Systems and methods for predicting blood glucose

Abstract

A method for predicting a blood glucose or blood glucose condition in an individual. The method includes receiving, at a current time, data associated with the individual, the data including (i) a set of one or more static data points each being generated or collected at an identical prior point in time occurring before the current time, and (ii) a series of sequential data points, each of the sequential data points being generated or collected within a prior window of time occurring before the current time; inputting the data associated with the individual into a model; and receiving from the model an indication of a risk of the individual developing hypoglycemia within a subsequent window of time after the current time.

Description

1 SYSTEMS AND METHODS FOR PREDICTING BLOOD GLUCOSE CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] The present application is based on and claims priority to U.S. Provisional PatentApplication No.63 / 675,784, filed on July 26, 2024 and titled “SYSTEMS AND METHODS FOR PREDICTING BLOOD GLUCOSE,“ and U.S. Provisional Patent Application No. 63,688,658, filed on August 29, 2024 and titled “SYSTEMS AND METHODS FOR PREDICTING BLOOD GLUCOSE,” all of which are incorporated herein by reference, in their entirety. TECHNICAL FIELD

[0002] The present disclosure relates generally to systems and methods for predictingblood glucose, and more particularly, to systems and methods for predicting blood glucose using static and sequential data, and a model. BACKGROUND

[0003] Abnormal blood glucose, like low blood glucose (i.e., hypoglycemia), can be amajor problem in hospitals, and can cause various complications for admitted patients. However, there are a multitude of different factors that can affect blood glucose levels during admission, which makes managing glucose levels difficult in patients. Thus, new and improved systems and methods are needed for controlling blood glucose and complications to patients. SUMMARY

[0004] According to some implementations of the present disclosure, a method forpredicting blood glucose in an individual comprises receiving, at a current time, data associated with the individual, the data including (i) a set of one or more static data points each being generated or collected at an identical prior point in time occurring before the current time, and (ii) a series of sequential data points, each of the sequential data points being generated or collected within a prior window of time occurring before the current time. The method further comprises inputting the data associated with the individual into a model. The method further comprises receiving from the model an indication of a risk of the individual developing hypoglycemia within a subsequent window of time after the current time.

[0005] According to some implementations of the present disclosure, a system for4912-3076-7444.4 065472-000953WOPT2 predicting blood glucose in an individual comprises at least one electronic interface, at least one memory device, and at least one processing device. The least one electronic interface is configured to receive data associated with the individual, the data including (i) a set of one or more static data points each being generated at an identical prior point in time occurring before a current time, and (ii) a series of sequential data points, each of the sequential data points being generated within a prior window of time occurring before the current time. The at least one memory device is configured to store the data associated with the individual. The at least one processing device is configured to implement a model that receives the data associated with the individual and outputs an indication of a risk of the individual developing hypoglycemia during a subsequent window of time after the current time.

[0006] According to some implementations, a method for predicting a blood glucosecondition is provided. In some aspects, the method includes receiving data associated with an individual, the data including (i) a set of one or more static data points each being generated or collected before a current time, and (ii) a series of sequential data points, each of the sequential data points being generated or collected within a prior window of time occurring before the current time. In some aspects, the method includes inputting the data associated with the individual into a model. In some aspects, the method includes receiving, from the model, an indication of a risk of the individual developing a blood glucose condition within a subsequent window of time of the current time. In some aspects, the method includes receiving, accessing, and / or generating theoretical sequential data and / or altered sequential data associated with a proposed intervention, and inputting the theoretical sequential data and / or altered sequential data into the model. In some aspects, the method includes receiving, from the model, an updated indication of risk of the individual developing a blood glucose condition within the subsequent window of time. In some aspects, one or more control signals may be generated based on the indication of a risk and / or updated indication of risk, and transmitted to a treatment apparatus. In some aspects, responsive to the transmitted one or more control signals, the treatment apparatus may initiate, modify, pause, adapt, and / or cease administration of a treatment.

[0007] The above summary is not intended to represent each implementation or everyaspect of the present disclosure. Additional features and benefits of the present disclosure are apparent from the detailed description and figures set forth below. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The disclosure, and its advantages and drawings, may be understood from the4912-3076-7444.4 065472-000953WOPT3 following description of representative embodiments together with reference to the accompanying drawings. The drawings depict only representative embodiments, and should not be considered as limitations on the scope of the various embodiments or claims.

[0009] FIG. 1 shows a block diagram of an example system for predicting hypoglycemia,according to aspects of the present disclosure;

[0010] FIG. 2 shows a flowchart setting forth steps of a method for predictinghypoglycemia, according to aspects of the present disclosure;

[0011] FIG. 3 shows an illustration showing a dataset used for training a model to predicthypoglycemia, according to aspects of the present disclosure;

[0012] FIG. 4A shows a graph illustrating a distribution of lowest blood glucose levelrecord per encounter for the dataset of FIG.3, according to aspects of the present disclosure;

[0013] FIG. 4B shows a graph illustrating overlayed density curves versus number of daysuntil a blood glucose level of <70 mg / dL, <60 mg / dL, and <50 mg / dL was measured per encounter for the dataset of FIG.3, according to aspects of the present disclosure;

[0014] FIG.5A shows a graph illustrating model area under precision recall curve (AUPR)for a long short-term memory (LSTM) model as a function of training epochs completed for both a training portion of the dataset of FIG.3 and a held-out validation portion of the dataset of FIG.3, according to aspects of the present disclosure;

[0015] FIG. 5B shows a graph illustrating model loss for an LSTM model as a function oftraining epochs completed for both a training portion of the dataset of FIG. 3 and a held-out validation portion of the dataset of FIG.3, according to aspects of the present disclosure;

[0016] FIG. 6 is an illustration showing an example timeline of a lookback window afterpatient admission during which sequential data is generated and / or collected, and prediction horizon following the lookback window during which the blood glucose level of the patient is predicted, according to aspects of the present disclosure;

[0017] FIG. 7 shows a flowchart setting forth steps of a method, according to aspects ofthe present disclosure;

[0018] FIG. 8A is an illustration showing an example timeline for collecting model input(baseline characteristics, medications, and laboratory measurements) and output (4h POC BG) variables over a course of a patient admission, according to aspects of the present disclosure;

[0019] FIG. 8B is an illustration showing an example analysis pipeline demonstratingextraction of variables with Lasso, assessment of statistical significance with MELR, and evaluation of novelty of medication-induced BG variability using Micromedex, according to aspects of the present disclosure; 4912-3076-7444.4 065472-000953WOPT4

[0020] FIG. 9A shows a graph indicating most administered medications determinedaccording to aspects of the present disclosure;

[0021] FIG. 9B shows a graph indicating an example relationship between a mean POCBG over 24h lookback window and a 4h POC BG, according to aspects of the present disclosure;

[0022] FIG. 10A shows a graph MSE versus log10(alpha) for multiple Lasso-CV folds,according to aspects of the present disclosure;

[0023] FIG. 10B shows a graph comparing Actual vs Lasso-predicted 4h POC BG,according to aspects of the present disclosure;

[0024] FIG. 10C shows a graph indicating variable coefficients assigned by Lasso,according to aspects of the present disclosure;

[0025] FIG. 10D shows a graph indicating mean absolute coefficient scores of top Lassovariables determined according to aspects of the present disclosure;

[0026] FIG. 11A shows a graph indicating 4 hr POC BG output versus ketorolac injecteddose, according to aspects of the present disclosure;

[0027] FIG.11B shows a graph indicating 4 hr POC BG output versus fentanyl intravenousdose, according to aspects of the present disclosure;

[0028] FIG. 11C shows a graph indicating 4 hr POC BG output versus rifaximin tab dose,according to aspects of the present disclosure;

[0029] FIG. 11D shows a graph indicating 4 hr POC BG output versus norepinephrineintravenous dose, according to aspects of the present disclosure;

[0030] FIG. 12 is an illustration depicting significant variables unique to a number oflookback windows, according to aspects of the present disclosure;

[0031] FIG. 13 is an illustration depicting an example architecture for an LSTM model, inaccordance with aspects of the present disclosure;

[0032] FIG. 14 is an illustration showing an example prediction framework, in accordancewith aspects of the present disclosure;

[0033] FIG. 15 shows graphs indicating temporal patterns of POC BG measurements, inaccordance with aspects of the present disclosure;

[0034] FIG. 16 shows graphs indicating model predictions for a test patient, in accordancewith aspects of the present disclosure;

[0035] FIG. 17 shows graphs indicating predictors and insulin dosing pattern precedinghypoglycemia for an example test patient, according to aspects of the present disclosure; and

[0036] FIG.18 shows graphs indicating daily performance metrics across multiple decision4912-3076-7444.4 065472-000953WOPT5 thresholds for a trained prediction model, in accordance with aspects of the present disclosure.

[0037] While specific implementations and embodiments are shown in the drawings anddescribed in detail herein by way of example, the present disclosure is susceptible to various modifications and alternative forms. Hence, it should be understood that the present disclosure is not intended to be limited to particular forms disclosed. Rather, the present disclosure covers all modifications, equivalents, and alternatives falling within the spirit and scope of the present disclosure as defined by the appended claims. DETAILED DESCRIPTION

[0038] Hypoglycemia and other blood glucose (BG)-related conditions can pose asignificant problem. In general, insulin is used to treat hyperglycemia (typically defined when BG exceeds 140 mg / dL) for hospitalized patients (with or without diabetes). Twenty percent of hospitalized patients receiving insulin have at least one hypoglycemic event (BG<70 mg / dL) and 7% have at least one severe hypoglycemic event (BG<50 mg / dL). Incidence of hypoglycemia in critically ill patients undergoing glycemic control (e.g., with a target glucose between 80 and 100 mg / dL) has been reported as high as 45%. Patients admitted for diabetes or hyperglycemia have on average two hypoglycemic events / week. In a healthcare setting, patients with hypoglycemia can often stay longer and have increased risk of death compared to those without hypoglycemia (adjusted OR=1.66, 95% CI 1.55-1.78). Mortality incidence is higher with increased severity of hypoglycemia. In some cases, a BG<40 mg / dL can be designated as a “never event,” and result in funding to be denied to a hospital per the Centers for Medicare and Medicaid Services.

[0039] Therefore, prevention of certain BG-related events, such as a first hypoglycemicevent, can be highly desirable, and can be critical to patient outcomes. While hospitals and other healthcare facilities can provide controlled environments, BG complexity and dynamics makes achieving euglycemia challenging using conventional techniques. This is because there are a multitude of factors affecting BG (e.g., diet, renal function, antihyperglycemic drugs and doses, cognitive status, postoperative stress, age, and so on), which can make predicting and treating certain BG-related conditions difficult, if not impossible. Further, such factors often fluctuate during hospitalization due to changes in acute illness, nutritional intake, evolving insulin needs, and so forth.

[0040] Yet despite these challenges, there are no widely accepted tools and methods forassessing risk and / or treating patients prone to or suffering from certain BG and BG-related conditions, like hypoglycemia, diabetes, and so forth. Rather, clinical resources are often 4912-3076-7444.4 065472-000953WOPT6 applied to patient cases after a BG-related event has already occurred. Further, provider confidence, knowledge gaps, complex workflows, and clinical inertia, may often impede effective inpatient glycemic management, and result in aggravated conditions or undesired complications. Therefore, improved technologies are needed, which can aid in identifying and treating BG and BG-related conditions.

[0041] Disclosed herein are systems and methods for predicting BG and BG-relatedconditions in individuals, such as patients admitted to hospitals or other healthcare facilities. In some implementations, disclosed systems and methods provide a technical approach that relies on static data generated at the time of admission (e.g., race and sex, which do not change through the course of the admittance), and / or sequential data generated after admission (e.g., while an individual is still admitted) to determine risk to an individual. In some implementations, a model can be used to process or analyze static data and / or sequential data from an individual, and generate a prediction indicative of a risk of the individual to developing a blood glucose or blood glucose condition, such as hypoglycemia, within a subsequent window of time.

[0042] As appreciated from details provided herein, the present approach solves a numberof technological problems associated with conventional systems and methods of disease treatment, and provides a number of advantages and improvements to prior diagnosis and treatment-related technologies. For instance, unlike conventional diagnosis and / or treatment systems, in some implementations, the present systems and methods allow for advanced warning of a BG or BG-related condition, or indication of a risk thereof, within a subsequent window of time. Further, unlike conventional diagnosis and / or treatment systems, in some implementations, the present systems and methods allow for advanced action based on predicted BG or BG-related conditions (e.g., initiating, modifying, pausing, adapting, and / or ceasing administration of a treatment). Providing advanced warning and treatment action (or recommendation for treatment) can aid in reducing and / or eliminating undesirable effects or clinical scenarios. For instance, the present approach can aid in reducing and / or eliminating the number hypoglycemic events experienced by a patient. Also, the present approach can provide significant advanced warning (e.g., up to 24 hours, or more), which can be more clinically useful because it allows for sufficiently advanced action (e.g., drug administration, nutrition, and so forth) to be effective. The present approach can also aid in improving outcomes, reducing admission times, reducing clinician fatigue, and so forth.

[0043] FIG. 1 illustrates a block diagram of a system 100, which can be used to aid inpredicting future blood glucose levels or blood glucose conditions, in accordance with aspects 4912-3076-7444.4 065472-000953WOPT7 of the present disclosure. The system 100 can include one or more processing devices 110, which can each include any one or more of a processor 112, a memory 114, a user output device 116 (e.g., a display), a user input device 118, and / or other components. In general, a processing device 110 may include and / or cooperate with any computing device, such as a personal computer, laptop, mobile device (e.g., smartphone, tablet, personal digital assistant (PDA), etc.), wearable device, mainframe, local server, remote servers (e.g., cloud-based server), and so forth.

[0044] The processor(s) 112 can carry out various processing tasks to operate theprocessing device 110. Generally, the processor 112 can include any number of suitable processing devices, such as one or more general purpose computing system, microprocessors, digital signal processors, micro-controllers, computer processing unit (CPUs), graphical processing units (GPUs), application specific integrated circuits (ASICs), programmable logic devices (PLDs) field programmable logic devices (FPLDs), programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), and so forth.

[0045] In some implementations, the processor(s) 112 may include a machine learningengine 119 configured to carry out steps in accordance with aspects of the present disclosure. For instance, in some implementations, the machine learning engine 119 may be configured to construct a machine learning model, as well as other models. In particular, a machine learning model may include a variety of architectures, inputs, and outputs, and may be trained in accordance with aspects of the present disclosure. By way of example, a machine learning model can include a long short-term memory (LSTM) model, a model that includes one or more LSTM blocks, a transformer model, a logistic regression model, or any other suitable model. Moreover, in some implementations, multiple models may be used, and / or features from multiple different models may be used.

[0046] In some implementations, a machine learning model generated using the machinelearning engine 119 may be configured to receive various static and / or sequential data, and output one or more outputs, indications, and / or parameters, such as indication of a risk or a probability that an individual will develop a blood glucose condition, such as hypoglycemia. To this end, the machine learning engine 119 may be configured to carry out training and / or validation processes to train the machine learning model using one or more training datasets and / or validation datasets (e.g., including static and / or sequential data). Training and / or validation may include optimizing various model parameters (e.g., number of units, number of layers, number of trials, weights, dropout rate, learning rate, batch size, activation function, and so forth) to achieve a predetermined model performance metric (e.g., precision, recall, F1 4912-3076-7444.4 065472-000953WOPT8 score, AUPR score, and so forth). Further, various training and / or optimization techniques may be used, such network weight initialization (e.g., using Glorot uniform, He normal, etc), optimizer to update model weights (e.g., Adam RMsprop, SGD, etc), regularization (e.g., L1 regularization, L2 regularization, etc), early stopping (e.g., to stop training when validation performance begins to degrade), gradient clipping (e.g., to prevent exploding gradients), and so forth. To this end, various training datasets and / or validation datasets may be obtained from the memory 114 and / or EHR system 120, as well as from other data locations or systems providing or generating such data.

[0047] In some implementations, data obtained from the EHR system 120, or elsewhere,may be pre-processed (e.g., the processor(s) 112 and / or machine learning engine 119) to generate various training datasets and / or validation datasets to be input in a model. For example, various input data may be generated through one or more pre-processing steps, such as combining duplicative variable names (e.g., glucose(POC) and glucose-POC), combining doses of a given medication, removing medication doses or lab values with missing or zero values, averaging lab or medication values, assigning indicator values (e.g., a “0” or “1” depending on whether a non-NPO diet value was present), and so forth. In some implementations, input data may be generated by using datapoints corresponding to selected timesteps (e.g., occurring over a predetermined lookback window), to create one or more input sequences.

[0048] In one non-limiting example, a machine learning model may be an LSTM modelwith an architecture includes a number of input layers configured to receive static and / or dynamic data, a concatenate layer configured to merge the inputs, one or more dense layers for final processing, and an output layer providing a classification. As illustrated in the non- limiting example of FIG. 13, in some implementations, the classification may be a binary classification indicating whether a BG metric (e.g., , a BG level measured in mg / dL, mmol / L, and so forth) is greater than and / or equal to a predetermined threshold, or a BG less than the predetermined threshold (e.g., a predetermined threshold blood glucose level). In some non- limiting examples, the predetermined threshold is approximately 70 mg / dL, or less, or more. During training, the LSTM model may be optimized using various approaches (e.g., KerasTuner Bayesian optimization technique). For instance, various hyper parameters (e.g., layers, units, dropout rate, learning rate, batch size, activation functions, and so forth) may be optimized using training dataset. To note, while FIG. 13 illustrates an LSTM model with a specific architecture and inputs, variations may be possible, including fewer, more, and / or different inputs, blocks, layers, outputs, and so forth. 4912-3076-7444.4 065472-000953WOPT9

[0049] Referring again to FIG. 1, the memory 114 can include machine-readableinstructions (e.g., executable by the one or more processors 112 or machine learning engine 119) for carrying steps of methods disclosed herein, and / or other methods. The memory 114 can also store various data and information, such as the static data and the sequential data. As such, the memory 114 can include any suitable memory device and / or machine-readable medium that is capable of storing, encoding, and / or carrying a set of instructions for execution, for example, by a processing device 110 and that cause the processing device 110 to perform and / or implement any of the features or steps discussed herein, including solid-state memories, optical media, magnetic media, random access memory (RAM), read only memory (ROM), a floppy disk, a hard disk, a CD ROM, a DVD ROM, flash memory, or other computer readable medium that is read from and / or written to by a magnetic, optical, or other reading and / or writing system that is coupled to the processing device, can be used for the memory or memories.

[0050] In general, the output device(s) 116 can be used configured to provide any data andinformation, in accordance with methods described herein. For instance, in some implementations, the output device 116 may provide a report, e.g., as generated by the processor(s) 112, via a display. As such, the output device 116 can include any known output technology, such as, but not limited to, display devices that use Liquid Crystal Display (LCD), Light Emitting Diode (LED), and other technologies The report provided by the output device 116 may have any form and include any information. For example, the report may include results of analysis, such as indication of risk or probability for an individual to develop a blood glucose condition (e.g., hypoglycemia). In some implementations, the report may include a warning, a recommendation, and so forth, as detailed further herein. In some implementations, the report, or portions thereof, may be stored in a memory 114, a patient record, and elsewhere.

[0051] The user input device(s) 118 can be configured to allow a user to interact with thesystem 100 for any suitable purpose, in accordance with methods described herein. For instance, in some implementations, the user input device(s) 118 may allow for initiating, pausing, or terminating the analysis by a model, inputting or adjusting one or more parameters of an analysis or model training, providing instructions, and so forth. As such, the input device 118 may include various input devices, such as any type of mouse, keyboard, microphone, camera, and so forth.

[0052] In some embodiments, the system 100 may also include an electronic health record(EHR) system 120, which may contain a number N of patient records 122. The EHR system 120 may be operated using various systems, devices, computers, servers, databases, and other 4912-3076-7444.4 065472-000953WOPT10 hardware. A patient record 122 can generally be any suitable type of electronic health record. As illustrated in FIG. 1, a patient record 122 may include various static data and / or the sequential data. For example, when the individual is admitted, various information about them (e.g., medical history, current measurements of various parameters, etc.) can be obtained and stored in a patient record 122 as static data. Then, as information and / or measurements are generated during the patient’s stay in the hospital (e.g., amounts and / or types of medication administered to the individual, blood glucose and / or other measurements, carbohydrate amounts consumed by the individual, laboratory results, etc.), this information and / or measurements can be stored in the patient record 122 as sequential data. In some implementations, a patient record 122 of a patient may include or receive (e.g., from a processing device 110) theoretical and / or altered sequential data that is associated with a proposed intervention for the patient. A processing device 110 (and / or other processing devices or processors) can access the static data and / or the sequential data stored in one or more patient records 122 for use, in accordance with methods described here.

[0053] As illustrated in FIG. 1, the EHR system 120 may be connected or connectable to aprocessing device 110. As such, the processing device 110 (and / or other processing devices or processors) may store data in and / or access data from the EHR system 122. For instance, in some implementations, the processing device 110 may utilize data (e.g., static data, sequential data, and so on) from a number of patient records 122 to train a machine learning model, according to aspects of the present disclosure. In some implementations, the processing device 110 may store prediction data, for instance, indicative of a predicted clinical condition or event, such that a clinician with access to the EHR system may be informed of such predicted condition or event.

[0054] In some embodiments, a processing device 110 may be connected or connectable,via wired and / or wireless communication, to a treatment apparatus 124, such as an apparatus configured to administer medication to a patient (e.g., an insulin pump, an infusion pump, etc). For instance, based upon a prediction of a clinical condition or event (e.g., a hypoglycemic condition or event, or risk thereof) determined in accordance to aspects of methods described herein, the processing device 110 and / or output device 116 may be configured to generate and transmit one or more control signals to the treatment apparatus 124. In some implementations, the control signal(s) may be generated based on a lookup table or reference, which may include various qualitative and / or quantitative information, data, values, and so forth. In one non- limiting example, a lookup table may include one or more indication of risk (e.g., one or more probabilities of developing a BG condition within a subsequent window of time), one or more 4912-3076-7444.4 065472-000953WOPT11 treatment parameters or settings, one or more demographic values or information (e.g., age, sex, weight, and so forth), and so on.

[0055] Responsive to the transmitted control signal(s), the treatment apparatus 124 mayinitiate, modify, pause, adapt, and / or cease administration of a treatment. For example, responsive to one or more transmitted control signal(s), a pump (e.g., insulin pump, infusion pump, etc) may initiate, pause, and / or cease administration of a medication (e.g., insulin, glucose, and so on). In some implementations, responsive to one or more transmitted control signal(s), one or more treatment parameters or settings on a pump, or other treatment apparatus 124, may be adjusted (e.g., basal rate, flow rate, insulin-to-carbohydrate ratio, bolus, time of administration, frequency of administration, and so forth).

[0056] While FIG. 1 shows the processing device 110 as being separate from the treatmentapparatus 124, in some embodiments, the processing device 110 and treatment apparatus 124 may be combined into a single treatment system or device. Also, while FIG. 1 shows an embodiment that includes an EHR system 120, in other embodiments, the system 100 need not include an EHR system 120, and may instead receive static data and / or the sequential data from another source. For instance, static and / or sequential data may be obtained as it is being recorded or acquired.

[0057] FIG. 2 shows a flowchart of a method 200 for predicting blood glucose or bloodglucose conditions in an individual, which may generally be a patient that has been admitted to a hospital or other healthcare facilities. The method 200 may be carried out using any suitable device, apparatus, or system, such as the system 100 described with reference to FIG. 1. In some implementations, steps of the method 200 may be implemented as instructions stored in non-transitory computer-readable media, as a program, firmware or software, and executed by various general-purpose, programmed or programmable computers, processors or other processing devices. In other embodiments, steps of the method 200 may be hardwired in an application-specific computer, server, processor, dedicated system, or module. Although the method 200 is illustrated and described as a sequence of steps, it is contemplated that the steps may be performed in any order or combination, need not include all illustrated steps, and may include additional steps.

[0058] Step 210 of the method 200 includes receiving, at a current time, static data andsequential data associated with one or more individuals. The current time may generally refer to a time at which blood glucose analysis is performed or a time immediately preceding a blood glucose altering treatment (e.g., an insulin dose administration). In general, static data received at step 210 may include a set of one or more static data points, and sequential data may include 4912-3076-7444.4 065472-000953WOPT12 one or more series of sequential data points (e.g., time-series of data points). In some implementations, a series of sequential data points may include at least two data points captured at separate times for each type or category of data being received (e.g., two blood glucose measurements, two amounts of carbohydrates consumed during a meal, etc.). As indicated by step 210, in some implementations, static and / or sequential data may be accessed from a data storage location, such as an EHR system as described with reference to FIG.1.

[0059] Generally, some or all static data may be generated and / or collected at an identicalprior point in time that occurred before the current time. For example, a prior point in time could be the time when the individual was admitted to the hospital. Some static data may include data points generated at the prior point in time (e.g., initial vital signs taken when the individual is admitted), while other static data may include data points collected at the prior point in time, but associated with some earlier time (e.g., the amount of some medication that the individual has taken within the past day or week).

[0060] The term “hospital” is used to refer to the facility where the individual has beenadmitted, but it is understood by those of skill in the art that the features herein are applicable to an individual admitted to any healthcare facility, even if that facility is not a hospital.

[0061] While static data is described herein as being generated and / or collected at a priorpoint in time, it may be understood that the term “prior point in time” need not refer to a single instantaneous moment, but may span a window of time. For example, if static data is generated when the individual is admitted, some or all static data may generally be generated and / or collected within a certain window of time (e.g., 10 minutes, 30 minutes, 1 hour, 2 hours, etc.), and inputted into a system (such as system 100) and / or an electronic health record (such as electronic health record 120). In general, however, some static data may include static data that is not of the same type as other static data. Thus, static data may generally not include a first measurement (e.g., a blood pressure measurement) generated when the individual first arrives at the hospital for admission, and a second measurement of the exact same type (e.g., another blood pressure measurement) taken 20 minutes later during the admission process.

[0062] Static data can include generally any suitable type of data. For example, static datacan include an age of the individual, a sex of the individual, a weight of the individual, a medical history of the individual, other types of measurements and / or information, or any combination thereof.

[0063] Sequential data are all collected and / or generated within a prior window of time thatoccurs before the current time. The prior window of time can also be referred to as the “lookback window.” For example, sequential data may include a current blood glucose level 4912-3076-7444.4 065472-000953WOPT13 of the individual, a current body temperature of the individual, a current pulse rate of the individual, a current respiration rate of the individual, a current blood pressure of the individual, a current blood oxygen saturation of the individual, an amount of carbohydrates consumed by the individual during a prior meal, other types of measurements and / or information, or any combination thereof.

[0064] In general, the prior window of time within which all sequential data is generatedand / or collected begins after the prior point in time when all static data is generated and / or collected. Thus, sequential data is generated and / or collected after the prior point in time when the static data are generated and / or collected (e.g., the time of admission). However, in some implementations, there may be some overlap. In some of these implementations, at least one of the sequential data points is generated and / or collected at the prior point in time when the static data point(s) are generated and / or collected (e.g., at the time of admission), such that the prior point in time (for static data) occurs during the prior window of time (for the sequential data). Moreover, there may be at least one data point that is part of both the set of static data and the set of sequential data.

[0065] The prior window of time may include any suitable length of time prior to thecurrent time when the analysis is to be performed. For example, the prior window of time can have a duration of about 6 hours, about 12 hours, about 18 hours, about 24 hours, about 36 hours, about 48 hours or more, between about 6 hours and about 12 hours, between about 6 hours and about 18 hours, between about 6 hours and about 24 hours, between about 6 hours and about 36 hours, between about 6 hours and about 48 hours or more, between about 12 hours and about 18 hours, between about 12 hours and about 24 hours, between about 12 hours and about 36 hours, between about 12 hours and about 48 hours or more, between about 18 hours and about 24 hours, between about 18 hours and about 36 hours, between about 18 hours and about 48 hours or more, between about 24 hours and about 36 hours, between about 24 hours and about 48 hours or more, between about 36 hours and about 48 hours or more, or any other suitable duration. The prior window of time may generally begin at a time that is equal to or after the prior point in time when the static data is generated and / or collected, and end at a time that is before or equal to the current time.

[0066] Similar to static data, sequential data can generally include any suitable type ofinformation and / or measurement, and can include any of the types of information referenced herein about static data. In some implementations, sequential data may include at least two sequential data or data points belonging to each of one or more categories or types of data, where the two sequential data or data points from each respective category were obtained at 4912-3076-7444.4 065472-000953WOPT14 separate times. This allows for analyzing and predicting the evolution of data or data points in each category. Example categories can include a first category associated with laboratory measurements (e.g., blood glucose measurements), a second category associated with food consumed by the individual or a diet of the individual (e.g., intake of carbohydrates, fats, proteins, liquids, and so forth, restriction(s) on food intake, administration by tube feeding, and so forth), and a third category associated with medication administered to the individual.

[0067] In some implementations, sequential data or data points associated with the firstcategory can include a first blood glucose measurement obtained at a first point in time with the prior window of time, and a second blood glucose measurement obtained at a second point in time within the prior window of time that is after the first point in time. The blood glucose measurement may be quantified using any suitable units, such as milligrams per deciliter (mg / dL).

[0068] In some implementations, sequential data or data points associated with the secondcategory can include a first amount of carbohydrates consumed by the individual at a first point in time with the prior window of time, and a second amount of carbohydrates consumed by the individual at a second point in time within the prior window of time that is after the first point in time. The amount of carbohydrates can be quantified using any suitable units, such as grams of carbohydrates or calories stemming from carbohydrates.

[0069] In some implementations, sequential data or data points associated with the thirdcategory can include a first amount of insulin administered to the individual at a first point in time with the prior window of time, and a second amount of insulin administered to the individual at a second point in time within the prior window of time that is after the first point in time. The sequential data or data points associated with the third category can alternatively include a first type of insulin administered to the individual at a first point in time with the prior window of time, and a second type of insulin administered to the individual at a second point in time within the prior window of time that is after the first point in time. In some implementations, the sequential data or data points associated with the third category can alternatively include a first type of insulin and an amount of the first type of insulin administered to the individual at a first point in time with the prior window of time, and a second type of insulin and an amount of the second type of insulin administered to the individual at a second point in time within the prior window of time that is after the first point in time.

[0070] The amounts of the insulin may be quantified using any suitable units, such asmilliliters (mL). The type of insulin could refer to the specific formulation of the insulin (e.g., 4912-3076-7444.4 065472-000953WOPT15 fast-acting insulin vs. intermediate-acting insulin vs. long-acting insulin) but could additionally or alternatively refer to the method of administration of the insulin (e.g., subcutaneous, intravenous, sliding scale, and so forth).

[0071] In general, sequential data can include any combination of data or data points fromany suitable category. In some implementations, the sequential data includes at least two sequential data points from each category generated and / or collected at two different points in time within the prior window of time. However, in some cases, data may be recorded only once in a lookback window, or not at all. For example, some individuals may have several laboratory measurements and medication recorded during the lookback window. Others may have none recorded. Yet an absence of medication or laboratory measurements may still be informative because, for instance, an individual that is not given medication may end up not having a hypoglycemic event. Further, even if data is recorded only once, it may still be informative in relation to other data or data points.

[0072] Referring again to FIG.2, step 220 of the method 200 includes inputting static dataor data points and sequential data or data points into a model. In some implementations, the model is a machine learning model, an LSTM model, a model that includes one or more LSTM blocks, a transformer model, a logistic regression model, a statistical model, or any other suitable model. Also, in some implementations, multiple models may be used. In other implementations, features from multiple different models may be additionally or alternatively be used.

[0073] In some implementations, the model used at step 220 includes separate input blocksto receive static data or data points and sequential data or data points. For instance, in one example, a first input block can receive various static data points, and may include one or more dense layers that operate on the static data points. A second input block can receive various sequential data points, and can include one or more cells (e.g., one or more LSTM cells) that can operate on the sequential data points. In this example, the model may include at least one intermediate dense layer connected to the first input block and the second input block, at least one batch normalization layer connected to the at least one intermediate dense layer, and a final layer connected to the at least one batch normalization layer. The final layer may include a sigmoid function that outputs the indication of the risk that the individual develops hypoglycemia within a subsequent window of time.

[0074] As illustrated, step 230 of the method 200 includes receiving, from the model, anoutput or indication of the risk of the individual developing a blood glucose condition, such as developing hypoglycemia, hyperglycemia, and so forth, within a subsequent window of time. 4912-3076-7444.4 065472-000953WOPT16 The output or indication may have any form, and may include any information and / or data. In some implementations, the providing an output indicative of the occurrence of a future hypoglycemia condition or event in the individual (e.g., a blood glucose measurement that satisfies a predetermined threshold blood glucose level indicative of hypoglycemia).

[0075] As used herein, “developing a blood glucose condition” generally refers toascertaining, obtaining, and / or measuring a blood glucose associated with a blood glucose condition, and / or with a state of being in the blood glucose condition. Also, “developing hypoglycemia” generally refers to ascertaining, obtaining, or measuring a blood glucose less than or equal to some threshold blood glucose level, and / or refers to a state of being hypoglycemic (e.g., having a blood glucose level that is less than or equal to some threshold blood glucose level). Further, “developing hyperglycemia” generally refers to ascertaining, obtaining, or measuring a blood glucose greater than or equal to some threshold blood glucose level, and / or refers to a state of being hyperglycemic (e.g., having a blood glucose level that is greater than or equal to some threshold blood glucose level).

[0076] Herein, subsequent window of time can be referred to as a “prediction horizon.”The subsequent window of time can have any suitable duration. For example, a prior window of time can have a duration of about 6 hours, about 12 hours, about 18 hours, about 24 hours, about 36 hours, about 48 hours or more, between about 6 hours and about 12 hours, between about 6 hours and about 18 hours, between about 6 hours and about 24 hours, between about 6 hours and about 36 hours, between about 6 hours and about 48 hours, between about 12 hours and about 18 hours, between about 12 hours and about 24 hours, between about 12 hours and about 36 hours, between about 12 hours and about 48 hours, between about 18 hours and about 24 hours, between about 18 hours and about 36 hours, between about 18 hours and about 48 hours, between about 24 hours and about 36 hours, between about 24 hours and about 48 hours, between about 36 hours and about 48 hours, or any other suitable duration.

[0077] In some implementations, a prior window of time (during which the sequential datapoints are generated and / or collected) has an identical duration as the subsequent window of time. In other implementations, a prior window of time has a longer duration than the subsequent window of time. In yet other implementations, a subsequent window of time has a longer duration than a prior window of time.

[0078] In some implementations, a subsequent window of time begins at a current time. Inother implementations, a subsequent window of time begins at some point after a current time. In general, after static data and sequential data have been input into a model, the subsequent 4912-3076-7444.4 065472-000953WOPT17 window of time may begin once an indication of risk is received from the model, or after the indication of risk is received from the model.

[0079] An indication of risk of an individual may be quantified in any suitable manner,and specified using various forms and formats. For example, in some implementations, an indication of risk or output of a model may include a probability value, a classification (e.g., that a probability that the individual’s blood glucose level satisfies a predetermined threshold blood glucose level at some point in the subsequent window of time), and so forth. Herein, “satisfying” the threshold blood glucose level could refer to the individual’s blood glucose level being less than the predetermined threshold blood glucose level, less than or equal to the predetermined threshold blood glucose level, greater than the predetermined threshold blood glucose level, or greater than or equal to the predetermined threshold blood glucose level. For example, the predetermined threshold blood glucose level is about 70 mg / dL. Hence, in some implementations, the output from a model can be a probability that the individual will have a blood glucose level that is less than 70 mg / dL at some point within the subsequent window of time, or less than or equal to 70 mg / dL at some point within the subsequent window of time.

[0080] In some implementations, the model may output a plurality of probabilities, forinstance, where each probability represents a likelihood that the individual’s blood glucose level will fall within a predetermined range of blood glucose levels at a point in the subsequent window of time. In some applications, each predetermined range of blood glucose levels may be associated with a blood glucose condition, such as hyperglycemia, hypoglycemia, and so forth, as well as other blood glucose conditions.

[0081] In some implementations, the method 200 further includes optional steps 240 and / or250. In step 240, theoretical and / or altered sequential data that is associated with a proposed intervention is generated and provided as input into the model. Such theoretical and / or altered sequential data may be used to test the efficacy of a variety of different medical interventions that could be undertaken by a healthcare provider to mitigate or avoid an undesired blood glucose level in the future.

[0082] Theoretical sequential data is additional sequential data obtained some time afterthe last sequential data or data point in the original sequential data that was input into the model at step 220. In this implementation, the entirety of the sequential data that has been input into the model at step 240 is associated with a new window of time that ends after the original prior window of time from which original sequential data was generated and / or collected. The new window of time may include the entirety of the original prior window of time and an additional amount of time. Alternatively, the new window of time may have the same duration as the 4912-3076-7444.4 065472-000953WOPT18 original prior window of time and may begin after the original prior window of time began. Altered sequential data is sequential data that is varied or modified from the initial or original sequential data that is input into the model. In this implementation, the entirety of the sequential data that has been input into the model at step 240 is associated with the same prior window of time as the original sequential data. In some implementations, a model may be configured to receive medication data and laboratory data as input.

[0083] Theoretical data and / or altered sequential data need not actually bemeasured / received, but instead may be what would be measured / received (and / or be an approximation of what would be measured / received) if a certain medical intervention were to be performed.

[0084] At step 250, an updated indication of the risk of the individual developing a bloodglucose condition, such as hypoglycemia, is received from the model. In general, the updated indication of the risk can be formed and / or formatted the same as the indication of risk received at step 230. However, the updated indication of the risk is indicative of the risk of the individual developing hypoglycemia, or other blood glucose condition, if the proposed intervention associated with the theoretical sequential data and / or the altered sequential data were to be performed. In some implementations, the subsequent window of time to which the updated indication of the risk applies is different from the subsequent window of time to which the original indication of risk applies. For example, if theoretical sequential data input into the model at step 240 is associated with a window of time that extends past the original prior window of time, then the subsequent window of time to which the updated indication of risk applies may be shifted forwarded in time. In other implementations, the subsequent window of time to which the updated indication of the risk applies is the same as the subsequent window of time to which the original indication of risk applies.

[0085] Thus, the model, in addition to being able to predict the risk of developing a bloodglucose condition, such as hypoglycemia, as described, can also be used to test out various medical interventions that can reduce that risk. For instance, different interventions can be tested by inputting different theoretical and / or altered sequential data into the model at step 240. In some implementations, the performance of steps 240 and 250 may be triggered by the original, or first indication of risk satisfying some threshold. For example, the original or first indication of risk indicating that the individual has a percentage chance or likelihood of having a hypoglycemic blood glucose level in the future that is greater than a threshold percentage chance can trigger steps 240 and 250 to test different interventions. In another example, the original or first indication of risk predicting that the blood glucose level of the individual in the 4912-3076-7444.4 065472-000953WOPT19 future may be less than some threshold blood glucose level can trigger steps 240 and 250 to test different interventions. In yet another example, the original or first indication of risk predicting that the blood glucose level of the individual in the future to be within a predetermined range of blood glucose levels can trigger steps 240 and 250 to test different interventions.

[0086] In some implementations, a report may be generated and provided in method 200.The report may be in any form, and include any data and information. For instance, the report may provide a warning and / or a recommendation for controlling blood glucose levels in accordance with risk. For example, the recommendation may include a recommended intervention, a recommended nutrition, a recommended drug dose, or drug dose adjustment, which may be determined, for example, based on the indication of risk, one or more output probability, and so forth, in accordance with aspects of the present disclosure. In some aspects, the report may include a predicted effectiveness of different interventions tested by inputting different theoretical and / or altered sequential, as described.

[0087] The steps of the method 200 may be carried out in any order, any number of times,such as intermittently, periodically, and / or subject to user input. In some implementations, steps of the method 200 may be carried out in a rolling prediction framework, whereby a number predictions (e.g., indications of risk, predicted blood glucose values, and so forth) may be carried out using a sliding window that shifts the prior window of time (e.g., lookback window) by a predetermined amount of time. As such, new or updated sequential data and / or static data associated with each updated prior window of time may be used to produce one or more updated predictions over one or more updated prediction horizons. For instance, in some implementations, steps of the method 200 may be performed on hourly, daily, weekly, monthly, and so forth.

[0088] Disclosed herein is an example study associated with the method 200 illustrated inFIG.2. In particular, a preliminary LSTM model was trained to make predictions on a cohort of 89,857 encounters and achieved a precision of 0.58 and recall of 0.76. To improve recall without sacrificing precision (i.e., such that the model correctly detects all hypoglycemia cases, but does not cause alert fatigue), focus was placed on optimizing precision and recall, and identifying top predictors of hypoglycemia according to the optimized LSTM model. It is recognized herein that an LSTM model with two parallel input blocks that combine static and sequential data allows for: 1) a model with clinical utility and 2) transparent predictions. One objective herein is to improve monitoring and care for patients at high risk of hypoglycemia, as well as other blood glucose conditions. 4912-3076-7444.4 065472-000953WOPT20

[0089] Some approaches have used machine learning (ML) and large electronic healthrecord (EHR) datasets to predict inpatient hypoglycemia as a binary output. While such work reported some improved performance over logistic regression (LR) models, no prior approach has used neural networks to predict inpatient hypoglycemia without continuous glucose monitoring (CGM) data. By contrast, it is recognized herein that neural networks have the potential to surpass ML models when given large amounts of data.

[0090] One study of about 33,000 admissions was designed to detect, not predict, ahypoglycemic event during an admission. Since the ML model utilized in that study was trained on data from an entire hospital admission, the model could not predict a future hypoglycemic event at any timepoint within the hospital stay. Also, despite high performance (precision of 0.88 and recall of 0.70), that study used data from an entire hospital stay, such as the average total daily insulin dose over the admission rather than the individual insulin doses leading up to the hypoglycemic event. Since a clinician would not have information downstream from a hypoglycemic event, the model was limited in applicability to a real-world setting. Additionally, averaging variables like insulin doses would not take advantage of the inter- dependence between individual doses that ultimately led to the event. That study paved the way for subsequent work aiming for predicting rather than detecting hypoglycemia.

[0091] Some studies addressed this issue by using a lookback window, a timeframe priorto a specified timepoint over which patient inputs to the model may be derived. However, such studies also used traditional ML methods that collapsed sequential data into static inputs, which limited precision and recall.

[0092] Another approach constructed a hypoglycemia prediction model that took invariables from a lookback window, either in the last 24 hours prior to an index BG or from admission to the index BG. The approach used information from the lookback window to predict whether a hypoglycemic event would occur at a certain prediction horizon. The study included about 184,000 admissions. The highest performing ML model reported a precision of 0.12 and recall of 0.73 for hypoglycemia prediction. Within the lookback window, sequential variables were separated into static inputs. For example, BG was separated into average BG, BG immediately before index BG, and the minimum BG. Additionally, the study imputed missing labs and vitals with the mean value for the entire admission. Yet imputation of missing BG values may misrepresent a clinical course of a patient. While the study was successful in applying a lookback window, it left room for improvement in terms of model precision.

[0093] Yet another approach used a lookback window from admission to an index BG anda prediction horizon of 24 hours after each index BG was used. That study included data from 4912-3076-7444.4 065472-000953WOPT21 about 55,000 admissions. The highest performing ML model reported a precision of 0.09 and a recall of 0.82. Sequential data was represented as static descriptive statistics, for example, nadir BG and peak BG since admission were used as model inputs rather than the sequence of BGs themselves. Although recall was higher compared to previous approaches, the model was still limited by low precision.

[0094] Yet another approach reported a precision of 0.72 and recall of 0.59 for predictionof inpatient hypoglycemia from about 64,000 admissions. The lookback window in that study included the time between admission and when the last data point was measured before the hypoglycemic event. The prediction horizon was the time between this last data point and the hypoglycemic event. Like other approaches, the study collapsed sequential data into static variables. For example, mean, standard deviation, inter-quartile range, and total range of laboratory data were used as inputs, rather than the data values themselves. Yet, a significant limitation of such previous ML study is that model generalization was not tested on future data, which potentially allows data leak that can produce elevated and unrealistic model performance.

[0095] Neural networks trained on outpatient CGM data collected from patients with type1 diabetes may be helpful. LSTMs outperform standard ML using outpatient CGM data from individuals with type 1 diabetes (T1D). Using CGM data from T1D to predict hypoglycemia, a LSTM outperformed a standard statistical timeseries model and logistic regression at a one hour prediction horizon. Another study found that support vector machines under-performed compared to a LSTM to predict hypoglycemia in T1D using inputs of meals, insulin and CGM. Lastly, a study used a LSTM model and about 27,000 days of matched CGM and insulin doses from 124 individuals with T1D. The dataset consisted of about 18,000 hypoglycemic events. The LSTM model, which was trained on series of previous BG values to predict the next BG, had an accuracy of 0.91 in predicting an impending hypoglycemic event. However, there does not appear to be any study that uses a neural network to predict inpatient hypoglycemia using non-CGM data. Patients rarely have CGM data when admitted to the hospital, so a model designed and trained on this level of information would likely not perform well using point-of- care testing (POCT) BG measurements. Using an LSTM model with time series and static inputs to predict whether an inpatient hypoglycemic event occurred may be beneficial. It is recognized herein that principles of model architecture optimization established in these studies may be used to predict hypoglycemia in hospitalized individuals with T1D, type 2 diabetes (T2D) or no diabetes. 4912-3076-7444.4 065472-000953WOPT22

[0096] Neural networks do not intrinsically indicate which input makes an accurateprediction. There is also a need to understand why a model makes a certain prediction to increase adoption by clinicians and provide actionable insights. For example, if hundreds of time series representing labs, medications, and diet orders that occurred during a patient’s stay in the hospital are input into the model, which variables were most predictive of the hypoglycemia? To answer this, a tool that “interprets” the model is needed.

[0097] Variable importance is sometimes used for model interpretation in ML inpatienthypoglycemia prediction. Variable importance plots may be used present top predictors. Typically designed for tree-based models, this approach calculates mean decrease in impurity after splitting by a variable (also known as Gini Impurity), a measure of misclassification lost when a predictor is used by the model. Variable importance, however, is limited to tree-based models and cannot rationalize predictions on an individual patient level. Shapley additive explanations (SHAP) is a method to explain model predictions specific to an individual or globally. Interpretability is an important aspect of the model to establish before developing a clinician-trusted tool that would be used in routine care.

[0098] In accordance with aspects of the present disclosure, one approach is to optimize anLSTM model to predict inpatient hypoglycemic events. In some implementations, EHR data may be used. One criteria of data inclusion may be any adult inpatient encounter in which at least one anti-hyperglycemic medication is administered.

[0099] By way of example, the following datasets have been extracted for each patient:past medical history, admission and discharge notes, demographic information, vital signs, social history, laboratory, inpatient medications, prior to admission medications, diet orders, and procedures. A baseline logistic regression (LR) model may be trained, against which an LSTM model may be trained.

[0100] An example dataset includes 56,957 unique adult inpatients with 103,860 uniqueencounters between May 14, 2014 and September 30, 2022, as shown in FIG. 3. The data indicates an average of ~2 encounters per patient. A total of 3,126,621 POCT BG values were recorded among all encounters. Mean length of stay among all encounters was 8.5 days, meaning an average of ~30 BG values per encounter. Approximately 20% (20,943 / 103,860) of unique encounters had at least one BG<70 mg / dL during their hospital stay and the median time to BG<70 mg / dL was 6.3 days.

[0101] Distribution of the lowest BG recorded per encounter is shown in FIG. 4A. Thelowest BG recorded per encounter was less than 100 mg / dL in ~60% (61,259 / 103,860) of encounters. FIG.4B shows overlayed density plots of days until a BG <70, <60 or <50 mg / dL 4912-3076-7444.4 065472-000953WOPT23 across all encounters. Most encounters had hypoglycemic events within the first few days of admission.

[0102] A preliminary LSTM model was trained using thirty time windows of medications,labs, diet and statics over 5 days to predict whether the admission would experience hypoglycemia. A total of 29 meds, 11 labs and 4 diets were tracked as sequential data and 39 as static data. Class weights were incorporated to address imbalanced classes. Performance of the model on the holdout test set was assessed to evaluate for overfitting. Neural networks were fit incrementally over an “epoch,” which represented one full pass through the data. FIGs.5A and 5B depict model area under the precision recall curve (AUPR) and loss, respectively, with each epoch. Model loss quantifies the difference between true and predicted values. FIGs.5A and 5B each illustrates two curves, one for performance on the training set and one for the held- out validation set. Rising AUPR and decreasing loss indicate learning. Overfitting can be seen if accuracy continues to rise in the train set but not in the validation set. As appreciated from FIGs. 5A and 5B, the data indicates that the model can learn quickly from limited data. The data also indicates that LSTM models can be trained, as described. The model was tested on a third set of held aside data and achieved precision of 0.42 and recall of 0.76.

[0103] A synthetic minority over-sampling technique (SMOTE) was used to mitigate classimbalance, which achieved a precision of 0.58 and recall of 0.76 on the same test set. Assuming 20% prevalence, R2Nagelkerke of 0.15, shrinkage factor of 0.9, the number of parameters should not exceed 1,000 for a sample size of 56,967 inpatients to minimize overfitting and ensure precise estimation. In some aspects, statistics between hypoglycemic vs control encounters may be obtained. For instance, encounters with at least one BG<70 mg / dL may be compared to encounters without BG<70 mg / dL events recorded during a hospital stay. Distribution of continuous variables may be checked for normality using the Shapiro-Wilk test to determine the appropriate test for comparison of patient characteristics between groups. A t-test (or ANOVA for 3+ groups) may be used if a variable is normally distributed; otherwise, a Wilcoxon rank sum test (or Kruskal-Wallis test for 3+ groups) may be used. A chi square test may be used for comparison of categorical variables between groups. Sex may be an input to the model. Counts and percentages of males and females in the hypoglycemia and control groups may be reported. Sex differences in glucose homeostasis have been demonstrated. It is recognized herein that if there is an impact of sex on the risk of hypoglycemia, a model can learn this.

[0104] In some aspects, data may be prepared for model training. Since errors, outliers,and missing data may be present with EHR data, and data cleaning may be performed. Outlier 4912-3076-7444.4 065472-000953WOPT24 data points may be visualized using Cook’s distances (distance should be close to 0 when no outliers), and entries that are erroneous and outside reasonable bounds may be removed. Categorical variables may be mapped to integers. An embedding layer may be used to map categorical variables into vectors, which may then be used as inputs to an LSTM model. For an LR model, categorical variables may be one-hot encoded. Continuous variables may be scaled to be in a range of 0 to 1 (MinMax normalization) so that the model is not biased toward the variables with higher magnitude values.

[0105] In some aspects, a neural network and prediction task may be constructed. Forinstance, static and sequential data are both likely informative, and an LSTM model may take both types of inputs simultaneously. Previous work using ML for diabetes complication prediction found that inputs of all available datasets combined had higher performance than each dataset individually. An LSTM model may be designed with two parallel input blocks that allow combining the static and sequential data. One dense layer may accept static data, and LSTM blocks may accept sequential data. Two layers may be combined with further dense layers with batch normalization, before a final layer may use a sigmoid activation function to predict a probability of hyperglycemia over various prediction horizons. A model, as described, may be implemented in keras in python with a tensorflow backend, for example.

[0106] In some aspects, model optimization may be performed. For instance, an LSTMmodel input may include data points occurring during a lookback window to determine whether a hypoglycemic event occurred during a prediction horizon (FIG.6). For example, a lookback window of 6 hours and a prediction horizon of 24 hours indicates that the model can predict whether a hypoglycemic event will occur within 24 hours using (i) patient data recorded within the last 6 hours and (ii) any static data recorded at admission. A prediction horizon may be selected to provide enough time for clinician assessment and intervention, and capture effects of variables on BG. Model performance using combinations of lookback windows (6, 12, 18, 24, 36, 48, or more hours) and prediction horizons (6, 12, 18, 24, 36, 48, or more hours) may be compared.

[0107] To assess model performance, data may be split into 10% test and 90% trainingdata. Transformations of the training set may be done separately from the test set to prevent data leak into training data. K-fold cross validation with five folds may be used to test various combinations of hyperparameters using the tree-structured parzen estimators (TPE) or random search methods implemented in hyperopt. Hyperparameter tuning may be used to optimize on F1 score. AUPR score or other performance metrics for the minority class, such as precision, recall, and F195% confidence interval (CI) may be calculated using bootstrapped resampling. 4912-3076-7444.4 065472-000953WOPT25 The ideal hyperparameters may be used to train a final model on the 90% training set before assessing generalization on the 10% test. The AUPR may be used as the primary performance metric for an LSTM model, but other model fit statistics may be analyzed, including precision, recall, and F1.

[0108] In some aspects, a baseline LR model may be developed. For instance, static modelinputs may be the same as those used in the LSTM model, but time series data may be converted to descriptive statistics as done previously in ML hypoglycemia studies. For example, labs may be input to the model as minimum, maximum and average values in the lookback window. The model output may be hypoglycemia (BG<70 mg / dL) vs no hypoglycemia. Variance inflation factor (VIF) may be used to identify variables affected by multicollinearity. Variables with a VIF >10 are strongly correlated and may be removed. The same model may be fit but using the reduced set of variables and optimize hyperparameters using a random search on AUPR. Variable coefficients, standard errors, p values and model performance metrics, F1, precision, recall, accuracy and AUPR, may be reported. AUPR 95% CI may be computed using bootstrapped resampling.

[0109] The present approach may provide optimized LSTM and LR models to predictinpatient hypoglycemia using EHR data. The combination of lookback window, prediction horizon, and hyperparameters leading to the highest AUPR may be selected for an LSTM model. The LR model using the same lookback window and prediction horizon as the LSTM model may be assessed for multicollinearity, then optimized on AUPR. AUPR score 95% Cis for the two models may be compared. If the Cis overlap, models are not significantly different.

[0110] Some factors that affect BG dynamics, such as carbohydrate intake and tube feedingstart / stop times, are sometimes inconsistently recorded in the EHR. However, previously published hypoglycemia prediction models found the contribution of carbohydrate intake, insulin and exercise to be lower than that of CGM data. The data used also contained timestamps for diet orders, which may help elucidate when caloric intake was restricted. In addition, if baseline patient characteristics are different between groups, differences may be mitigated to prevent models learning only from overrepresented demographic categories. Potential solutions may include removing discrimination towards some groups in the original dataset or adding a penalization term in the objective function. Further, class imbalance (hypoglycemia group compromises 20% of encounters) may be addressed by using performance metrics less likely to be biased by a majority class (e.g., F1 score), as well as using SMOTE or class weights in the model. 4912-3076-7444.4 065472-000953WOPT26

[0111] Model interpretation can be important for bolstering clinician trust. Specifically, ifclinicians understand and agree with predictors of a model, the model is more likely to be used in practice. In some implementations, SHAP may be used on a per-individual level to identify top predictors and magnitudes that contribute to decisions made by an LSTM, as described. A package designed to approximate SHAP values for deep learning called DeepExplainer or GradientExplainer may be used. Additionally, DeepExplainer may return important predictors based on a time before the hypoglycemic event, so any changes in predictors between various timepoints before an event may be assessed. Force plots, which show “forces” that balance each other to result in the actual model prediction, may be used to explain predicted hypoglycemia probabilities on a per-individual level. For instance, SHAP force plots may be used to explain predictions on a per-individual level from an optimized LR model, as described.

[0112] In some aspects, predictors between different models may be compared. Forinstance, to determine whether an LSTM model, in accordance with the present disclosure, is effective in capturing complexity of patient factors driving hypoglycemia, and thus clinically advantageous, a degree of similarity between predictors deemed important by different models (e.g., LSTM and LR) may be assessed. For example, top ten, or more, or less, predictors from each model’s per-individual interpretation may be compared to determine common predictors, predictors considered important by LSTM but not LR, predictors considered important by LR but not LSTM, or a combination thereof. Unique predictors and respective counts in each of these three categories may be graphically displayed.

[0113] The present disclosure recognizes that more unique factors may be consideredimportant for prediction of hypoglycemia using an LSTM model compared to an LR model. Once implemented, the model may provide important predictors to clinicians for alerting a patient with high risk of hypoglycemia. Knowledge of modifiable risk factors, such as insulin dose and absence of diet orders, would motivate intervention by the clinician. Non-modifiable risk factors would still be useful as they have the potential to motivate institutional change. For example, if gender and socioeconomic status were found to play an integral role in hypoglycemia risk across many patients, changes in hospital insulin dosing guidelines that better tailor therapy to these more sensitive groups may be warranted.

[0114] Although neural networks can be more robust to multicollinearity compared tolinear models, deep learning (e.g., using DeepExplainer) may be performed multiple times (e.g., 100 times) using different random seeds, and Shapley values may be averaged over the 4912-3076-7444.4 065472-000953WOPT27 trials. Such approach may be used to filter out correlated variables. Uncorrelated variables may have a high Shapley value across trials and remain among top predictors.

[0115] In accordance with aspects of the present disclosure, a flowchart setting forth stepsof a method 700 for predicting blood glucose or blood glucose condition in an individual, is illustrated in FIG.7. For instance, as described, steps of the method 700 may be used to predict the occurrence of future hypoglycemia in the individual (e.g., a blood glucose measurement that satisfies a predetermined threshold blood glucose level indicative of hypoglycemia). Steps of the method 700 may be carried out using any suitable systems, devices, and so forth, such as the system described with reference to FIG.1. In some implementations, steps of the method 700 may be implemented as instructions stored in non-transitory computer-readable media, as a program, firmware, or software, and executed by various general-purpose, programmed or programmable computers, processors or other processing devices. In other embodiments, steps of the method 700 may be hardwired in an application-specific computer, server, processor, dedicated system, or module. Although the method 700 is illustrated and described as a sequence of steps, it is contemplated that the steps may be performed in any order or combination, need not include all illustrated steps, and may include additional steps.

[0116] The method 700 may begin at step 710 with receiving static data and sequential dataassociated with an individual. As described, static data may include data or data points generated and / or collected at the same time or time point before a current time (e.g., time when the individual was admitted to the hospital), while sequential data may include data or data points generated and / or collected within some prior window of time occurring before the current time (e.g., lookback window). As described, in some of implementations, at least one of the sequential data or data points may be generated and / or collected at a prior point in time when the static data or data points are generated and / or collected (e.g., at the time of admission), such that the prior point in time (for static data or data points) occurs during the prior window of time (for the sequential data points). Moreover, there may be at least one data point that is part of both the set of static data points and the set of sequential data points.

[0117] The static data and sequential data may then be inputted into a model. As described,the model may include a variety of architectures, inputs, and outputs. In some implementations, as indicated by step 720, the model may be a machine learning model that is trained in accordance with aspects of the present disclosure. By way of example, the machine learning model can be an LSTM model, a model that includes one or more LSTM blocks, a transformer model, a logistic regression model, or any other suitable model. Moreover, in some implementations, multiple models may be used, and / or features from multiple different models 4912-3076-7444.4 065472-000953WOPT28 may be used. In some implementations, the model may be configured to receive medication data and laboratory data as input.

[0118] An indication of the risk of the individual developing a blood glucose condition(e.g., hypoglycemia, hyperglycemia, and so forth) within a subsequent window of time (e.g., prediction horizon) may then be received at step 730 from the model. As described, the indication may have any form and format, and may include a variety of information and / or data. For instance, the indication of risk outputted by the model may be in the form of one or more probabilities, one or more classifications, and so forth. For instance, in some implementations, an outputted probability may be indicative of whether the individual’s blood glucose level satisfies a predetermined threshold. In one example, an outputted probability may indicate whether blood glucose exceeds the predetermined threshold. In another example, an outputted probability may indicate whether blood glucose is equal to the predetermined threshold. In yet another example, an outputted probability may indicate whether blood glucose is less than the predetermined threshold. As described, one or more outputted probabilities may be associated with a subsequent window of time. In some implementations, the model may be additionally, or alternatively, based on inputted static data and / or sequential data, output one or more prediction data, such as one or more predicted values for blood glucose within the subsequent window of time.

[0119] In some aspects, an indication of risk, one or more probabilities, as well as otherdata and information may be generated and provided in a report. In some aspects, the report may indicate a recommendation for controlling blood glucose levels in accordance with risk. To this end, the recommendation may include a recommended dose, or dose adjustment, determined, for example, based on the indication of risk, one or more output probability, and so forth.

[0120] Referring again to FIG. 7, in some implementations, theoretical sequential dataassociated with a proposed intervention may be inputted into the model, as indicated by step 740. As described, theoretical sequential data may include additional sequential data obtained or generated after the last sequential data or data point in the original sequential data that was input into the model at step 720. In some implementations, altered sequential data associated with the proposed intervention may be alternatively or additionally inputted into the model at step 740. As described, altered sequential data may include sequential data that is modified from, or different with respect to, the initial or original sequential data that is input into the machine learning model at step 720. The output generated by the machine learning model may then be updated, by using theoretical and / or altered sequential data, thereby providing an 4912-3076-7444.4 065472-000953WOPT29 updated or second indication of risk. Such updated or second indication of risk may then be provided or reported using systems and devices, in accordance with the present disclosure. In some implementations, based on the first indication or risk, second indication of risk, or both, a treatment of the individual may be initiated, adapted, modified, and / or ceased.

[0121] Steps of the method 700 may be carried out any number of times. For instance, insome implementations, steps of the method 700 may be carried out to perform a rolling prediction framework, whereby a number predictions (e.g., indications of risk, predicted blood glucose values, and so forth) may be carried out using a sliding window that shifts the prior window of time (e.g., lookback window) by a predetermined amount of time. As such, new or updated sequential data and / or static data associated with each updated prior window of time may be used to produce updated predictions over updated prediction horizons. In some implementations, steps of the method 700 may be performed on hourly, daily, weekly, monthly, and so forth.

[0122] As described, a multitude of factors can affect a hospitalized individual’s BG,making hyperglycemia and other BG conditions difficult to predict and manage. While there are some established medications that affect BG, such as antihyperglycemics, there are likely undiscovered associations between medications and BG variability. Therefore, among other advantages of the present approach, identifying medications impacting BG may offer important benefits for glycemic management and patient safety.

[0123] As detailed below, in another implementation of the present disclosure, EHR datafrom about 103,839 inpatient encounters collected over 8 years within a large, urban health system was used to extract over 500 medication administrations, laboratory measurements, and clinical predictors of BG. Associations were detected using regression models, in accordance with the present approach, and evaluated for novelty against a comprehensive adverse drug event database. In one example, a BG prediction model was developed and applied to predict 4-hr point of care BG, producing an R2 value 0.35 and a mean squared error of 0.20 between true and predicted BG using a test set. Among results, 39 statistically significant factors were found to affect BG. Also, 28 medications, including 12 medications not previously documented, were found to affect BG. Other factors associated with BG included labs, social history, prior diagnosis of type 2 diabetes, and demographics.

[0124] Mechanistic studies provide evidence in support of unexpected medications’ effecton BG, such as increase in insulin secretion (sodium bicarbonate) and inhibition of insulin binding (heparin). Further investigation of these potential candidates may enhance clinical utility of these findings. The present study uniquely identified medication predictors of glucose 4912-3076-7444.4 065472-000953WOPT30 homeostasis using a large EHR dataset combined with a BG prediction model. Inclusion of more confounding variables, larger datasets, and prospective validation, may further improve results.

[0125] Individuals with diabetes are at a significantly increased risk of hospital admission,with rates ranging from two to six times higher than those without diabetes. Uncontrolled BG in hospital is associated with increased morbidity and mortality. Inpatient hypoglycemia, or BG<70mg / dL that occurs during hospitalization, has been reported in 20% of hospitalized insulin-treated patients, and is the most common adverse event associated with inpatient treatment of diabetes. While hypoglycemia can cause troublesome acute symptoms, such as confusion, impaired vision, and seizures, it can also increase the risk of chronic complications, length of stay and all-cause mortality. Inpatient hyperglycemia, defined as a BG>140mg / dL in hospitalized patients, can cause acute (e.g., diabetic ketoacidosis) and long term (e.g. retinopathy, nephropathy, cardiovascular disease) complications, which are both associated with increased risk of mortality.

[0126] The dangers associated with dysglycemia highlight the importance ofunderstanding mechanisms involved in glucose homeostasis. Numerous pharmacologic agents are understood to cause hyperglycemia, including steroids and beta-blockers, through a variety of mechanisms such as impaired insulin secretion or direct effects on beta cell proliferation. However, many mechanisms underlying glucose alterations due to drugs are not well understood. Management of BG in hospitals is further complicated by the dynamic nature of medication dosing and the unpredictable responses of patients. Additionally, confounding may be present between patient risk factors and blood glucose. For example, a drug administered to an individual with diabetes may be reported to cause hyperglycemia, but it is unclear whether such correlation is due to the disease or drug.

[0127] Some approved medications can have unexpected effects, and some positivedevelopments have been made in drug re-purposing research. For instance, bumetanide, which is a diuretic, has been recognized as a potential therapy for Alzheimer’s. In another approach, a computational pipeline engaged in a large type 2 diabetes mellitus (T2DM) genome-wide association study with outpatient electronic health records (EHR) to investigate potential drugs for T2DM repurposing and found that angiotensin-converting enzyme inhibitors and calcium channel blockers can lower glucose levels. Also, an open-source repository with drug targets and genes specific to diabetes screened around 1500 drugs on 20 proteins, and identified five drugs with repurposing potential to treat diabetes. Therefore, it is recognized herein that a data- 4912-3076-7444.4 065472-000953WOPT31 guided approach using EHR data may uncover drugs previously not established to cause BG alterations.

[0128] Therefore, in one approach of the present disclosure, a large, comprehensive EHRdataset was extracted from hospitalized patients who received at least one anti-diabetic medication, which is a risk factor for dysglycemia. The 500 most frequently prescribed inpatient medications were combined with relevant past medical history, demographics, social history, and laboratory measurements readily available in the EHR. A least absolute shrinkage and selection operator (Lasso) linear regression model was used to filter initial variables using highest model coefficients. Accounting for repeated encounters among patients, a mixed effects linear regression (MELR) model was fit and tested on selected features from Lasso. It is recognized herein that MELR modeling can accommodate inter-patient variability that a standard linear model may not effectively capture, and allow for interpretability through statistical significance testing lacking with ML models. Using EHR data from a controlled inpatient environment where medication administrations are being reliably recorded, this study utilized a computationally driven strategy to reveal associations between medications and point-of-care (POC) blood glucose (i.e., POC BG). As such, the present approach provides a first study to evaluate predictive performance of several hundred medications without prior filtering based on prior knowledge of medications known to affect BG.

[0129] As a data source, retrospective EHR data was extracted for patients admitted to thehospital between May 2014 and September 2022 within the Cedars-Sinai Health System, which includes a tertiary medical center, Cedars Sinai Medical Center (CSMC), and community hospitals, Marina Del Rey Hospital and Huntington Hospital. Inclusion criteria consisted of adult inpatients (i.e., at least 18 years old) who received at least one anti-hyperglycemic medication during their admission. Encounters to Labor & Delivery, Emergency Department, or with a length of stay (LOS) less than 24 hours were excluded. The following EHR datasets were extracted for each encounter: past medical history, demographics, social history, laboratory, and inpatient medications. Cedars-Sinai IRB approval was obtained for this study.

[0130] A data pre-processing approach was also performed. Time-dependent variables, orvariables that change during admission, included labs and medications, were pre-processed and analyzed. Static variables, or variables that do not change during inpatient stay, included past medical history, demographic data, and social history. In some implementations, missing lab values and medication doses were removed. Also, medications were filtered to include the most administered medications, such as 500 of the most administered out of the initial 5,267 medications. In some aspects, to ensure inclusion of common anti-hyperglycemic medications 4912-3076-7444.4 065472-000953WOPT32 in a model, subcutaneous (SQ) insulins (lispro, regular, glargine and NPH / Regular 70-30), intravenous (IV) regular insulin, and dosage forms of commonly prescribed oral anti- hyperglycemic medications (sitagliptin, glimepiride, glipizide, glipizide XR, glyburide and pioglitazone) were added to the top 500, if not already been included in the top 500 list. This selection resulted in 508 unique medications with 31,336,833 administrations recorded from 56,930 patients with 103,839 encounters.

[0131] In one implementation, eleven labs were selected of the available 4,094 labs,namely POC glucose, random glucose, hemoglobin A1c, C-reactive protein, creatinine, non- African American estimated glomerular filtration rate (eGFR), African American eGFR, AST, ALT, albumin, L-lactate. The resultant labs dataset had 7,556,267 lab values from 56,913 patients with 103,812 encounters across the eleven labs. A sum of all doses for each medication was calculated during the lookback window and were used as inputs to a model. Mean, minimum, maximum and last value of POC BG and mean of all remaining labs were calculated and used as inputs. EHR Dataset Number of Model Variables variables Past Medical 12 Liver failure (K72), CKD (N18), T1DM (E10), History (ICD-10 T2DM (E11), malignancy (C80.1), anemia code) (D64.9), CHF (I50), hypothyroidism (E03.0), hyperthyroidism (E05), pregnancy (Z33.1), Static hypoglycemia (E16.2), pain (R52) Demographic 14 Ethnic group, sex, race, smoking tobacco last use status, IDU last use status, alcohol last status, admission diagnoses: CHF, sepsis, GIB, nausea / vomiting, altered mental status, AKI, ESRD + / - dialysis, pain Social History 4 Age, BMI, hospital location (ICU vs non-ICU) Time- Laboratory 14 Mean POC BG, minimum POC BG, last POC dependent BG, maximum POC BG, mean hemoglobin A1c, mean C-reactive protein, mean creatinine, mean eGFR non-African American, mean eGFR African American, mean random glucose, mean AST, mean ALT, mean albumin, mean L- lactate Inpatient 508 Top 500 most administered medications + anti- Medications hyperglycemic medications Table 1. EHR datasets, number of variables per dataset, and variables extracted from each dataset that were included in the study.

[0132] The demographic dataset included 103,760 encounters. Social history and pastmedical history included 56,958 and 56,929 patients, respectively. In one implementation, past medical history and admission diagnoses were filtered for co-morbidities previously described to be predictive of BG variability. The five parent datasets (past medical history, 4912-3076-7444.4 065472-000953WOPT33 demographics, social history, laboratory, and inpatient medications) were merged on shared unique patient encounter numbers. Categorical variables were one-hot encoded, and all continuous variables were log2 transformed. Encounters with age, BMI, and laboratory outside of physiologic ranges for labs were removed, resulting in a final 76,232 encounters with 587 variables across 45,070 patients used during downstream feature selection. Table 1 above lists all static and time-dependent variables selected for downstream analysis.

[0133] A statistical analysis was also performed. To assist with convergence of the MELRmodel, feature selection was performed. In one implementation, 80% percent of encounters were used to generate a training dataset, and the remaining 20% were used to generate a testing dataset. Feature selection was performed using the 80% training dataset, which was further split into 80% training and 20% validation datasets to avoid data leak into the testing dataset during feature selection. Since out of many predictors to a model, only a subset can have a non- zero effect, a Lasso regression, a linear model with a regularization term, was utilized to perform feature selection. A regularization parameter of Lasso can drive less informative variables to have coefficients of zero and help identify features most outcome predictive. Hence, in one implementation, a Lasso 5-fold cross-validation determined a regularization parameter, alpha, among ten values between 10e-4 to 10 based on the lowest MSE mean over each fold. Using alpha, a trained model was then applied to predict POC BG in the validation dataset and performance was evaluated by comparing predicted BG with true value using R2 (coefficient of determination), mean squared error (MSE) and root mean squared error (RMSE).

[0134] Medication-induced BG alteration may be immediate (e.g. steroids) or take weeks(e.g. hydrochlorothiazide). Therefore, lookback windows of varying lengths were utilized to determine which medications are important when their doses are collected over longer or shorter periods. In some implementations, Lasso models were constructed using input variables collected over lookback windows of 12h, 24h, or 48h. Performance for the 24h lookback period was highest (R2=0.355) compared to 12h (R2=0.320) and 48h (R2=0.337) for the Lasso model, and thus was used for the analysis presented below.

[0135] To ensure robustness and reproducibility in coefficient ranking, a Lasso model withan optimized alpha was run using RepeatedKFold with 5 splits, 5 repeats and a maximum iteration limit of 10,000 on the training dataset. The mean absolute value of extracted coefficients (25 total coefficients per variable) was then calculated for each variable and sorted to identify the 50 variables with the highest mean absolute coefficient values. The distribution 4912-3076-7444.4 065472-000953WOPT34 of coefficient values for these variables was visualized using a boxplot to assess stability of the calculated coefficients across folds / splits.

[0136] The top 50 ranked predictors of POC BG obtained from the Lasso model wereinputted into a multivariate MELR model to assess statistical significance in predicting POC BG between variable and predictor with other variables were held constant. It is appreciated herein that MELR can be suitable for datasets with multiple observations per subject, whereas linear regression assumes observations are independent. Due to repeated encounters per patient, MELR was used to predict a 4h POC BOG. Fixed effects were data collected in the past 12, 24 or 48h, with random effects being set to the unique patient identifier to control for correlations among variables within the same patient. Benjamini-Hochberg (B-H) multiple testing correction was applied to p-values and variables with an adjusted p value <0.01 were considered to have statistically significant associations with BG. While not feasible to include every potential interaction term, MELR was repeated with clinically relevant interaction terms selected from the EHR. Variables that retained statistical significance using 12, 24, and 48h lookback windows were compared to determine differences in medications predictive of POC BG over different lookback time courses. Statistical analysis was performed in Python version 3.9.12 using the following packages (versions): sklearn (1.4.2), statsmodels (0.13.5), scipy (1.11.4), and seaborn (0.12.2). The Micromedex is a comprehensive evidence-based resource used by clinicians. The ‘Adverse Effects In-Depth Answers’ section was reviewed for documentation of BG alteration properties, which includes up to date post-marketing surveillance data and case studies.

[0137] FIG. 8A depicts an example timeline over which model input and output variableswere collected during a patient admission. Variables that do not change over the course of the hospital stay (demographics, past medical history and social history) were collected at the time of admission. Variables that change over the course of the admission (medications and laboratory) were collected over 12, 24, and 48h lookback windows. POC BG was predicted after a 4h prediction horizon. FIG. 8B is an illustration showing an example data analysis pipeline, beginning with several hundred variables across datasets, selection of top 50 variables after Lasso, and finally selection of variables with a B-H adjusted p-value less than 0.01. These variables were evaluated for prior documentation of BG variability in Micromedex. Variable Summary statistics for all encounters Age (years) 68.8 (58.1-78.8)Race Caucasian 48,685 (63.9%) Black or African American 13,833 (18.1%) 4912-3076-7444.4 065472-000953WOPT35 Asian 6,078 (8.0%) Other 5,831 (7.6%) Native Hawaiian or Other Pacific Islander 257 (0.33%) American Indian or Alaska Native 218 (0.29%) Declined / Missing / Unknown 1,330 (1.7%) Ethnic Group Non-Hispanic 59,955 (78.6%) Hispanic 14,755 (19.4%) Declined / Missing / Unknown 1,522 (2.0%) Sex Female 33,658 (44.2%) Male 42,533 (55.8%) Unknown 40 (0.05%) BMI (kg / m2) 27.0 (23.2-31.6)PMH T2D 53,040 (69.6%) T1D 7,717 (10.12%) CKD 31,223 (41.0%) History hypoglycemia 6,611 (8.7%) Liver failure 2891 (3.8%) Pregnancy 222 (0.29%) Labs (during 24h lookback) Mean POC BG (mg / dL) 153.8 (125.3-194.5) Mean random BG (mg / dL) 136.5 (109.0-178.5) Mean Hemoglobin A1c (%) 7.0 (6.0-8.8) Mean Creatinine (mg / dL) 1.1 (0.8-1.7) Admission location Non-ICU 69,949 (91.8%) ICU 6,283 (8.2%) Table 2. Summary statistics across patient encounters. Continuous variables are shown as median and IQR. Categorical variables are shown as counts and percent.

[0138] Baseline characteristics of 76,232 encounters from 45,070 patients across threehospitals in a large urban health system collected over 8 years (2014 to 2022) used for model training and evaluation is summarized in Table 2. The median age was 69 years old, with 64% being Caucasian, 18% Black or African American, and 19% of encounters. Forty-four percent of all encounters were female, and the median body mass index (BMI) was 27kg / m2. Most encounters had T2DM (70%) and nearly half of encounters had CKD (41%). Ten percent had T1DM and 9% had hypoglycemia listed as a past medical history prior to the encounter of interest. The median and inter-quartile range (IQR) of the mean POC BGs collected for each encounter during the 24h lookback period were 154 mg / dL and 125-195 mg / dL, respectively. Eight percent of encounters were admitted to the ICU.

[0139] The top 10 administered medications, filtered out of 508 medications, are displayedin FIG.9A. Three out of the ten medications are antihyperglycemics (insulin lispro SQ, insulin 4912-3076-7444.4 065472-000953WOPT36 glargine SQ and regular insulin IV. A Log2(x+1) transformation was applied to all variables to meet a linear regression assumption of linearity between predictors and responses, reduce impact of outliers, and enhance interpretability of coefficients. As described, medications and labs recorded during the lookback window were included in a model as inputs. Hence, input data is sparse since medications were recorded on a small subset of individuals during the lookback window. Log2(x+1) transformation was specifically chosen since zero values can be meaningful; one was added to retain the zero values. Overlayed distributions of log transformed input and output POC BGs is shown in FIG. 9B with a Spearman correlation coefficient of 0.60, indicating that prior recorded BGs are highly predictive of the next BG.

[0140] As described, a Lasso-based feature selection was performed. The 80% dataset(60,985 encounters) from the original dataset was further split to 80% training (48,788 encounters) and 20% validation datasets (12,197 encounters). Lasso 5-fold cross-validation on the training dataset was used to fit a Lasso model. This model was used to predict on the validation dataset, yielding R2 score 0.355, MSE 0.197, and RMSE 0.444. An optimal alpha, or alpha with the lowest mean MSE over the five folds, was 0.001 using the training dataset (FIG.10A). Actual vs predicted mean logged POC BG were linearly correlated (FIG.10B). Lasso L1 regularization prioritized the most informative features while setting coefficients of less important predictors to zero (FIG.10C). Using Lasso 5-fold cross-validation and only the training dataset, the model was trained to predict POC BG at the 4h prediction horizon. To ensure reproducibility in selection of top ranked coefficients, the Lasso model with optimized alpha was run using RepeatedKFold with 5 splits and 5 repeats, and a maximum iteration limit of 10,000 on the training data. The 50 variables with the highest mean absolute coefficient values were selected for downstream analysis. Distribution of coefficient values calculated across folds / spits for the top 25 variables were stable (FIG. 10D). The list of 50 selected variables and their mean absolute value coefficient scores are shown in STable 1 of EXAMPLE A.

[0141] To address potential confounders, control for inter-patient variability, and assessstatistical significance of coefficients, a MELR model expressing 4h POCT BG as a linear function of 50 Lasso-selected input variables with a random intercept was employed for each patient. The MELR model was used to predict on held-out 20% test dataset (i.e., dataset not used for feature selection). Resultant performance metrics were R20.347, MSE 0.197, and RMSE 0.443. B-H false-discovery rate correction was applied due to multiple testing and adjusted p-values for the 50 variables, coefficient, standard error and 95% confidence interval (CI) are shown in STable 2 of EXAMPLE A. Eleven variables ranked in the top 50 variables 4912-3076-7444.4 065472-000953WOPT37 by Lasso predictive of 4h BG did not meet the B-H adjusted p-value cut off 0.01 (STable 3 of EXAMPLE A). Of remaining 39 variables (Table 3), 28 were medications and eleven were a combination of labs, social history, past medical history and demographics. More than half (16 / 28) of the medications are recognized to alter BG, including insulins, nateglinide, metformin and steroids. All steroids were positively correlated with BG. Two anti- hyperglycemics (insulin lispro and aspart) were positively correlated with a 4h POC BG while all others were negatively correlated. Mean POC BG and mean random BG had the lowest adjusted p-values and were positively correlated with output BG. Mean L-lactate and mean Cr were negatively correlated with BG. Age, BMI, Asian race and T2DM were positively correlated with BG while Black / African American race was negatively correlated. Twelve variables met the adjusted p-value cut-off of <0.01 and were not recorded as causing alterations in BG in Micromedex. Of those twelve, medications negatively correlated with BG were sodium bicarbonate 8.4% IV syringe, ketorolac 30mg / ml inj, lorazepam 2mg / ml inj and iohexol 350mg iodine / ml IV and positively correlated medications included sterile water for irrigation, lidocaine 10mg / mL (1%) inj, rifaximin 550mg oral tablets, heparin 100units / ml IV, levetiracetam 100mg / ml oral solution, fluconazole 100mg oral tablets, norepinephrine in D5W 8mg / 250ml IV, and fentanyl citrate in D5W 5000mcg / 250ml IV.4912-3076-7444.4 065472-000953WOPT38Medication No Anesthetic Lidocaine PF 10mg / mL Inj* 8.66E-08 0.017 0.003 0.011 0.023Medication No Antibiotic Rifaximin 550mg PO Tabs* 2.84E-06 0.007 0.001 0.004 0.009Medication No Heparin Heparin Porcine PF 100 units / ml IV 0.00011 0.004 0.001 0.002 0.006Medication No NSAID Ketorolac 30mg / ml Inj* 0.00016 -0.011 0.003 -0.017 -0.006Medication No BenzodiazepineLorazepam 2mg / ml Inj* 0.00026 -0.026 0.007 -0.04 -0.013Medication No Anticonvulsant Levetiracetam 100mg / ml PO Soln 0.00033 0.006 0.002 0.003 0.009Medication No Antifungal Fluconazole 100mg PO Tabs 0.00043 0.008 0.002 0.003 0.012Medication No Catecholamine Norepinephrine in D5W 8mg / 250ml IV 0.0034 0.005 0.002 0.002 0.008Medication No Opioid Fentanyl Citrate in D5W 5000mcg / 250ml IV 0.0094 0.008 0.003 0.002 0.014Medication No Contrast Iohexol 350mg Iodine / ml IV 0.0094 -0.004 0.001 -0.007 -0.001Table 3. Thirty-nine variables with an adjusted p-value <0.01 according to a MELR model. EHR dataset, whether the medication was documented in Micromedex as having an effect on BG, the drug class, B-H adjusted p-value, coefficient, standard error and 95% CI are listed. The “*” character in Table 3 above indicates that the respective variable was significant across 12h, 24h, and 48h lookback windows.

[0142] The model was run with the same 50 variables plus inclusion of interaction termsto differentiate the effects of the underlying disease processes known to cause dysglycemia from the medication in altering BG. Interactions assessed included rifaximin 550mg oral tablets and history of liver failure, norepinephrine in D5W 8mg / 250ml IV and admission for sepsis, ketorolac 30mg / ml inj and admission for pain, and fentanyl citrate in D5W 5000mcg / 250ml IV and admission for pain. No interaction term had an adjusted p-value less than 0.01, and each of the medications alone retained significance after inclusion of the four interaction terms in the model. This indicates diseases presence did not differentiate the effect of the medication on the outcome (STable 4 of EXAMPLE A). FIGs. 11A-11D illustrate relationships between ketorolac (FIG. 11A), fentanyl (FIG. 11B), rifaximin (FIG. 11C), and norepinephrine (FIG. 11D), and 4h POC BG across patient encounters. The data points are colored based on whether a selected disease was present in that encounter. In this manner, various groups can be differentiated, and any potential differences can be observed in the relationship between, for example, norepinephrine dose and output BG value for patients with and without sepsis (FIG.11D).

[0143] Lastly, differences in variables that were statistically significant for various lengthof lookback windows were observed (FIG.12). Adjusted p-value, coefficient, standard error, and 95% CI for 12h lookback window (STable 5 of EXAMPLE A) and 48h lookback window (STable 6 of EXAMPLE A) are presented in summary tables. Ten variables were found to be significant using a 12h lookback window, four using a 24h lookback window, and eight using a 48h lookback window. Trends were observed in medication half-life affecting statistical significance based on lookback window duration. For example, short half-life steroid, hydrocortisone, was found to be unique to a 12h lookback window, and long-acting insulin glargine SQ was found to be unique to a 48h lookback window. 4912-3076-7444.4 065472-000953WOPT39

[0144] The present approach provides a first study that identifies medication predictors ofglucose homeostasis alterations using a large EHR dataset in conjunction with a BG prediction model. Incorporation of well-established BG-altering variables in the model served as controls, and showed significant correlations with BG. For example, mean POC BG, mean random BG and last POC BG were highly significant and positively correlated with output BG, and retained significance across different lookback windows. Studies show that prior BG is important in predicting future BG across outpatient continuous glucose monitoring and inpatient EHR data. Additionally, age, BMI, history of T2DM were significantly positively correlated with output BG across all three lookback windows. Age and obesity have been established through mechanistic and population studies to lead to metabolic dysregulation and insulin resistance and to be predictive of A1c.

[0145] Results described herein indicate that Asian race was positively correlated with 4hBG while the opposite was true for African American race. Similar trends in BG among races have been demonstrated. For instance, after a glucose tolerance test, African Americans had a lower glucose AUC (computed using the fasting, 30 minute and 2-hour plasma BG) and Asians had higher glucose AUC compared to Caucasians.

[0146] Several diabetes medications utilized to lower BG were significant, includinginsulin regular IV, insulin lispro SQ, insulin aspart SQ, and metformin across all lookback windows. Noteworthy, all diabetes medications that were significant had a negative coefficient, except short-acting insulins lispro and aspart. These medications have a peak response within 30-90 minutes, duration of 3-4 hours and elimination half-life of approximately 1 hour, so their BG lowering effect is diminished at the four-hour mark. Medications in classes of beta- blockers, statins, and parenteral nutrition were observed to be significantly correlated with BG across different lookback windows, all of which have been established as affecting BG variability.

[0147] One medication not documented in Micromedex as affecting BG, but found hereinto significantly reduce BG, was sodium bicarbonate IV, an alkalizing agent used in metabolic acidosis conditions. Bicarbonate’s effect on glucose dynamics has been unclear, and little research has been conducted to understand its role. Metabolic acidosis may be a contributor of insulin resistance and the National Health and Nutrition Examination Survey (NHANES) found that individuals with lower sodium bicarbonate levels had higher insulin resistance. In pancreatic beta cells, higher concentrations of bicarbonate extracellularly led to increased glucose-induced insulin secretion. However, in ambulatory individuals without diabetes treated with oral sodium bicarbonate over three months, there was no significant effect on fasting 4912-3076-7444.4 065472-000953WOPT40 plasma glucose. The route of administration of sodium bicarbonate and timing of BG measurement may be important factors in sodium bicarbonate’s effect on BG variability.

[0148] Another medication not been previously established but found herein tosignificantly alter BG, was rifaximin, a poorly absorbed antibiotic used commonly for hepatic encephalopathy. Two studies with very small sample sizes investigated rifaximin’s effect on insulin sensitivity. One study included 12 subjects with obesity or metabolic syndrome administered rifaximin or control over 12 weeks and found worsening insulin sensitivity through oral glucose tolerance tests and euglycemia clamping before and after treatment. Another study that included 15 individuals with nonalcoholic steatohepatitis found that rifaximin increased insulin resistance after 12 weeks. Changes in gut microbiota induced by rifaximin use may be linked to insulin resistance, but larger clinical trials are needed.

[0149] Heparin IV, an anticoagulant, was significantly positively correlated with BG.Although not documented in Micromedex as BG altering, endogenous heparin’s effect on glucose homeostasis has been studied. Heparin is a naturally occurring glycosaminoglycan and humans with diabetes and diet-induced obesity mice were shown to have higher blood heparin levels. Heparin inhibits insulin binding and PI3K / Akt / GLUT4 pathway, impairing skeletal muscle glucose uptake and leading to insulin resistance. This knowledge may assist clinicians with diabetic patients receiving heparin and provide a target for diabetes treatment.

[0150] There was limited evidence for other drugs found herein to significantly alter BG.However, available studies generally agree with findings described herein. For instance, the present study found fentanyl IV to be positively correlated with BG. There is some evidence that opiates may affect insulin secretion, and fentanyl inducing inhibition of glucose stimulated insulin release in rat beta islet cells. Also, for ketorolac injections, found herein to be negatively correlated with 4h BG, there was only one study that examined ketorolac effect acutely on BG. Specifically, 95 cholecystectomy patients received either ketorolac infusion or saline infusion perioperatively, and the study reported that BG concentrations were not different between the groups during the 24 hours following surgery.

[0151] Lorazepam injection was also found herein TO be negatively correlated with 4hBG. Evidence of benzodiazepine effect on BG is sparse, and results may indicate effects on BG is dependent on timing of BG measurement. For instance, one case report linked lorazepam use to hypoglycemia, and a study of 300 patients with different types of drug poisonings found patients with benzodiazepine overdoses were more likely to have hypogylycemia. However, a study on long term benzodiazepine administration using the NHANES dataset reported no significant association between benzodiazepines and prediabetes prevalence. 4912-3076-7444.4 065472-000953WOPT41

[0152] Comparison of lookback windows, as described herein, allows for identification ofvariables that differ in how rapidly they affect BG. While medium and long half-life steroids were significant regardless of the lookback window length, hydrocortisone, which is a short half-life steroid, was unique to a 12h window. Insulin glargine was significant using a 48h lookback window only, which may be attributed to the inability to detect its effect in only a 12h period. Atovaquone, an anti-malarial drug, was significant only using a 48h lookback. Atovaquone was one of five drugs selected as having high potential for anti-diabetes activity by a comprehensive, web-based drug repurposing tool, molecular property diagnostic suite (MPDS).

[0153] The present framework provides an unbiased approach for finding medicationpredictors of BG. That is, rather than hand-picking known BG predictors, 500 most administered medications in the hospital were included, and a Lasso model was allowed to perform feature selection. Both critically and non-critically ill patients were included, as well as all types of insulin. Previous studies excluded critically ill patients due to varying timing of BG measurement and routine use of IV regular insulin compared to non-critically ill patients. A mixed effects model was also used, which can handle repeated measures by including random effects to account for correlations between encounters of the same patient. The model utilized herein had comparable performance to a linear model that used a 24-hour moving average of inpatient BG measurements collected from EHRs to predict the next BG; R2 was 0.45 using all observations, and performance varied based on glycemic variability category (very high glycemic variability R20.14 to low glycemic variability R20.65).

[0154] It is recognized herein that there can be various confounders in a hospital setting,including surgery, stress and infection, which may affect a patient’s BG response. Additionally, while drug interactions that may cause an alteration in BG were not included herein, they may be included in some implementations. It is also recognized that EHR data can include missing data, erroneous entries, and lack of documentation of relevant variables, such as carbohydrate intake. The linear regression method employed herein did not use time series data and did not model non-linear interactions between medications and BG. However, a study using linear, cubist, random forest and K-nearest neighbor models to predict the next BG using previous BGs (collected using various moving average and rolling regression windows) plus demographic, laboratory, insulin doses, and steroid doses found no difference in performances (R2 CI’s overlapped). Randomized controlled trials where patients are matched between groups may assess any effect of drugs identified in this study. Large scale inpatient EHR data 4912-3076-7444.4 065472-000953WOPT42 may bring to light potential diabetes therapy candidates for further testing or drugs for use in individuals with hyper- or hypoglycemia.

[0155] Analysis performed as described herein identified significant predictors of glucosehomeostasis alterations and revealed an impact of various medications. Various expected correlations (e.g., insulin and metformin), and various unexpected correlations (e.g., heparin, sodium bicarbonate, fentanyl, and rifaximin) were found. Moreover, results described herein highlight that medications can have significant effects on glucose levels.

[0156] In yet another example, and to address critical gaps in conventional approaches, anLSTM-based model was developed for predicting inpatient hypoglycemia within a 24-hour horizon using point-of-care blood glucose (POC BG) data and structured clinical variables from an EHR. This study advances diagnostics and treatment fields in several ways. First, model performance was directly compared to expert clinician predictions on the same patient cohort—an evaluation rarely conducted in prior work. Second, fairness was assessed by evaluating performance across demographic subgroups, recognizing that even well-performing models can propagate or exacerbate healthcare disparities if subgroup performance is not evaluated. Third, interpretability was emphasized by using Shapley Additive explanations (SHAP), providing patient-level explanations of model predictions, helping clinicians understand not only the estimated risk of hypoglycemia but also clinical factors contributing to that risk. Finally, longitudinal external validation was performed using daily EHR extracts from a live clinical environment to ensure that the model utilized maintains accuracy and generalizability over time. By integrating real-time prediction, patient-level interpretability, fairness evaluation, and ongoing external validation, the present model provided a robust foundation for clinically actionable decision support aimed at improving inpatient glycemic safety.

[0157] The method included analysis of de-identified EHR data from all inpatients ≥18years old with a length of stay ≥24 hours who received at least one anti-hyperglycemic medication and were admitted to Cedars-Sinai Health System between 2014 and 2025. Individuals admitted to Labor and Delivery and who were pregnant were excluded. The dataset included ~160,000 inpatients across Cedars Sinai Medical Center (CSMC; 886 beds), a large, academic medical center, and two affiliate community hospitals: Marina Del Rey Hospital (MDRH; 133 beds) and Huntington Health (HH; 619 beds). There were 14,999 medications, 5,808 labs, 104 diet orders, 34,713 ICD-10 codes, and 13 socio-demographic features available in the dataset. 4912-3076-7444.4 065472-000953WOPT43

[0158] Ground truth labels were generated by assigning a label every 4 hours starting from4 hours after the first POC BG level until the last POC BG for each admission, based on whether a BG<70mg / dL occurred within the subsequent 24 hours from each prediction time. As shown in the example of FIG.14, a rolling prediction framework for inpatient hypoglycemia may be utilized. In this example, predictions were generated every 4h using a sliding window. At each prediction time, the model used a lookback (LB) window of time-series and static data to assess risk of BG<70mg / dL within a prediction horizon (PH). In particular, the PH extended 24h beyond each prediction point (e.g., current time) to identify whether a BG<70mg / dL event would occur. In the example of FIG.14, a red indicator with a value of “60” is used to show that 60mg / dL fell within the 4th PH, and was assigned a positive output label. For patients hospitalized fewer than 5 days, the LB window is zero-filled to maintain uniform input dimensions. Each prediction output was in the form of a binary classification, indicated by a value of “1” if a BG<70mg / dL occurs in the subsequent 24 hours, and a value of “0” otherwise. Other outputs, labels, or indicators (e.g., “positive,” “negative”, and so forth), may be possible. For patients hospitalized fewer than 5 days, the LB window was zero-filled to maintain uniform input dimensions. The 24-hour prediction window was selected based on CSMC endocrinologist consensus as a clinically actionable timeframe for intervention. However, longer or shorter prediction windows may be possible.

[0159] Predictions were updated every 4 hours as new clinical data became available. Inparticular, the model generated a prediction in just 0.07 seconds using only CPU resources, making frequent updates computationally feasible. Since predictions were performed every 4 hours, the number of predictions per admission exceeded the number of actual hypoglycemic events, i.e., a single event was associated with multiple positive prediction labels. For example, across all admissions, 6,426,254 prediction windows were generated and 4.4% of predictions were labeled positive, indicating significant class imbalance.

[0160] The model was configured to receive four time-series inputs that includedmedications, laboratory values, diet orders, and meal intake percentage, alongside one static input that included socio-demographic data, although any combination of inputs, as well as other inputs, may be possible. Features were selected in consideration of inpatient medications and labs associated with BG variation. To evaluate input utility, a curated list of 33 medications (insulin, common anti-hyperglycemics, and others associated with BG), 13 lab tests (e.g., Cr, lactate), 4 diet types, and percentage of meals consumed were compared to broader sets of 140 most administered medications, 102 frequently ordered labs, and 61 diet orders. The input set was selected based on superior performance and clinical relevance. Static inputs included 65 4912-3076-7444.4 065472-000953WOPT44 demographic and comorbidity variables. Missing lab and medication data were removed, resulting in only a 1% loss of encounters. Variables with <20% missingness were assessed for association with BG using chi-square tests; ethnicity and sex were significantly associated with BG (p < 0.01) and treated as missing not at random using one-hot encoding, while remaining variables were assumed missing at random and imputed using mode imputation. The temporal resolution of the model was optimized by evaluating various lookback windows; a 5-day window segmented into 30 timesteps (4 hours each) provided the best performance in this example. Over each timestep, summary values were computed: medication doses were summed, labs were averaged, diet type were encoded, and % meals consumed were totaled. If an input was absent during a 4-hour bin, a zero value was inserted for that interval.

[0161] Data was stratified by hospital admission and split into training (64%), validation(16%), and test (20%) sets to ensure no overlap of admissions across splits. A bidirectional LSTM architecture was built using TensorFlow / Keras to predict inpatient hypoglycemia within a 24-hour horizon. Model training was performed on a high-performance computing cluster equipped with NVIDIA A10080GB GPUs. Input data consisted of five modalities: time-series of medications (33 features), labs (13), diet (4), meals (1), and static variables (65). Each time- series input was passed through a dedicated branch composed of up to three bi-directional LSTM layers, each followed by batch normalization and dropout. The number of LSTM layers, units per layer, and dropout rates were all treated as tunable hyperparameters. Static inputs were processed through a dense ReLU-activated layer with dropout. The outputs of all branches were concatenated, passed through an additional dense layer with ReLU activation and dropout, and then fed to a final sigmoid-activated output layer for binary classification.

[0162] To address class imbalance, two strategies were applied during training: (1) classweighting (CW) to increase the penalty for misclassified positive cases, and (2) a custom weighted binary cross-entropy (WBCE) loss function to assign different penalties to positive versus negative classes during backpropagation. Class weights were empirically set to 11.3 for positive cases and 0.5 for negative cases based on the inverse frequency in the training set. Additionally, a focal loss function was implemented for sensitivity analysis.

[0163] Model hyperparameters—including the number of LSTM layers, units per layer,dropout rates, and learning rate—were tuned using KerasTuner’s Bayesian Optimization to maximize F1 score on the validation set. Batch sizes of 1024, 2048, and 4096 were assessed in the search. All models were trained for up to 10 epochs with early stopping (patience = 3), using 10 hyperparameter tuning trials, and evaluated on the held-out test set. The best- performing model was selected. The model’s final output layer produced probabilities using a 4912-3076-7444.4 065472-000953WOPT45 sigmoid activation function, and these were converted to binary classification using thresholds of 0.5, 0.6, 0.7, 0.8, or 0.9 to evaluate tradeoffs between precision and recall. Bootstrapped performance metrics (95% confidence intervals [CI]) were generated using 200 bootstrap samples (50% of the test set each) and compared using the Mann–Whitney U-test.

[0164] To evaluate the performance of simpler architecture models, a dense neural networktrained on temporally flattened inputs was developed. Time-series features (medications, labs, diet, and meals) were reshaped from 3D (samples × timesteps × features) to 2D by concatenating all timepoints along the feature axis, then combined with static features. This vector was passed to a feedforward network with two ReLU-activated dense layers and dropout for regularization. Hyperparameters including layer size, dropout rate, and learning rate were optimized using Bayesian optimization via Keras Tuner. CW was applied to address imbalance. Final performance was assessed on the test set using bootstrapped precision, recall, and F1 score across multiple thresholds. Logistic regression (LR) and XGBoost models were trained using the same flattened inputs. Each model’s hyperparameters were optimized independently. The LR and XGBoost models used CW to address class imbalance. Both models were evaluated across decision thresholds (0.5-0.9). Performance was estimated using 100 bootstrap iterations on the held-out test set, with 95% CIs calculated for precision, recall, and F1.

[0165] To evaluate model fairness across demographic and clinical subgroups, binarymasks were defined for each subgroup using the test static features matrix. These included age (≤75 vs. >75), sex, comorbidities (CKD, T1DM, T2DM), insurance type, race / ethnicity, and ICU status. Using these masks, F1 scores were computed separately for each subgroup by comparing the model’s predicted labels with the ground truth labels. To assess whether differences in model performance between subgroups were statistically significant, a bootstrapping procedure was implemented. For each predefined subgroup pair, F1 scores were sampled 1,000 times with replacement from the respective subgroup indices. The mean F1 difference and corresponding 95% CI from the bootstrapped distribution were calculated. Statistical significance was determined by whether the 95% CI excluded zero. Following model selection, a probability calibration was conducted on a held-out post-test calibration set (10% of data). Poor calibration can lead to disparities in care by misrepresenting risk. Calibration was performed using isotonic regression, and performance was quantified using the Brier score. Reliability plots stratified by sensitive attributes (e.g., race, sex, age) were generated to ensure that predicted probabilities aligned with observed hypoglycemia risk across all subgroups.

[0166] SHAP was used to interpret model outputs at the patient level. A GradientExplainerwas fitted on 5000 background samples drawn from the validation set. For each test set patient, 4912-3076-7444.4 065472-000953WOPT46 individualized input features—time-series of labs, meds, diets, meals, and static characteristics—were passed to the explainer to compute SHAP values. These values were flattened across time and ranked to identify the top clinical features driving each prediction, allowing clinicians to understand and act on the factors contributing to a patient’s hypoglycemia risk. To assess generalizability over time, longitudinal external validation was performed using daily clinical data extracts from Cedars-Sinai’s production EHR system. Each daily cohort was prospectively scored using the best-performing LSTM model , and prediction outputs were logged alongside observed blood glucose values. For each date, labs, medication administrations, diet orders, meal intake, demographics, and past medical history were extracted. The model was loaded from a saved H5 file and applied to data from each day using a consistent 5-day lookback window to generate risk probabilities for hypoglycemia within the subsequent 24-hour window. Predictions were matched to next-day POC BG values, with hypoglycemia defined as any BG <70 mg / dL. Each day's predictions were evaluated using multiple thresholds (0.5–0.9) to compute F1 score, precision, recall, and confusion matrix components. Model inputs were pre-processed using trained imputers and encoders from the model training phase.

[0167] To evaluate clinical utility, model predictions were compared to those made byexpert clinicians for a held-out cohort of 140 patients. Clinicians were given access to each patient’s EHR summary up to the prediction time and asked to estimate the risk of hypoglycemia in the next 24 hours. Model performance (F1 score, precision, recall) was benchmarked against these expert predictions to assess added value and feasibility for deployment.

[0168] Panel (A) of FIG. 16 shows a histogram of POC BG <70 mg / dL by hour of day.Panel (B) of FIG.16 shows stacked bar plots with annual percentage of POC BG values in four categories: <70 (hypoglycemia, red), 70-100 (blue), 100-180 (green), and >180 mg / dL (hyperglycemia, yellow), from 2014 to 2025. Labeling inside each segment indicates the proportion of total readings per year. Panel (C) of FIG.16 shows a histogram of all POC BG values. Models were trained and tested on 143,124 inpatient admissions collected between May 14, 2014 and March 26, 2025 with 4.8 million POC BG measurements, averaging ~30 POC BGs per admission. Hypoglycemia (BG<70mg / dL) occurred in 19% of admissions, with 2.8 events per admission. There was an average of 23 hypoglycemic events experienced by 8 patients per day. An analysis of temporal patterns of POC BG values may indicate the frequency of hypoglycemia peaks between 6-9 AM (Panel (A) of FIG. 15), the annual distribution of POC BG values across four BG categories indicates stable proportions of 4912-3076-7444.4 065472-000953WOPT47 hypoglycemia and hyperglycemia over time (Panel (B) of FIG.15), and most POC BG readings fall between 80-200 mg / dL(Panel (C) of FIG. 15). The proportion of BGs <70 mg / dL out of all recorded BG values per year remained consistently low, ranging from 1.3% to 1.9% annually.

[0169] Among all admissions, 10% had T1DM, 78% T2DM, and the remainder receivedantihyperglycemic agents without a formal diabetes diagnosis. The average age was 66.5 years, body mass index (BMI) 27.8 kg / m2, 43% were female, 59% Caucasian, 18% Black or African American (AA), and 9% Asian. Additionally, 58% had chronic kidney disease (CKD), 10% had prior hypo-glycemia, 50% had Medicare insurance, and 44% had commercial insurance. 10% were admitted to the ICU, and 88% of admissions occurred at CSMC, with 8% at HH and 4% at MDRH. Table 4 presents patient characteristics across the train, validation, test, daily external validation and clinician prediction cohorts. Demographic and clinical features are well-balanced across splits. Most patients were admitted to Cedars-Sinai Medical Center (CSMC), with smaller proportions from Marina Del Rey Hospital (MDRH) and Huntington Health (HH). The clinician prediction cohort serves as the subset used for comparing model performance to clinician assessments.

[0170] A total of 6,426,254 prediction windows were generated across all admissions.Consistent with the rarity of inpatient hypoglycemia, only 4.4% of prediction windows were labeled positive, reflecting significant class imbalance and the low incidence of inpatient hypoglycemia in the dataset (Table 4). The class distribution remained stable across splits.Table 4. Distribution of hypoglycemia and control labels across the training, validation, and test sets. La-bels were generated at 4-hour intervals based on whether a hypoglycemic event (BG <70 mg / dL) occurred within the subsequent 24 hours.

[0171] To identify an optimal temporal context for hypoglycemia prediction, modelperformance was evaluated across lookback (LB) windows of 2, 3, 5, and 7 days, each segmented into 4-hour intervals (Table 5), although other LB windows and segmentations may be possible . For each window length, models were trained using both class weighting (CW) and weighted binary cross-entropy (WBCE) to address class imbalance, with a fixed batch size of 4096. The model with 1 LSTM layer with 112 units and dropout 0.5, and learning rate 4912-3076-7444.4 065472-000953WOPT48 0.0003 achieved the highest mean F1 score of 0.302 (95% CI: 0.298-0.306), with precision of 0.230 and recall of 0.441 when evaluated on a held-out test set at a decision threshold of 0.7, offering an overall balance between sensitivity and specificity.

[0172] These results indicate the ability to provide 15 alerts / day, with 3.3 false positivesper true alert, and opportunity to prevent 3.5 hypoglycemia events per day (44% of the 8 anticipated hypoglycemic patients / day). The model had a Brier Score of 0.146. Both shorter and longer LB windows (e.g., 2, 3, or 7 days) resulted in lower F1 scores, potentially due to insufficient context or dilution of predictive signal from older data. These results indicate that a 5-day LB window with CW can provide for real-time prediction of hypoglycemia.Table 5. Model Performance Comparison Using Temporally Flattened vs Sequential Inputs. Classification performance across four models trained on the same inpatient data. The logistic regression (LR), XGBoost, and dense neural network (Dense) models were trained on temporally flattened 2D inputs that ignore sequential structure, while the best LSTM model was trained on 3D sequential inputs preserving time dependencies. For each model, the best- performing decision threshold (DT), optimized hyperparameters (HPs), and resulting F1 score, precision, recall with 95% confidence intervals, and p value with respect to the best LSTM are shown.

[0173] Table 5 compares the performance of models trained on flattened 2D inputs, whichdiscarded temporal structure with that of the highest F1-achieving bidirectional LSTM. A dense neural network, logistic regression (LR) and XGBoost models achieved F1 scores of 0.18, 0.15 and 0.26, respectively. In contrast, the best-performing LSTM model, which retained time- dependent patterns in the data, achieved a significantly higher F1 score of 0.30, highlighting the added value of modeling temporal dynamics in clinical prediction tasks. 4912-3076-7444.4 065472-000953WOPT49

[0174] To assess the relative contribution of each input modality, separate models weretrained using only a single input source using a 5-day LB window and 4-hour intervals (Table 6). All models used a batch size of 4096 and CW. Among individual inputs, laboratory values and medications yielded the highest F1 scores (0.214 and 0.217, respectively), with moderate precision and recall. In contrast, models using only % meals consumed (F1 = 0.069), diet orders (F1 = 0.093), or static history (F1 = 0.153) performed poorly, highlighting their limited predictive value in isolation. Notably, the best-performing configuration was achieved when all input modalities were combined (F1 = 302; Table 5), underscoring that a multimodal strategy may be used to capture complementary signals for hypoglycemia prediction, as well as other predictions.Table 6. Model performance using single input types compared to all inputs. Performance of LSTM models trained on individual input modalities. Inputs included medications, laboratory values, percent meals consumed, diet orders, and static patient history. For each model, the best-performing decision threshold (DT), optimized hyperparameters (HPs), and resulting F1 score, precision, and recall with 95% confidence intervals are shown.

[0175] An expanded input dataset that included the top 140 most administered inpatientmedications, 102 most ordered inpatient labs, 61 diet orders, % meal consumed and 65 static features was compared to the one curated based on clinical relevance and prior studies used up to now (33 medications, 13 labs, 4 diet types, % meal consumed and 65 static features). Both were evaluated using CW or WBCE, with a 5-day LB window and were trained with batch 4912-3076-7444.4 065472-000953WOPT50 sizes of either 2048 or 4096. The expanded input set, despite including more variables, performed worse overall. The best F1 achieved was 0.268 using WBCE and batch size 2048, with a higher recall of 0.459 but lower precision of 0.189 compared to the optimal model trained with the curated input dataset. These findings suggest that expanding the feature set may introduce noise rather than signal. Therefore, a clinically-informed curated input set for predictive modeling of inpatient hypoglycemia, and other conditions, may be suitable.

[0176] Results for the highest F1 achieving LSTM model using data from a representativepatient are illustrated in FIG.16. Specifically, Panel (A) of FIG.16 illustrates a binary outcome label used for training, where a label^of “1” was assigned starting 24h prior to each BG<70mg / dL event. Panel (B) shows actual POC BG values for this test patient. Panel (C) shows the model’s predicted probability of BG<70mg / dL. A threshold of 0.8 (horizontal red line in Panel (C)) triggered an alert beginning on 12 / 26, one day before the first event. As shown, the patient experienced one hypoglycemic event on 01 / 02 / 2015 (9:07 AM). To capture risk before these events, the model was trained using binary outcome labels that switch to “1” beginning 24 hours before each event (Panel (A) of FIG. 16). Model predictions were generated every 4 hours starting from 4 hours after the first POC BG measurement. As shown in Panel (B) of FIG.16, the actual POC BG measurements confirmed the hypoglycemic event on 01 / 02 / 2015. Panel (C) of FIG. 16 displays a predicted probability of hypoglycemia over time. Using a decision threshold of 0.7 led to the best F1, with the model issuing a correct alert starting on 01 / 01 / 2015, one day before the first event, flagging the event day. A false alert was also produced on 01 / 06 / 2015 when the BG decreased rapidly and approached 70 mg / dL but did not fall below it.

[0177] Understanding clinically relevant factors that contribute to the predictedhypoglycemia risk enhances confidence in model accuracy and resilience to confounding influences. To enhance clinical interpretability, SHapley Additive exPlanations (SHAP) with GradientExplainer was used, and more specifically, the positional SHAP (PoSHAP) concept was used to interpret time-series. PoSHAP computes each feature's contribution to a model prediction at each timestep, i.e., how much each feature pushes the model output up or down relative to a baseline. Positive values indicate increased risk, while negative values indicate decreased risk of hypoglycemia. For every 4-hour prediction time, the top 10 contributing features were recorded across all input types. This enabled consistent identification of significant predictors and assessment of their modifiability in clinical workflows. The same test patient as in FIG. 16 was to understand influential predictors of the first hypoglycemia prediction made the day before the event (FIG.17). In this test patient, insulin glargine given 4912-3076-7444.4 065472-000953WOPT51 at timesteps -2 (timestep 0 is most recent relative to prediction), -5 and -8 were the most influential modifiable contributors to this patient’s increased (positive SHAP values) hypoglycemia risk (Panel (A) of FIG. 17). The test patient’s insulin glargine doses can be visualized in Panel (B) of FIG. 17, showing the doses of 60, 85, and 65 units given at -2, -5, and -8 timesteps prior to the prediction made at the end of timestep 0.

[0178] Daily performance of the best-trained LSTM hypoglycemia prediction model wasevaluated across multiple decision thresholds over a 3-week period, following the end of the original training and testing dataset. Graphs in FIG. 18 show model performance metrics calculated daily from 6 / 26 / 2025 to 7 / 14 / 2025 across five decision thresholds (namely 0.5, 0.6, 0.7, 0.8, 0.9), applied to adult inpatients receiving antihyperglycemic medications with a length of stay >24h. Cumulative performance values for each threshold are included in the legends. This simulation aimed to assess how the model performs in real-time clinical settings using daily, most recent EHR cohorts of Cedars-Sinai adult inpatients with a hospital length of stay>24 hours who were receiving at least one antihyperglycemic medication. When the bestLSTM model was deployed daily during this period, it maintained performance to that achievedduring training at a 0.7 threshold (F1 score of 0.30), supporting the ability to generalize to daily cohorts and potential for real-world, EHR-integrated deployment. As appreciated from FIG. 18, a threshold of 0.5 consistently achieved the highest recall (global recall: 0.59), but also resulted in the most false positives (FP sum = 479). Conversely, higher thresholds (e.g., 0.8 and 0.9) improved precision (0.36 and 0.38, respectively) but reduced recall and the number of true positives (TP sum = 26 and 3, respectively). True negatives had higher counts at stricter thresholds due to reduced alerting. False negatives increased at higher thresholds, with a steep rise at 0.9 (FN sum = 124), indicating more missed events.

[0179] Table 7 summarizes pairwise comparisons of model F1 scores across demographicand clinical subgroups. No significant performance differences were observed between males and females, Asian and White patients, or Hispanic and Non-Hispanic patients. A small but statistically significant difference was noted between age groups, with the model performing slightly better in patients ≤75 years (F1 = 0.305) compared to those >75 years (F1 = 0.294). Differences in model performance emerged across clinical subgroups. The model performed better in patients with CKD (F1 = 0.310 vs 0.278), T1DM (F1 = 0.353 vs 0.304 for T2DM), and those in the ICU (F1 = 0.321 vs 0.298 for non-ICU). Hypoglycemia rates were higher in these populations (i.e., 10.0% T1DM vs 4.8% T2DM, 5.8% ICU vs 4.3% non-ICU, and 5.3% CKD vs 3.2% non-CKD), suggesting that higher incidence may aid the model in learning more informative patterns and generate more true positives, thereby improving F1 score. Similarly, 4912-3076-7444.4 065472-000953WOPT52 Black patients showed modestly higher F1 scores (0.320 vs 0.295 for White), which aligns with their higher baseline hypoglycemia rate (5.5% vs 4.2%). This performance difference may be clinically interpretable, as higher event rates enhance model learning and detection. Overall, the model demonstrates stronger performance in groups with greater hypoglycemia incidence, which may reflect improved detection performance in higher-risk groups. While some differences in performance may exist, these largely reflect underlying clinical and physiological variation, rather than systematic model bias.Table 7. Fairness Evaluation Across Demographic and Clinical Subgroups. Pairwise comparisons of best LSTM model F1 scores across demographic and clinical subgroups are shown to assess potential disparities in performance. For each pair, the F1 score in Group 1 and Group 2 is reported along with the F1 score difference (Group 1 - Group 2), and the 95% confidence interval (CI) of that difference derived via bootstrapping (1,000 iterations). Asterisks (*) indicate statistically significant differences where the 95% CI does not include zero.

[0180] Further, in a cohort of 140 encounters collected after model training / testing, ourmodel outperformed clinical experts in hypoglycemia prediction. The mean clinician F1 score of was 0.267, precision 0.167 and recall 0.667.

[0181] As appreciated from description herein, the present disclosure provides a numberof advantages, benefits, and improvements over conventional approaches. For instance, beyond aiding in reducing or eliminating real harm to individuals, the present approach provides the ability to predict hypoglycemia, as well as other BG-related conditions, and with 4912-3076-7444.4 065472-000953WOPT53 significant impact on clinical workflow and financial implications. For example, assuming a conservative real cost of $10,000 per event and opportunity cost (potential increase in revenue related to bed occupancy) of $6,500, the ability to predict and / or prevent 33% events per day (or more based on results obtained using systems and methods described), could have a total financial impact approaching $16 million per year while only costing 25% of one pharmacist’s time (~$50,000 / year). Further, extrapolating such 33% detection rate to all hospital beds in the US (~900,000) and assuming 1% of beds experience hypoglycemia daily with a total financial impact of $16,500 per bed, results in a potential yearly total US financial impact of $18 billion.

[0182] One or more elements or aspects or steps, or any portion(s) thereof, from one ormore of any of the claims below can be combined with one or more elements or aspects or steps, or any portion(s) thereof, from one or more of any of the other claims or combinations thereof, to form one or more additional implementations and / or claims of the present disclosure.

[0183] While the present disclosure has been described with reference to one or moreparticular embodiments or implementations, those skilled in the art will recognize that many changes may be made thereto without departing from the spirit and scope of the present disclosure. Each of these implementations and obvious variations thereof is contemplated as falling within the spirit and scope of the present disclosure. It is also contemplated that additional implementations according to aspects of the present disclosure may combine any number of features from any of the implementations described herein. EXAMPLE A

[0184] The disclosure may be further understood from the following description ofexemplary implementations. It should be understood that the present disclosure is not intended to be limited to particular forms disclosed, and the present disclosure is susceptible to various modifications and alternative forms. The present disclosure covers all modifications, equivalents, and alternatives falling within the spirit and scope of the present disclosure as defined by the appended claims. 4912-3076-7444.4 065472-000953WOPT54 STable 1 variable mean absolute value coeficientPage 1 of 1355 variable mean absolute value coeficientPHENYLEPHRINE_IN_0_9__NACL_100_MG_250_ML__0_4_MG_ML__IV_S 0.001857288 Page 2 of 1356 variable mean absolute value coeficientPage 3 of 1357 variable mean absolute value coeficientAMIODARONE_IV_INFUSION__1_MG_ML_ 0.000516079 Page 4 of 1358 variable mean absolute value coeficientPage 5 of 1359 variable mean absolute value coeficientPage 6 of 1360 variable mean absolute value coeficientPage 7 of 1361 variable mean absolute value coeficientPage 8 of 1362 variable mean absolute value coeficientPage 9 of 1363 variable mean absolute value coeficientPage 10 of 1364 variable mean absolute value coeficientPage 11 of 1365 variable mean absolute value coeficientPage 12of 1366 variable mean absolute value coeficient FENTANYL_CITRATE_IN_0_9__NACL_5000_MCG_250_ML__20_MCG_ML_ 0 HEPARIN_IN_D5W_25000_UNIT_250_ML_FOR_CRRT_NX_STAGE_PUMP 0Page 13of 1367 STable 2 variable unadjusted_p_value adjusted_p_value coeficient standard error 95CI_lower 95CI_upper Intercept 0 02.1790.056 2.07 2.289 P 3.60E-258 9.37E-257 0.429 0.013 0.405 0.454 1.59E-1232.76E-1220.1120.0050.1030.121 6.56E-58 8.52E-57 0.073 0.005 0.064 0.081 5.48E-380.0730.0060.0620.0853.31E-12 1.51E-11 0.04 0.006 0.029 0.052Asian 0.000631929 0.000938865 0.024 0.007 0.01 0.037 2.85E-080.0140.0030.0090.019 3.49E-12 1.51E-11 0.015 0.002 0.011 0.02 6.57E-050.0001219690.0230.0060.0120.034 7.01E-06 1.46E-05 0.018 0.004 0.01 0.026 6.56E-070.0110.0020.0070.016 0.011016739 0.013639772 -0.011 0.004 -0.02 -0.003 0.0604114830.0628279420.0170.009-0.0010.034 0.007007816 0.00911016 -0.006 0.002 -0.011 -0.002 1.90E-08-0.0070.001-0.01-0.005 5.43E-20 3.53E-19 0.008 0.001 0.006 0.009 8.49E-050.000152167-0.0110.003-0.017-0.006 0.000152255 0.000263908 -0.026 0.007 -0.04 -0.013 6.42E-050.0001219690.0080.0020.0040.0120.052309663 0.055512295 -0.009 0.005 -0.018 0 NATEGLINID 0.0026731570.00375687-0.0060.002-0.01-0.002 FAT_EMULS 0.037023382 0.040108664 0.008 0.004 0 0.016 QUETIAPINE 0.0688376870.070187445-0.0050.003-0.0110 0.016982544 0.019197658 -0.009 0.004 -0.017 -0.002-0.004 0.075516981 0.075516981 -0.007 0.004 -0.015 0.001 0.0002086540.0003382310.0070.0020.0030.011 1.06E-06 2.39E-06 0.007 0.001 0.004 0.009 0.0058854010.0080537060.0080.0030.0020.014 0.006337433 0.008449911 0.006 0.002 0.002 0.01 0.0010264120.0014825950.0050.0020.0020.008 0.000296496 0.000453465 0.008 0.002 0.003 0.012 5.51E-050.0001102190.0040.0010.0020.006 0.014385539 0.016758462 0.004 0.001 0.001 0.006 0.0074802530.00948715 -0.0040.001-0.007-0.001_ 0.023063483 0.025517045 -0.003 0.001 -0.006 0 LEVETIRACE 0.0001631430.000273659 0.0060.0020.0030.009 EPINEPHRIN 0.014502515 0.016758462 -0.009 0.003 -0.015 -0.002 Page 1of 268 variable unadjusted_p_value adjusted_p_value coeficient standard error 95CI_lower 95CI_upper CEFAZOLIN_ 0.014078346 0.016758462 0.0030.001 0.0010.005 Group Var 5.44E-06 1.18E-05 0.003 0.002 Page 2of 269 STable 3Page 1 of 170 STable 4 variable unadjusted_p_value adjusted_p_value coeficient standard error 95CI_lower 95CI_upper Intercept 0 02.1760.056 2.066 2.286 P 4.68E-258 1.38E-256 0.429 0.013 0.405 0.454 2.29E-1244.51E-1230.1120.0050.1030.122 1.61E-58 2.37E-57 0.073 0.005 0.064 0.082 5.20E-380.0730.0060.0620.0851.42E-12 7.63E-12 0.041 0.006 0.03 0.052Asian 0.000557792 0.000940278 0.024 0.007 0.01 0.037 5.02E-080.0140.0030.0090.019 1.08E-11 5.31E-11 0.015 0.002 0.011 0.019 6.83E-050.0001439560.0230.0060.0120.034 9.29E-06 2.28E-05 0.018 0.004 0.01 0.026 8.10E-070.0110.0020.0070.016 0.008634384 0.012425089 -0.012 0.004 -0.021 -0.003 0.0670925390.0761242270.0160.009-0.0010.034 0.006383004 0.00941493 -0.006 0.002 -0.011 -0.002 1.81E-08-0.0070.001-0.01-0.005 7.84E-20 5.78E-19 0.008 0.001 0.006 0.009 0.0002932090.000524223-0.0110.003-0.017-0.005 0.000131597 0.000267731 -0.026 0.007 -0.04 -0.013 6.18E-050.0001401940.0080.0020.0040.0120.052671357 0.063420614 -0.009 0.005 -0.018 0 NATEGLINID 0.0026182370.004291-0.0060.002-0.01-0.002 FAT_EMULS 0.035964032 0.04420579 0.008 0.004 0.001 0.016 QUETIAPINE 0.0696328710.077515837-0.0050.003-0.0110 0.016190988 0.020766701 -0.009 0.004 -0.017 -0.002-0.004 0.074169922 0.081037507 -0.007 0.004 -0.015 0.001 0.0002044160.0003884310.0070.0020.0030.011 2.61E-05 6.16E-05 0.008 0.002 0.004 0.012 0.0030631480.0048844790.0090.0030.0030.015 0.005398973 0.008167676 0.006 0.002 0.002 0.01 0.003266360.0050714530.0050.0020.0020.008 0.000373003 0.00064727 0.007 0.002 0.003 0.012 6.49E-050.000141850.0040.0010.0020.006 0.012952919 0.017368686 0.004 0.001 0.001 0.007 0.009323170.013096835 -0.0040.001-0.007-0.001_ 0.021377637 0.026835757 -0.003 0.001 -0.006 0 LEVETIRACE 0.0002106750.000388431 0.0060.0020.0030.009 EPINEPHRIN 0.012613743 0.017307228 -0.009 0.003 -0.016 -0.002 Page 1of 271 variable unadjusted_p_value adjusted_p_value coeficient standard error 95CI_lower 95CI_upper CEFAZOLIN_ 0.0143038650.018753957 0.0030.0010.0010.005 PMH_Liver_ 0.054611842 0.064441973 0.022 0.011 0 0.044 RIFAXIMIN_ 0.0961070890.103096696-0.0050.003-0.010.001 Admit_Seps 0.913042024 0.913042024 0.001 0.011 -0.02 0.023 NOREPINEP 0.263229280.2724653950.0070.006-0.0050.019 Admit_Pain 0.061216797 0.070819432 -0.01 0.005 -0.02 0 KETOROLAC 0.8408943460.8553925240.0010.007-0.0130.016 FENTANYL_ 0.188416147 0.198509869 -0.017 0.013 -0.043 0.008 Group Var0.0030.002Page 2of 272 STable 5Page 1 of 173 STable 6Page 1 of 1

Claims

74 CLAIMS WHAT IS CLAIMED IS:

1. A method for predicting blood glucose in an individual, the method comprising: receiving, at a current time, data associated with the individual, the data including (i) a set of one or more static data points each being generated or collected at an identical prior point in time occurring before the current time, and (ii) a series of sequential data points, each of the sequential data points being generated or collected within a prior window of time occurring before the current time; inputting the data associated with the individual into a model; and receiving from the model an indication of a risk of the individual developing hypoglycemia within a subsequent window of time after the current time.

2. The method of claim 1, wherein the prior point in time when each of the one or more static data points was generated occurs before the prior window of time when each of the sequential data points was generated.

3. The method of claim 1, wherein the prior point in time when each of the one or more static data points was generated occurs during the prior window of time when each of the sequential data points was generated.

4. The method of claim 1, the prior window of time ends at the current time or before the current time, and has a duration of about 6 hours, about 12 hours, about 18 hours, about 24 hours, about 36 hours, about 48 hours or more, between about 6 hours and about 12 hours, between about 6 hours and about 18 hours, between about 6 hours and about 24 hours, between about 6 hours and about 36 hours, between about 6 hours and about 48 hours or more, between about 12 hours and about 18 hours, between about 12 hours and about 24 hours, between about 12 hours and about 36 hours, between about 12 hours and about 48 hours or more, between about 18 hours and about 24 hours, between about 18 hours and about 36 hours, between about 18 hours and about 48 hours or more, between about 24 hours and about 36 hours, between about 24 hours and about 48 hours, or between about 36 hours and about 48 hours or more.

5. The method of claim 1, wherein the series of sequential data points includes at least two sequential data points from each of one or more categories of data, and wherein the at least two 4912-3076-7444.4 065472-000953WOPT75 sequential data points of each respective category were obtained at a separate point in time within the prior window of time.

6. The method of claim 5, wherein the one or more categories of data comprises (i) a first category associated with laboratory measurements, (ii) a second category associated with a diet of the individual, (iii) a third category associated with medication administered to the individual, or (iv) any combination of (i)-(iii).

7. The method of claim 6, wherein the at least two sequential data points associated with the fist category includes a first blood glucose measurement obtained at a first point in time within the prior window of time, and a second blood glucose measurement obtained at a second point in time within the prior window of time, the second point in time occurring after the first point in time.

8. The method of claim 6 or claim 7, wherein the at least two sequential data points associated with the third category includes a first amount of carbohydrates consumed by the individual at a first point in time within the prior window of time, and a second amount of carbohydrates consumed by the individual at a second point in time within the prior window of time, the second point in time occurring after the first point in time.

9. The method of any one of claims 6 to 8, wherein the at least two sequential data points associated with the second category includes a first amount of insulin administered to the individual at a first point in time within the prior window of time, and a second amount of insulin administered to the individual at a second point in time within the prior window of time, the second point in time occurring after the first point in time.

10. The method of any one of claims 6 to 8, wherein the at least two sequential data points associated with the second category includes a first type of insulin administered to the individual at a first point in time within the prior window of time, and a second type of insulin administered to the individual at a second point in time within the prior window of time, the second point in time occurring after the first point in time.

11. The method of any one of claims 6 to 8, wherein the at least two sequential data points associated with the second category includes a first type of insulin and an amount of the first 4912-3076-7444.4 065472-000953WOPT76 type of insulin administered to the individual at a first point in time within the prior window of time, and a second type of insulin and an amount of the second type of insulin administered to the individual at a second point in time within the prior window of time, the second point in time occurring after the first point in time.

12. The method of claim 10 or claim 11, wherein the first type of insulin and the second type of insulin are identical.

13. The method of any one of claims 10 to 12, wherein one or both of the first type of insulin and the second type of insulin is fast-acting insulin, intermediate-acting insulin, or long- acting insulin.

14. The method of any one of claims 1 to 13, wherein the one or more static data points includes an age of the individual, a sex of the individual, a weight of the individual, a medical history of the individual, or any combination thereof.

15. The method of any one of claims 1 to 14, wherein the individual is a patient currently admitted to a healthcare facility, and wherein the prior point in time when the static data was generated is an admission time of the individual.

16. The method of claim 15, wherein at least one of the sequential data points generated within the prior window of time was generated at the admission time.

17. The method of claim 15, where all of the sequential data points were generated after the admission time.

18. The method of any one of claims 1 to 17, wherein the output received from the model is indicative of a probability that the individual will develop hypoglycemia within a subsequent window of time that begins at or after the current time.

19. The method of claim 18, wherein the subsequent window of time has a duration of about 6 hours, about 12 hours, about 18 hours, about 24 hours, about 36 hours, about 48 hours or more, between about 6 hours and about 12 hours, between about 6 hours and about 18 hours, between about 6 hours and about 24 hours, between about 6 hours and about 36 hours, between 4912-3076-7444.4 065472-000953WOPT77 about 6 hours and about 48 hours, between about 12 hours and about 18 hours, between about 12 hours and about 24 hours, between about 12 hours and about 36 hours, between about 12 hours and about 48 hours, between about 18 hours and about 24 hours, between about 18 hours and about 36 hours, between about 18 hours and about 48 hours, between about 24 hours and about 36 hours, between about 24 hours and about 48 hours or more, or between about 36 hours and about 48 hours.

20. The method of claim 18 or claim 19, wherein the probability that the individual will develop hypoglycemia within the subsequent window of time is a probability that a blood glucose level of the individual will satisfy a threshold blood glucose level at any point in time during the subsequent window of time.

21. The method of claim 20, wherein the threshold blood glucose level is 70 mg / dL.

22. The method of claim 20 or claim 21, wherein the probability that the blood glucose level of the individual will satisfy the threshold blood glucose level at any point in time during the subsequent window of time includes a probability that the blood glucose level of the individual may be greater than 70 mg / dL at any point in time during the subsequent window of time, a probability that the blood glucose level of the individual may be greater than or equal to 70 mg / dL at any point in time during the subsequent window of time, a probability that the blood glucose level of the individual may be less than 70 mg / dL at any point in time during the subsequent window of time, a probability that the blood glucose level of the individual may be less than or equal to 70 mg / dL at any point in time during the subsequent window of time, or any combination thereof.

23. The method of any one of claims 1 to 22, wherein the model is a machine learning model.

24. The method of claim 23, wherein the model is a long short-term memory (LSTM) model, or a model that includes one or more LSTM blocks.

25. The method of claim 23 or claim 24, wherein the machine learning model includes a first input block that receives the static data and a second input block that receives the series of 4912-3076-7444.4 065472-000953WOPT78 sequential data points, the first input block including at least one dense layer, the second input block including one or more LSTM cells.

26. The method of claim 25, wherein the machine learning model includes at least one intermediate dense layer connected to both the first input block and the second input block, at least one batch normalization layer connected to the intermediate dense layer, and a final layer connected to the batch normalization layer, the final layer including a sigmoid function that outputs a probability that the individual will develop hypoglycemia during the subsequent window of time.

27. The method of any one of claims 1 to 22, wherein the model is a transformer model, a logistic regression model, or a statistical model.

28. The method of any one of claims 1 to 27, further comprising generating one or more control signals based on the indication of risk, and transmitting the one or more control signals to a treatment apparatus.

29. The method of any one of claims 28, wherein responsive to the one or more control signals transmitted, the treatment apparatus may initiate, modify, pause, adapt, and / or cease administration of a treatment.

30. A method for predicting a blood glucose condition in an individual, the method comprising: receiving data associated with an individual, the data including (i) a set of one or more static data points each being generated or collected before a current time, and (ii) a series of sequential data points, each of the sequential data points being generated or collected within a prior window of time occurring before the current time; inputting the data associated with the individual into a model configured to receive the set of one or more static data points and the series of sequential data points; and receiving, from the model, an indication of a risk of the individual to develop a blood glucose condition within a subsequent window of time of the current time; 4912-3076-7444.4 065472-000953WOPT79 31. The method of claim 30, further comprising receiving theoretical sequential data, altered sequential data, or both associated with a proposed intervention, and inputting the theoretical sequential data, the altered sequential data, or both into the model.

32. The method of claim 30 or 31, further comprising receiving, from the model, an updated indication of risk of the individual developing a blood glucose condition within the subsequent window of time.

33. The method of any one of claims 30 to 32, wherein one or more control signals may be generated based on the indication of a risk, the updated indication of risk, or both, and transmitted to a treatment apparatus.

34. A system for predicting blood glucose in an individual, the system comprising: at least one electronic interface configured to receive data associated with the individual, the data including (i) a set of one or more static data points each being generated at an identical prior point in time occurring before a current time, and (ii) a series of sequential data points, each of the sequential data points being generated within a prior window of time occurring before the current time; at least one memory device configured to store the data associated with the individual; and at least one processing device configured to implement a model that is configured to receive the data associated with the individual and output an indication of a risk of the individual developing hypoglycemia during a subsequent window of time after the current time.

35. The system of claim 30, wherein the at least one processing device is configured to implement the method of any one of claims 2 to 33. 4912-3076-7444.4 065472-000953WOPT

Images