days of hundreds or more human subjects includes vital signs 404 for the first day ("Day 1") for all variables (e.g., including heart rate, respiratory rate, etc.). The known sample also includes vital signs for the second day ("Day 2"), which includes one variable 408 and a set 406 of all vital signs except one variable 408 (in this case, heart rate (HR)). It should be understood that Day 1 and Day 2 do not necessarily refer to information obtained on the same day, but rather information obtained across different (e.g., non-overlapping) time periods, whether on the same calendar date or on different calendar dates. Data for Day 1 also does not necessarily occur before data for Day 2. Data for Day 1 and Day 2 from the same subject should be used together; rather than data for Day 1 from one subject and data for Day 2 from another subject being used together.

[0133] The day 1 vital signs 404 are applied to an encoding neural network 410. The encoding neural network 410 produces a cardiac profile vector 412 that is applied to a function generator network 414. The function generator network 414 generates weights 416 that are used to set weight values ​​in a generated estimation neural network 418.

[0134] The day 2 vital signs 406 with all variables except one (HR in this case) are applied to the generated estimation network 418 to produce an estimate of the retained variable 420. The actual HR value is compared to the estimated HR 421 by a difference function 422 to produce an error 423. The error 423 is back-propagated to the encoder neural network 410 and the function generator network 414 to further train these networks.

[0135] The encoding neural network 410, the function generator network 414, and the estimation neural network 418 are relatively deep neural networks, and in one specific example are convolutional neural networks (CNNs). CNNs are optimized feature finders and extract important intrinsic features in a noisy background to perform the best efficient classification or estimation of the holdout variables, such as Day 2. The feature extraction and mapping capabilities of CNNs utilize linear algebra implemented by input values, intermediate nonlinear or linear activation functions at "nodes" in the network, and "weights" between these nodes to modulate the activation values. The final feature extraction and mapping capabilities of the trained neural network are tied to the learned weights.

[0136] Throughout the day (or some other time period), a person performing an ADL exhibits dynamic changes in these vital signs due to effort. For example, depending on cardiopulmonary fitness, an individual's response spike may be higher / lower, delayed / undelayed; and recovery from effort may be normal / extended. Absolute effort levels may vary between individuals. These are hints that an interpretable metric needs to be detected and mapped to cardiopulmonary fitness. The data includes vital signs or samples 402.

[0137] The encoding neural network 410 is configured to learn the relationships between the vital signs from the "Day 1" data and generate a cardiac profile vector that characterizes the cardiopulmonary dynamics of the individual determined from the Day 1 data.

[0138] The function generator neural network 414 is configured to generate network parameters (such as weights and / or biases) of another neural network as its output. It takes the cardiac profile vector as input.

[0139] The generated estimation neural network 418 estimates the HR for the data of day 2 from the input of the data of day 2 excluding the HR. Its weights are dynamically generated by the function generator neural network 414 (based on the cardiac profile it receives from the encoding neural network 410). The generated estimation neural network 418 does not independently "learn" its own weights. Therefore, the generated estimation neural network 418 produces an estimate of the excluded variable (e.g., heart rate) 420 after inputting the day 2 vital signs 406. The difference operator 422 takes the difference between the actual value 408 and the estimated value 420 of the excluded variable to produce an error or cost function value 423.

[0140] Only the cost function 423 is used to train the encoding neural network 410 and the function generator neural network 414. The generated estimation neural network 418 is not trained; it obtains the weights and biases of its node-node connections dynamically generated by the function generator neural network 414. However, it is the output of the estimation neural network 418 that creates the cost function value 423, i.e., the error between the estimated vital sign 420 and the actual measured value 408 of the vital sign. These errors 423 are not propagated back to the estimation neural network 418, but are propagated back to the function generator neural network 414 and the encoding neural network 416.

[0141] In various aspects, training 414 of the encoder neural network 410 and the function generator neural network proceeds as follows.

[0142] For training, iterate over the following:

[0143] (1) Input the data 404 of day 1 into the encoding neural network 410. Create a cardiac profile vector 412.

[0144] (2) The cardiac profile vector 412 is input to the function generator neural network 414, which generates weights and / or biases for the estimation neural network 418.

[0145] (3) The data 406 for day 2 (excluding the vital signs to be estimated) is input to the estimation neural network 418, which outputs 420 the estimated values ​​of the excluded vital signs.

[0146] (4) Determine error 423 as the difference between estimated vital signs 420 and known day 2 vital signs 408.

[0147] (5) The error 423 is back-propagated into the function generator neural network 414 and the encoding neural network 410 to adjust their internal weights.

[0148] The training improves the skill of the encoding neural network 410 and the function generator neural network 414 in creating weights that allow the estimation neural network 418 to effectively output accurate vital sign estimates. The data from Day 1 is utilized in the encoding neural network 410. In these aspects, the encoding neural network 410 and the function generator neural network 414 do not "see" the data from Day 2, except through the indirect process of error back propagation. Thus, the encoding neural network 410 is effectively trained to generalize a set of features (cardiac profile) that represent general cardiopulmonary function.

[0149] In various aspects, the set of encoding neural network 410 / function generator neural network 414 is trained using day pairs from a population of individuals. The example individuals should span the range of health states expected to be seen in use. The day pairs do not need to be in sequence, nor do they need to be adjacent calendar dates - but the day pairs need to be from the same individual and need to be close enough in time to represent individuals in substantially the same health state. For example, using day 1 from a heart failure patient that is 6 months earlier in time than day 2 from that patient may be problematic because the patient's health may have changed over that 6 month time frame.

[0150] To summarize the above method, the training of the encoding neural network 410 and the function generator neural network 414 is completed using a cost function. Once the network architecture is established (e.g., multiple layers, nodes in each layer, interconnectivity), the weights between the connected nodes are randomly initialized. As described above, an example is input into the network, the output of the network is compared with the known label or value associated with the input example, and the cost is determined using the cost function (output error compared to the known true label / value), and the cost (error) is propagated back through the network, the weights in the network at each layer are adjusted, and convergence testing is performed. When sufficiently converged, the weights are frozen, and specifically the trained encoding neural network (the weights in its weight table are frozen) is deployed as part of an estimator of functional capacity, whose regression training is combined with Figure 3 318 in are described.

[0151] HR is included as an input to the cardiac profile generated by the neural network 410 by training on day 1 data, and this is then generalized by using the learned relationship to estimate HR for day 2 data (where HR was not included as an input). In this way, the error on data not seen by the system is minimized. A cost function - the error between the estimated HR and the actual HR - is generated using data from the "unseen" day (day 2 data 406) to improve generalization.

[0152] It will be appreciated that the generated estimation neural network 418 can be configured to estimate all of the vital sign inputs for day 2, and not just the HR. (A penalty for inaccuracy is of course imposed on the upstream function generator neural network 414 and encoding neural network 410). This will be done on a round-robin basis, leaving out one vital sign and estimating it using the generated estimation neural network 418, and then comparing the estimate for that excluded vital sign to the actual vital sign measurement. Since the generated estimation neural network 418 architecture is fixed, and the weights generated by the function generator neural network 441 need to work on each vital sign on a round-robin basis to make the error measurements, the architecture must accommodate all inputs and outputs, yet each variable needs to be kept in turn to allow for inferred estimates. This is accomplished by: (1) having the full input vital sign tensor (all minutes, all vital signs), but masking the vital signs to be excluded by setting their values ​​to zero; (2) having the network output estimates for all vital signs, but using only the masked out vital signs at the input in the cost calculation.

[0153] Go to Figure 5A , describes another example of training the neural network of the present invention. In this example, the training of the encoding neural network and the function generator neural network is combined with the training neural network to be used as a functional capacity estimator for VO2Max estimation as shown. As described below, the cost function is a combination of the error from the vital sign estimate and the VO2Max estimate on day 2. As before, the goal is to not require a large number (e.g., tens of thousands) of VO2Max results.

[0154] Figure 5A Many components and Figure 4 Like numbered elements are identical and perform the same functions as described above with respect to Figure 4The same functions as described by the same elements in . These include a known sample 502, which includes vital signs 504 for day 1 for all variables (e.g., including heart rate, respiratory rate, etc.). The known sample also includes vital signs 506 for day 2, except for one retained variable 508 (heart rate (HR) in this case; and in the round-robin method, the retained variables for each round). It should be understood that day 1 and day 2 do not necessarily refer to information obtained on the same day, but rather information obtained across different (e.g., non-overlapping) time periods, whether on the same calendar date or on different calendar dates, but close enough in time to represent substantially the same physiological function of the person from whom the information was obtained. The data for day 1 does not necessarily appear at a time before day 2.

[0155] The day 1 vital signs 504 are applied to an encoding neural network 510. The encoding neural network produces a cardiac profile vector 512 that is applied to a function generator network 514. The function generator network 514 generates weights 516 that are used to set weight values ​​in a generated estimation network 518.

[0156] The day 2 vital signs 506 with all but one variable (HR in this case) are applied to the generated estimation network 518 to produce an estimate 521. The actual HR value is compared to the estimated HR value 521 by a difference function 522 to produce an error 524 estimated for day 2.

[0157] Figure 5A The system includes a VO2Max estimation neural network 530. The cardiac profile vector 512 is input to the VO2Max estimation neural network 530 along with the individual's physical attributes 531 to produce an estimated VO2Max 532 that is applied to a difference operator 534. The difference operator 534 takes the difference between the estimate 532 and an actual VO2Max sample 536 (measured with a "gold standard" reference method and temporally associated with the physiology of the human subject represented in the data on Day 1 and Day 2) to produce an error output 538. A cumulative cost function operator 540 combines the output 538 with the output 524 to produce a total error 542, which is applied to the training encoding neural network 510, the function generator neural network 514, and the VO2Max estimation neural network.

[0158] Training samples representing a large number of individuals will not have VO2Max results. A relatively small number of samples will have VO2Max results. When a VO2Max result is not available for a sample, the error in VO2Max will be set to zero so that there is no driver for neural weight adjustments from that aspect of the system.

[0159] Once the encoding neural network 510, the function generator neural network 514, and the VO2Max estimation neural network 530 are fully trained, only the encoding neural network 510 and the VO2Max estimation neural network 530 need to be deployed for VO2Max estimation on new subject data. Other regressors (such as linear regression decision trees, etc.) as mentioned above can be substituted for the VO2Max estimation neural network 530. Figure 6 See the architecture.

[0160] exist Figure 5B An alternative embodiment for training the encoding neural network 510 together with the VO2Max estimation neural network 530 is shown in FIG. Figure 5B Parts and Figure 5A The components in are the same with minor modifications. Figure 5A In the example, the encoding neural network 510, the function generator neural network 514, and the VO2Max estimation neural network 530 are all trained together using the total error 542. Figure 5B , this is broken down into two training phases. First, the encoder neural network 522 and the function generator neural network are trained together using the error 524 output from the difference function 510 for the estimate for day 2. Once this training is complete, in the second phase, the encoder neural network 510 is trained together with the VO2Max estimation neural network 530 (and without involving the function generator neural network 514) using the error output 538 produced by the difference operator 534 when comparing the estimated VO2Max 532 and the actual VO2Max samples 536. Thus, in this case, the encoder neural network 510 is trained to a state (part of a system that minimizes the estimate for day 2 of the vital signs being maintained), and then from that state as a starting point, is further trained as part of a combination that minimizes the error in estimating VO2Max.

[0161] Reference now Figure 6 , describes an example of a system for estimating a functional capacity value (in this case VO2Max). The 24-hour vital signs 602 (or vital signs for some other period of time) of a monitored human subject are applied to a trained encoding neural network 604. The encoding neural network 604 has been trained, for example, as described with respect to Figure 4 , Figure 5A and Figure 5B Applying vital signs 602 to trained encoding neural network 604 produces cardiac profile vector 606.

[0162] Physical attribute values ​​608 may also be used by the VO2Max network 610. In addition to the cardiac profile as an input to the VO2Max estimation regression, the VO2Max regression is also performed with physical attribute values ​​(height, weight, gender, age, body fat, etc.). In some aspects, these additional inputs are critical to the success of the VO2Max algorithm: an individual's VO2Max score includes their cardiorespiratory fitness as well as their physical attributes (such as the size of their lungs). Information about their cardiorespiratory fitness is captured by the cardiac profile, but the subject's physical attributes must be provided in addition to make an accurate VO2Max estimate.

[0163] Vector 606 is applied to a trained VO2Max network 610, which produces an estimated VO2Max 612. VO2Max network 610 may use Figure 5A or Figure 5B The cost function described here or trained in some other way.

[0164] Therefore and instead of measuring a large number (e.g., thousands) of VO2Max results (which may be too time consuming and / or expensive), different values ​​(e.g., heart rate) are estimated, many examples of which are readily and cost-effectively available.

[0165] As mentioned, heart rate is an example of a variable to be monitored. In some aspects, the neural network of the present method is trained to estimate heart rate from other vital signs, where the cost function is the mean square error between the estimated heart rate and the actual heart rate. Compared with capturing functional capacity metrics (such as VO2Max) in a clinical laboratory environment, a large amount of heart rate and other vital sign data from many human subjects can be easily and cheaply obtained. With the method provided herein, there is no need to calculate variables (e.g., VO2Max) for all those subjects, because the neural network is now trained to simply be good at estimating heart rate from other data rather than VO2Max. In doing so, the neural network combination provided herein will effectively extract the same valuable features from the input data to do a good job in HR estimation, just as in order to do a good job in VO2Max estimation.

[0166] While the HR estimated from the neural network may be useful in itself, it alone cannot regress the VO2Max samples in a useful way. What is obtained from the encoding neural network is an intermediate result that represents the "features" that the encoding neural network has identified as important for estimating HR. It is assumed that these features are equally important for estimating the measurement of cardiopulmonary performance and therefore the measurement of VO2Max. Features can be found upstream of the HR estimation output of the generated estimation neural network, at a higher layer, at the output of the encoding neural network, as the activation of the nodes in the layer. The node layer provides a vector of activation values ​​that indicates the presence or absence of the features that the neural network has learned to extract. The activation layer (for example, set somewhere between 40 and 100 nodes in the architecture) includes a cardiac profile vector of the individual from which the input data is obtained. In other examples, the activation layer includes about one thousand nodes. However, it will be understood that any number of nodes can be used. As is typical in deep neural networks, these features themselves are not necessarily suitable for direct interpretation-but include vectors of values ​​that still represent key attributes of cardiopulmonary behavior in ADL vital signs.

[0167] For example, a cardiac profile vector of 100 node activations at a critical middle layer can then be routinely regressed to estimate VO2Max based on regression using only tens (50-200) samples. This is far less than the 10,000-1,000,000 samples required to train a neural network to estimate VO2Max directly. Furthermore, the cardiac profile can be easily regressed against other outputs (such as health category or disease progression stage) instead of VO2Max.

[0168] Once the collection of networks described above is trained to generate cardiac profile vectors, these vectors can be regressed for classes of interest. A smaller set of class discoveries for an individual can be used with the present invention than would be necessary if a typical neural network were trained directly on the classes. Regression for classes of interest can be performed by a variety of techniques: neural network classifiers, decision trees / random forests / gradient boosting, support vector machines, and K-nearest neighbors. Other examples are possible.

[0169] In deployed use and in one example, the system accepts 24 hours of input of vital signs along with physical attribute values ​​for the individual and estimates VO2Max represented by the dynamics of the vital signs over that period. The estimation mode utilizes both population learning (training data from many people, regressed to a "ground truth" VO2Max result) and a personalized cardiac profile generated from the individual's ADL vital signs to map to the individual's instantaneous estimated VO2Max.

[0170] Similarly, the cardiac profile vector may be regressed against categories, such as the New York Heart Association (NYHA) functional classification; the Global Initiative for Chronic Obstructive Lung Disease (GOLD) criteria for classification of severity of airflow limitation in COPD; or other cardiopulmonary fitness category rating systems. It may also be regressed against other continuous variables commonly measured in clinical laboratory settings, such as forced expiratory volume in 1 second (FEV1), residual lung function capacity (FRC), residual lung capacity (RV), maximum heart rate, VE / VCO2 slope, heart rate at anaerobic and aerobic thresholds, cardiac output, stroke volume, ejection fraction, and predicted 6-minute walk test (6MWT) distance. The same principle of the present invention applies to all cases: since the encoding neural network is trained using a large number of training samples across a population of subjects to generate a cardiac profile vector that encapsulates the basic features of the learned cardiopulmonary control system (interaction of vital signs in ADL), the final regression or training of the second stage machine learning tool for mapping the profile to the desired category or continuous variable can be performed with much less training data.

[0171] In one use case, a continuous day-by-day estimate of an individual's VO2Max can be used to track an individual's cardiorespiratory fitness. This can be particularly useful for patients with chronic diseases who are at risk of evolving from a stable state to an exacerbated disease state requiring hospitalization and acute care within a few days or weeks. By tracking the estimated VO2Max over time, a drop can be detected, which may be an early warning sign that the patient is deteriorating and may soon require acute care. Clinicians can intervene earlier to avoid acute care events.

[0172] In another use case, clinical trials may be conducted to determine the effects of a drug, (implanted) device, or other procedure or intervention on the health of the subjects under study. An estimated VO2Max can be used to replace conventional health measures, such as a true VO2Max test (too expensive to perform multiple times on subjects in a study) or a 6-minute walk test (6MWT) (too crude and inaccurate), and provide a more accurate and more frequent assessment of cardiorespiratory fitness status. This can be a state-of-the-art technique for evaluating the efficacy of a drug or device between a control and test group in a trial, and can provide more data samples for each subject to improve the extraction of the effect of the intervention, i.e., improve "extracting the signal from the noise".

[0173] In another use case, an insurance company interested in ensuring that its insured members are receiving adequate preventive treatment for evolving disease conditions that could be expensive if not managed could assign sensor patches to its members to be worn for a week and to upload data for analysis. In this case, the cardiac profile could be regressed against health categories previously determined from training samples. Once the insured member has worn the patch and uploaded data for analysis, the present invention can match the individual to the category indicated by the cardiac profile. This can identify members of the insured population who previously had undiagnosed chronic conditions and help get the condition managed.

[0174] Reference now Figure 7 , describes an example of a cardiac profile vector 700. Vector 700 includes multiple node output values ​​702, which have numerical values ​​(e.g., real numbers or integers). These values ​​are generated by means of an activation function 704, which represents the activation level of a node 706 at the output layer of the encoding neural network. The activation function can take the form of a function known in the art, such as a sigmoid, inverse tangent, hyperbolic tangent, rectified linear unit (ReLU), leaky ReLU, exponential linear unit (ELU), etc. The vector can be generated by the encoding neural network described herein. The number of such nodes (and therefore the length of vector 700) can be on the order of tens or hundreds of nodes, and in one embodiment thereof is 100 nodes. Each such node 706 receives inputs 708 from other nodes at higher or upstream layers of the network, which represent the activation values ​​of those nodes. The incoming activation values ​​are multiplied by the learned weights 712 to produce a number that is summed at each node 706. In this way, the incoming activation value from any given upstream node can be amplified or attenuated by the weight on that connection as it contributes to the overall sum of the incoming activations for the node 706. A bias value 714 is also added to this sum for each node 706. As the neural network is trained by back-propagation of error, the learned weights 712 and biases 714 are updated. Adjustments to these weights and biases improve the performance of the neural network based on the task through which the error is measured.

[0175] In various aspects, field 702 represents an intermediate result encoding features that the neural network has identified as important for estimating a variable such as HR. It is assumed that these features are also important for estimating cardiorespiratory performance and therefore VO2Max or other functional capacity variable or category of measurement. As mentioned, the features can be found upstream of the HR estimation output of the neural network, at a higher layer, as activations of nodes in that layer. This layer of nodes will provide a vector of activation values, which will represent the presence or absence of the features that the neural network has learned to extract. The activation layer (set somewhere between 40 and 100 nodes in the architecture) is the "cardiac profile" vector 700 of the individual from which the input data is obtained.

[0176] Reference now Figure 8 , describes a method for training a neural network for estimating functional capacity (e.g., VO2Max). It will be appreciated that Figure 8 The method presented in represents one iteration of the entire training process, and Figure 8 The method described in is repeated many times (e.g., in one example, 100,000 times) to train the neural network. In practice and in a first iteration, several hundred duration data pairs are selected from multiple subjects and performed Figure 8 Then, in the second iteration, several hundred other data pairs are selected and the Figure 8 After multiple (e.g., 100,000) iterations, the entire process is complete.

[0177] At step 852, data from four individuals from a training population is assembled. The data includes data 802 from a first person, data 804 from a second person, data 806 from a third person, and data 808 from a fourth person. A random pair of duration data is selected from each data set 802, 804, 806, and 808. Specifically, a first random pair of duration data (810, 812) is selected from the first person data 802; a second random pair of duration data (814, 816) is selected from the second person data 804; a third random pair of duration data (818, 820) is selected from the third person data 806; and a fourth random pair of duration data (822, 824) is selected from the first person data 808. As mentioned, there may be other pairs, but only four pairs are shown for visualization purposes.

[0178] It will be appreciated that each of these pairs includes a first duration and a second duration. The first duration and the second duration need not be consecutive in time; for some pairs, the first duration occurs (in time) before the second duration, and for other pairs, the first duration occurs (in time) after the second duration. In this example, the first duration data includes data 812, 816, 818, and 822. The second duration data includes data 810, 814, 820, and 824.

[0179] At step 854, the data for the first duration (812, 816, 818, and 822) is passed to an encoding neural network 870 (which is the same as the other encoding neural networks described herein). This responsively generates cardiac profile vectors 871 (corresponding to the first person), 872 (corresponding to the second person), 873 (corresponding to the third person), and 874 (corresponding to the fourth person). As discussed elsewhere herein, the vectors may be vectors of real numbers representing the cardiopulmonary characteristics of the person.

[0180] At step 856, vectors 871, 872, 873, and 874 are applied to a function generator network 876, which generates weights used by an estimation neural network 877. Estimation network 877 is actually a first estimation network 878 (with parameters tailored for vector 871), a second estimation network 879 (with parameters tailored for vector 872), a third estimation network 880 (with parameters tailored for vector 873), and a fourth estimation network 881 (with parameters tailored for vector 874).

[0181] At step 858, the data of the second duration (810, 814, 820, and 824) is further processed so that at least one variable (e.g., heart rate) is excluded from the data. The data may be labeled (or identified in some other manner) and data associated with the excluded variable removed.

[0182] At step 860, the data (810, 814, 820, and 824) for the second duration, now with the excluded variable removed, is sent to an estimation neural network 877, which creates estimates of the excluded variable (e.g., heart rate) for each of the four individuals. More specifically, data 810 (with the excluded variable removed) is applied to a first estimation network 878 (with parameters customized by vector 871), data 814 (with the excluded variable removed) is applied to a second estimation network 879 (with parameters customized by vector 872), data 820 (with the excluded variable removed) is applied to a third estimation network 880 (with parameters customized by vector 873), and data 824 (with the excluded variable removed) is applied to a fourth estimation network 881 (with parameters customized by vector 874).

[0183] At step 862, the cost (mean square error between the excluded variable and the estimated excluded variable) is determined. This can be performed for each individual and the total error is calculated as the sum.

[0184] At step 864, the error calculated at step 862 is back-propagated to the function generator neural network 876 and the encoding neural network 870. The propagation in these networks adjusts the weights in these networks so that the cost (error) produced by these networks is reduced.

[0185] Reference now Fig. 9, describes an example of obtaining training data. Cardiopulmonary variables 902, 904, 906, and 908 are collected over time and may represent vital sign values ​​determined at a common sampling rate (such as once per minute or once every 15 seconds, as two examples). Such cardiopulmonary variables are recorded together. The total number of vital signs may of course be more than just four. The recorded cardiopulmonary variables from individuals 910, 912, 914, and 916 form the training data, with the total data contribution from each person varying in the order of days to weeks. Return to reference Figure 8 , the cardiopulmonary variables are collected as data sets for a first person 802, a second person 804, a third person 806, and a fourth person 808 among many subjects. In total, the training data consists of data from hundreds of people who have completed multiple weeks of data.

[0186] Reference now Fig.10 , describes an example of an architecture for an encoding neural network. The vital signs 1002 (e.g., from Figure 8 The first duration referred to in ( ) is applied to the first layer 1020 of the encoding neural network. The first layer includes rows and columns, where each row represents the value of a particular vital sign over time, and where the "columns" of data across the vital sign inputs are consistent in time. The input layer forms a 2D "image" of sorts, where the dimensions are (1) vital sign times (2) time. The architecture of the network is a convolutional neural architecture, and therefore, includes a set of node-to-node weights (e.g., in the filtering mode of Figure 7 ) is convolved with the input "image" by applying it to a region 1018 of the input and then sliding it to the next region of the input to apply it there, and so on, thereby convolving it with all the input node activations. Region 1018 should contain all variables (all rows) and may contain multiple consecutive columns (e.g., time spans). The set of weights that make up the filter pattern may be different from each other, but their values ​​are maintained during the convolution. It can be seen that a convolution operation is performed on the first region 1018, where the set of weights is applied to the corresponding input layer values ​​in region 1018, and the weighted values ​​are summed to form entries 1004 in the second layer 1022. The regions in the second layer are also convolved with different filter patterns specific to that layer, and summed to obtain entries 1006 in the third layer 1024. This continues until the fully connected layer 1026. At the fully connected layers, in contrast to the convolutional layers, all upstream nodes are fully connected to all downstream nodes and have their own individual weights (as known to those skilled in the art), so that no filtering pattern of weights is used at these layers. The cardiac profile vector 1028 is the output of two or more fully connected layers.

[0187] Reference now Fig.11, we will describe in more detail the typical convolution operations used in encoding neural networks. Fig.10 The layered layer 1020 is represented as a series of rows and columns in a table as the values ​​of the input data. The weight table 1104 includes Fig.10 The filtering pattern described is applied to region 1106 of layer 1020. The weight table has as many rows (variables) as layers 1020, but does not extend across all columns (all time samples) of layer 1020, but only covers a certain continuous span of time. The value is obtained by multiplying the weight from the weight table with the corresponding value in the layer. The weighted values ​​are then summed to provide an output value at the corresponding node of the next layer.

[0188] In this example, row 1, column 1 from the weight table is multiplied by row 1, column 1 of layer 1020. Row 1, column 2 from the weight table is multiplied by row 1, column 2 of layer 1020. Then, row 2, column 1 from the weight table is multiplied by row 2, column 1 of layer 1020. This process continues as shown to produce a weighted sum (e.g., Fig.11 The values ​​shown are 1), where there is a weighted sum for the area filtered by the filter mode.

[0189] Then, the weight table is slid to the right by one column (one time sample) and the process is repeated to obtain the value 2, as Fig.11 As shown. This is the process of convolving the filter pattern (weight table) across the input "image" to produce the output value for each step. This process is the same not only for the input layer (input "image"), but also for deeper convolutional layers. Note that the weights are tabulated filter patterns that do not change when the filter pattern is convolved across layers. In all respects, the weights are static for each training iteration. The weights in the filter pattern weight table are updated (via the back-propagated error function as described herein) to reduce the error. Once training is complete, the weights do not change.

[0190] Reference now Fig.12 , depicts an example of a function generator neural network. A cardiac profile vector 1202 is a fully connected input layer to the first layer 1204 of the network. Connections 1206 are shown as generating outputs 1208, 1210, 1212, and 1214, whose values ​​will be used as weights and biases for the estimation network ("generated weights"). The function generator network is fully connected at each layer; that is, unlike in a convolutional layer, the weights connecting each upstream node to each downstream node are independent.

[0191] Reference now Fig.13 , showing the application of the second duration data to the estimation neural network. The second duration data 1302 (from a person's vital signs excluding one variable (such as heart rate)) is applied to the first layer 1304 of the estimation neural network. Fig.10 As depicted, layer 1304 includes an "image" with variables that make up the rows and span the columns. The estimation neural network also includes convolutional layers. The weights (from Fig.12 The weights from weight table 1306 (provided as from Fig.12 The output weights 1208 of Fig.12 The output bias 1210 of ) is applied in the convolution process with region 1308 as described above to produce entries 1310 in the second layer 1312. Similarly, the weights from weight table 1314 (provided as from Fig.12 The output weights 1212) and corresponding biases 1315 (provided as output from Fig.12 The output deviation 1214 of the function generator neural network is applied with the region 1316 as described above to produce entries 1318 in the third layer 1320. Layer 1320 is an estimate 1322 of the value of the excluded vital sign parameter (e.g., heart rate) over the entire second duration. The estimate 1322 is compared with the actual value of the heart rate 1324 to produce a cumulative error (average, sum, or other statistical measure), which is used to back-propagate the error to the encoder neural network and the function generator neural network, as described elsewhere herein. The error adjusts the weights of these networks (e.g., in the weight tables of the filter patterns used in the convolutional layers, or in the node-to-node connections in the fully connected layers).

[0192] Reference now Fig.14 , depicts an example of applying generated weights 1402 (from a function generator neural network) to a weight table in an estimation neural network. Generated weights 1402 may be a vector of values ​​that populate weight table 1404. The first value is applied to row 1, column 1, the second value is applied to row 2, column 1, the third value is applied to row 3, column 1, and so on.

[0193] This time, the convolution method (the above Fig.10 and Fig.11 The weights are applied to layer 1406 of the estimation neural network (described above). These are convolved as already described to form entries 1408 in the second layer 1410. When the second duration data (modified to exclude the variable) is input to the estimation neural network, this continues to produce an estimate of the excluded variable.

[0194] Reference now Fig.15 , describes an example of a VO2Max estimation network 1500. Fig.15As shown, a cardiac profile 1502 is input into a VO2Max estimation network 1500 along with an enhanced physical property value 1504. The VO2Max estimation network 1500 is composed of several fully connected layers 1506, which ultimately form an estimated VO2Max value 1508.

[0195] It will be appreciated that other neural network architectures commonly used in the art for image and sequence analysis may also be used in place of or in combination with the convolutional neural networks described in the encoding neural network and the estimation neural network. In other embodiments, one or both of the encoding neural network and the estimation neural network may take the form of a causal relationship preserving network, such as a recurrent neural network, a temporal convolutional network, or a transformer (attention) network, etc.

[0196] Preferred embodiments of the present disclosure are described herein, including the best mode known to the inventors. It should be understood that the illustrated embodiments are exemplary only and should not be taken as limiting the scope of the appended claims.

Claims

1. A method for improving a person's cardiopulmonary health, the method comprising: configuring the encoding neural network to accept as input the time series of the selected plurality of cardiopulmonary variables and to generate a multi-element cardiac profile vector; configuring a function generator neural network to generate weights when supplied with the cardiac profile vector as input; configuring an estimator neural network to accept as input a time series of a subset of the selected plurality of cardiopulmonary variables, the subset omitting at least one excluded variable, and to generate an estimate of the at least one excluded variable, wherein the estimator neural network is configured using the weights output by the function generator neural network as weights for connections between nodes in the estimator neural network; connecting the encoder neural network and the function generator neural network in series and iteratively training the encoder neural network over a plurality of iterations to produce a trained encoder neural network and iteratively training the function generator neural network to produce a trained function generator neural network, the training using an error metric comprising a difference between the at least one excluded variable and the estimated value of the at least one excluded variable generated by the estimator neural network, wherein for each of the plurality of iterations and for each of the plurality of training human subjects, inputting into the encoding neural network training data of a first duration comprising a continuous time series of the selected plurality of cardiopulmonary variables from a training human subject, providing input to said estimation neural network training data comprising a continuous time series of said selected plurality of cardiopulmonary variables from said training human subjects for a second duration that does not overlap in time with said first duration, and wherein said estimation neural network provides said estimated value for said at least one excluded variable; configuring a functional capacity estimator to accept as input a cardiac profile vector generated by said trained encoding neural network and to output an estimate of a functional capacity metric, training the functional capacity estimator using a plurality of matched examples to produce a trained functional capacity estimator, each of the matched examples comprising a cardiac profile vector matched to a corresponding measured functional capacity metric of the person to enable regression of an input of the cardiac profile vector to an estimate of the functional capacity metric; deploying the trained encoding neural network and the trained functional capacity estimator to accept as input a duration of monitored data from a monitored human subject, the monitored data comprising a continuous time series of the selected plurality of cardiopulmonary variables, generating a current cardiac profile vector for the monitored human subject, and applying the current cardiac profile vector to the trained functional capacity estimator to generate a current estimate of a functional capacity metric; as well as Based on the current estimate of the functional capacity metric, performing one or more actions selected from the group consisting of: displaying the current estimate along with previous estimates in a time series to a clinician for reviewing health changes in the monitored human subject; comparing the current estimate to a previous estimate and testing for changes, triggering an alert to a clinician to investigate the health of the monitored human subject; comparing said current estimate to a previous estimate and testing for changes, sending a questionnaire to said monitored human subject; comparing the current estimate to a previous estimate and testing for changes, making an entry in the medical record of the monitored human subject indicating a change in the health of the monitored human subject; comparing the current estimate with estimates from other monitored human subjects to quantify the health effects between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial; and Controlling the operation or setting parameters of a medical device associated with the monitored human subject.

2. The method of claim 1, wherein the functional capacity is VO2Max.

3. The method of claim 1, wherein the time series of a selected plurality of cardiopulmonary variables is obtained by sensors non-invasively affixed to a person's skin.

4. The method of claim 1, wherein the time series of the selected plurality of cardiopulmonary variables is obtained by a sensor, and the sensor is selected from the group consisting of: a camera, an electrocardiogram sensor, a light sensor, an accelerometer, an RF energy sensor, a temperature sensor, and a gyroscope sensor.

5. The method of claim 1, wherein the functional capacity estimator generates the current estimate of the functional capacity metric using one of the following: a linear regression technique, a decision tree, a random forest estimate, or a gradient boosting model.

6. The method of claim 1, wherein the functional capacity estimator comprises estimating a neural network.

7. The method of claim 1, wherein the cardiopulmonary variable comprises a measure of physical activity or exercise.

8. The method of claim 1, wherein the cardiopulmonary variables include one or more of the following: heart rate, temporal heart rate variability, respiratory rate, tilt angle, atrial fibrillation probability, and sample time difference.

9. The method of claim 1, wherein the current cardiac profile vector comprises a plurality of real numbers, each of the real numbers being associated with one or more of the cardiopulmonary variables.

10. A system for improving a person's cardiopulmonary health, the system comprising: a sensor for obtaining monitored data from a monitored human subject for a duration of time, the monitored data comprising a continuous time series of a selected plurality of cardiopulmonary variables, the sensor being deployed with the monitored human subject; a control circuit coupled to the sensor, the control circuit being configured to: - receiving monitored data for said duration; - applying the monitored data of the duration to a trained encoding neural network and responsively creating a current cardiac profile vector for the monitored human subject; - generating a current estimate of a functional capacity metric from the current cardiac profile vector by applying the current cardiac profile vector to a trained functional capacity estimator; wherein the trained functional capacity estimator is created using a plurality of matched examples to produce the trained functional capacity estimator, each of the matched examples comprising a cardiac profile vector matched to a corresponding measured functional capacity metric of the person to enable regression of an input of the cardiac profile vector to an estimate of the functional capacity metric; and wherein one or more actions are performed based on the current estimate of the functional capacity metric, the one or more actions being selected from the group consisting of: The control circuit displays the current estimate together with the previous estimate in a time series on a display screen to a clinician for reviewing health changes in the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, triggering and transmitting an electronic alert to a clinician to investigate the health of the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, the control circuitry transmitting an electronic questionnaire to the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, making an entry in the monitored human subject's electronic medical record indicating a change in the monitored human subject's health; The control circuit compares the current estimate with estimates from other monitored human subjects to quantify health effects between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial; and The control circuit transmits electronic control signals that control the operation or set parameters of a medical device associated with the monitored human subject.

11. The system according to claim 10, comprising: an encoding neural network configured to accept as input a time series of a selected plurality of cardiopulmonary variables and generate a multi-element cardiac profile vector; a function generator neural network configured to generate weights when supplied with the cardiac profile vector as input; an estimation neural network configured to accept as input a time series of a subset of the selected plurality of cardiopulmonary variables, the subset omitting at least one excluded variable, and to generate an estimate of the at least one excluded variable, wherein the estimation neural network is configured using the weights output by the function generator neural network as weights for connections between nodes in the estimation neural network; wherein the encoding neural network and the function generator neural network are connected in series and trained through a plurality of iterations, the training of the encoding neural network producing the trained encoding neural network, and the training of the function generator neural network producing the trained function generator neural network, the training of the encoding neural network and the function generator neural network using an error metric, the error metric comprising a difference between at least one excluded variable and the estimated value of the at least one excluded variable generated by the estimator neural network, Wherein for each of the plurality of iterations: inputting into the encoding neural network training data of a first duration comprising a continuous time series of the selected plurality of cardiopulmonary variables from a training human subject, Training data comprising a continuous time series of the selected plurality of cardiopulmonary variables from the training human subjects for a second duration that does not overlap in time with the first duration is provided as input to the estimation neural network, and wherein the estimation neural network provides the estimated value of the at least one excluded variable.

12. The system of claim 10, wherein the functional capacity is VO2Max.

13. The system of claim 10, wherein the sensor is selected from the group consisting of: a camera, an electrocardiogram sensor, a light sensor, an accelerometer, an RF energy sensor, a temperature sensor, and a gyroscope sensor.

14. The system of claim 10, wherein the cardiopulmonary variables include measures of physical activity or exercise.

15. The system of claim 10, wherein the cardiopulmonary variables include one or more of: heart rate, temporal heart rate variability, respiratory rate, tilt angle, atrial fibrillation probability, and sample time difference.

16. The system of claim 10, wherein the current cardiac profile vector comprises a plurality of real numbers, each of the real numbers being associated with one or more of the cardiopulmonary variables.

17. A system for improving a person's cardiopulmonary health, the system comprising: an encoding neural network configured to accept as input a time series of a selected plurality of cardiopulmonary variables and generate a multi-element cardiac profile vector; a function generator neural network configured to generate weights when supplied with the cardiac profile vector as input; an estimation neural network configured to accept as input a time series of a subset of the selected plurality of cardiopulmonary variables, the subset omitting at least one excluded variable, and to generate an estimate of the at least one excluded variable, wherein the estimation neural network is configured using the weights output by the function generator neural network as weights for connections between nodes in the estimation neural network; a control circuit configured to connect the encoder neural network and the function generator neural network in series and iteratively train the encoder neural network through a plurality of iterations to produce a trained encoder neural network and iteratively train the function generator neural network to produce a trained function generator neural network, the training using an error metric comprising a difference between the at least one excluded variable and the estimated value of the at least one excluded variable generated by the estimation neural network, wherein for each of the plurality of iterations, inputting into the encoding neural network training data of a first duration comprising a continuous time series of the selected plurality of cardiopulmonary variables from a training human subject, providing input to said estimation neural network training data comprising a continuous time series of said selected plurality of cardiopulmonary variables from said training human subjects for a second duration that does not overlap in time with said first duration, and wherein said estimation neural network provides said estimated value for said at least one excluded variable; a functional capacity estimator configured to accept as input a cardiac profile vector generated by the trained encoding neural network and to output an estimate of a functional capacity metric, wherein the control circuit is further configured to train the functional capacity estimator using a plurality of matched examples to produce a trained functional capacity estimator, each of the matched examples comprising a cardiac profile vector matched to a corresponding measured functional capacity metric of the person to enable regression of an input of the cardiac profile vector to an estimate of the functional capacity metric; wherein said trained encoding neural network and said trained functional capacity estimator are deployed to accept as input a duration of monitored data from a monitored human subject, said monitored data comprising a continuous time series of said selected plurality of cardiopulmonary variables, to generate a current cardiac profile vector for said monitored human subject, and to generate a current estimate of a functional capacity metric from said current cardiac profile vector; as well as wherein one or more actions are performed based on the current estimate of the functional capacity metric, the one or more actions being selected from the group consisting of: The control circuit displays the current estimate and the previous estimate in a time series on a display screen to a clinician for reviewing health changes in the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, triggering and transmitting an electronic alert to a clinician to investigate the health of the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, the control circuitry transmitting an electronic questionnaire to the monitored human subject; the control circuitry comparing the current estimate to a previous estimate and testing for changes, making an entry in the monitored human subject's electronic medical record indicating a change in the monitored human subject's health; The control circuit compares the current estimate with estimates from other monitored human subjects to quantify health effects between at least one control group receiving a first intervention and a test group receiving a second intervention in a clinical trial; and The control circuit transmits electronic control signals that control the operation or set parameters of a medical device associated with the monitored human subject.

18. The system of claim 17, wherein the functional capacity is VO2Max.

19. The system of claim 17, wherein the time series of a selected plurality of cardiopulmonary variables is obtained via sensors non-invasively affixed to a person's skin.

20. The system of claim 17, wherein the time series of the selected plurality of cardiopulmonary variables is obtained via a sensor selected from the group consisting of: a camera, an electrocardiogram sensor, a light sensor, an accelerometer, an RF energy sensor, a temperature sensor, and a gyroscope sensor.

21. The system of claim 17, wherein the functional capacity estimator is implemented by the control circuit and is configured to generate a current estimate of the functional capacity metric using one of: a linear regression technique, a decision tree, a random forest estimate, or a gradient boosting model.

22. The system of claim 17, wherein the functional capacity estimator comprises an estimating neural network.

23. The system of claim 17, wherein the cardiopulmonary variables include measures of physical activity or exercise.

24. The system of claim 17, wherein the cardiopulmonary variables include one or more of: heart rate, temporal heart rate variability, respiratory rate, tilt angle, atrial fibrillation probability, and sample time difference.

25. The system of claim 17, wherein the current cardiac profile vector comprises a plurality of real numbers, each of the real numbers being associated with one or more of the cardiopulmonary variables.

Images