Predictive funnels for generating hypoglycemic and hyperglycemic alerts based on continuous glucose monitoring data
A personalized glucose prediction model with a multi-range alarm strategy improves the accuracy of hypoglycemia and hyperglycemia predictions, addressing the limitations of existing methods by personalizing models and considering multiple prediction ranges, enhancing patient safety and insulin delivery.
Patent Information
- Application Number
- JP2023579318
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-11-02
- Filing Date
- 2022-11-01
- Publication Date
- 2024-11-13
AI Technical Summary
Existing methods for predicting hypoglycemic and hyperglycemic events in type 1 diabetes management are limited by the lack of personalized models that effectively handle inter- and intra-subject variability, and conventional approaches often focus on single prediction ranges without considering clinical impact, leading to inaccurate alarms and suboptimal patient care.
A personalized glucose prediction model using linear predictive models, combined with a cost function that accounts for clinical impact, and an alarm strategy that considers multiple prediction ranges, is employed to improve the accuracy of hypoglycemia and hyperglycemia predictions, utilizing a decision support system that includes a continuous analyte monitoring system, display device, and machine learning algorithms to generate tailored alerts.
The proposed method significantly enhances the accuracy of hypoglycemia and hyperglycemia predictions, reducing false alarms and improving patient safety by personalizing models and considering multiple prediction ranges, thereby optimizing insulin delivery and patient response.
Smart Images

Figure 2024541786000001_ABST
Abstract
Description
[Technical field]
[0001] (CROSS REFERENCE TO RELATED APPLICATIONS) This application claims priority to and the benefit of U.S. Provisional Patent Application No. 63 / 263,433, filed November 2, 2021, which is assigned to the assignee hereof and is expressly incorporated by reference in its entirety herein as if fully set forth below and for all applicable purposes. [Background technology]
[0002] Technological advances provided by continuous glucose monitoring (CGM) devices [1] and portable pumps for continuous subcutaneous insulin infusion (CSII) [2] have significantly improved the quality of life of subjects with type 1 diabetes (T1D). As pointed out in a recent report on artificial intelligence (AI) applications for diabetes management [3], the combined use of CGM devices, insulin pumps, and dedicated mobile applications [4] has brought the possibility to record different kinds of information, e.g., CGM data, insulin, diet, physical activity, and self-reported life events. This information allows the development of advanced AI-enabled decision support systems (DSS), composite tools that implement multiple software modules to assist patients in the decision-making process.
[0003] One of the key elements that can be embedded in an advanced DSS is a prediction module. Indeed, by knowing in advance whether blood glucose (BG) is approaching potentially harmful values, patients can take preventive actions to mitigate or avoid critical episodes such as hypoglycemia (i.e., BG below 70 mg / dL), significantly improving T1D management [5]-
[10] . Several research efforts have investigated BG prediction
[11] and many literature studies have focused on the challenge of predicting hypoglycemic episodes
[12] . In these efforts and studies, hypoglycemia prediction was addressed either by classification-based or regression-based approaches. Classification-based approaches consist of developing binary classifiers
[13] , i.e., algorithms that generate only two types of possible outputs: “impending hypoglycemia” or “no hypoglycemia predicted.” Instead, regression-based methods are two-step procedures that predict future glucose concentrations via regression as a first step and then issue an alarm if the predicted value falls below a suitable threshold (usually, but not necessarily, 70 mg / dL). In these efforts and studies, the predicted glucose concentrations used in the first step of the regression-based methods were obtained by using either linear predictors
[14] -
[16] or nonlinear methods
[17] -
[23] . Interestingly, one method has yet to be demonstrated to be superior to the other. [Brief description of the drawings]
[0004] So that the above-listed features of the present disclosure can be understood in detail, a more particular description, briefly summarized above, can be made by reference to embodiments, some of which are illustrated in the drawings. It should be noted, however, that the attached drawings merely illustrate certain exemplary embodiments of the present disclosure and therefore should not be considered as limiting its scope, since the description may be allowed for other equally valid embodiments. [Figure 1]1 illustrates aspects of an exemplary decision support system that may be used in connection with implementing embodiments of the present disclosure. [Diagram 2] FIG. 1 conceptually illustrates an exemplary continuous analyte monitoring system including an exemplary continuous analyte sensor with sensor electronics, in accordance with certain aspects of the present disclosure. [Diagram 3] 2 illustrates example inputs used by a prediction module of the decision support system of FIG. 1 according to some embodiments disclosed herein. [Figure 4] 1 is an exemplary workflow for training a machine learning model to predict future glucose values according to some embodiments disclosed herein. [Figure 5A] 1 illustrates a Clarke error grid, according to certain embodiments disclosed herein. [Figure 5B] 1 illustrates an example penalty function based on a Clarke error grid, according to certain embodiments disclosed herein. [Figure 5C] 1 illustrates an example glucose prediction for a machine learning model trained using glucose-specific mean squared error, according to some embodiments disclosed herein. [Figure 6] 1 is an exemplary workflow illustrating the use of predictive funnels, in accordance with certain embodiments of the present disclosure. [Figure 7] FIG. 7 is an exemplary flow diagram illustrating a method for generating alerts using the predictive funnel of FIG. 6 in accordance with certain embodiments of the present disclosure. [Figure 8] 1 is an exemplary plot illustrating the operation of a predictive funnel in accordance with certain embodiments of the present disclosure. [Figure 9] 1 is an exemplary workflow illustrating the use of predictive funnels to control an automatic insulin delivery (AID) system, according to certain embodiments disclosed herein. [Figure 10]FIG. 11 is a block diagram illustrating a computing device configured to perform the operations of FIGS. 6, 7, and / or 8 in accordance with certain embodiments disclosed herein.
[0005] For ease of understanding, wherever possible, the same reference numerals have been used to designate identical elements that are common to each of the figures. It is contemplated that elements disclosed in one embodiment may be beneficially utilized in other embodiments without specific recitation. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0006] The following description details a machine learning-based approach to predict hypoglycemic and / or hyperglycemic events based on glucose predictions generated using a glucose prediction model, such as a personalized glucose prediction model. Using a personalized prediction model allows for the handling of large inter- and intra-subject variability, one of the major challenges in glucose prediction. The glucose prediction model may be a linear prediction model, which allows for a high degree of personalization. Linear prediction models provide strong convergence results and allow for statistical characterization
[24] . Furthermore, linear prediction models may be used with computationally convenient algorithms, such as the Kalman predictor
[25] . Finally, although metabolic physiology is nonlinear
[26] , linear strategies have proven capable of capturing the essential dynamics
[14]
[16]
[27]
[28] and remain challenging competitors to nonlinear approaches
[29]
[30] .
[0007] The following detailed description provides improved techniques for both (a) training a predictive model to predict hypoglycemia and hyperglycemia, and (b) an alarm generation strategy (e.g., a rule-based model) that receives the output of the predictive model as input. The predictive model can be trained using a cost function previously introduced in
[31] that can take into account the clinical impact of prediction errors, thus enabling the identification of more effective models for predicting future hypoglycemic episodes. The alarm generation strategy does not focus on a single prediction horizon, but rather considers multiple prediction horizons simultaneously, thereby accounting for the expected decrease in accuracy of predictions from the predictive model as the prediction horizon increases.
[0008] Exemplary Decision Support System Including a Prediction Module FIG. 1 illustrates an exemplary decision support system 100 for implementing improved predictive model training techniques and improved alert generation strategies, according to certain embodiments. The decision support system 100 is configured to provide decision support to a user 102 (individually referred to herein as a user and collectively referred to herein as users) using sensor data provided by a continuous analyte monitoring system 104 including at least a continuous glucose sensor. The user may be a patient, or in some cases, a caregiver for the patient, in certain embodiments. In certain embodiments, the decision support system 100 comprises the continuous analyte monitoring system 104, a display device 107 running an application 106, a decision support engine 114, a user database 110, a historical records database 112, a training server system 140, and the decision support engine 114, each of which is described in more detail below. The training server system 140 may implement improved techniques for training predictive models, while a prediction module 116 in the decision support engine 114 uses the predictive model to implement improved alert generation strategies.
[0009] As used herein, the term "analyte" is a broad term used in its ordinary sense, including but not limited to referring to a substance or chemical component within a biological fluid (e.g., blood, interstitial fluid, cerebrospinal fluid, lymphatic fluid, or urine) that can be analyzed. In the example described below, the analyte is glucose in a user's bloodstream. However, the concentration of any analyte that can be measured or any time-varying value can be predicted using the techniques described herein. For example, analytes can include naturally occurring substances, man-made substances, metabolites, and / or reaction products. Analytes measured by the device and method may include, but are not limited to, potassium, glucose, acarboxyprothrombin; acylcarnitines; adenine phosphoribosyltransferase; adenosine deaminase; albumin; alpha-fetoprotein; amino acid profile (arginine (Krebs cycle), histidine / urocanic acid, homocysteine, phenylalanine / tyrosine, tryptophan); androstenedione; antipyrine; arabinitol enantiomers; arginase; benzoylecgonine (cocaine); biotinidase; biopterin; c-reactive protein; carnitine; carnosinase; CD4; ceruloplasmin; chenodeoxycholic acid; chloroquine; cholesterol; cholinesterase; conjugated 1-beta hydroxy-cholic acid; cortisol. ;creatine kinase;creatine kinase MM isoenzyme;cyclosporine A;d-penicillamine;deethylchloroquine;dehydroepiandrosterone sulfate;DNA (acetylation polymorphism, alcohol dehydrogenase, alpha 1-antitrypsin, glucose-6-phosphate dehydrogenase, hemoglobin A, hemoglobin S, hemoglobin C, hemoglobin D, hemoglobin E, hemoglobin F, D-Punjab, beta-thalassemia, hepatitis B virus, HCMV, HIV-1, HTLV-1, MCAD, RNA, PKU, Plasmodium vivax, 21-deoxycortisol);desbutylhalofantrine;dihydropteridine reductase;diphtheria / tetanus antitoxin;erythrocyte arginase;erythrocyte protoporphyrin;esterase D;fatty acids / acylglycines;free beta-human chorionic gonadotropin;Free erythrocyte porphyrin; free thyroxine (FT4); free triiodothyronine (free tri-iodothyronine, FT3); fumarylacetoacetase; galactose / gal-1-phosphate; galactose-1-phosphate uridyltransferase; gentamicin; glucose-6-phosphate dehydrogenase; glutathione; glutathione peroxidase; glycocholate; glycosylated hemoglobin; halofantrine; hemoglobin variants; hexosaminidase A; human erythrocyte carbonic anhydrase I; 17-α-hydroxyprogesterone; hypoxanthine phosphoribosyltransferase; immunoreactive trypsin; lactate; lead; lipoproteins ((a), B / A-1, β); lysozyme; mefloquine; netilmicin; phenobarbitone; phenytoin; phytanic acid / pristanic acid; progesterone; prolactin; prolidase; purine nucleoside phosphorylase; quinine; reverse triiodothyronine (reverse tri-iodothyronine, rT3); selenium; serum pancreatic lipase; sisomicin; somatomedin C; specific antibodies recognizing any one or more of the following, which may include adenoviruses, antinuclear antibodies, anti-zeta antibodies, arboviruses, pseudorabies virus, dengue virus, dracunculiasis, tapeworm, Entamoeba histolytica, enteroviruses, giardiasis, Helicobacter pylori, hepatitis B virus, herpes virus, HIV-1, IgE (atopic disease), influenza virus, Leishmania donovani, Leptospirosis, measles / mumps / rubella, Mycobacterium leprae, Mycobacterium pneumoniae plasma, myoglobin, Onchocerca volvulus, parainfluenza virus, malaria parasite, poliovirus, Pseudomonas aeruginosa, respiratory syncytial virus, Rickettsia (tsutsugamushi disease), Schistosoma mansoni, Toxoplasma gondii, Treponema pallidum, Trypanosoma cruzi / rangeli, vesicular stomatitis virus, Wuchereria bancrofti, yellow fever virus); specific antigens (Hepatitis B virus, HIV-1); succinylacetone; sulfadoxine; theophylline; thyrotropin (TSH); thyroxine (T4); thyroxine-binding globulin; trace elements; transferrin; UDP-galactose-4-epimerase; urea;Examples of analytes include uroporphyrinogen I synthase; vitamin A; white blood cells; and zinc protoporphyrin. Salts, sugars, proteins, fats, vitamins, and hormones naturally occurring in blood or interstitial fluids can also constitute analytes in certain embodiments. Ions are charged atoms or compounds that can include (e.g., sodium, potassium, calcium, chloride, nitrogen, or bicarbonate). Analytes can be naturally occurring in biological fluids, such as metabolites, hormones, antigens, antibodies, ions, and the like. Alternatively, the analyte may be introduced into the body or may be exogenous, such as contrast agents for imaging, radioisotopes, chemical agents, fluorocarbon-based synthetic blood, challenge agent analytes (e.g., introduced for the purpose of measuring the increase and / or decrease in the rate of change of the concentration of the challenge agent analyte or other analyte in response to the introduced challenge agent analyte), or exogenous insulin; glucagon, ethanol; cannabis (marijuana, tetrahydrocannabinol, hashish); inhalants (nitrous oxide, amyl nitrite, butyl nitrite, chlorohydrocarbons, hydrocarbons); cocaine (crack cocaine); stimulants (amphetamines, methamphetamines, Ritalin, Silurt, Preluzine, Didrex, Prestate, Boranil, Sandrex, Pregin;Antidepressants (barbiturates, tranquilizers such as methaqualone, Valium, Librium, Miltown, Serax, Equuanil, Tranxine);Hallucinogens (phencyclidine, lysergic acid, mescaline, peyote, psilocybin);Nucleic acids (heroin, codeine, morphine, opium, meperidine, Percocet, Percodan, Tasionex, fentanyl, Darvon, Talwin, Lomotil);Synthetic narcotics (fentanyl, meperidine, amphetamine, methamphetamine, and phencyclidine analogues, e.g., ecstasy);Anabolic steroids;The analyte may be a drug or pharmaceutical composition, including, but not limited to, nicotine and nicotine. Metabolites of drugs and pharmaceutical compositions are also contemplated analytes. Analytes such as neurochemicals and other chemicals produced in the body, such as ascorbic acid, uric acid, dopamine, noradrenaline, 3-methoxytyramine (3MT), 3,4-dihydroxyphenylacetic acid (DOPAC), homovanillic acid (HVA), 5-hydroxytryptamine (5HT), and 5-hydroxyindoleacetic acid (FHIAA), as well as intermediates of the citric acid cycle, can also be analyzed;
[0010] In certain embodiments, the continuous analyte monitoring system 104 is configured to continuously measure one or more analytes and transmit the analyte measurements to an electronic medical records (EMR) system (not shown in FIG. 1 ). An EMR system is a software platform that enables electronic entry, storage, and maintenance of digital medical data. EMR systems are commonly used throughout hospitals and / or other care facilities to document clinical information about patients over time. EMR systems organize and present data in a manner that assists clinicians, for example, in interpreting health conditions and providing ongoing care, scheduling, billing, and follow-up. Data contained in the EMR system may also be used to generate reports for the patient's clinical care and / or disease management. In certain embodiments, the EMR may be in communication (e.g., over a network) with the decision support engine 114 to perform the techniques described herein. Communication may occur via a variety of networked data configurations, including but not limited to Web API protocols, HL7, FHIR, EDI, XML, CDA, and the like.
[0011] These data communication configurations may be sent directly to the EMR or may be sent through one or more intermediate systems, including but not limited to an interface engine, before entering the EMR system and then being displayed. Patient data communicated to the EMR through any of these other means may be matched to patient records via probabilistic matching, manual human matching, or an EMPI or MPI (Master Patient Index, Electronic Master Patient Index) to ensure that data entry into one system matches patient information in another system. Data from the analyte device may also be matched with data from alternative devices or systems before being entered into the EMR, or data may be sent back to our historical records database 112, user database 110, and / or decision support engine.
[0012] These data transfers enable the system to perform optimized decision support via means described herein. In particular, as described herein, the decision support engine 114 may retrieve data associated with a user, use the retrieved data as input to one or more trained models, and output predictions. In some cases, the EMR may provide the decision support engine 114 with data to be used as input to one or more models. Additionally, in some cases, the decision support engine 114 may provide a prediction to the EMR after making the prediction.
[0013] In certain embodiments, the continuous analyte monitoring system 104 is configured to continuously measure one or more analytes and transmit the analyte measurements to the display device 107 for use by the application 106. In some embodiments, the continuous analyte monitoring system 104 transmits the analyte measurements to the display device 107 via a wireless connection (e.g., a Bluetooth connection). In certain embodiments, the display device 107 is a smartphone. However, in certain other embodiments, the display device 107 may instead be any other type of computing device, such as a laptop computer, a smart watch, a tablet, or any other computing device capable of running the application 106. In some embodiments, the continuous analyte monitoring system 104 and / or the analyte sensor application 106 transmit the analyte measurements to one or more other individuals interested in the patient's health (e.g., a family member or a physician for real-time treatment and care of the patient). The continuous analyte monitoring system 104 may be described in more detail with respect to FIG. 2.
[0014] Application 106 is a mobile health application configured to receive and analyze analyte measurements from analyte monitoring system 104. In particular, application 106 stores information about the user, including the user's analyte measurements, in a user profile 118 associated with the user for processing and analysis, and for use by decision support engine 114 to provide decision support output to the user (e.g., alerts, recommendations, guidance, signals for controlling an insulin pump / pen, etc.).
[0015] The decision support engine 114 refers to a set of software instructions having one or more software modules, including the prediction module 116. In certain embodiments, the decision support engine 114 executes entirely on one or more computing devices in a private or public cloud. In such embodiments, the application 106 communicates with the decision support engine 114 over a network (e.g., the Internet). In some other embodiments, the decision support engine 114 executes partially on one or more local devices, such as the display device 107, and partially on one or more computing devices in a private or public cloud. In some other embodiments, the decision support engine 114 executes entirely on one or more local devices, such as the display device 107. As described in more detail herein, the decision support engine 114 may provide decision support output to a user via the application 106. The decision support engine 114 provides the decision support output based on information contained in the user profile 118.
[0016] The user profile 118 may include information collected about the user from the application 106. For example, the application 106 provides a set of inputs 128 that include analyte measurements received from the continuous analyte monitoring system 104 that are stored in the user profile 118. In certain embodiments, the inputs 128 provided by the application 106 include other data in addition to the analyte measurements received from the continuous analyte monitoring system 104. For example, the application 106 may obtain the additional inputs 128 via manual user input, one or more other non-analyte sensors or devices, other applications running on the display device 107, etc. Non-analyte sensors or devices may include, but are not limited to, one or more of an insulin pump, an electrocardiogram (ECG) sensor or heart rate monitor, a blood pressure sensor, a sweat sensor, a respiration sensor, a thermometer, a sensor or device provided by the display device 107 (e.g., an accelerometer, a camera, a global positioning system (GPS), a heart rate monitor, etc.), or other user accessory (e.g., a smart watch), or any other sensor or device that provides relevant information about the user. The input 128 of the user profile 118 provided by the application 106 is described in further detail below with respect to FIG.
[0017] The prediction module 116 of the decision support engine 114 is configured to process the set of inputs 128 to predict future analyte levels, risk of adverse events (e.g., hypoglycemia and hyperglycemia), and / or determine whether to provide a decision support output, and the content of such output, based on the inputs 128. As described below, various types of decision support outputs can be provided, such as alerts, decision support recommendations, control signals for controlling the operation of a medication delivery device (e.g., an insulin pump or pen), etc.
[0018] The user profile 118 may also include demographic information 120, disease progression information 122, and medical information 124. In certain embodiments, such information may be provided via user input or may be obtained from a particular data store (e.g., electronic medical record (EMR), etc.). In certain embodiments, the demographic information 120 may include one or more of the user's age, body mass index (BMI), ethnicity, gender, etc. In certain embodiments, the disease progression information 122 may include information regarding the user's disease, such as diabetes, or whether the user has a history of hyperkalemia, hypokalemia, hyperglycemia, hypoglycemia, etc. In certain embodiments, the information regarding the user's disease may also include the length of time since diagnosis, stage of the disease, level of disease control, level of compliance with disease management therapy, other types of diagnoses (e.g., heart disease, obesity), or measures of health (e.g., heart rate, exercise, stress, sleep, etc.), etc. In certain embodiments, the disease progression information 122 may be provided as the output of one or more predictive algorithms and / or trained models based on analyte sensor data generated, for example, via the continuous analyte monitoring system 104.
[0019] In certain embodiments, the medical information 124 may include information regarding the amount, frequency, and type of medication taken by the user. In certain embodiments, the amount, frequency, and type of medication taken by the user are time-stamped and correlated with the user's analyte levels, thereby indicating the effect the amount, frequency, and type of medication had on the user's analyte levels. In certain embodiments, the medical information 124 may include information regarding the ingestion of one or more drugs known to control and / or improve glucose homeostasis. The one or more drugs known to control and / or improve glucose homeostasis may include medications for lowering blood glucose levels, such as insulin, including fast-acting and long-acting insulins, other glucose-lowering medications, such as metformin, and the like. As described in more detail below, the decision support system 100 may be configured to use the medical information 124 to determine optimal insulin doses to be prescribed for different users. In particular, the decision support system 100 may be configured to identify one or more optimal insulin doses based on the patient's health, the patient's current condition, and / or the effectiveness of the insulin dose.
[0020] In certain embodiments, the user profile 118 is dynamic because at least a portion of the information stored in the user profile 118 may be modified over time and / or new information may be added to the user profile 118 by the decision support engine 114 and / or the application 106. Thus, the information in the user profile 118 stored in the user database 110 provides an up-to-date repository of information related to the user.
[0021] User database 110, in some embodiments, refers to a storage server operating in a public or private cloud. User database 110 may be implemented as any type of data store, such as a relational database, a non-relational database, a key-value data store, a file system, including a hierarchical file system, etc. In some example implementations, user database 110 is distributed. For example, user database 110 may include multiple persistent storage devices that are distributed. Additionally, user database 110 may be replicated such that the storage devices are geographically distributed.
[0022] The user database 110 includes user profiles 118 associated with multiple users who similarly interact with the applications 106 executing on the display devices 107 of other users. The user profiles stored in the user database 110 are accessible to the applications 106 as well as the decision support engine 114. The user profiles in the user database 110 may be accessible to the applications 106 and the decision support engine 114 via one or more networks (not shown). As described above, the decision support engine 114, and more specifically the prediction module 116 of the decision support engine 114, may fetch inputs 128 from the user database 110 and generate decision support outputs, which may then be stored as application data 126 in the user profiles 118.
[0023] In certain embodiments, the user profiles 118 stored in the user database 110 may also be stored in the historical records database 112. The user profiles 118 stored in the historical records database 112 may provide a repository of current and historical information for each user of the application 106. Thus, the historical records database 112 essentially provides all data related to each user of the application 106, with the data stored according to an associated timestamp. The timestamp associated with the information stored in the historical records database 112 may identify, for example, when information related to a user was obtained and / or updated.
[0024] Additionally, the historical records database 112 may maintain time series data collected for users over a period of time, including for users who use the continuous analyte monitoring system 104 and application 106. For example, analyte data for a user who used the continuous analyte monitoring system 104 and application 106 for five years to manage the user's health may have the time series analyte data associated with the user maintained for the five years.
[0025] Additionally, in certain embodiments, historical record database 112 may include data about one or more patients who are not users of continuous analyte monitoring system 104 and / or application 106. For example, historical record database 112 may include information (e.g., user profiles) about one or more patients who have not been previously diagnosed with diabetes, as analyzed, for example, by a health care physician (or other known method), as well as information (e.g., user profiles) about one or more patients who have been previously diagnosed with diabetes (of various types and stages), as analyzed, for example, by a health care physician (or other known method). The data stored in historical record database 112 may be referred to herein as population data.
[0026] The data associated with each patient stored in the historical records database 112 may provide time series data collected over the patient's disease lifetime, where the disease may be diabetes. For example, the data may include information associated with the patient during the lifetime of the disease, including information about the patient before they were diagnosed with diabetes, and information related to co-morbid diseases associated with diabetes. Such information may be indicative of the patient's symptoms, the patient's physiological condition, the patient's glucose level, the patient's insulin level, the status / condition of one or more of the patient's organs, the patient's habits (e.g., activity level, food consumption, etc.), prescribed medications, etc.
[0027] In another example, the data may include information associated with the patient during the lifetime of the disease, including information about the patient before they were diagnosed with diabetes, hyperglycemia, or hypoglycemia, and information related to diabetes that has progressed and / or regressed in the patient, as well as information related to other diseases such as hyperglycemia, hypoglycemia, renal disease, hypertension, cardiac conditions and diseases, or similar diseases that are co-morbid in association with diabetes. Such information may indicate the patient's symptoms, the patient's physiological state, the patient's glucose level, the patient's potassium level, the patient's lactate level, the patient's insulin level, the status / condition of one or more organs of the patient, the patient's habits (e.g., activity level, food consumption, etc.), prescribed medications, medication compliance, etc. throughout the lifetime of the disease.
[0028] Although shown as separate databases for conceptual clarity, in some embodiments, the user database 110 and the historical records database 112 may operate as a single database. That is, historical and current data associated with users of the continuous analyte monitoring system 104 and application 106, as well as historical data associated with patients who were not previous users of the continuous analyte monitoring system 104 and application 106, may be stored in a single database. The single database may be a storage server operating in a public or private cloud.
[0029] As discussed above, the decision support system 100 is configured to diagnose, stage, treat, and assess a user's risk of diabetes using a continuous analyte monitoring system 104 including at least a continuous glucose sensor. For example, the decision support engine 114 may be configured to collect information associated with the user in a user profile 118 stored in the user database 110 and perform analyses thereon to (1) predict future analyte levels (e.g., glucose levels), (2) predict risk of adverse events such as hypoglycemia and hyperglycemia, (3) generate alarms based on the predicted risk of adverse events, and / or (4) provide treatment recommendations. In certain embodiments, the user's glucose metrics may include glucose level, glucose level rate of change, glucose trend, average glucose, glucose management indicator (GMI), glycemic variability, time in range (TIR), glucose clearance rate, and the like.
[0030] The user profile 118 may be accessible to the decision support engine 114 via one or more networks (not shown) to perform such analyses. In certain embodiments, the decision support engine 114 is configured to provide real-time and / or non-real-time decision support regarding diabetes to the user and / or others, including but not limited to healthcare providers (HCPs), the user's family members, the user's caregivers, researchers, and / or other individuals, systems, and / or groups supporting care or learning from the data.
[0031] In certain embodiments, the decision support engine 114 may utilize a prediction module 116 that includes one or more trained machine learning models for (1) predicting future analyte levels (e.g., glucose levels), (2) predicting risk of adverse events such as hypoglycemia and hyperglycemia, and / or (3) generating decision support output based on the predicted future analyte levels and / or predicted risk of adverse events. In the illustrated embodiment of FIG. 1, the prediction module 116 may utilize trained machine learning models provided by a training server system 140. Although shown as separate servers for conceptual clarity, in some embodiments, the training server system 140 and the decision support engine 114 may operate as a single server. That is, models may be trained and used by a single server, or may be trained by one or more servers and deployed for use on one or more other servers. In certain embodiments, models may be trained on one or many virtual machines (VMs) running at least in part on one or many physical servers in relational and / or non-relational database formats.
[0032] The training server system 140 is configured to train the machine learning models using training data, which may include data (e.g., from user profiles) associated with one or more patients (e.g., users or non-users of the continuous analyte monitoring system 104 and / or application 106) who have previously been diagnosed with (1) not being diabetic or (2) having various stages of diabetes. The training data may be stored in the historical records database 112 and may be accessible to the training server system 140 via one or more networks (not shown) for training the machine learning models. The training data may also include user-specific data, in some cases, for users over time.
[0033] Training data refers, for example, to a characterized and labeled dataset. In certain embodiments, the dataset may be a population-based dataset that includes a plurality of historical data records. Each historical data record in the dataset may include information corresponding to a different user profile stored in the user database 110. Furthermore, each historical data record may be characterized and labeled. In machine learning and pattern recognition, a feature is an individual measurable characteristic or feature. Generally, the feature that best characterizes the patterns in the data is selected to create a predictive machine learning model. Data labeling is the process of adding one or more meaningful and informative labels to provide context to data for learning by the machine learning model.
[0034] As an illustrative example, each relevant characteristic of a user reflected in a corresponding historical data record may be a feature used in training a machine learning model. Such features may include features associated with the user's demographic information (e.g., age, gender, etc.), time-stamped analyte measurements (e.g., time-stamped glucose measurements), time-stamped meal intake information, time-stamped insulin administration information (e.g., administered insulin dosages). Each historical data record may be labeled according to what the corresponding model is trained to predict.
[0035] The model is then trained by the training server system 140 using the characterized and labeled training data. In certain embodiments, the features of each historical data record may be used as input to a machine learning model, and the generated output may be compared to the label associated with the corresponding historical data record. The model may calculate a loss based on the difference between the generated output and the provided label. This loss is then used to modify the internal parameters or weights of the model. By iteratively processing each historical data record corresponding to each historical patient, in certain embodiments, the model may be iteratively refined to generate accurate predictions.
[0036] As shown in FIG. 1, the training server system 140 deploys these trained models to the decision support engine 114 for use during runtime. For example, the decision support engine 114 may obtain a user profile 118 associated with the user, use information in the user profile 118 as input to the trained models, and output a prediction of future analyte levels (e.g., glucose levels) and / or risk of hypoglycemia and / or hyperglycemia. The decision support engine 114 may then use the prediction to generate a decision support output. For example, if the prediction indicates a hypoglycemic or hyperglycemic event, the decision support output may include an alert (i.e., an alarm) generated using an alert generation model as described below. Generating the alert may include causing a display device (e.g., display device 107) to output a human-perceptible alert, such as a visual message, an audible alert, or a tactile alert (e.g., with a haptic device), which may indicate, for example, the predicted event and / or the patient's glucose level.
[0037] In certain embodiments, along with or instead of the alert, the decision support output may also include decision support recommendations for the patient to exercise, consume carbohydrates, administer insulin (e.g., change the infusion rate), etc. In certain embodiments, along with or instead of the alert, the decision support output may further include a signal to an insulin pump to modify the amount of insulin being or to be administered to the user.
[0038] The decision support output may be provided to the user (e.g., through application 106), to the user's caregiver (e.g., a parent, relative, guardian, teacher, nurse, etc.), to the user's physician, or to any other individual interested in the user's health, possibly with a view to improving the user's health, such as by achieving a recommended treatment.
[0039] In certain embodiments, the user's own data is used to personalize one or more models that may have been initially trained based on population data. For example, a model (e.g., trained using population data) may be deployed for use by the decision support engine 114 to predict the user's future analyte levels. After making a prediction using the model, the decision support engine 114 may be configured to obtain the user's actual analyte value and calculate a loss between the prediction and the actual glucose value, which loss may be used to retrain the model. Thus, the model may be continually retrained and personalized using the calculated loss between the prediction and the actual glucose value as input to the model to personalize the model for the user. Additional details regarding training one or more models herein are described in more detail below.
[0040] 2 illustrates an analyte monitoring system 200 including an exemplary continuous analyte sensor system 104, a non-analyte sensor 206, a medical device 208, and multiple display devices 210, 220, 230, and 240 in accordance with certain aspects of the present disclosure. The components of analyte monitoring system 200 are configured to operate to continuously monitor one or more analytes of a user in accordance with certain aspects of the present disclosure.
[0041] The continuous analyte monitoring system 104 in the illustrated embodiment comprises a sensor electronics module 204 and one or more continuous analyte sensors 202 (individually referred to herein as a continuous analyte sensor 202 and collectively referred to herein as continuous analyte sensors 202) associated with the sensor electronics module 204. The sensor electronics module 204 may wirelessly communicate (e.g., directly or indirectly) with one or more of display devices 210, 220, 230, and 240. In certain embodiments, the sensor electronics module 204 may also wirelessly communicate (e.g., directly or indirectly) with one or more medical devices, such as a medical device 208 (individually referred to herein as a medical device 208 and collectively referred to herein as medical devices 208), and / or one or more other non-analyte sensors 206 (individually referred to herein as a non-analyte sensor 206 and collectively referred to herein as a non-analyte sensor 206).
[0042] In certain embodiments, the continuous analyte sensor 202 may include a sensor for detecting and / or measuring an analyte. The continuous analyte sensor 202 may be a multi-analyte sensor configured to continuously measure two or more analytes, or a single analyte sensor configured to continuously measure a single analyte as a non-invasive, subcutaneous, transdermal, transdermal, and / or intravascular device. In certain embodiments, the continuous analyte sensor 202 may be configured to continuously measure an analyte level of a user using one or more measurement techniques, such as enzymatic, chemical, physical, electrochemical, spectrophotometric, polarimetric, calorimetric, iontophoretic, radiometric, immunochemical, etc. In certain aspects, the continuous analyte sensor 202 provides a data stream indicative of the concentration of one or more analytes in the user. The data stream may include a raw data signal, which is then converted to a calibrated and / or filtered data stream used to provide an estimated analyte value to the user.
[0043] In certain embodiments, the continuous analyte sensor 202 may be a multi-analyte sensor configured to continuously measure multiple analytes in the user's body. For example, in certain embodiments, the continuous multi-analyte sensor 202 may be a single multi-analyte sensor configured to measure two or more of glucose, insulin, lactate, ketones, pyruvate, and potassium in the user's body.
[0044] In certain embodiments, one or more multi-analyte sensors may be used in combination with one or more single-analyte sensors. As an illustrative example, a multi-analyte sensor may be configured to continuously measure potassium and glucose, and in some cases may be used in combination with an analyte sensor configured to measure only lactate levels. Information from each of the multi-analyte and single-analyte sensors may be combined to provide diabetes decision support using the methods described herein.
[0045] The continuous analyte sensor 202 may be implemented as a continuous glucose monitor (CGM). Some examples of continuous glucose monitors include glucose monitoring sensors. In some embodiments, the glucose monitoring sensor is an implantable sensor such as described with reference to U.S. Pat. No. 6,001,067 and U.S. Patent Publication No. 2011-0027127-A1. In some embodiments, the glucose monitoring sensor is a transcutaneous sensor such as described with reference to U.S. Patent Application Publication No. 2006-0020187-A1. In some embodiments, the glucose monitoring sensor is a dual electrode analyte sensor such as described with reference to U.S. Patent Publication No. 2009-0137887-A1. In still other embodiments, the glucose monitoring sensor is configured to be implanted in a host blood vessel or outside the body, such as the sensor described in U.S. Patent Publication No. 2007-0027385-A1. These patents and publications are incorporated herein by reference in their entirety.
[0046] As used herein, the term "continuous" can mean fully continuous, semi-continuous, periodic, etc. Such continuous monitoring of an analyte is advantageous in the diagnosis and staging of disease, given that continuous measurements provide continuously up-to-date measurements as well as information regarding trends and rates of analyte change over continuous periods of time. Such information can be used to make more informed decisions in the assessment of glucose homeostasis and treatment of diabetes.
[0047] In certain embodiments, the sensor electronics module 204 includes electronic circuitry associated with measuring and processing the continuous display device sensor data, including predictive algorithms associated with processing and calibrating the sensor data. The sensor electronics module 204 can be physically connected to the continuous analyte sensor 202 and can be integral (non-releasably attached) or releasably attached to the continuous analyte sensor 202. The sensor electronics module 204 can include hardware, firmware, and / or software that enables measurement of the display device levels via the continuous analyte sensor 202. For example, the sensor electronics module 204 may include a potentiostat, a power supply for providing power to the sensor, other components useful for signal processing and data storage, and a telemetry module for transmitting data from the sensor electronics module to one or more display devices. The electronics can be affixed to a printed circuit board (PCB) or the like and can take a variety of forms. For example, the electronics can take the form of an integrated circuit (IC), such as an application-specific integrated circuit (ASIC), a microcontroller, and / or a processor.
[0048] The display devices 210, 220, 230, and / or 240 are configured to display displayable sensor data, including analyte data, that may be transmitted by the sensor electronics module 204. Each of the display devices 210, 220, 230, or 240 may include a display, such as a touchscreen display 212, 222, 232, / or 242, for displaying the sensor data to a user and / or accepting input from a user. For example, a graphical user interface (GUI) may be presented to the user for such purposes. In some embodiments, the display device may include other types of user interfaces, such as a voice user interface, instead of or in addition to a touchscreen display, for communicating the sensor data to a user of the display device and / or accepting user input. The display devices 210, 220, 230, and 240 may be examples of the display device 107 shown in FIG. 1 used to display the sensor data to a user of FIG. 1 and / or to receive input from a user.
[0049] In some embodiments, one, some, or all of the display devices are configured to display or otherwise communicate the sensor data as it is communicated from the sensor electronics module (e.g., in a data package sent to the respective display device), without any additional anticipated processing required for calibration and real-time display of the sensor data.
[0050] The display devices may include custom display devices designed specifically to display a particular type of displayable sensor data associated with the analyte data received from the sensor electronics module. In certain embodiments, the display devices may be configured to provide alerts / alarms based on the displayable sensor data. Display device 210 is an example of such a custom device. In some embodiments, one of the display devices is a display device 220 representing a smartphone, e.g., a mobile phone configured to display a graphical representation of continuous sensor data (e.g., including current and historical data) using a commercially available operating system (OS). Other display devices may include a display device 230 representing a tablet, a display device 240 representing a smartwatch, a medical device 208 (e.g., an insulin delivery device or a blood glucose meter), and / or a desktop or laptop computer (not shown).
[0051] Because different display devices have different user interfaces, the contents of the data package (e.g., amount, format, and / or type of data displayed, alarms, etc.) can be customized (e.g., programmed differently by the manufacturer and / or end user) for each particular display device. Thus, in certain implementations, multiple different display devices can wirelessly communicate directly with the sensor electronics module (e.g., the on-skin sensor electronics module 204 physically connected to the continuous analyte sensor 202) during a sensor session to enable multiple different types and / or levels of display and / or functionality associated with the displayable sensor data.
[0052] In certain embodiments, one or more of the display devices may have a user interface, which may include various interfaces such as a liquid crystal display (LCD), a vibrator, an audio transducer (e.g., a speaker), a backlight (not shown), etc., for presenting UI features. Components with such user interfaces may provide controls for interacting with a user (e.g., a host). One or more UI features may enable, for example, toggles, menu selections, option selections, status selections, yes / no responses to on-screen questions, an "off" function (e.g., for an alarm), an "acknowledge" function (e.g., for an alarm), a reset, etc. The UI features may also provide, for example, visual data output to the user. An audio transducer (e.g., a speaker) may provide an audible signal in response to the triggering of a particular alert, such as a current and / or predicted condition. In some example implementations, the audible signals may be differentiated by tone, volume, duty cycle, pattern, duration, etc. In some example implementations, the audible signal may be configured to be muted (eg, acknowledged or turned off) by pressing one or more buttons.
[0053] As mentioned above, the sensor electronics module 204 may be in communication with a medical device 208. The medical device 208 may be a passive device in some exemplary embodiments of the present disclosure. For example, the medical device 208 may be an insulin pump for administering insulin to a user. For various reasons, it may be desirable for such an insulin pump to receive and track glucose, potassium, lactate, insulin, ketones, and / or pyruvate values transmitted from the continuous analyte monitoring system 104, and the continuous analyte sensor 202 is configured to measure glucose, insulin, potassium, lactate, ketones, and / or pyruvate.
[0054] Additionally, as discussed above, the sensor electronics module 204 can also communicate with other non-analyte sensors 206. The non-analyte sensors 206 may include, but are not limited to, altimeter sensors, accelerometer sensors, temperature sensors, respiration rate sensors, sweat sensors, and the like. The non-analyte sensors 206 may also include monitors such as heart rate monitors, ECG monitors, blood pressure monitors, pulse oximeters, caloric intake, and medication delivery devices. One or more of these non-analyte sensors 206 may provide data to the decision support engine 114, which is described further below. In some aspects, a user may manually provide a portion of the data for processing by the training server system 140 and / or the decision support engine 114 of FIG. 1.
[0055] In certain embodiments, the non-analyte sensor 206 may be combined in any other configuration, such as, for example, combined with one or more continuous analyte sensors 202. As an illustrative example, a non-analyte sensor, such as a heart rate sensor, may be combined with a continuous analyte sensor 202 configured to measure glucose to form a glucose / heart rate sensor that is used to transmit sensor data to the sensor electronics module 204 using a common communications circuit. As another illustrative example, a non-analyte sensor, such as a heart rate sensor and / or an ECG sensor, may be combined with a multi-analyte sensor 202 configured to measure glucose and potassium to form a glucose / potassium / heart rate sensor that is used to transmit sensor data to the sensor electronics module 204 using a common communications circuit.
[0056] In certain embodiments, a wireless access point (WAP) may be used to link one or more of the continuous analyte monitoring system 104, the multiple display devices, the medical device 208, and / or the non-analyte sensors 206 together. For example, the WAP 138 may provide Wi-Fi and / or cellular connectivity among these devices. Near Field Communication (NFC) and / or Bluetooth may also be used between the devices shown in the diagram 200 of FIG. 2.
[0057] Figure 3 illustrates exemplary inputs used by the decision support system of Figure 1, in accordance with certain embodiments disclosed herein. In particular, Figure 3 provides a more detailed view of the exemplary inputs introduced in Figure 1. Figure 3 illustrates exemplary inputs 128 on the left, decision support engine 114 including application 106 and prediction module 116 in the center, and alerts 130 on the right. In certain embodiments, alerts 130 may correspond to and describe events predicted using prediction module 116.
[0058] The application 106 obtains input 128 via one or more channels (e.g., manual user input, sensors / monitors, other applications running on the display device 107, EMR, etc.). As previously mentioned, in certain embodiments, the input 128 may be processed by the prediction module 116 and / or the decision support engine 114 to output a decision support output (e.g., alert 130). As previously mentioned, the alert 130 is only one example of a type of decision support output that may be provided to a user. For example, in certain embodiments, the input 128 may be used by the decision support engine 114 to provide additional or alternative decision support outputs to the user, such as decision support recommendations, signals to an insulin delivery device (e.g., an insulin pump or pen), etc.
[0059] Input 128 may include a glucose measurement (g(k)) at a given time t(k), meal intake information (CHO(k)) at a given time t(k), and amount of exogenous insulin information (I(k)) administered at a given time t(k). Note that the times at which glucose is measured, the times at which the meal is consumed, and the times at which exogenous insulin is administered may be different.
[0060] The glucose measurements g(k) may be measured by and received from at least a continuous glucose sensor (or at least a multi-display device sensor configured to measure glucose) that is part of the continuous analyte monitoring system 104. The meal intake information CHO(k) may include information about one or more of the meals, snacks, and / or beverages, such as one or more of the size, content (number of milligrams (mg) of potassium, glucose, lactate, carbohydrates, fat, protein, etc.), order of consumption, and time of consumption. In certain embodiments, the meal intake information CHO(k) may be provided to the user by manual input, by providing a photo via an application configured to recognize the type and amount of food (e.g., the potassium and glucose / carbohydrate content of the food), and / or by scanning a barcode or menu. In various examples, the meal size may be manually input as one or more of calories, amount (e.g., "3 cookies"), menu item (e.g., "Royale with Cheese"), and / or food exchange (1 fruit, 1 dairy product). In some examples, the dietary intake information CHO(k) may be received via a convenient user interface provided by the application 106.
[0061] In certain embodiments, the dietary intake information CHO(k) entered by the user may relate to glucose consumed by the user. Glucose for consumption may include any natural or artificial food or beverage containing glucose, dextrose, or carbohydrates (e.g., glucose tablets, bananas, or bread, etc.).
[0062] The exogenous insulin information I(k) may include information regarding the user's insulin delivery. In particular, input related to the user's insulin delivery may be received via a wireless connection on the smart pen, via user input, and / or from an insulin pump. The insulin delivery information may include one or more of insulin volume, delivery time, etc. Other parameters such as insulin action time, insulin activity rate, or duration of insulin action may also be received as input.
[0063] In certain embodiments, the time may be provided as an input, such as, for example, the time of day, or the time from a real-time clock, etc. For example, in certain embodiments, the glucose measurements g(k), meal intake information (CHO(k)), and exogenous insulin information I(k) may be time-stamped to indicate the date and time the information was received for the user.
[0064] User input of any of the above-mentioned inputs 128 may be made via a user interface of the display device 107 of Figure 1. As noted above, in certain embodiments, the prediction module 116 determines whether to provide a decision support output and the content of such output (e.g., whether to generate an alert 130 and the content of the alert 130) based on the inputs 128.
[0065] Exemplary methods and systems for providing predictive decision support output using a predictive module FIG. 4 illustrates an example workflow 400 for training a machine learning model 402 configured to predict future glucose measurements based on inputs 128. The machine learning model 402 may be trained using the inputs 128 for an individual user or a population of users. The machine learning model 402 may be used by the prediction module 116 to obtain predicted future glucose measurements to determine whether to generate a decision support output (e.g., an alert 130) and the content of such output (e.g., whether to generate an alert 130 and the content of the alert 130, as described in more detail below). The training of the machine learning model 402 may be performed by the training server system 140 or by a software module within the prediction module 116 itself. The workflow 400 may be performed continuously or periodically to update the machine learning model 402 over time.
[0066] For a given time t(k), the input 128 for that time t(k) or a window preceding time t(k) is processed using the machine learning model 402 to generate a predicted glucose measurement for a subsequent time t(k+1).
[0067]
number
[0068] Then, the predicted glucose measurement for time t(k+1)
[0069]
number
[32] and summarized below. The value g(k+1) may be obtained by interpolation, extrapolation, curve fitting, or other techniques based on glucose measurements taken at times other than the time corresponding to g(k+1). The output of the loss function 404 is then input to a training algorithm 406. The training algorithm 406 then updates one or more parameters of the machine learning model 400 according to the output of the loss function 404 such that the machine learning model 402 is trained to predict glucose measurements based on past values of the input 128.
[0070] The machine learning model 402 can be a linear regression-based model or other types of machine learning models, such as a nonlinear machine learning model. The machine learning model 402 is a multi-input, single output (MISO) model. The MISO model presents a number of degrees of freedom that can be selected to implement the machine learning model 402, such as model class (autoregressive with exogenous input (ARX); autoregressive moving average with exogenous input (ARMAX); autoregressive integrated moving average with exogenous input (ARIMAX); output-error (OE); Box-Jenkins (BJ)) and model complexity (number of parameters estimated). The test results summarized below were obtained using an ARIMAX model with the Bayesian information criterion (BIC) used as a method for selecting the model order (see
[24] ,
[29] ,
[30] ).
[0071] To estimate the model parameters of the machine learning model 402, a prediction error method (PEM) may be used to minimize the one-step-ahead prediction error. In particular, we designate the model parameters as θ and implement the loss function 404 as an MSE cost function. Thus, the estimated model parameters
[0072]
number
[0073]
number
[0074]
number
[0075] 5A, 5B, and 5C, in the improved approach, the estimated model parameters
[0076]
number
[32] .
[0077]
number
[0078] The gMSE metric is suggested by the Clarke error grid (CEG)
[36] shown in Fig. 5A and describes the clinical impact of prediction errors. The gMSE metric is obtained by increasing the MSE value according to a penalty function in the case of overestimation of hypoglycemia and underestimation of hyperglycemia. For example, the gMSE value can be obtained by increasing the MSE value by 250% in the case of glucose overestimation during hypoglycemia and by 200% in the case of glucose underestimation during hyperglycemia. By doing so, overestimation in hypoglycemia is penalized more than underestimation in the same area, since the former is clinically more dangerous, and overestimation in hypoglycemia may prevent the detection of an episode or induce an optimal assessment of its severity, leading to inappropriate treatment. Symmetric reasoning applies to the case of hyperglycemia, but since hypoglycemia is considered to be more dangerous than hyperglycemia, in the first case the MSE is increased (e.g., by 250%) more than in the second case (e.g., only by 200%).
[0079] The increase in MSE is done in such a way that gMSE preserves two fundamental mathematical properties of the original metric: smoothness and convexity, since these properties simplify the optimization involved in parameter estimation. In particular, the penalty function is modified by g(k+1) and g(k+1) as described in
[32] to achieve a smooth surface, as shown in Figure 5B.
[0080]
number
[0081] Referring to FIG. 5C, a machine learning model 402 trained using gMSE as a loss function 404 will output predictions that are more likely to be accurate in the region near and above the hyperglycemic threshold (e.g., 180 mg / dl) compared to a model trained using only MSE, because errors in this region were penalized more during training. Overestimated errors are favored in this region (near and above the hyperglycemic threshold) compared to underestimated errors, because overestimated errors are less likely to cause clinically impactful damage. Similarly, a machine learning model 402 trained using gMSE as a loss function n 404 will output predictions that are more likely to be accurate in the region near the hypoglycemic threshold (e.g., 70 mg / dl) with respect to MSE, because errors in this region were penalized more during training. Underestimated errors are favored in this region (near the hypoglycemic threshold) compared to overestimated errors, because underestimated errors are less likely to cause clinically impactful damage.
[0082] 6 illustrates a workflow 600 that includes a predictive funnel evaluator 602 for determining the appropriateness of generating a decision support output, such as an alert 130, based on glucose values predicted using a machine learning model 402. Although the workflow 600 is shown and described with reference to a machine learning model 402 trained using gMSE, other machine learning models trained using MSE or other loss functions may be used in place of the machine learning model 402 while still achieving at least some of the benefits of the predictive funnel evaluator 602.
[0083] Several conventional approaches proposed in the literature for generating alerts using predicted glucose values focus on a single prediction and rarely consider the prediction accuracy in detecting the crossing of the hypoglycemic threshold
[14] ,
[15] . In contrast, the predictive funnel implemented by workflow 600 considers multiple predictions in different prediction ranges simultaneously, while also considering the prediction uncertainty.
[0084] In particular, the machine learning model 402 is used to construct a predictive filter 604, such as a Kalman predictor, using standard procedures described in
[26] . The predictive filter 604 receives input 128 for a time value t(k) and obtains a predicted glucose measurement value from the machine learning model 402.
[0085]
number
[0086]
number
[0087] Next, prediction
[0088]
number
[0089]
number
[0090] 7 illustrates an example method 700 implemented by the predictive funnel evaluator 602. The method 700 calculates a predictive funnel for a time value t(k).
[0091]
number
[0092]
number
[0093] The method 700 may include, at step 702, deriving a confidence interval from the uncertainties received from the prediction filter 604. For example, for a given prediction
[0094]
number
[0095]
number
[0096]
number
[0097] The method 700, in step 704, determines a minimum number N of confidence intervals that are entirely above the hyperglycemic threshold (e.g., 180 mg / dl). pred This includes evaluating whether the pred is PH from 1 max For hypoglycemia alarms, experiments conducted by the inventors have shown that N pred It has been found that using a value of 1 for provides adequate results, however the benefits of the predictive funnel evaluator 602 may be achieved with other values as well. If the condition of step 704 is met, the method 700 includes predicting a hyperglycemic event and / or invoking the generation of a hyperglycemic alert 130 in step 706.
[0098] The method 700, in step 708, determines the number of confidence intervals that are completely below the hypoglycemic threshold (e.g., 70 mg / dl) to be N pred This may include evaluating whether the pred The value of may be the same or different for the evaluations of steps 704 and 708. If the condition of step 708 is met, the method 700 includes predicting a hypoglycemic event and / or invoking generation of a hypoglycemic alert 130 in step 710. If neither of the conditions of steps 704 and 708 are met, the method ends in step 712 without predicting a hypoglycemic or hyperglycemic event and invoking generation of an alert 130.
[0099] The decision support engine 114 may generate a hypoglycemic or hyperglycemic alert 130 by causing the display device 107 to provide a human-perceptible alert, such as a visual message, an audible alert, or a tactile alert (e.g., using a haptic device). The decision support engine 114 may alternatively or additionally generate instructions to another device, such as a fitness tracker, a smartwatch, or other wearable computing device, to generate a visual, audible, or tactile alert.
[0100] If a hypoglycemic or hyperglycemic event is predicted, in certain embodiments, additional or alternative decision support outputs may be provided to the patient. For example, the decision support output may include decision support recommendations for the patient to exercise, consume carbohydrates, administer insulin, etc. In certain embodiments, the decision support output may further include a signal to an insulin pump to modify the amount of insulin being or to be administered to the user, as described below with respect to FIG.
[0101] FIG. 8 is a plot that visualizes the prediction funnel. As is evident from FIG. 8, the funnel increases in height with the temporal distance from the current glucose measurement, i.e., g(k). This increase in height corresponds to an increase in the size of the confidence interval for the predicted glucose measurement with the temporal distance from g(k). Thus, the probability that the confidence interval will be entirely above the hyperglycemic threshold or entirely below the hypoglycemic threshold decreases with the temporal distance to the future time t(k+i) corresponding to a given prediction g(k+i|k). Thus, the prediction funnel has the advantage of evaluating multiple prediction ranges while at the same time taking into account increasing uncertainty and increasing prediction ranges.
[0102] Various modifications to the predictive funnel evaluator 602 may be made while still achieving at least some of the benefits described above.
[0103] First, various types and numbers of machine learning models may be used instead of the machine learning model 402 described above. For example, a non-linear physiological model may be used instead of the ARIMAX machine learning model. For some alternative machine learning model types, it may not be feasible to use the prediction filter 604 to obtain predictions more than one time step ahead. Instead, multiple machine learning models may be trained, each with a different prediction range and providing a corresponding confidence interval or uncertainty metric that can be used to obtain a confidence interval.
[0104] Some types of alternative machine learning models do not inherently provide confidence intervals or uncertainties that can be used to derive confidence intervals. Thus, confidence intervals can be calculated independently of one or more machine learning models by estimating the prediction error variance for the training set at each forecast horizon and using the estimated prediction error variance for each forecast horizon to obtain a confidence interval at each forecast horizon.
[0105] Some alternatives to using the ARIMAX machine learning model 402 with a Kalman filter as the predictive filter 604 include at least (a) a nonlinear physiological model of glucose-insulin dynamics or other nonlinear black-box model (e.g., a deep neural network (DNN), a convolution neural network (CNN), etc.) combined with (b) either an extended Kalman filter, an unscented Kalman filter, or a particle filter.
[0106] In another alternative, an ensemble approach is used in which multiple models of any of the above types are fitted to a training set and a distribution of prediction errors for the training set is calculated for each of the multiple models. The distribution of prediction errors for each model can then be used to weight the predictions of each model and calculate a new estimate for each model, for example via a linear combination. The new estimates of the models can then be combined to obtain an ensemble prediction by selecting the new estimate with the lowest prediction error, averaging the new estimates, or by another approach. The error variance of the ensemble prediction can be obtained from the training data set or can be derived from the distribution of prediction errors for each model. The predictions at different prediction horizons and the error variance of the ensemble can then be used to build a prediction funnel as described above.
[0107] Referring to FIG. 9 , the illustrated workflow 900 may be used to control delivery of insulin and / or performance of other remedial actions instead of, or in addition to, generating a human-perceivable output in response to an alert 130.
[0108] In workflow 900, the exogenous insulin I(k) delivered by insulin pump 902 is calculated by an automatic insulin delivery (AID) algorithm 904 based on inputs 906, which may include glucose measurements g(k), meal intake information CHO(k), and possibly other values 908 received from a user or from any of the sensors, such as any of the non-analyte sensors 206 described herein above. The AID algorithm 904 may be any AID algorithm known in the art.
[0109] The control signal from the AID algorithm 904 instructing the insulin pump 902 to provide a particular rate of insulin delivery may be received by an override module 910. The override module 910 refers to a set of software instructions that may be provided as part of the decision support engine 114. The override module 910 may pass the control signal from the AID algorithm 904 to the insulin pump 902, suppress the control signal entirely, or modify the control signal to change the insulin delivery rate relative to the insulin delivery rate instructed by the AID algorithm 904. The override module 910 may take as input the decision 606 output by the predictive funnel evaluator 602. The decision 606 may indicate whether hypoglycemic and hyperglycemic events have been predicted, as described above.
[0110] In a first implementation, in the event of a predicted hypoglycemic event, the override module 910 completely inhibits insulin delivery by the insulin pump 902 until the decision 606 no longer indicates a hypoglycemic event.
[0111] In a second implementation, the override module 910 receives the predictive funnel, i.e., the confidence interval for each glucose prediction, from the predictive filter 604 and adjusts the amount of insulin infusion ordered by the AID algorithm 904 according to the predictive funnel. For example, the override module 910 may reduce the insulin infusion by 50% if the predictive funnel is less than 100 mg / dl, reduce the insulin infusion by 75% if the funnel is less than 80 mg / dl, and reduce the insulin infusion by 100% if the predictive funnel is less than 70 mg / dl.
[0112] In a third implementation, the value of I(k) used to control the infusion by the insulin pump 902 is selected using the workflow 600. For example, let I(k1) be the insulin infusion rate selected by the AID algorithm 904. The override module 910 selects the set of insulin amounts I(k1)+δj (j=1 to P, where P is the number of insulin doses and δ j A new insulin infusion I(k2) may be selected by generating a new insulin infusion rate I(k2) based on the calculated insulin infusion rate I(k1) (where the value of g(k) is selected to provide a range of insulin infusions that includes I(k1). Multiple predictive funnels may be obtained according to the workflow 600, each predictive funnel including the g(k) and CHO(k) and the value I(k1)+δ j The value of I(k2) selected by the override module 910 to control insulin delivery by the insulin pump 902 is calculated using one of the values I(k1)+δ where the corresponding predictive funnel is completely above 70 mg / dl. j If none of the predicted funnels are completely above 70 mg / dl, the override module 910 may output instructions on the display device 107 to consume a predetermined amount of carbohydrates, for example, 10, 15, or 20 grams.
[0113] The second or third implementation may be substituted for or used in conjunction with providing a decision support recommendation for output by the display device 107. For example, in the case of a predicted hypoglycemic event, the override module 910 instructs the display device 107 to output a decision support recommendation to consume a predetermined amount of carbohydrates, for example 15 grams. The prescribed amount of carbohydrates may be fixed or determined based on a predicted funnel, for example 10 grams if the predicted funnel is partially below 70 mg / dl but all above 50 mg / dl, 15 grams if the funnel is entirely below 70 mg / dl and partially below 50 mg / dl, and 20 grams if the funnel is entirely below 50 mg / dl.
[0114] The prescribed amount of carbohydrates may also be determined using the workflow 600. For example, multiple predictive funnels may be obtained according to the workflow 600, with each predictive funnel calculated using g(k) and I(k) and one of multiple alternative values for CHO(k), such as a range of alternative values that includes the current value of CHO(k). The prescribed amount of carbohydrates selected by the override module 910 may be the amount when the corresponding predictive funnel is generally above 70 mg / dl.
[0115] Test results The study results were obtained using improved training of the prediction model described with respect to Figures 4-5C and the alarm generation model described with respect to Figures 6-8. The study was conducted using data collected in a multicenter clinical trial (NCT02137512) aimed at evaluating the long-term use of a hybrid closed-loop insulin delivery system (artificial pancreas) developed at the University of Virginia
[32] . The study and all experimental procedures were approved by the local IRB / ethics committee. Fourteen (type 1 diabetes) T1D individuals participated in a 5-month main phase testing the 24 / 7 use of the artificial pancreas system. Glucose data were recorded using DexCom G4® sensors (DexCom, Inc., San Diego, CA, USA) with a 5-minute sample time, and insulin was delivered with a Roche Accu-Check Spirit Combo® insulin pump (Roche Diabetes Care, Inc., Indianapolis, IN, USA). These two off-the-shelf medical devices communicated using Bluetooth with the main system component, Diabetes Assistant (DiAs)
[33] , a platform based on Android smartphones in charge of managing the data flow, providing a friendly user interface and hosting the control algorithms. Individuals were instructed to manually deliver the appropriate amount of insulin for every meal by inserting all carbohydrate intakes into the system. Further details of the experiment are described in
[32] .
[0116] The test data was pre-processed to handle the different nature of experimental data (CGM (Continuous Glucose Monitor), insulin information, and carbohydrate intake), which raises several technical issues related to device synchronization, the integrity of the stored data, and the reliability of the information provided by the patient. First, to solve the synchronization issue, all signals were synchronized with the T S The data were aligned to the same time grid, equally sampled at 0.5 min = 5 min. Secondly, a part of the data of 14 consecutive days, including only incomplete data parts lasting less than 30 min, was selected for the purpose of this study, and then divided into a training set (first 7 days) and a test set (last 7 days). Finally, for missing information, the incomplete data parts lasting less than 30 min were filled differently depending on whether the gap occurred in the training or in the test data, as described below. According to this data pre-processing step, three subjects were excluded from the analysis because they lacked sufficient acceptable data. Finally, for missing data, in the training set, cubic spline interpolation was used to fill the remaining missing samples contained in the time series. In fact, the training set is available globally during model training, and non-causal techniques can be used. Instead, for the test set, glucose prediction should be performed only in real time. For this reason, the gaps were closed using a zero-order hold applicable in real time, using only past samples.
[0117] Dealing with experimental data of different nature (CGM, insulin information, and carbohydrate intake) raises several technical issues related to device synchronization, completeness of stored data, and reliability of patient-provided information. First, to solve the synchronization problem, all signals were aligned to the same time grid equally sampled with TS=5 min. Second, a part of the data of 14 consecutive days, which is suitable for the purpose of this experiment, containing only incomplete data parts lasting less than 30 min, was selected and then divided into a training set (first 7 days) and a test set (last 7 days). Finally, for missing information, the incomplete data parts lasting less than 30 min were filled in differently depending on whether the gap occurred in the training or in the test data, as described below. According to this data pre-processing step, three subjects were excluded from the analysis because sufficient acceptable data could not be found. Finally, for missing data, in the training set, cubic spline interpolation was used to fill the remaining missing samples in the time series. In fact, the training set is globally available during model training, and non-causal techniques can be used. Instead, for the test set, glucose predictions can only be performed in real time, for this reason we closed the gap using a zero-order hold applicable in real time, using only past samples.
[0118] According to the definition proposed in a consensus paper by a group of experts in the field
[39] , a hypoglycemic episode (HE) occurs at time t(k) if BG falls below 70 mg / dL for at least 15 minutes. The episode lasts until BG remains above this threshold for at least 15 minutes. A true positive (TP) occurs if HE occurs at time t(k) and an alarm is generated before the event, precisely within a window from 60 minutes to 5 minutes before t(k). According to this definition, only alarms relatively close to HE are considered correct, alarms that are too far in the past are not counted as TP. A false positive (FP) occurs when an alarm is generated at time t(k) but no hypoglycemia occurs in the following 60 minutes. A false negative occurs when HE occurs at time t(k) but no alarm is generated in the preceding 60 minutes. A late alarm is defined as an alarm until time t(k) or 15 minutes later. Since they are not timely, they are not counted as TP. Conversely, late alarms increment the FN count. Nevertheless, late alarms do not increment the FP count because they cannot be considered as errors.
[0119] Once the events have been labeled as TP, FP, and FN, we evaluate the state-of-the-art and the proposed approach using the following metrics:
[0120]
number
[0121] Precision is the proportion of correct alarms over the total number of alarms raised. Recall, also known as sensitivity, is the proportion of correctly detected hypoglycemic events over the total number of events. The F1 score is the harmonic mean of the two aforementioned metrics. As the dataset is highly unbalanced, we also evaluated the average number of FPs generated by the algorithm per day (FP / day). Finally, the results include the time gain (TG) of the hypoglycemic alarm as the time between when the alarm is raised by the algorithm and the onset of the HE. According to the definition of TP, the maximum achievable TG is 60 minutes and the minimum is 5 minutes. Due to the limited number of hypoglycemic events, the values of the hypoglycemic prediction metrics were obtained by considering all hypoglycemic events of the different subjects as they belong to a unique time series. The results are then expressed as a single value for all considered metrics except TG, which is expressed as the mean and standard deviation (SD) of the time gain of all detections.
[0122] Table 1 shows the hypoglycemia prediction performance of several predictive algorithms. The first row of the table reports the baseline performance achieved by the state-of-the-art algorithm (single PH, MSE), considering only one prediction horizon (PH=30 min), using an individualized model, minimizing MSE identified. The last row of the table reports the performance achieved by the predictive models proposed herein, including various aspects of the model (e.g., the use of gMSE for model identification and a predictive funnel-based strategy).
[0123] [Table 1]
[0124] To clarify the contribution of each aspect of the prediction model to the final performance, Table 1 also reports the performance achieved including one modification in time (gMSE + single PH and MSE + prediction funnel). Furthermore, different values of the height parameter are investigated. We first focus on the single PH strategy and investigate the impact of PH on the prediction performance of the state-of-the-art methods by evaluating three possible PHs: PH = 30, 45, and 60 min. The best results are achieved with PH = 30 min, which is actually commonly adopted in the literature. In particular, the larger the PH, the higher the TG, but the worse the P, R, and FP / day. For example, comparing the state-of-the-art methods at PH = 30 min and PH = 60 min, we see that TG increases from 15 to 20 min (average values), but P drops from 43% to 36%, R from 95% to 76%, and FP / day increases from 0.77 to 0.82. By introducing the use of gMSE instead of MSE, both precision and recall are improved with respect to the state-of-the-art methods. This improvement applies for all considered values of PH. For example, considering PH=30 min, the use of gMSE increases P from 43% to 44% and R from 95% to 100%, while FP / day and TG are almost the same.
[0125] The impact of an improved alarm strategy (predictive funnel-based instead of using a single PH) was also investigated. In particular, two different approaches to tuning the parameter m were considered. In both cases, patient-specific values of m were considered, but in the first approach, m was set to obtain a recall similar to that achieved by the state-of-the-art (MSE+single PH, PH=30 min). As a second alternative approach, m was chosen in each patient to maximize F1 in that patient's training set. The first approach is denoted in Table 1 as Tuning 1 and the second approach as Tuning 2.
[0126] The improvements achieved by the predictive funnel can be clearly seen in Tuning 1, which provides higher precision, higher F1 and fewer FP / day over the state of the art. P increases from 43% to 51%, F1 increases from 59% to 65% and FP / day decreases from 0.77 to 0.59. This improvement is achieved while retaining a similar recall (R from 95% to 91%). The performance of Tuning 2 is more difficult to interpret, as it chooses a different trade-off between precision and recall. Specifically, it gives up some recall in favor of better precision and fewer FP / day, resulting in an overall improvement in the F1 score.
[0127] Finally, combining gMSE with the use of predictive funnels outperforms the state-of-the-art. Here again, this can be clearly seen using Tuning 1, which achieves similar recall and TG as the state-of-the-art (above 95%), but with higher precision, F1 score, and fewer false positives. P goes from 43% to 51%, F1 goes from 59% to 69%, and false positives per day go from 0.77 to 0.59. By adopting a slightly more conservative approach proposed in Tuning 2, precision and false positives can be further improved (P=65%, FP / day=0.29, i.e., about once every 3 days), at the expense of a lower recall (R=88%). This new trade-off provides a better F1 score that reaches 75%.
[0128] The preceding description provides a new approach for predicting hypoglycemic events based on an individualized linear black-box model identified by using an ad-hoc cost function designed to take into account the clinical impact of prediction errors and an improved alarm strategy that considers the entire prediction funnel. Results show that the model identified via gMSE minimization provides better hypoglycemia prediction performance than the MSE-based model. Furthermore, results show that the new alarm generation model based on the prediction funnel improves hypoglycemia detection thanks to the possibility of utilizing multiple PHs. The adoption of both proposed improvements gives the best performance.
[0129] It is important to note that the definition of a hypoglycemic episode (HE) used to evaluate the predictive performance was formulated by a panel of international experts in a consensus paper
[39] . HE occurs when BG is below 70 mg / dL for at least 15 min. The same episode continues until BG remains above this threshold for at least 15 min. According to this definition, a hypoglycemic episode does not exist if BG is below the hypoglycemic threshold for only one or two time samples. If an alarm is issued in these cases, hereafter called a "quasi-hypoglycemic episode" (qHE), the alarm is counted as an FP.
[0130] Since FP errors caused by qHE may be considered less clinically relevant than other FPs, it is interesting to investigate how many false positives are of this kind. While in the state-of-the-art method (MSE + single PH method, PH = 30 min), 26% of the recorded FPs were associated with qHE, this percentage rises to 34% in the newly proposed algorithm (gMSE + predictive funnel), further supporting the superiority of the proposed algorithm. By discarding these events from the FP count, as is frequently done in the literature, the state-of-the-art method reduces FP / day from 0.77 to 0.62 and increases precision from 43% to 54%, whereas the proposed method reduces FP / day from 0.29 to 0.21 and increases P from 65% to 73%.
[0131] Another consequence of the adopted definition of HE is that even a non-predictive hypoglycemia detector based simply on the BG trace crossing the 70 mg / dL threshold may give a false positive alarm (FP / day=0.2). According to our definition of TP, BG generates only late alarms (an event is accurately detected whenever it starts at TG=0 min). As a consequence, TP counts are constrained to 0, and both recall and precision are necessarily 0 (i.e., P=R=0).
[0132] The study results show the benefit of including both model training using gMSE and the use of a prediction funnel to determine the appropriateness of generating an alert when performing personalized hypoglycemia prediction. The study results show an improvement over previous contributions in the field
[10] ,
[14] ,
[16] ,
[17] ,
[19] ,
[20] ,
[22] -
[24] . These previous contributions should be interpreted with caution. Different definitions of events can significantly affect the final metrics (FP / day, P, R, F1, TG), as discussed above for qHE, and seemingly minor modifications in the definition of HE have a non-negligible impact on FP / day and accuracy. Furthermore, different validation datasets may be collected in very different conditions (highly controlled clinical trials vs. real life) introducing further confounding factors.
[0133] Taking this caveat into account, the obtained results are comparable to most of the other literature studies (based on both linear and nonlinear regression). The authors of
[10] ,
[19] ,
[20] reached recall rates of about 93%, 93%, and 86%, respectively, which are comparable or slightly better than the recall rate of the approach described herein with R = 88%, at the expense of lower precision (about 24%, 38%, and 62%), while the approach described herein achieved P = 65%. Similarly,
[14] ,
[24] achieved a remarkable recall rate of R = 100%, but at the expense of a very large number of FPs (more than 1 FP per day). The authors of
[16] ,
[17] ,
[23] showed similar recall rates (89%, 94%, and 94%) to those obtained using the approach described herein, with better precision (78%, 90%, and 77%). However, the authors adopted a more permissive HE definition. For example,
[23] considered as TP are also alarms issued after BG passes the hypoglycemic threshold, whereas we consider them as FN. In
[16] , performance was evaluated using controlled inpatient data. The authors of
[22] showed a slightly poorer recall (86%) and did not report any metrics related to false alarms.
[0134] 10 is a block diagram illustrating a computing device 1000 configured to predict hypoglycemic and hyperglycemic events, according to certain embodiments disclosed herein. Although shown as a single physical device, in embodiments, the computing device 1000 may be implemented using a virtual device and / or across several devices, such as in a cloud environment. As shown, the computing device 1000 comprises a processor 1005, a memory 1010, a storage 1015, a network interface 1025, and one or more I / O interfaces 1020. In the embodiment shown, the processor 1005 retrieves and executes programming instructions stored in the memory 1010, as well as stores and retrieves data present in the storage 1015. The processor 1005 generally represents a single CPU and / or GPU, multiple CPUs and / or GPUs, a single CPU and / or GPU with multiple processing cores, etc. The memory 1010 is included to generally represent random access memory (RAM). Storage 1015 may be any combination of disk drives, flash-based storage devices, and the like, and may include fixed and / or removable storage devices, such as fixed disk drives, removable memory cards, cache, optical storage, network attached storage (NAS), or a storage area network (SAN).
[0135] In some embodiments, input and output (I / O) devices 1035 (keyboard, monitor, etc.) can be connected via I / O interface 1020. Additionally, via network interface 1025, computing device 1000 can be communicatively coupled to one or more other devices and components, such as user database 110 and / or historical record database 112. In particular embodiments, computing device 1000 is communicatively coupled to other devices via a network, which may include the Internet, a local network, etc. The network may include wired connections, wireless connections, or a combination of wired and wireless connections. As shown, processor 1005, memory 1010, storage 1015, network interface 1025, and I / O interface 1020 are communicatively coupled by one or more interconnects 1030. In particular embodiments, computing device 1000 represents a display device 107 associated with a user. In particular embodiments, as discussed above, display device 107 may include a user's laptop, computer, smartphone, etc. In a particular embodiment, computing device 1000 is a server running in a cloud environment.
[0136] In the illustrated embodiment, the storage 1015 includes a user profile 118. The memory 1010 includes a decision support engine 114, which itself includes a prediction module 116. The decision support engine 114 is executed by the computing device 1000 to perform operations of the workflow 400 of FIG. 4, the workflow 600 of FIG. 6, the method 700 of FIG. 7, and / or the workflow 900 of FIG. 9.
[0137] Additional Considerations The methods disclosed herein include one or more steps or actions for achieving the method. These method steps and / or actions may be interchanged with one another without departing from the scope of the claims. In other words, unless a specific order of steps or actions is specifically specified, the order and / or use of specific steps and / or actions may be modified without departing from the scope of the claims.
[0138] As used herein, a phrase referring to "at least one of" a list of items refers to any combination of those items, including single members. By way of example, "at least one of a, b, or c" is intended to cover a, b, c, ab, ac, bc, and abc, as well as any combination of multiples of the same element (e.g., aa, aaa, aab, aac, abb, acc, bb, bbb, bbc, cc, and ccc, or any other permutation of a, b, and c).
[0139] The foregoing description is provided to enable those skilled in the art to practice the various embodiments described herein. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other embodiments. Thus, the claims are not intended to be limited to the embodiments shown herein, but are to be given the full scope consistent with the claim language, in which reference to an element in the singular is not intended to mean "one and only one" unless specifically stated otherwise, but conversely is intended to mean "one or more". The term "some" refers to one or more, unless specifically stated otherwise. All structural and functional equivalents to the elements of the various embodiments described throughout this disclosure that are now known or that will become known in the future to those skilled in the art are expressly incorporated herein by reference and are intended to be encompassed within the scope of the claims. Furthermore, nothing disclosed herein is intended to be dedicated to the public, regardless of whether such disclosure is expressly recited in the claims. No claim element is to be construed under the provisions of 35 U.S.C. 112(f) unless the element is expressly recited using the phrase "means for," or, in the case of a method claim, the element is recited using the phrase "step for."
[0140] While various examples of the present invention have been described above, it should be understood that the examples are presented by way of example only and not by way of limitation. Similarly, various diagrams may depict example architectural or other configurations for the present disclosure, which configurations are implemented to aid in understanding features and functionality that may be included in the present disclosure. The present disclosure is not limited to the illustrated example architecture or configurations, but may be implemented using various alternative architectures and configurations. Furthermore, while the present disclosure has been described above with respect to various illustrative examples and aspects, it should be understood that various features and functions described in one or more of the individual examples are not limited in their applicability to the particular example for which they are described. On the contrary, those features and functions may be applied, alone or in any combination, to one or more of the other examples of the present disclosure, whether or not such an example is described, and whether or not such features are presented as part of a described example. Thus, the breadth and scope of the present disclosure should not be limited by any of the illustrative examples described above.
[0141] All references cited herein are incorporated herein by reference in their entirety. To the extent that publications and patents or patent applications incorporated by reference conflict with the present disclosure contained herein, the present specification is intended to supersede and / or take precedence over any such conflicting material.
[0142] Unless otherwise defined, all terms (including technical and scientific terms) are to be given their ordinary and customary meanings to those of ordinary skill in the art and are not to be limited to any special or customized meaning unless expressly defined as such herein.
[0143] Terms and phrases used in this application, and variations thereof, unless expressly stated otherwise, particularly in the appended claims, are to be construed as open ended, as opposed to limiting. As an example of the foregoing, the term "including" should be read to mean "including without limitation," "including, but not limited to," and the like. The term "comprising" as used herein is synonymous with "including," "containing," or "characterized by," and is inclusive or open ended and does not exclude additional, unrecited elements or method steps. The term "having" should be interpreted as "having at least." The term "including" should be interpreted as "including, but not limited to." The term "examples" is used to provide illustrative examples of items in the description, not an exhaustive or exclusive listing of such examples. Adjectives such as "known," "usual," "standard," and terms of similar import should not be construed as limiting the items being described to a given period of time or to items available at a given time, but rather should be read to embrace known, conventional, or standard technology that may be available or known at any time now or in the future, and words such as "preferably," "preferred," "desired," or "desirable," and similar imports should not be understood to imply that a particular feature is critical, essential, or even essential to the structure or function of the invention, but rather are merely intended to highlight alternative or additional features that may or may not be utilized in a particular example of the invention. Similarly, a group of items connected with the conjunction "and" should not be read as requiring that any and all of those items be present in the group, but rather should be read as "and / or" unless expressly stated otherwise.Similarly, a group of items joined with the conjunction "or" should not be read as requiring mutual exclusivity among the groups, but rather as "and / or" unless expressly stated otherwise.
[0144] As used herein, "comprising" is synonymous with "including," "containing," or "characterized by" and is inclusive or open-ended and does not exclude additional or unrecited elements or method steps.
[0145] All numerical values expressing amounts of raw materials, reaction conditions, and the like used herein should be understood to be modified in all instances by the term "about." Accordingly, unless otherwise indicated, the numerical parameters set forth herein are approximations that may vary depending upon the desired properties sought to be obtained. At the very least, and not as an attempt to limit the application of the doctrine of equivalents to the scope of any claims in any application claiming priority to this application, each numerical parameter should be construed in light of the number of significant digits and ordinary rounding approaches.
[0146] Moreover, although the foregoing has been described in some detail using illustrations and examples for purposes of clarity and understanding, it will be apparent to those skilled in the art that certain changes and modifications may be practiced. Thus, the descriptions and examples should not be construed as limiting the scope of the invention to the specific examples and embodiments described herein, but rather as covering all modifications and alternatives that come within the true scope and spirit of the invention.
[0147] References The following documents are incorporated herein by reference in their entirety: [1]M.Vettoretti,G.Cappon,G.Acciaroli,A.Facchinetti,and G.Sparacino,「Continuous Glucose Monitoring:Current Use in Diabetes Management and Possible Future Applications,」Journal of Diabetes Science and Technology,vol.12,no.5,pp.1064-1071,2018。 [2]B.McAdams and A.Rizvi,「An Overview of Insulin Pumps and Glucose Sensors for the Generalist,」Journal of Clinical Medicine,vol.5,no.1,p.5,2016。 [3]I.Contreras and J.Vehi,「Artificial intelligence for diabetes management and decision support:Literature review,」Journal of Medical Internet Research,vol.20,no.5,2018。 [4]C.Luca,C.Giacomo,S.Giovanni,and F.Andrea,「A new integrated platform for gathering and managing multivariable and multisensor data in diabetes clinical studies,」in 2020 International Conference on e-Health and Bioengineering(EHB),2020,pp.1-4。 [5]K.Turksoy,J.Kilkus,I.Hajizadeh,S.Samadi,J.Feng,M.Sevil, C.Lazaro,N.Frantz,E.Littlejohn,and A.Cinar,「Hypoglycemia Detec-tion and Carbohydrate Suggestion in an Artificial Pancreas,」Journal of Diabetes Science and Technology,vol.10,no.6,pp.1236-1244,2016。 [6]N.Camerlingo,M.Vettoretti,S.Del Favero,G.Cappon,G.Sparacino,and A.Facchinetti,「A real-time continuous glucose monitoring-based algorithm to trigger hypotreatments to prevent / mitigate hypoglycemic events,」Diabetes Technology and Therapeutics,vol.21,no.11,pp.644-655,2019。 [7]J.P.Shivers,L.Mackowiak,H.Anhalt,and H.Zisser,「」Turn It Off!」:Diabetes Device Alarm Fatigue Considerations for the Present and the Future,」Journal of Diabetes Science and Technology,vol.7,no.3,pp.789-794,2013。 [8]G.McGarraugh,「Alarm characterization for continuous glucose monitors used as adjuncts to self-monitoring of blood glucose,」Journal of Diabetes Science and Technology,vol.4,no.1,pp.41-48,2010。 [9]X.Mo,Y.Wang,and X.Wu,「Hypoglycemia prediction using extreme learning machine(ELM)and regularized ELM,」2013 25th Chinese Control and Decision Conference,CCDC 2013,pp.4405-4409,2013。
[10] K.Turksoy,E.S.Bayrak,L.Quinn,E.Littlejohn,D.Rollins,and A.Cinar,「Hypoglycemia early alarm systems based on multivariable models,」Industrial and Engineering Chemistry Research,vol.52,no.35,pp.12 329-12 336,2013。
[11] S.Oviedo,J.Veh’ l ,R.Calm,and J.Armengol,「A review of personalized blood glucose prediction strategies for T1DM patients,」International Journal for Numerical Methods in Biomedical Engineering,vol.33,no.6,pp.1-21,2017。
[12]
[12] O.Mujahid,I.Contreras,and J.Vehi,「Machine learning techniques for hypoglycemia prediction:Trends and challenges,」Sensors(Switzerland),vol.21,no.2,pp.1-21,2021。
[13] H.Efendic,H.Kirchsteiger,G.Freckmann,and L.Del Re,「Short-term prediction of blood glucose concentration using interval probabilistic models,」2014 22nd Mediterranean Conference on Control and Automation,MED 2014,pp.1494-1499,2014。
[14] J.Yang,L.Li,Y.Shi,and X.Xie,「An ARIMA Model with Adaptive Orders for Predicting Blood Glucose Concentrations and Hypoglycemia,」IEEE Journal of Biomedical and Health Informatics,vol.23,no.3,pp.1251-1260,2019。
[15] M.Eren-Oruklu,A.Cinar,D.K.Rollins,and L.Quinn,「Adaptive system identification for estimating future glucose concentrations and hypoglycemia alarms,」Automatica,vol.48,no.8,pp.1892-1897,2012。
[16] M.Eren-Oruklu,A.Cinar,and L.Quinn,「Hypoglycemia prediction with subject-specific recursive time-series models,」Journal of Diabetes Science and Technology,vol.4,no.1,pp.25-33,2010。
[17] C.Toffanin,S.Del Favero,E.M.Aiello,M.Messori,C.Cobelli,and L.Magni,「Glucose-insulin model identified in free-living conditions for hypoglycaemia prevention,」Journal of Process Control,vol.64,pp.27-36,2018。
[18] J.Veh’ l ,I.Contreras,S.Oviedo,L.Biagi,and A.Bertachi,「Prediction and prevention of hypoglycaemic events in type-1 diabetic patients using machine learning,」Health informatics Journal,vol.26,no.1,pp.703-718,2020。
[19] D.Dave,D.J.DeSalvo,B.Haridas,S.McKay,A.Shenoy,C.J.Koh,M.Lawley,and M.Erraguntla,「Feature-Based Machine Learning Model for Real-Time Hypoglycemia Prediction,」Journal of Diabetes Science and Technology,2020。
[20] M.Gadaleta,A.Facchinetti,E.Grisan,and M.Rossi,「Prediction of Adverse Glycemic Events From Continuous Glucose Monitoring Signal,」IEEE Journal of Biomedical and Health Informatics,vol.23,no.2,pp.650-659,2019。
[21] M.Frandes,B.Timar,and D.Lungeanu,「A risk based neural network approach for predictive modeling of blood glucose dynamics,」Studies in Health Technology and Informatics,vol.228,pp.577-581,2017。
[22] K.S.Eljil,G.Qadah,and M.Pasquier,「Predicting hypoglycemia in diabetic patients using data mining techniques,」2013 9th International Conference on Innovations in Information Technology, IIT 2013,pp.130-135,2013。
[23] E.I.Georga,V.C.Protopappas,D.Ardig`o,D.Polyzos,and D.I.Fotiadis,「A glucose model based on support vector regression for the prediction of hypoglycemic events under free-living conditions,」Diabetes Technology and Therapeutics,vol.15,no.8,pp.634-643,2013。
[24] E.Daskalaki,K.Norgaard,T.Z¨uger,A.Prountzou,P.Diem,and S.Mougiakakou,「An early warning system for hypo- glycemic / hyperglycemic events based on fusion of adaptive prediction models,」Journal of Diabetes Science and Technology,vol.7,no.3,pp.689-698,2013。
[25] L.Ljung,System identification-Theory for the user.PTR Prentice Hall,1987。
[26] P.J.Kim and L.Huh,Kalman Filter for Beginners:with MATLAB Examples.MathWorks,2011。
[27] C.Dalla Man,R.A.Rizza,and C.Cobelli,「Meal simulation model of the glucose-insulin system,」IEEE Transactions on Biomedical Engineering,vol.54,no.10,pp.1740-1749,2007。
[28] D.A.Finan,H.Zisser,L.Jovanovic,W.C.Bevier,and D.E.Seborg,「Identification of Linear Dynamic Models for Type 1 Diabetes:a Simulation Study,」in IFAC Proceedings Volumes,vol.39,no.2.IFAC,2006,pp.503-508。
[29] M.Messori,M.Ellis,C.Cobelli,P.D.Christofides,and L.Magni,「Improved postprandial glucose control with a customized Model Predictive Controller,」Proceedings of the American Control Conference,pp.5108-5115,2015。
[30] F.Prendin,S.Del Favero,M.Vettoretti,G.Sparacino,and A.Facchinetti,「Forecasting of glucose levels and hypoglycemic events:Head-to-head comparison of linear and nonlinear data-driven algorithms based on continuous glucose monitoring data only,」Sensors,vol.21,no.5,2021。
[31] S.Faccioli,A.Facchinetti,G.Sparacino,G.Pillonetto,and S.Del Favero,「Linear Model Identification for Personalized Prediction and Control in Diabetes,」IEEE Trans Biomed Eng.Online ahead of print,2021,pMID:34347589。
[32] S.Del Favero,A.Facchinetti,and C.Cobelli,「A Glucose-Specific Metric to Assess Predictors and Identify Models,」Journal of Diabetes Science and Technology,vol.6(2),p.A30,2012。
[33] B.Kovatchev,P.Cheng,S.M.Anderson,J.E.Pinsker,F.Boscari,B.A.Buckingham,F.J.Doyle,K.K.Hood,S.A.Brown,M.D.Breton,D.Chernavvsky,W.C.Bevier,P.K.Bradley,D.Bruttomesso,S.Del Favero,R.Calore,C.Cobelli,A.Avogaro,T.T.Ly,S.Shanmugham,E.Dassau,C.Kollman,J.W.Lum,and R.W.Beck,「Feasibility of Long-Term Closed-Loop Control:A Multicenter 6-Month Trial of 24 / 7 Automated Insulin Delivery,」Diabetes Technology and Therapeutics,vol.19,no.1,pp.18-24,2017。
[34] P.Keith-Hynes,B.Mize,A.Robert,and J.Place,「The diabetes assistant: A smartphone-based system for real-time control of blood glucose,」Electronics (Switzerland),vol.3,no.4,pp.609-623,2014。
[35] S.Bittanti,Model Identification and Data Analysis.John Wiley & Sons,Inc.,2019。
[36] W.L.Clarke,D.Cox,L.A.Gonder-Frederick,W.Carter,and S.L.Pohl,「Evaluating clinical accuracy of systems for self-monitoring of blood glucose,」Diabetes Care,vol.10,no.5,pp.622-628,1987。
[37] C.P’erez-Gand’ la,A.Facchinetti,G.Sparacino,C.Cobelli,E.G’omez,M.Rigla,A.de Leiva,and M.Hernando,「Artificial neural network algorithm for online glucose prediction from continuous glucose monitoring,」Diabetes technology & therapeutics,vol.12,no.1,pp.81-88,2010。
[38] J.Daniels,P.Herrero,and P.Georgiou,「A Multitask Learning Approach to Personalised Blood Glucose Prediction,」IEEE J Biomed Health Inform,2021。
[39] T.Danne,R.Nimri,T.Battelino,R.M.Bergenstal,K.L.Close,J.H.DeVries,S.Garg,L.Heinemann,I.Hirsch,S.A.Amiel,R.Beck,E.Bosi,B.Buckingham,C.Cobelli,E.Dassau,F.J.Doyle,S.Heller,R.Hovorka,W.Jia,T.Jones,O.Kordonouri,B.Kovatchev,A.Kowalski,L.Laffel,D.Maahs,H.R.Murphy,K.Norgaard,C.G.Parkin,E.Renard,B.Saboo,M.Scharf,W.V.Tamborlane,S.A.Weinzimer,and M.Phillip,「International consensus on use of continuous glucose monitoring,」Diabetes Care,vol.40,no.12,pp.1631-1640,2017。
[40] GFFranklin, JDPowell, MLWorkman et al., Digital control of dynamic systems. Addison-Wesley Reading, MA, 1998, vol.3. [Explanation of symbols]
[0148] 100 Decision Support Systems 102 users 104 Continuous Analyte Monitoring Systems 104 Analyte Monitoring Systems 104 Continuous Analyte Sensor System 106 Applications 106 Analyte Sensor Applications 107 Display Devices 110 User Database 112 History Record Database 114 Decision Support Engine 116 Prediction Module 118 User Profile 120 Demographic Information 122 Disease progression information 124 Medical Information 126 Application Data 128 inputs 130 Alerts 130 Hyperglycemia Alert 130 Hypoglycemia Alert 140 Training Server System
Claims
1. 1. A method for managing a user's blood glucose level, the method comprising: receiving, by a computing device, a glucose measurement of the user from a sensor; receiving, by the computing device, one or more values for one or more additional inputs related to the blood glucose level of the user; processing, by the computing device, the glucose measurements and one or more values for the one or more additional inputs to obtain a plurality of predicted glucose values, each predicted glucose value of the plurality of predicted glucose values corresponding to a different prediction range and having a corresponding uncertainty; generating, by the computing device, a confidence interval for each predicted glucose value of the plurality of predicted glucose values based on each predicted glucose value and the corresponding uncertainty; generating, by the computing device, a decision support output in response to determining that the confidence interval satisfies a threshold condition.
2. The method of claim 1 , wherein the one or more additional inputs include dietary intake information.
3. The method of claim 1 , wherein the one or more additional inputs include an amount of exogenous insulin infused to the user.
4. The method of claim 1 , wherein the one or more additional inputs include meal intake information and an amount of exogenous insulin administered to the user.
5. The method of claim 1 , wherein the glucose measurements and one or more values are processed using one or more machine learning models.
6. The method of claim 5 , wherein the one or more machine learning models include a predictive filter.
7. The method of claim 6 , wherein the predictive filter is a Kalman filter.
8. 8. The method of claim 7, wherein the Kalman filter is based on a predictive machine learning model trained to use a glucose-specific mean squared error (gMSE) loss function.
9. 10. The method of claim 8, wherein the predictive machine learning model is an Autoregressive Integrated Moving Average with Exogenous Inputs (ARIMAX) machine learning model.
10. 2. The method of claim 1, wherein the threshold condition is that a minimum number of the confidence intervals are either (a) entirely above a hyperglycemic threshold, or (b) entirely below a hypoglycemic threshold.
11. 11. The method of claim 10, wherein for each predicted glucose value, generating the confidence interval based on the each predicted glucose value and the corresponding uncertainty of the each predicted glucose value comprises scaling the uncertainty.
12. The method of claim 1 , wherein the decision support output comprises a human perceptible alert.
13. The method of claim 1 , wherein the decision support output comprises a signal sent to an insulin delivery device to cause the insulin delivery device to alter an amount of insulin infused to the user.
14. The method of claim 1 , wherein the decision support output comprises human perceptible instructions for consuming carbohydrates.
15. 1. A glucose monitoring system for managing a user's blood glucose level, the glucose monitoring system comprising: a glucose sensor system configured to generate one or more glucose measurements for the user; one or more processing devices; and one or more memory devices coupled to the one or more processing devices, the one or more memory devices, when executed by the one or more processing devices, providing the one or more processing devices with: receiving glucose measurements g(k) of the one or more measurements from the glucose sensor system over a time t(k); receiving meal intake information CHO(k) and an amount of insulin injected I(k) for said user; g(k), CHO(k), and I(k) are processed to obtain a number of predicted glucose values. [0019] from [0020] (Here, pH max is an integer greater than 1) and multiple uncertainties σ 2 (k+1|k) to σ 2 (k+PH max |k); and Multiple confidence intervals ##EQU00021## (i=1 to PH max generating a predictive funnel by calculating generating a decision support output in response to determining that the confidence interval satisfies a threshold condition.
16. The glucose monitoring system of claim 15 , wherein the glucose sensor system is a continuous glucose monitor.
17. 16. The glucose monitoring system of claim 15, wherein the g(k), CHO(k), and I(k) are processed using one or more machine learning models including a Kalman filter.
18. 20. The glucose monitoring system of claim 17, wherein the Kalman filter is constructed based on a predictive machine learning model trained using a glucose-specific mean squared error (gMSE) loss function.
19. 16. The glucose monitoring system of claim 15, wherein the threshold condition is that a minimum number of the confidence intervals are either (a) entirely above a hyperglycemic threshold, or (b) entirely below a hypoglycemic threshold.
20. the decision support output being: Human perceptible alerts, instructions to modify the amount of insulin injected into the user; or 16. The glucose monitoring system of claim 15, including at least one of human perceptible instructions for consuming carbohydrates.
