Method and apparatus for detecting conditions from physiological data

Abstract

The present invention relates to a computerized system for measuring and / or detecting a response or condition in a human being based on data obtained from a wearable sensor worn in a natural free-living situation. Based on the measurement and / or detection, various actions may be taken. Physiological data may be obtained and instructions to modify the vaccine composition and / or dosage may be sent to the vaccine manufacturer.

Description

[Technical Field]

[0001] [CROSS-REFERENCE TO RELATED APPLICATIONS] This application claims priority to U.S. Provisional Application No. 63 / 392,218, filed July 26, 2022, the contents of which are incorporated herein by reference in their entirety.

[0002] [Technical field] The present invention relates generally to the detection and treatment of inflammatory processes in humans. [Background technology]

[0003] Inflammatory responses can occur in humans and are considered a biological response of the human immune system, characterized by the initiation of signaling pathways. Inflammatory responses can be triggered by a variety of factors, including pathogens, damaged cells, harmful compounds, vaccines, and other immune system activators. In some cases, inflammatory responses are considered a negative condition (e.g., as a result of disease), while in other cases, inflammatory responses may be considered a positive outcome (e.g., as a beneficial result of vaccination). [Brief explanation of the drawings]

[0004] [Figure 1] FIG. 1 shows a block diagram of a system for detecting and counteracting inflammatory processes in humans, according to various embodiments of the present invention.

[0005] [Figure 2] FIG. 2 shows a flowchart of an approach for detecting and counteracting inflammatory processes in humans, according to various embodiments of the present invention.

[0006] [Figure 3] FIG. 3 shows a flowchart of a general approach for generating inflammatory markers using residuals, according to various embodiments of the present invention.

[0007] [Figure 4] FIG. 4 shows a flowchart of an approach for detecting and counteracting inflammatory processes in humans, according to various embodiments of the present invention.

[0008] [Figure 5] FIG. 5 is a diagrammatic representation of the safety monitoring system of the present invention used for remote monitoring of a patient in a home environment.

[0009] Elements in the figures are illustrated for simplicity and clarity and are not necessarily drawn to scale. For example, the dimensions and / or relative locations of some elements in the figures may be exaggerated relative to other elements to facilitate an understanding of various embodiments of the present invention. Additionally, elements that are understood to be common, but useful or required in commercially feasible embodiments, may not be shown so as to obscure the view of these various embodiments of the present invention. While certain acts and / or steps may be described or illustrated in a particular order, those skilled in the art will understand that such a specific order is not required. DETAILED DESCRIPTION OF THE INVENTION

[0010] The approaches described herein provide a computerized system for measuring and / or detecting adverse or adverse conditions (e.g., inflammatory responses) in humans based on data acquired from wearable sensors in natural, everyday life situations. Based on the measurement and / or detection, various countermeasures can be taken. "Inflammation" or "inflammatory response" refers to a biological response of the immune system characterized by the initiation of signaling pathways that can be triggered by a variety of factors, including pathogens, damaged cells, harmful compounds, vaccines, and other immune system activators. It will also be understood that these approaches can be used to detect positive effects, such as the efficacy of vaccines, in humans. In other words, the approaches provided herein can be used to determine physiological changes associated with inflammation in humans, and whether those changes are positive or negative can be characterized.

[0011] Measuring side effects or conditions, such as inflammatory responses, in humans is useful for a variety of reasons. To name a few, measuring inflammatory responses is useful for safety monitoring (e.g., in clinical care or clinical trial settings), when patients or participants return home after receiving treatment, or in recognizing and addressing side effects that manifest through immune system activation. Measuring inflammatory responses is also useful in detecting the reactogenicity of vaccines, thereby assessing their efficacy or degree of action (e.g., immunogenicity) and safety across a representative population. Additionally, measuring inflammatory responses is useful in determining whether a medical intervention is producing a desired response or whether patients who appear to be responding poorly should receive increased therapy or additional booster treatments.

[0012] Measuring positive or negative side effects or conditions, such as inflammatory response, is also useful in home monitoring of patients vulnerable to infection, such as sepsis or infection at the site of recent surgery. Measuring inflammatory response is also useful for identifying the overall course of inflammatory conditions, such as autoimmune diseases like rheumatoid arthritis, and an individual's response to a therapeutic intervention.

[0013] The approaches provided herein use sensors to acquire data from humans. In one aspect, the sensor data from sensors includes continuous vital sign data and statistical derivatives thereof, such as heart rate, respiratory rate, core body temperature, skin temperature, and blood pressure, measured from wearable sensors throughout activities of daily living. The sensor data is collected and transmitted to a computerized platform, such as the cloud, where it is integrated into analyses of the patient's inflammatory response.

[0014] The approach provided herein generates an individualized (or "personalized") baseline model of tandem vital sign behavior, typically based on one to several days of vital signs collected from an individual. "Tandem vital sign behavior" refers to the characteristics of the interactions or correlations between multiple vital signs over time through activities of daily living. The baseline model can be used to control for natural individual differences across populations that may interfere with the detection of inflammatory response signals. This baseline model represents how an individual's vital signs change in tandem based on unconstrained activities of daily living. In one embodiment, the model is an autoassociative model or autoencoder, constrained to recognize joint patterns of an individual's tandem vital sign fluctuations and estimate expected vital sign behavior, as described in more detail herein.

[0015] In many of these embodiments, physiological data is collected from at least one wearable sensor worn by the patient during a pre-treatment period. An individualized prognostic model is created based on the patient's collected pre-treatment physiological data. The model can estimate expected behavior of physiological variables in response to receiving new physiological data from the at least one wearable sensor worn by the patient.

[0016] Additional physiological data is collected from at least one wearable sensor worn by the patient during a post-treatment period. An estimate of an expected value of the post-treatment physiological data is generated using an individualized estimation model. The post-treatment physiological data is compared to the estimate to determine whether a predefined effect pattern exists based, at least in part, on the comparison. If a predefined effect pattern exists, an action is determined and executed. The action can be one or more of: triggering an electronic questionnaire to be prompted to the patient; providing instructions to the patient to perform a measurement; instructing the patient to contact a clinician; triggering a ticket in a call center system to queue a call for the patient; creating a prompt in an app on the patient's phone to contact a clinician; providing instructions to the patient related to effect triage; sending a control signal to control a medical device related to the patient's treatment; sending instructions to a vaccine manufacturer to modify vaccine composition and / or dosage; or sending instructions to a clinician to modify treatment composition and / or dosage. Other example actions are also possible.

[0017] In some instances, the effect is an adverse side effect. In other instances, the effect is a positive one, such as an indication that a vaccine is effective.

[0018] In yet other examples, the physiological data includes heart rate data, respiration rate data, core body temperature data, skin temperature data, and activity data. Other examples are possible.

[0019] In another aspect, the personalized prediction model is trained using data collected from the patient while they are in a free-living physiological state where their inflammatory state is stable and not expected to change. In yet other aspects, the comparison is made by determining the residual between the estimate and the physiological data. In some aspects, the residuals are combined into a single score as a scalar index.

[0020] In yet another aspect, the personalized estimation model comprises a neural network or a system of neural networks. Other exemplary models and modeling approaches are possible.

[0021] In another example of these embodiments, a system for monitoring the effects of pharmacological therapy on a patient includes at least one wearable sensor and control circuitry, wherein the at least one wearable sensor is configured to be worn by the patient during a pre-treatment period and to collect physiological data from the patient.

[0022] The control circuitry is coupled to the at least one wearable sensor and is configured to create an individualized estimation model based on physiological data collected pre-treatment of the patient, the model capable of estimating physiological variables in response to receiving new physiological data from the at least one wearable sensor worn by the patient, and the at least one wearable sensor collects additional physiological data from the at least one wearable sensor worn by the patient during a post-treatment period.

[0023] The control circuitry is further configured to generate an estimate of the post-treatment physiological data using the individualized estimation model, compare the post-treatment physiological data measured by the sensor to the estimate, determine when a predefined effect pattern exists based at least in part on the comparison, and determine and execute an action if the effect pattern exists. The action may include one or more of the following: triggering a prompt to present an electronic questionnaire to the patient, instructing the patient to take a measurement, instructing the patient to contact a clinician, triggering a ticket in a call center system to queue a call for the patient, creating a prompt in the patient's phone app to contact a clinician, providing instructions to the patient related to effect triage, sending control signals to control medical equipment related to the patient's treatment, or sending instructions to a vaccine manufacturer to modify the vaccine composition and / or dosage. Other example actions are possible.

[0024] In another aspect, the baseline model is trained using data collected from an individual while they are in a free-living physiological state where their inflammatory state is stable and not expected to change. For safety monitoring in clinical trials or clinical care, training data is collected before drug administration or other therapeutic interventions where inflammatory responses are a concern. For detecting vaccine activity, training data is collected before vaccination. For detecting infectious diseases, training data is preferably collected before surgery or immediately after hospital discharge, when the infection has not yet progressed to a systemic response. In the context of autoimmune diseases, training data is collected before the initiation of a new treatment that may reduce systemic inflammation.

[0025] In some aspects, at least one circadian cycle is utilized to train or personalize the model, preferably using several days' worth of data. After training or personalization from the initial data, the model switches from operating in a learning (or training) mode to operating in a monitoring mode, where it is able to generate estimates of vital signs in response to receiving new data from the sensors. The estimated vital signs are compared to the actual measured vital signs, and a residual, the measured value minus the estimated value, is calculated. This residual can be negative if the estimated value is higher than the measured value and positive if the estimated value is lower. Furthermore, the residuals can be combined into a single score as a scalar indicator of the change in vital sign behavior compared to behavior during baseline modeling. In some aspects, this score ranges from 0 to 1, with 0 indicating that vital sign behavior is similar or identical to that initially learned and 1 indicating that vital sign behavior has clearly changed from what was learned.

[0026] In another aspect, the approach provided herein relies on residual patterns or signatures indicative of an inflammatory response. Multiple patterns may be tested simultaneously, alternating between them, to explore how the inflammatory response manifests within the signature. At the same time, all signatures must satisfy other conditions. Alternatively, it is possible to define a signature indicative of a reduced inflammatory response. This is applicable to individuals who already have inflammation during the baseline period. In another aspect, the approach provided herein identifies the persistence of one or more signatures or signatures within a time window as characteristic of an inflammatory response. Transient residual values, or any pattern or signature that does not persist, are not considered to represent a true physiological response.

[0027] Analytics associated with this approach may be aggregated from many individuals to aid in drug and vaccine development or presented to clinicians, patients, and other clinical decision makers to influence patient status. Such actions may include assisting in ensuring symptom confirmation for patients remotely from the clinical setting. In another example, an action may be automatically sending a survey to the patient via a smartphone provided to the patient as part of the system, prompting the patient to answer questions that confirm or deny symptoms. In yet another example, an action may be having the patient perform another measurement with a manually operated device to provide further data to confirm or deny suspected inflammation. In yet another example, an action may be to intervene or self-manage the patient to appropriately mitigate the cause of the inflammation, whether it be an infectious agent or an autoimmune response, such as taking an antibiotic or a drug that suppresses the immune response. In yet another example, an action may be controlling a call center ticket queue system to create a new ticket, with the call center staffed by trained clinicians who can contact the patient. If it is difficult to find a time for a patient and clinician to call, yet another example could involve activating or displaying a button on the patient's smartphone app that the patient can click when they are available to talk to automatically initiate contact with a call center clinician trained to handle their condition.

[0028] Wearable sensors include any device capable of recording signals representative of heart rate, temperature, activity, respiration rate, and other behavioral and physiological parameters. Statistical derivatives of these values, such as heart rate variability, may also be used to generate baseline models. Exemplary sensors include a chest-worn adhesive patch capable of collecting continuous electrocardiogram (ECG) signals, accelerometry, and temperature data.

[0029] In general, there is clinical interest and value in monitoring patients for signs of acute immune system activation in at-risk populations. Additionally, in some individuals, there is value in detecting evidence of chronic inflammation diminished by therapeutic interventions to reduce inflammation. In some respects, this approach is directed toward situations where it is important to detect changes in systemic inflammation that may be life-threatening (e.g., cytokine release syndrome or infection) or useful for optimizing the efficacy and safety of treatments (e.g., vaccines or biologics), that can occur at any time, occur systematically, and occur with a wide range of severity, and for which ongoing monitoring is necessary or beneficial. In this context, "acute" refers to a time period measured in hours to weeks, as opposed to long-term inflammatory processes measured in months to years.

[0030] One example of this is immunotherapy, a cancer treatment that harnesses a patient's own immune system to attack cancer. One example of immunotherapy is CAR-T cell therapy, in which a patient's own T cells are modified to carry special receptors called chimeric antigen receptors (CARs). These receptors help the T cells recognize and attack cancer cells.

[0031] More specifically, in the first step, the patient donates blood to collect T cells. In the second step, the T cells are genetically engineered in the laboratory to encode a tumor-specific CAR. In the third step, the CAR-T cells are expanded. In the fourth step, the manufactured CAR-T cells are verified for safety and purity and then cryopreserved. In the fifth step, the CAR-T cells are infused into the patient. In the sixth step, the CAR-T cells seek out, bind to, and destroy cancer cells.

[0032] One of the side effects of immunotherapy for cancer is cytokine release syndrome. Cytokines are small proteins that transmit messages that regulate the body's immune response. However, in cytokine release syndrome, the excessive release of cytokines can have harmful effects on organs and, in severe cases, can be fatal. Cytokine release syndrome (CRS) can develop approximately 3 to 14 days after immunotherapy administration. Because of this potential risk, treatment has traditionally been performed in a clinical setting, requiring patients to remain hospitalized throughout the treatment period. This has resulted in extremely high treatment costs. The approach provided herein advantageously allows patients to return home with a monitoring system that detects early signs of CRS, allowing clinicians to respond immediately if detected. This reduces costs compared to conventional systems.

[0033] Typical known symptoms of CRS include fever, decreased blood pressure, and decreased SpO2, which occur at different stages of the acute phase. Early on, an increased heart rate may be observed, and changes in heart rate variability and breathing may also occur.

[0034] Using the approach provided herein, prior to administration of CAR-T cells (step 5 above), patients are provided with a wearable sensor kit and wear the sensor at home 24 hours a day for approximately 1-4 weeks. This preparation can occur prior to step 1 above, the timing of which may depend on other steps taken to treat or prepare for cancer.

[0035] The data collected during this period, prior to step 5 above, is used to train a personalized baseline model. The patient then presents to the clinic to receive treatment. The CAR-T cells are infused into the patient. The patient returns home with a wearable sensor and possibly a manual measurement device such as a blood pressure monitor. The data collected from the patient is monitored and evaluated according to the approaches described herein. The system sweeps the received data for one or more signatures of CRS or a general inflammatory response, as described in more detail below.

[0036] If a detection occurs, the results are provided to the clinician via one of several notification methods, such as sending an alert to a web-enabled clinical portal. Push notifications may also be sent to other medical monitoring systems. Push notifications can also be sent via text message or to the clinician's email address. A push notification to the smartphone can also be sent to the patient, instructing them to contact their clinical care provider or automatically connecting them to the clinician's call center by clicking a newly generated button within the app. A push notification can be sent to the patient via their smartphone, asking them to answer survey questions about their symptoms, which are then distributed or sent to the clinician. A push notification can also be sent to the patient on their smartphone, asking them to manually measure their temperature, SpO2, or blood pressure and enter the results into the smartphone app, which then distributes or sends the data to the clinician.

[0037] Another example of the use of the approach provided herein is vaccination. Vaccination activates innate immune mechanisms and triggers the synthesis of proinflammatory cytokines, which are critical for initiating antigen-specific adaptive immune responses. This physical phenomenon of inflammation (called reactogenicity) has traditionally been tracked solely through symptomatic investigations. Limited studies directly measuring inflammatory blood biomarkers have not only confirmed the wide variability of inflammatory responses between individuals, but also revealed a strong correlation between this response and systemic symptoms and humoral immune responses. Due to the lack of a scalable method for measuring an individual's vaccine response, for most individuals, the ultimate measure of the adequacy of vaccine-induced immune protection depends on the breakthrough infection and its severity.

[0038] Furthermore, objective evidence of individual inflammatory responses to vaccines could aid in the design of safer and more tolerable vaccines.The limitations of subjective surveys, the current gold standard for safety tracking, were highlighted in an analysis of reported adverse events in placebo-controlled COVID-19 vaccine trials, which showed that more than 50% of reported systemic adverse events could be attributed to "nocebo" reactions.

[0039] For vaccine-focused use cases, the approaches provided herein include a monitoring system, as described below. In some aspects, it is anticipated that individuals will begin monitoring prior to vaccination and continue for 5-10 days after vaccination.

[0040] In clinical trials for vaccine development, incorporating wearable sensors can better assess the variability of inflammatory responses to different doses of vaccine across different age groups in large populations; this information can inform optimal dosing strategies to optimize efficacy and safety. Variability among individual participants in cohort studies can be assessed by monitoring before vaccine administration, establishing a multivariate baseline during daily activities. The vaccine is administered. Monitoring then continues over the expected period of response. The degree of individual inflammatory response is assessed using the methods described below and determined by comparing baseline and post-vaccination periods. Results are then statistically analyzed.

[0041] Identifying each individual's unique response after vaccination involves many known and unknown factors that may influence an individual's response. Known factors include age, sex, comorbidities, and conditions or treatments that specifically affect the immune system. For individuals in whom the expected inflammatory response is not identified, further testing (e.g., measuring blood levels of vaccine-induced antibodies) or a booster vaccination may be scheduled. For some individuals who do not experience any symptoms after vaccination, appropriate changes detected by the monitoring system can be reassuring.

[0042] Another example of a situation in which this approach could be deployed is in the early detection of signs of sepsis or other infections. The primary function of the immune system is to defend the body against viral and bacterial pathogens. This response causes inflammation, which, if severe enough, manifests as fever. However, fever (e.g., a temperature above 38°C) is a lagging indicator of infection and is merely a measurement of a single inflammation-induced change on a population basis.

[0043] For many people, such as those with weakened immune systems, detecting infection as early as possible can be lifesaving. For others, early detection can prevent readmission and significant morbidity. For those at increased risk, such as those after surgery, especially transplant surgery, active surveillance for the first signs of immune system activation and inflammation may trigger earlier testing with blood cultures and more intensive monitoring in medical facilities.

[0044] In this use case, a baseline of multivariate vital sign behavior is established from unconstrained activity of daily living monitoring data when the patient is known not to have sepsis. Monitoring is then continued using the same sensors during periods of concern for sepsis risk, and the baseline model is used to compare the monitoring data, which, when compared to the multivariate baseline, reveals a signature of early signs of inflammation associated with sepsis, as described below. When an alert is triggered, clinical staff can intervene with appropriate treatment to prevent the onset of full-blown sepsis.

[0045] Operational and implementation aspects of the present approach are described below. The inflammatory process affects measurable vital signs in a signature manner. For example, increased body temperature, increased heart rate (HR), increased respiratory rate (RR), and decreased heart rate variability (HRV) may be manifested as a result of inflammation.

[0046] While it is possible to detect inflammation retrospectively using long-term averages, detection of inflammatory processes is advantageous when they unfold on a timescale of hours, which is very urgent. Over such time frames, long-term averages are less useful. The challenge is to detect relative changes in one or more of the above (a) for an individual patient, and (b) against the background of normal fluctuations due to free-living activities and behaviors.

[0047] The approach presented here establishes individualized multivariate dynamic baselines to detect relative abnormalities from the normal dynamic behavior of these vital signs. These approaches then measure whether the residuals are positive or negative, providing evidence of the aforementioned signatures. For example, it can determine whether the measured HR or RR is higher than it should be, whether the body temperature is higher than it should be, or whether the HRV is lower than it should be.

[0048] These changes are prioritized and monitored in some aspects. For example, HRV is the most important factor, and this change is considered by some to be the earliest evidence of inflammation specificity. Increased body temperature appears somewhat delayed, respiratory rate (RR) is subject to some degree of voluntary control and is difficult to measure reliably, and heart rate is not specific to inflammation (i.e., other factors may be responsible) and can become abnormally elevated. Therefore, physiological data on these four components, particularly HRV, are utilized. While all these vital signs may provide evidence, some of these approaches focus on detecting at least a relative decrease in HRV estimated by multivariate models.

[0049] Vital signs used in this approach include those that characterize the performance of the cardiopulmonary system. Multiple vital signs can be used, as described below.

[0050] Heart rate (HR) may be used. HR is usually expressed as beats per minute, but it can also be determined beat-by-beat by measuring the time interval between QRS peaks on an ECG (electrocardiogram) and then converted to an average HR over a time window, such as one minute, to generate the number of beats per minute. The average heart rate per minute can be obtained by determining the trimmed average of the one-minute beat-to-beat intervals and then inverting that value to obtain the number of beats per minute. This can also be done from a PPG waveform or ballistocardiogram, to name a few.

[0051] Respiratory rate (RR) can also be used. RR can be measured from several sources, including through respiratory sinus arrhythmia, a phenomenon in which the heart rate slows and speeds up slightly with each breath, by an accelerometer in an adhesive patch, or chest wall movement measured by measuring the envelope of the ECG QRS peak, which fluctuates with changes in the volume of gas in the thoracic cavity.

[0052] Total activity (ACT) can also be used, measured as the standard deviation of the magnitude of the vector of accelerometer deviations, although other methods of quantifying movement are also available, such as "activity counting," a method long used to quantify movement with vibration sensors.

[0053] Heart rate variability (HRV) is also available. HRV can be measured in several ways, using pulse as a measure of heart rate within a certain time window. HRV may include the standard deviation of "normal" RR intervals (SDNN), the standard deviation of all RR intervals (SDRR), the standard deviation of the average NN intervals every 5 minutes within a 24-hour time window (SDANN), the root mean square of the difference between consecutive RR intervals (RMSSD), the spectral frequency band of HRV, or the Poincaré HRV dispersion parameter.

[0054] Skin temperatures (TEMPs) are typically measured by a skin-side thermistor that is part of the sensor device, but skin temperatures can also be used. In addition, body core temperatures (TEMPc) can also be used. TEMPc is estimated based on two thermistors, one facing the skin side of the sensor and the other facing away from the sensor. Body core temperatures are calculated as a function of heat flux entering from one side and leaving from the other side, both with thermistors.

[0055] Changes in the ECG waveform can also be used, for example the QT interval, measured as the time difference from the onset or peak of the QRS complex of the ECG to the end of the T wave of the ECG, or by a surrogate such as the time interval to the peak of the T wave.

[0056] Because multivariate modeling of relationships between vital signs is facilitated by simultaneously acquired "snapshots," it is optimal for vital sign statistics over a time window (e.g., average heart rate, maximum activity, 95th percentile temperature, etc.) to be provided over a common window that is assigned the same timestamp. In other words, each input to downstream modeling has a vector with data corresponding to each vital sign at the point in time represented by the vector. Examples include vectors of 1-minute average values ​​for HR, HRV, RR, ACT, TEMPc, and QT.

[0057] As described elsewhere herein, the present approach uses a personalized estimation model. In some aspects, the personalized estimation model is any multivariate model that outputs an estimate of the expected value of one or more vital sign values ​​measured from a patient. Such models may be generated by decision tree / random forest function approximators, similarity-based models, neural networks, etc.

[0058] Some aspects require training data, collected from patients prior to monitoring for adverse events such as CRS or sepsis. Once training is complete, the personalized model is used to generate estimates in response to newly measured data inputs from the patient. The inputs include new measurements of new multivariate vital signs data from wearable sensors. The output from the model includes estimates of the inputs, which are used to generate residuals, i.e., measurements minus estimates, as described below.

[0059] Similarity-Based Modeling (SBM) is a self-associating or self-encoding pattern reconstruction method. At any given time, the set of measurements from all variables can be thought of as an input pattern (similar to pixels in an image). The self-associating estimation process reconstructs the input pattern based on the learned patterns used to generate the SBM model. SBM essentially attempts to reproduce the input pattern based on information accumulated in the training data. This reproduction is only possible if a linear combination of the training data can be identified that fits the input pattern. This is possible if the input pattern is representative of the multivariate behavior of the training data. If the input pattern is not representative of the behavior of the training data, one or more of the estimated pattern elements may not closely match the corresponding input element. In effect, the estimate reflects how SBM evaluates each pattern element based on the information contained in the model's training data and the input pattern. The residual pattern (the difference between the input and estimated pattern) reveals where the pattern deviates from the training data. These differences are often subtle and accumulate over time to drive the decision-making process.

[0060] The mathematical foundation of the SBM approach focuses on the application of "similarity operations" to pairs of observation vectors and the manipulation of a "state" matrix D, which stores a set of past training vectors (input patterns). The number of columns in the state matrix D is equal to the number of representative training vectors (M), and the number of rows is equal to the number of data sources contained in each vector (L). At some point n j The set of measurements taken at j ), then

number

[0061] where x i (n j ) is the time nj are measurements from data source i at , and the state matrix D is expressed as:

number

[0062] The result of a similarity operation on two observation vectors is a similarity score (a scalar value). This similarity operation is nonlinear, but can be extended as a matrix operation that computes a scalar similarity score for each combination of two vectors stored in two matrices of appropriate dimensions. Thus, given an input vector (or pattern) x containing a single measurement from each data source L, in Given the corresponding estimated data source value x est is determined using equations (3)-(5). The estimates of each variable in the input are used to form a "reconstructed" input pattern using a linear combination of the training vectors in the state matrix D.

number

[0063] where w is a set of weighting factors derived from the following equation:

number

number

[0064] Similarity operations are symbols

number

[0065] A localization feature is provided. The localized SBM approach relies on dynamically generating the D matrix to characterize only the local behavior of the system at a particular time. The idea is to select a subset of data that defines the state matrix D(t) most relevant to the current input pattern from a much larger superset matrix H that characterizes the entire dynamic range of the monitored system. Estimates are then generated based only on these selected currently relevant vectors, as shown in equations (6)-(8) below. This process is repeated for each new input vector. Thus, the weighting coefficients are

number

number

[0066] Here, F(·,·) in equation (6) is the input vector x inThis is the process of selecting relevant vectors based on (t) and the reference data matrix H. Finally, the residual vector at time t is given by

number

[0067] Random forest (RF) approaches are also available. In a "round-robin" configuration, they behave like autoencoding or autoassociative models, with M RFs each producing a single output for M variables, collectively representing M estimates for the M inputs. Each RF is trained to individualize to a patient using data obtained from the patient's pretreatment data collection. Neural networks can also be used to implement the individualized models provided here. When trained with sufficient examples, neural networks are highly effective at nonlinear pattern recognition, nonlinear encoding of key features from raw data, and approximating inference functions. In practice, the challenge for the use case envisioned here is that even if sensors are fitted to the patient during the pretreatment period, not enough raw vital sign data is collected to train a neural network from scratch to generate personalized estimates of the patient's vital signs. Thus, in one approach, a first neural network is developed as an "encoding" neural network that receives as input a time series of a first portion of vital sign data from a given person over one or more days and outputs a multi-element encoding vector representing that person's personalized cardiopulmonary behavior. An ensemble of one or more estimation neural networks is then developed that receives as input the encoding vector and a second portion of vital sign data from that person (data that does not overlap with the first portion) and estimates the expected value of the second portion of vital sign data. This combination includes individual neural networks that estimate one vital sign using all other vital signs as inputs. For example, one member of the combination estimates heart rate using activity, HRV, respiration, temperature, and the encoding vector as inputs, while a second member of the combination estimates HRV using activity, heart rate, respiration, temperature, and the encoding vector as inputs. In this way, each member of the combination estimates a vital sign that is not present in the inputs.

[0068] These networks are pre-trained using vast amounts of vital sign data from at least tens to hundreds of individuals before being deployed as individualized estimators. Each training sample comprises a first and a second portion of vital sign data from a single individual in the training set. The first portion of data is input to an encoding neural network to generate an encoding vector. This encoding vector and the second portion of data are input to a combination of estimation networks, each of which generates a respective vital sign estimate as output, where these estimates were not present as input, i.e., inferentially estimated. These inferential outputs are compared to the actual, known vital signs from the second portion to determine an estimation error, which is back-propagated through a cost function to train both the encoding network and the combination of estimators. By repeating this training process for many matched first and second portions of data, the encoding network becomes efficient at generating encoding vectors that accurately encode the cardiopulmonary behavior of any person whose data is input to the encoding network based on the input data, and each of the estimation networks becomes efficient at inferring estimated vital signs for each of the second portions of data from a person when given an encoding vector representing that person and that second portion of data from that person as input. The input to the encoding network can be structured, for example, as one minute of vital sign data over several days, and the input to the estimation network can be structured as the encoding vector from the first network and a series of vital sign data over a time window for timely monitoring of an inflammatory response. For example, this time window could be a three-hour window or a series of 24-hour windows with a sliding three-hour time window. For clarity, however, pre-training these networks does not include people undergoing inflammatory responses.Rather, the pre-training involves data from a person where the first and second parts of the data are expected to be physiologically consistent. In other words, these networks are trained to accurately estimate what the second part of the data will be, given the encoding of the first part.

[0069] Once pre-trained, these neural networks effectively perform personalized estimation as follows: Pre-treatment data from a patient is input into the encoding network to generate an encoding vector. This is the individualization step; this encoding vector tunes the estimation network to estimate the patient's specific expected vital sign behavior. Incidentally, when a window of post-treatment vital sign data excluding heart rate is input to a member of the heart rate estimation pair, the expected heart rate is estimated based on the pre-treatment encoding vector and the other non-heart rate vital sign data for the post-treatment time window. This estimated heart rate is then compared to the actual heart rate measured from the sensor, generating the aforementioned residual for heart rate. Similar processing is performed for other vital signs in other members of the pair. In this way, all vital sign residuals (as changes from normal health) needed to detect or quantify an inflammatory response are provided by the neural network on an individualized basis.

[0070] The approach presented here utilizes residuals. Residuals are each measured vital sign value minus its corresponding estimated value. For example, if average heart rate over one minute is used as a monitoring parameter, simultaneously measured vital sign values ​​are input to a model, which outputs a combined predicted or expected heart rate for the person, taking into account the other vital sign values. The heart rate residual is then the measured heart rate minus the estimated heart rate.

[0071] The residuals thus constitute time series data in themselves, one time series for each vital sign in the model. The value and sign (positive or negative) of the residuals have specific meanings.

[0072] For example, a high residual (positive, >0) means that the measured value is higher than the normal physiological state predicted by the model for that patient. A low residual (<0) means that the measured value is lower than the normal physiological state predicted for that patient. This residual is then examined by the detection rules described below.

[0073] The individual residuals, plus the normalized magnitude of the residuals at each measurement time point, may be combined to produce a single scalar index of overall vital sign behavior disturbance, such as the Multivariate Health Index described in U.S. Patent No. 8,620,591, the contents of which are incorporated herein in their entirety, and referred to herein as the "Multivariate Change Index (MCI)."

[0074] In summary, an individualized baseline model is used to generate estimates for at least a portion of the training dataset, and the relative magnitude of the residuals is utilized to scale the distribution of estimated variances typical of residuals from a baseline based on normal (undisturbed) vital sign behavior. As new data observations are input during monitoring mode, the generated residuals are compared to this distribution, generating a single scaled score ranging from 0 to 1 related to the likelihood that the residual vector is a member of the expected distribution. If the residual vector falls well within the typical distribution (characterized by the training data scaling results), the MCI output is close to 0 (low likelihood of vital signs exhibiting aberrant behavior). If the residual vector is frequently located in the tails of the multidimensional distribution, the MCI is closer to 1 (high likelihood of aberrant behavior). In this way, the MCI is an indicator of possible disturbances in vital sign behavior, regardless of the patient's underlying activity (e.g., whether sleeping, watching television, running, etc.).

[0075] Thus, the MCI is a time series, with a data point for each observed input of a vital sign estimated by the individualized model. The value of the MCI itself may be used as part of the rule pattern provided by the approach described herein.

[0076] Other methods for providing a single scalar index for a multidimensional vector of residual data are available as alternatives to the MCI described above. For example, Euclidean distance is a well-known method for measuring the distance of a point in a multidimensional space from a distribution of points. It can be used as follows: (a) normalize the residuals of the training sample (without inflammation) to a mean of 0 and a standard deviation of 1; (b) identify the "center" in the multidimensional space of the residuals from the training sample; (c) define a range of distances from that center, assigning the maximum probability of normality (set as 1) to a distance of 0 and the minimum probability (set as 0) to a maximum distance equal to a multiple of the standard deviation (distant distances greater than the maximum are truncated); (d) normalize each residual vector of the test observations and calculate their distances to map the probability of normality to a range between 0 and 1. While Euclidean distance assumes a spherical distribution, other techniques, such as Mahalanobis distance, may potentially be more suitable for correlated variables such as the vital signs used herein. However, the principle of mapping the multivariate residual vector of test data to a distribution expected for normal training data and representing how likely the residual vector is to be abnormal on a scale between 0 and 1 is similar to the MCI approach. All of these single scalar change index generation methods are considered when applied to residuals based on the predicted values ​​of a personalized model as signals used in the rules of the present invention. Where "MCI" is used herein to detect or quantify inflammatory responses, it should be understood that other single scalar indexes can be used instead in the present invention.

[0077] Based on the approach provided herein, the concept of a rule pattern (e.g., implemented as computer code) for detecting inflammation is to identify a signature of inflammation in the residuals of time-series data. As described herein, the characteristic pattern includes one or more relative changes based on the individualized physiological data and the individual's free-living environment. For example, these relative changes may include lower-than-predicted HRV, higher-than-predicted HR, higher-than-predicted body temperature, and / or higher-than-predicted respiration rate.

[0078] For example, during inflammation, it is reasonable to expect that heart rate will increase above what would normally be expected. The personalized model provides the ability to filter out raw heart rate changes due to normal daily activities and detect heart rate increases above normal values, which are detected as positive residual values. The same is true for respiratory rate.

[0079] Furthermore, rules can be gated by the magnitude of MCI, which serially quantifies the overall degree of disruption in the vital sign system. When measuring vital signs in a free-living environment, it is difficult for individualized models to isolate changes to one vital sign or another, and models can suffer from "spillover," where deviations in one variable affect the estimation of another, making the latter appear abnormal. Minor or small deviations in the above vital signs may not be reliable signals of inflammation; adding the requirement that there be at least minimal overall disruption between vital signs ensures more reliable detection of MCI.

[0080] Rules for detecting patterns may also include non-residual signals, for example, core body temperature simply exceeding a threshold; human body temperature is typically maintained within a narrow range for homeostasis regardless of daily activities (it may decrease during sleep and increase during exercise, but these changes are relatively small). Alternatively, the absolute value of MCI may be tested as part of the rule.

[0081] In general, the elements of a rule to detect any given pattern can be selected from several factors, including the directionality of the residual (e.g., positive or negative), the magnitude of the residual, the magnitude of the MCI, the duration or persistence of the pattern, the absolute vital signs, and the filters applied to the condition, to name a few.

[0082] The output of a rule is a modification of the time series data, e.g., MCI, that changes its value when the rule is true or false, or includes the triggering of an alert when the rule is met. Other examples are possible.

[0083] In an example of the operation of the rule, the rule in operation checks for a negative HRV residual (a drop in HRV relative to predicted value) in the presence of sufficiently large MCI.

[0084] In another example of rule operation, the rule verifies that the MCI is high enough and the HRV residual is not positive, and given its gating conditions, checks for either a positive HR residual, a positive RR residual, a positive temperature residual, or a negative HRV residual.

[0085] In yet another example of rule operation, the rule converts the MCI value to a zero value when the rule is false (no inflammation pattern detected), e.g., the MCI is converted to an inflammation-MCI (iMCI), where the value is retained if inflammation is recognized and set to zero if inflammation is not recognized. In this way, the MCI is specialized for inflammation, and the "iMCI" essentially becomes a univariate marker for inflammation.

[0086] In yet another example of rule operation, the rule processing compares the iMCI to a threshold and triggers a detection alert if the iMCI exceeds the threshold.

[0087] In yet another example of rule operation, the rule uses a window of iMCI values ​​to create an average within that window, smooth the signal, and use the smoothed signal for quantification or to trigger detection.

[0088] In yet another example of rule operation, the rule performs additional transformations on the iMCI value, "holding" the maximum value and then holding that value until a new maximum value is detected or a sufficiently low value is detected (e.g., keeping the signal high until the signal drops by at least a certain amount, indicating a clear decrease in evidence of inflammation).

[0089] 1, an example of a system 100 for determining an inflammatory response in a human and taking action thereon is shown. The system includes a first human 102, a second human 104, and a third human 106. A first sensor 108 is worn by the first human 102, a second sensor 110 is worn by the second human 104, and a third sensor 112 is worn by the third human 106. An electronic network 114 is connected to the sensors 108, 110, 112 and the control circuitry 116.

[0090] A memory device 118 is connected to the control circuitry 116. The memory device 118 stores a first individualized estimation model 120, a second individualized estimation model 122, and a third individualized estimation model 124. The electronic network 114 is also connected to a machine 126, an electronic device 128, and a manufacturing equipment 130. Other devices and systems may be connected to the network 116.

[0091] First person 102, second person 104, and third person 106 are, by way of example, patients. These people may be undergoing treatment for a disease or may be interested in monitoring their health status. In some instances, these people may also be participating in clinical research studies.

[0092] The first sensor 108, the second sensor 110, and the third sensor 112 may be any type of sensor or sensor arrangement (including multiple sensors) wearable by the person 102, 104, 106. These sensors may be located at convenient locations on the human body, such as the torso or wrist. In some aspects, a sensor patch is used that is attached to the torso and worn continuously by people to measure multiple values. These values ​​include heart rate and HRV from a single-lead ECG, movement quantification (activity) from a three-axis accelerometer, respiration rate from movement leads or amplitude / frequency modulation of HR, and skin temperature at the sensor site. As previously mentioned, these vital signs can be calculated from the ECG and three-axis accelerometer waveform data. As an example, the ECG may be sampled at 125 Hz, and the accelerometer sampling rate may be 15 Hz or higher. Other examples are possible.

[0093] Electronic network 114 may be any type of electronic communications network or combination of networks, such as a wireless network, the Internet, a local area network, a wide area network, or the like.

[0094] The control circuitry 116 refers to any type of processing device, processor, or controller, including, for example, electronic controllers, microcontrollers, servers, or any type of microprocessor. The control circuitry 116 may include internal memory in which computer instructions for performing the functions, rules, and other operations described herein are stored. The control circuitry 116 and memory 118 may be located in a central processing unit or central call center.

[0095] Memory 118 refers to any type of memory device, database, or combination of devices, such as read-only memory (ROM), random access memory (RAM), programmable ROM, or combinations of these and other types of electronic memory, to name a few.

[0096] The first individualized estimation model 120, the second individualized estimation model 122, and the third individualized estimation model 124 are configured to estimate physiological variables in response to new physiological data from sensors 108, 110, 112 worn by humans 102, 104, 106. The individualized estimation models 120, 122, 124 are multivariate models that provide as output an estimate of the expected value of one or more vital sign values ​​given measured sensor values ​​from the patient. These models may be generated by decision tree / random forest function approximators, similarity-based models, neural networks, etc., to name a few.

[0097] Machine 126 is, by way of example, a medical device used by one or more of humans 102, 104, 106. For example, machine 126 may dispense medication, monitor humans 102, 104, or 106, or perform other functions. Electronic control signals 125 are sent to machine 126 by control circuitry 116 to control machine 126 and / or aspects of its operation. Control signals 125 may activate or deactivate machine 126 or control operating parameters (e.g., the speed, amount of medication dispensed, operation of a display on a screen of machine 126, how often the machine monitors humans 102, 104, 106, to name a few).

[0098] Examples of electronic device 128 include a smartphone, laptop, personal computer, and mobile phone. Electronic signal 127 may be an electronic message, control signal, or other type of electronic signal sent from control circuitry 116. For example, electronic device 128 may be used by one of humans 102, 104, and 106 to display alerts, instructions, or other information to humans 102, 104, and 106. In other examples, electronic device 128 may be used by a medical professional (e.g., a doctor, nurse, hospitalist, or therapist) treating humans 102, 104, and 106. While only one electronic device 128 is shown, multiple devices may be present and operated by different individuals or institutions. Electronic device 128 may also communicate with control circuitry 116 and, via control circuitry 116, with other devices or systems connected to network 114. Electronic device 128 itself may include a processing unit or control circuitry to perform the described operations.

[0099] In one example, the manufacturer 130 is a vaccine manufacturer. The control circuitry 116 may send control signals or other electronic instructions 129 to the manufacturer 130. These instructions 129 automatically prompt or inform the manufacturer 130 to modify its processes. In one example, the manufacturer 130 is a vaccine manufacturer, and the electronic instructions 129 prompt the manufacturer 130 to modify the composition, dosage, or other characteristics of the vaccine (or drug) administered to the humans 102, 104, 106. In some embodiments, this may automatically control the machinery or processes (at the manufacturer's 130 facility) that create the vaccine or drug. In other examples, the electronic instructions include messages to the manufacturer 130 containing alerts or suggested or proposed modifications to the vaccine or drug. The manufacturer 130 may utilize electronic receivers, transmitters, transceivers, memories, databases, servers, processors, control circuits, displays, computers, and other electronic devices (and combinations of these devices) to perform these functions.

[0100] Before treatment of the humans 102, 104, 106, the models 120, 122, 124 are trained. If the models are neural networks, it is understood that the physical structure of these neural networks is altered. For example, the weights, layer structure, and other structures of the neural networks are altered from an initial structure or state to a second structure or state. It is also understood that the structures of the resulting trained models 120, 122, 124 are unique from one another. Each trained model 120, 122, 124 is trained using only data from the corresponding patient. The trained model 120 is trained only with data from the human 102, the trained model 122 is trained only with data from the human 104, and the trained model 124 is trained only with data from the human 106. As such, these models are completely different, unique, and individualized to each unique human and are not considered general-purpose computational resources. Applying the same data to different models does not necessarily produce the same or similar results.

[0101] During the monitoring or execution phase (after the training phase), the control circuitry 116 applies incoming data from a particular human 102, 104, 106 to the model 120, 122, 124 corresponding to that patient. In some aspects, the control circuitry 116 selects or retrieves from memory 118 the appropriate model 120, 122, 124 for the particular data it receives or processes. For example, the data may indicate the human from whom the data originated, or a technician may electronically inform the control circuitry 116 of the origin of particular data, allowing the control circuitry 116 to select the correct model.

[0102] Applying the data to the model generates estimates of that data. These estimates may be for one or more parameters, such as heart rate, heart rate variability, respiration rate, activity level, or body temperature. These estimates are compared to the actual data (for each parameter) to obtain the difference. This difference is the residual, which is determined over time and analyzed by control circuitry 116 for patterns, as described elsewhere herein. The results of this analysis are used by control circuitry 116 to determine and take action. These actions may include controlling machine 126, sending instructions to electronic device 128, and / or sending instructions to manufacturer 130, as described above. While only machine 126, electronic device 128, and manufacturer 130 are shown here, other devices may be controlled or notified by control circuitry 116. It will also be understood that communication between control circuitry 116 and these devices may be bidirectional, i.e., these devices may send instructions or other electronic information to control circuitry 116 for other purposes.

[0103] An example of an approach for detecting an inflammatory response in a human is shown with reference to Figure 2. In step 202, physiological data is collected from at least one wearable sensor attached to the patient during a pre-treatment interval.

[0104] In step 204, an individualized estimation model is created based on physiological data collected prior to the patient's treatment. The model is capable of estimating physiological variables in response to receiving new physiological data from at least one wearable sensor worn by the patient. The individualized estimation model is a multivariate model that provides as output an estimate of the expected value of one or more vital sign values ​​measured from the patient. Such models may be generated by decision tree / random forest function approximators, similarity-based models, neural networks, etc., to name a few.

[0105] In step 206, additional physiological data is collected during the post-treatment interval from at least one wearable sensor worn by the patient. The sensor may be any type of human-wearable sensor or sensor arrangement (including multiple sensors) and may be located at any convenient location on the human body, such as the torso or wrist.

[0106] In step 208, an estimate of post-treatment physiological data is generated using the individualized estimation model. In some aspects, the estimates are multiple individual estimates for parameters such as heart rate, heart rate variability, respiratory rate, activity, and temperature.

[0107] In step 210, the post-treatment physiological data is compared to the estimated values ​​and a determination is made, at least in part, based on the comparison, as to when a predefined effect pattern is present. A difference may be determined between each parameter or variable.

[0108] In step 212, if a predefined effect pattern exists, an action is determined and executed. The action may include one or more of: launching an electronic questionnaire to prompt the patient; providing instructions to the patient to take a measurement; providing instructions to the patient to contact their clinician; registering a ticket in a call center system to call the patient; creating a prompt on the patient's smartphone app to contact the clinician; providing instructions to the patient related to effect triage; sending control signals to control medical equipment related to the patient's treatment; sending instructions to the vaccine manufacturer to modify the vaccine composition and / or dosage. Other example actions are possible.

[0109] Referring to FIG. 3, an overview of determining and utilizing residuals based on the approach provided herein is shown. It will be appreciated that the steps illustrated in connection with FIG. 3 may be implemented as computer instructions executed on a processing unit or control circuit. In step 302, vital sign values ​​from sensor readings are received from a monitored patient. In step 304, the multivariate vital sign observations are filtered to remove irrelevant or poor quality data or activity states that may confound inflammation detection. Various criteria may be used to determine whether the data is irrelevant.

[0110] An individualized estimation model 306 (trained to be individualized for a particular person as described herein) receives the filtered data (filtered in step 304). As described elsewhere herein, the individualized estimation model is a multivariate model that provides as output an estimate of the expected value of one or more vital sign values ​​measured from the patient. Such models may be generated by decision tree / random forest function approximators, similarity-based models, neural networks, etc., to name a few.

[0111] The data (filtered in step 304) may comprise heart rate data, HRV data, respiration rate data, core body temperature data, skin temperature data, and activity data obtained from sensors placed on the human. Other examples are possible.

[0112] The data (filtered in step 304) is applied to the model 306, which in response generates an estimate in step 307. The estimate generated in step 307 represents what the model thinks the data should be. In some aspects, individual estimates are derived from heart rate data, HRV data, respiration rate data, core body temperature data, skin temperature data, and activity data.

[0113] A difference operator 308 takes the difference between the input data (filtered in step 304) and the estimates (generated in step 307) to generate residuals 310. Residuals 310 encode how much each measured vital sign differs from its predicted value based on the patient's pre-treatment data. In some aspects, separate differences are obtained for heart rate data, HRV data, respiration rate data, core body temperature data, skin temperature data, and activity data.

[0114] The residuals 310 may also be combined in step 312 as a single scalar value of overall change, i.e., a time-series change indicator (as described elsewhere in MCI and alternative approaches). In step 314, a pattern representing vital sign disturbances due to inflammation is applied to the residuals. This pattern may, in some aspects, be implemented as a complex rule set (e.g., which may be physically implemented as computer code or software), with the change indicator and measured vital sign values ​​used in rule evaluation to augment the residuals 310. In some aspects, the change indicator may be useful for capturing sufficiently significant overall change, and the measurements may play a role in identifying general indicators of inflammation (e.g., temperature). Because the residuals, change indicators, and measurements are all time-series data, application of the rule set to detect inflammation patterns also results in time-series data, indicating instantaneous inflammatory markers or biomarkers in step 316.

[0115] Given that inflammation is ongoing and may not consistently manifest in vital signs, it may be important to integrate momentary inflammatory biomarkers over time (step 318) to generate a signal that provides ongoing evidence of an inflammatory response. The integration may be statistical within a time window or may be a process of latching (e.g., using or comparing a threshold, as described elsewhere) combined with area-under-curve (AUC) integration, performed in step 320. The results determined in step 320 are used to determine various actions in step 322, which are described elsewhere.

[0116] Referring to Figure 4, one example of an approach for determining the impact of inflammation and taking action based on that impact is shown. Steps 402-410 are part of the model training or learning process, and the remaining steps are part of the monitoring process (which occurs after the training or learning process is complete).

[0117] In step 402, one or more sensors are attached to the patient and worn to acquire data. In some aspects, a torso-mounted sensor patch is used, worn continuously by the patient to enable measurement of multiple variables. These variables may include heart rate and HRV from a single-lead ECG, quantification of movement (activity) from a 3-axis accelerometer, respiration rate from a derivative of movement, amplitude / frequency modulation of HR, and / or skin temperature at the sensor site.

[0118] Vital signs are calculated from the ECG and 3-axis accelerometer waveform data in step 404. By way of example, the ECG may be sampled at 125 Hz and the accelerometer sampling rate may be 15 Hz or higher.

[0119] Heart rate may be calculated beat-by-beat. Respiration rate may be calculated at a sampling rate of 5 seconds, although higher or lower sampling rates are acceptable. Activity may be calculated as a vector of motion (vibration) from all three axes in any unit of measurement known in the art, typically at 1 Hz. Temperature measurements are typically performed at 1 Hz, although lower sampling rates corresponding to the expected rate of change in the human body and surrounding environment are acceptable. In general, all variables should be measured at a rate of at least once per minute, with the aforementioned higher sampling rates being recommended where possible.

[0120] In step 406, vital signs (e.g., HR, RR, HRV, activity, temperature) are statistically summarized in one-minute windows. In some aspects, this is a 10% trimmed mean of the one-minute values. However, other statistics such as median, N-percentile, mode, maximum or minimum are also acceptable and may be used.

[0121] In step 408, a signal quality index (SQI) is used to determine when a vital sign is usable for further processing due to motion artifacts or temporary degradation of the signal-to-noise ratio. In some aspects, the signal quality index may be scaled from 0 (e.g., indicating an inadequate or unacceptable signal) to 1 (e.g., indicating a perfect signal), and the ECG waveform may be used for evaluation. To provide such an SQI, a deep neural network may be reliably trained to output an SQI that correlates highly with human expert assessment of the usability of the ECG trace. The SQI input may, for example, include a 10-second ECG window with approximately 7 to 30 heartbeats (depending on the heart rate), evaluated at 5-second intervals (e.g., 5-second overlap). Other methods of generating an SQI using an ECG waveform are known to those skilled in the art. If an SQI ranging from 0 to 1 is obtained, the threshold for excluding data from processing may be set to <0.8. Other examples are possible.

[0122] In addition to SQI, activity level may also act as a filter for determining whether or not data should be used for processing. An absolute activity level value may be applied so that data above that value will not be used. This method provides a simple means of eliminating irrelevant data and can be set at a level of activity that the monitored person rarely exhibits during the day, preventing the loss of significant data needed for processing. The activity threshold for excluding data from processing can be set at a unit level of movement equivalent to brisk walking.

[0123] After the one-minute windowed vital sign averages are filtered using SQI and activity thresholds, we are left with HR, RR, HRV, activity and temperature samples collected from the monitored person and used to train individualized baseline models.

[0124] In step 410, as described elsewhere herein, many approaches can be used to create a reference model (individualized estimation model) that, after training, can generate estimates for comparison with measurements. In some aspects, similarity-based models (SBMs) are used. Such models can be generated from as few as one or two days of continuous minute-by-minute data (1440 samples per day) from a human circadian cycle, although more days of data can be used if the use case allows. For example, if a patient has a scheduled treatment at a later date, a week's worth of samples can be collected to build an individualized model.

[0125] Once the model is trained, it is transitioned or utilized in a monitoring mode in step 412, generating estimates in response to the input of each new sample of minute-by-minute multivariate vital signs data obtained from the monitored individual. This process is performed continuously, applying the filters described above, across all activities of daily living. As multivariate readings of HR, RR, HRV, activity, and temperature are obtained, they are input into the model, which generates estimates of their expected minute-by-minute values.

[0126] In step 414, residuals are generated by subtracting the estimated values ​​from the minute-by-minute measurements. For example, if the minute-by-minute heart rate is estimated to be lower than the estimated value, the residual will be positive, indicating that the actual heart rate is higher than expected. Residuals are generated for each vital sign parameter, and the residuals occur at the same sampling rate of once per minute.

[0127] In step 416, the residuals are combined into a multivariate change index (MCI) to quantify the overall variability between the set of measured and estimated vital signs. The method for calculating the MCI is described elsewhere. In some aspects, the MCI is scaled between 0 (meaning no abnormal behavior) and 1 (meaning high confidence in abnormal behavior). The MCI is generated at a sampling rate of 1 minute for each input observation.

[0128] Using the available minute-by-minute vital signs, residuals, and MCI as inputs, an inflammation-specific detection pattern is applied in the form of a rule. In step 418, the rule outputs an inflammation-specific modified MCI, referred to herein as iMCI. As an example, the rule logic follows these steps:

[0129] Initialize iMCI=0

[0130] If ((HRV RES <=0) and (Temp RES >=0) and (MCI>=0.1)) then:

[0131] If ((HR RES >0.5) or (RR RES >0.5) or (Temp RES >0.2) or (HRV RES <-0.025)) then:

[0132] iMCI=Set MCI

[0133] where:

[0134] HR RES is measured in beats per minute.

[0135] RR RES is measured in breaths per minute.

[0136] Temp RES is measured in degrees Fahrenheit.

[0137] HRV RES is measured in seconds.

[0138] This rule essentially zeros the MCI value unless the evidence is characteristic of inflammation, making the iMCI specific to inflammation. The first part of this rule gates on relatively high HRV, relatively low temperature, or generally insufficient overall change (MCI). In other words, this rule essentially looks for sufficient disturbance to be quantified by an MCI, unless that disturbance is due to higher than expected HRV or lower than expected temperature. Once this initial gating condition is met, the second part of the rule looks for evidence of possible inflammation in the form of higher than expected HR, higher than expected RR, higher than expected temperature, or lower than expected HRV, and determines the 1-minute inflammation MCI (M_iMCI) in step 420. If one or more of these types of evidence are present at levels above their respective thresholds, the iMCI is set to the value of the MCI.

[0139] M_iMCI is also a time series with a sampling rate of 1 minute. In step 422, the iMCI values ​​are summed or averaged within a window to check for persistence of evidence of inflammatory physiology disturbances and also smooth the signal from transient noise. In some aspects, this window is 3 hours, with averaging evaluated every 15 minutes (2 hours 45 minutes overlap). Other statistical characterizations of the windowed iMCI are possible.

[0140] In another aspect, in step 422, if the time series value continually exceeds a "latching" threshold, the one-minute iMCI time series is "latched" (frozen) at its highest value until the value falls sufficiently below a certain threshold and the iMCI is unlatched from its high value and frozen at a lower value. Then, in step 424, a three-hour sum or average of the iMCI is made of the latched version of the iMCI.

[0141] Latching can be illustrated with the following integer example, as shown in the first column below: Given a time series of instantaneous values: the latching threshold is 6, and there is a persistence condition of 2 samples. After the first sequence of 6 and 7, the sequence will latch to its highest value, as shown in the second column below, until the original sequence drops below the value 6 for at least two consecutive samples and is unlatched. [4,6,5,5,6,7,8,7,6,7,5,3,1,2,1] [4,6,5,5,6,7,8,8,8,8,8,8,1,2,1]

[0142] By itself, the 1-minute iMCI, the 3-hour windowed (summed or averaged) iMCI, or the latched iMCI with a 15-minute sample rate can serve as a quantification index for suspected inflammatory responses. This may be useful for quantifying trends or assessing changes before and after treatment in clinical trials. In patient drug trials, changes in the windowed iMCI or latched iMCI before and after treatment administration can be examined to compare the responses of different cohorts, for example, treatment and control groups. The 1-minute iMCI, 1-minute latched iMCI, windowed iMCI, or windowed latched iMCI can be aggregated over time (e.g., days) to quantify the overall relative inflammatory response in a patient or study participant's physiology.

[0143] In step 426, an action is taken. For example, if it is useful to detect inflammation, such as an acute inflammatory episode or cytokine release syndrome, and escalate the notification, the windowed or latched iMCI is compared to a threshold value. If the value exceeds the threshold, an acute inflammation level is recognized, an alert is sent to the clinician and / or patient, and steps are taken to mitigate the acute response. In some aspects, a notification is triggered when the windowed or latched iMCI value exceeds 0.25. This threshold helps distinguish acute inflammation levels (which are different from those requiring intervention) from expected inflammation levels (which do not require triage). This approach is useful, for example, when cancer treatments that harness the power of the human immune system are used. With such treatments, moderate inflammation levels are expected as the cancer treatment is working, but there is also a risk of acute episodes of inflammation, such as a potentially fatal cytokine "storm."

[0144] In yet another aspect, the threshold used to escalate or send an alert for acute inflammation is applied to the difference between windowed or latched iMCI samples: consecutive 15-minute samples of a 3-hour windowed value are checked for a sudden rise; if the rise is at least as large as the threshold, an alert is triggered, regardless of the absolute value.

[0145] Referring to FIG. 5, a safety monitoring system for patients undergoing medical procedures (e.g., patients who may develop undesirable cytokine release syndrome as a side effect, which may be dangerous to the patient) is shown. In some aspects, the system provides a safety net for patients leaving an acute care facility and moving to a home environment away from the acute care facility. The system includes a wearable torso patch sensor 505 that transmits data via Bluetooth wireless (or other communication technology or protocol) to a smartphone 510 carried by the patient. The sensor data is transmitted by an app (e.g., a software application) on the smartphone and transmitted over a public data infrastructure 515, which may be the Internet or other mobile data infrastructure, to a data center (also referred to as a central call center) that includes a processor unit (or control circuitry) 520 to which data storage memory 525 is connected. This central call center or data center may or may not be associated with a physician's office, hospital, acute care facility, hospital complex, clinic, or other healthcare facility, and may be a data processing center in a different geographic location from the healthcare facility. Other locations are also possible. The processor 520 and associated data storage memory 525 may be physically located at a central call center, but may also be located elsewhere, such as in the cloud or at another medical facility.

[0146] The patient's pre-treatment sensor data is used to train a personalized multivariate inferential model 530, which is stored in memory 525 and accessible to processor 520 for computational steps performed by the processor 520, including generating expected values ​​for the data measured by the sensors 505 and comparing them to generate univariate scores such as residuals and MCI, which are in turn used to detect signatures that escalate the patient's inflammatory process.

[0147] Buttons on the smartphone 510 may be used to initiate actions. When a button press or actuation is detected, the processor 520 sends instructions to an app on the smartphone, prompting actions for the patient embodied in on-screen button functions: button 540, which displays a survey on the smartphone to confirm or deny symptoms associated with inflammation; button 545, which automatically connects the patient to a clinician responsible for the patient's management, including a call center with trained staff for these tasks; and button 550, which allows the patient to take additional measurements using home-provided equipment, such as a thermometer to measure core body temperature, a blood pressure monitor, or a pulse oximeter, and report the results. In some aspects, as described, buttons 540, 545, and 550 are displayed on the smartphone 510 screen. The user can press or activate these buttons by touching or pressing the screen in the appropriate area where the button appears (e.g., using a finger, cursor, or stylus). Alternatively, buttons 540, 545, and 550 may be physical buttons on the smartphone 510, and the user may physically depress them.

[0148] More specifically, during some operations, processor 520 obtains sensor data 570, applies it in step 572 to generate an estimate, uses a difference operator 574 to subtract sensor data 570 from the estimate (obtained in step 572) to obtain a residual, and determines MCI as described herein in step 576. Rules are used to identify inflammation patterns in step 578 and actions to take in step 580. These steps are described in more detail elsewhere herein.

[0149] With respect to button 540, for example, pressing the button displays a question on the screen of smartphone 510. The question may include multiple pages through which the patient paginates (each page displayed on a separate screen). The patient may answer the question as they paginate through the question. In one example, depending on the nature and format of the question, the patient may type information or select checkboxes. In another example, the patient may verbally enter information into smartphone 510. That is, when a question is presented, the patient verbally states the answer, a microphone in smartphone 510 receives the information, and the information is received by a processor or control circuitry in the smartphone for further processing.

[0150] For button 545, for example, a patient may be identified by a patient number, code, and / or other suitable identifier. When button 545 is pressed, the app sends an electronic message containing the code to a call center (or other facility). In this manner, pressing button 545 may establish a telephone, video, or other electronic communications connection or link with a central call center. Each patient (via the patient number or code) may be associated with a particular clinician. Thus, in one example, the central call center's control circuitry 520 (associated with the clinician) may receive the patient code or number, identify the clinician associated with that number or code, and establish a link between the patient and the associated clinician.

[0151] The central call center stores separate and distinct procedures for contacting or reaching out to each different clinician. Each of these procedures is stored as a separate program or computer instruction (e.g., stored in memory 525) and executed (e.g., by control circuitry 520) as needed. For example, for a first clinician, a text message may be automatically generated instructing the clinician to call the patient. In another example, a call (or other electronic communication) to the clinician may be automatically initiated, and upon the clinician's approval, a direct telephone link or connection may be established between the clinician and the patient. It will be appreciated that execution of such computer instructions may automatically control electronic communications equipment by configuring switches, routing messages between networks and their components, e.g., gateways, and other various electronic devices within infrastructure 515. Once the appropriate clinician is identified, the central call center control circuitry 520 identifies, retrieves, and executes the appropriate program to contact the clinician.

[0152] To verify the patient's identity, a security or verification procedure may be performed automatically or manually at a central call center. For example, the patient code may be checked against a database of valid patient codes by the control circuitry 520. In another aspect, security questions may be presented to the patient on a smartphone before processing continues. In other words, the patient's correct answers to these questions are required to continue the process.

[0153] More specifically, the patient's answers are transmitted to a central call center, where control circuitry 520 compares the patient's answers with those stored in the central call center (e.g., memory 525). A correct match establishes a link and / or exchanges information. In the event of a mismatch, processing operations related to that patient may be interrupted. In this case, the program attempting to establish the link may be stopped, or an electronic message may be sent to smartphone 510 informing the patient that they did not pass the security check.

[0154] In yet another example, biometric information may indicate the patient's identity. In some aspects, the patient places a finger on the screen of the smartphone 510 (or other fingerprint capturing device connected to the smartphone 510), and the information is transmitted to a central call center. The control circuitry 520 compares the fingerprint to fingerprints stored on file at the central call center.

[0155] In yet another example, initiation of a process may involve obtaining a patient's insurance information. The insurance information may be stored in a central call center in some examples, but may be stored in an external source in other examples. If necessary, an electronic connection may be automatically established (e.g., automatically by control circuitry 520) between the central call center and the external source to access, obtain, or query the insurance information. The insurance information may be used to determine whether or how to proceed with the process. In some examples, depending on the nature of the insurance information, the process and the program that executes it may be electronically stopped by control circuitry 520 at the central call center. Infrastructure 515 may be used to establish the connections necessary to obtain information from external sources.

[0156] For button 550, the report may be entered manually or obtained from the measurement device via Bluetooth by a smartphone app. In some aspects, the patient may first establish a communication link between the smartphone 510 and the additional device. For example, the smartphone may attempt to detect the presence of the additional device in a setup mode. This process may be accomplished using a number of different protocols, whereby a link is established upon detection. For example, the smartphone 510 may send a signal to the additional device, which responds. Alternatively, the additional device may send a signal that is detected by the smartphone 510. Once the link is established, the smartphone 510 may receive information from the additional device periodically or continuously.

[0157] Measurements and questions answered by the patient are automatically uploaded to a data center via an app, and the detected inflammation patterns and corresponding patient feedback are notified to clinicians. Using this system offers the significant advantage of continuously monitoring patients in a home environment without the expense of full-day hospitalization in an acute care facility. Furthermore, personalized, model-based early warning signs of escalating inflammation, coupled with automated actions provided to patients, make clinicians' decision-making in the event of escalating detections more actionable and improve the timing of interventions, if necessary.

[0158] The approaches provided in this paper are beneficial for a variety of reasons. For example, they are personalized, offering greater sensitivity and specificity. These approaches learn from data how a person's physiology functions, and if inflammation impacts that physiology, the changes will be noticeable.

[0159] Furthermore, the approach presented here is objective: it does not rely on subjective perceptions such as symptom reporting.

[0160] Additionally, the approach presented here does not rely specifically on the patient actively performing tasks to collect data; the patient simply wears the sensor. In contrast, other methods that attempt to detect severe inflammatory responses include asking patients to periodically take their temperature. Relying on patients to do something at regular intervals is always problematic.

[0161] The approach uses a personalized model tailored to a specific patient and also looks for evidence of side effects as patterns in the difference between the values ​​predicted by the model and the values ​​actually measured by the patient's sensors.

[0162] This approach provides patients with instructions for managing side effects, which may include answering questions about symptoms or severity, visiting a clinic for testing or treatment, taking corrective medications, and / or taking manual measurements with other devices.

[0163] Models can be characterized as either (a) "dynamic," meaning that the input data to the model influences how estimates are obtained, or (b) "multivariate models." The "estimates" are static mean values, and deviations from those static values ​​are considered "individualized" residuals and are examined for inflammation "signatures" (significant variances).

[0164] Those skilled in the art will recognize that various other modifications, changes, and combinations are possible in connection with the above-described embodiments without departing from the scope of the present invention, and that such modifications, changes, and combinations are within the concept of the present invention.

Claims

1. 1. A method for monitoring the effect of pharmacological treatment in a patient, comprising: collecting physiological data during a pre-treatment period from at least one wearable sensor worn by the patient; creating an individualized estimation model based on physiological data collected prior to treatment of the patient, the individualized estimation model capable of estimating physiological variables in response to new physiological data received from the at least one wearable sensor worn by the patient; collecting additional physiological data from the at least one wearable sensor worn by the patient during a post-treatment period; generating an estimate of post-treatment physiological data using the individualized estimation model; comparing the post-treatment physiological data with the estimate and determining whether a predefined pattern of effect exists based at least in part on the comparison; determining and executing an action if the predefined effect pattern exists, the action comprising: triggering an electronic question to be prompted to the patient; providing instructions to the patient to take the measurement; providing the patient with instructions to contact the patient's clinician; triggering a ticket in a call center system to queue a call to said patient; generating a prompt in an app on the patient's smartphone to contact a clinician; providing instructions to said patient relating to triage of said effects; transmitting control signals to control medical equipment associated with the treatment of the patient; sending instructions to the vaccine manufacturer to modify the vaccine composition and / or dosage; method.

2. the effect is a side effect, The method of claim 1.

3. said effect being an indication of vaccine efficacy; The method of claim 1.

4. the physiological data comprises heart rate data, respiratory rate data, core body temperature data, skin temperature data, and activity data; The method of claim 1.

5. the individualized prediction model is trained using data collected from the patient during a free-living physiological state in which the patient's inflammatory state is expected to be stable and unchanged; The method of claim 1.

6. the comparing comprises determining a residual between the estimate and the physiological data. The method of claim 1.

7. the residuals are combined into a single score, the score being a scalar index; The method of claim 6.

8. the individualized estimation model comprises a neural network; The method of claim 1.

9. 1. A system for monitoring the effect of pharmacological treatment in a patient, comprising: at least one wearable sensor worn by the patient during a pre-treatment period, the at least one wearable sensor configured to collect physiological data from the patient; a control circuit coupled to the at least one wearable sensor, the control circuit creating an individualized prediction model based on physiological data collected prior to treatment of the patient, the model capable of estimating physiological variables in response to receiving new physiological data from the at least one wearable sensor worn by the patient; the at least one wearable sensor collecting additional physiological data from the at least one wearable sensor worn by the patient during a post-treatment period; The control circuit further comprises: generating an estimate of the physiological data after treatment using the individualized estimation model; comparing the post-treatment physiological data with the estimate and determining whether a predefined pattern of effect exists based at least in part on the comparison; and If the predefined effect pattern exists, providing instructions to said patient relating to triage of effects; transmitting control signals to control medical equipment associated with the treatment of said patient; determining and performing one or more actions of: sending instructions to the vaccine manufacturer to change the composition and / or dosage of the vaccine; system.

10. the effect is a side effect, The system of claim 9.

11. said effect being an indication of vaccine efficacy; The system of claim 9.

12. the physiological data comprises heart rate data, respiratory rate data, core body temperature data, skin temperature data, and activity data; The system of claim 9.

13. the control circuitry trains the individualized prediction model using data collected from the patient during a free-living physiological state in which the patient's inflammatory state is expected to be stable and unchanged; The system of claim 9.

14. the control circuitry is configured to perform the comparison by determining a residual between the estimate and the physiological data; The system of claim 9.

15. The control circuitry is configured to combine the residuals into a single score, the score being a scalar index. The system of claim 14.

16. the individualized estimation model comprises a neural network; The system of claim 9.

Images