Hit event identification system
The LSTM-AE system addresses the challenge of diagnosing HIT after cardiac surgery by learning normal platelet patterns and detecting deviations, enhancing diagnostic accuracy and timeliness through improved recall, precision, and F1-Score.
Patent Information
- Application Number
- PCT/CA2026/050663
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-05-06
- Filing Date
- 2026-05-06
- Publication Date
- 2026-11-12
AI Technical Summary
Current computing systems struggle to accurately diagnose heparin-induced thrombocytopenia (HIT) after cardiac surgery due to confusion with postoperative thrombocytopenia, leading to delayed or missed diagnoses, with existing scoring systems like the 4T score, HEP score, and Lillo-Le Louet score performing modestly in this setting.
A Long Short-Term Memory Auto Encoder (LSTM-AE) based architecture is developed to identify HIT events as anomalies in platelet count time series data by learning normal patterns and detecting deviations, using preprocessing techniques like interpolation and standardization, and generating alerts when reconstruction errors exceed a predefined threshold.
The LSTM-AE system improves the timeliness and accuracy of HIT diagnosis by providing better recall, precision, and F1-Score compared to statistical methods, enabling early identification and alerting clinicians to potential HIT cases.
Smart Images

Figure CA2026050663_12112026_PF_FP_ABST
Abstract
Description
HIT EVENT IDENTIFICATION SYSTEMCROSS-REFERENCE TO RELATED APPLICATION
[0001] This patent application claims priority to United States Provisional Patent Application No. 63 / 800,701, filed on May 6, 2025, and titled “HIT EVENT IDENTIFICATION SYSTEM”, the entire contents of which are hereby incorporated by reference.FIELD
[0002] The present disclosure relates to the identification of Heparin-Induced Thrombocytopenia (HIT) events, and in particular, to Long Short-Term Memory Auto Encoder (LSTM-AE) -based architecture for the identification of HIT events as anomalies in platelet count time series data from cardiac surgery patients.BACKGROUND
[0003] Every year, approximately 2 million people worldwide undergo cardiac surgery mostly with heparin, a parenteral anticoagulant that preserves hemostasis during cardiopulmonary bypass. A rare but dangerous complication of heparin is heparin-induced thrombocytopenia (HIT), an idiosyncratic drug reaction that paradoxically causes a highly prothrombotic syndrome resulting in thrombosis in 50 -75% of patients, limb loss in 10.6% and death in 21.8% . HIT is rare in patients undergoing major cardiac surgery; however, early diagnosis is critical to avoid complications including thrombosis and death.
[0004] Establishing the diagnosis of HIT is particularly challenging after cardiac surgery since the thrombocytopenia caused by HIT is easily confused by existing computing systems with the expected postoperative thrombocytopenia related platelet consumption and dilution after cardiac surgery. As a result, the data provided by a computing system results in a diagnosis of HIT that is often delayed or missed in this setting. Current computing systems execute scoring systems for HIT such as the 4T score and the HEP score, which have modest ability to distinguish HIT from non-HIT in the setting of cardiac surgery. Even the Lillo-Le Louet (LLL) score which was specifically designed for use post cardiac surgery performs modestly in this population.
[0005] In some cases, anomaly detection is defined as finding certain patterns within the data that differ from expected behavior. These anomalies typically involvenumerical deviations from the normal range or are distinguishable from other samples in terms of their distribution or patterning. Generating a statistical model for a problem as an anomaly for its detection or identification has been used in the healthcare domain. For example, anomaly detection has been used in various health problems, including pandemic detection (COVID-19), viral pneumonia and Parkinson 's disease.
[0006] The background herein is included solely to explain the context of the disclosure. This is not to be taken as an admission that any of the material referred to was published, known, or part of the common general knowledge as of the priority date.SUMMARY
[0007] To improve the ability to diagnose HIT, the present application discloses a description of the normal pattern of platelet count levels, monocyte levels and neutrophil levels after major cardiac surgery and identified deviations from those patterns in patients with HIT. A description of these deviations was used to develop a computing system that combines anomaly detection and machine learning to improve the timeliness of HIT diagnosis. In some cases, the computing system rapidly identifies patients at risk of HIT in a post-cardiac surgery setting through an alert system. This Long Short-Term Memory Auto Encoder (LSTM-AE) -based architecture identifies HIT events as anomalies in platelet count time series data from cardiac surgery patients. This method performs better than statistical methods such as percentile-based detection (PHD) and ARIMA, with better recall, precision, and F1-Score.
[0008] In at least a broad aspect, a computing system for generating an alert for heparin-induced thrombocytopenia (HIT), the system comprising: a memory storing instructions; and one or more processors operatively coupled to the memory, wherein the instructions, when executed by the one or more processors, cause the system to: receive a time-series of platelet count data for a patient; process the time-series of platelet count data using a pre-trained autoencoder model to generate a reconstructed time-series of platelet count data, wherein the autoencoder model was trained on a training dataset of platelet count time-series from a cohort of patients without HIT; calculate a reconstruction error based on a difference between the time-series of platelet count data for the patient and the reconstructed time-series of platelet countdata; and generate a HIT alert in response to determining that the reconstruction error exceeds a predefined threshold.
[0009] In some cases, the pre-trained autoencoder model is a Long Short-Term Memory (LSTM) autoencoder model.
[0010] In some cases, the LSTM autoencoder model comprises an LSTM encoder configured to generate a latent representation of the time-series of platelet count data, and an LSTM decoder configured to generate the reconstructed timeseries of platelet count data from the latent representation.
[0011] In some cases, the instructions further cause the system to generate the time-series of platelet count data by selecting a minimum daily platelet count value for each day within a post-operative monitoring period.
[0012] In some cases, the instructions further cause the system to preprocess the time-series of platelet count data by imputing one or more missing values using interpolation.
[0013] In some cases, the instructions further cause the system to preprocess the time-series of platelet count data by standardizing a length of the time-series to a uniform temporal length by padding or truncating the time-series.
[0014] In some cases, the patient is a patient who has undergone cardiac surgery, and the time-series of platelet count data corresponds to a pre-defined postsurgery period.
[0015] In some cases, generating the HIT alert comprises transmitting the HIT alert to a client computer system associated with a clinician or to an electronic health record (EHR) system.
[0016] In some cases, the predefined threshold is determined by optimizing an F1-score that balances precision and recall metrics for HIT detection.
[0017] In another broad aspect, a computer-implemented method is provided for generating an alert for heparin-induced thrombocytopenia (HIT), the method comprising: receiving, at a computing system, a time-series of platelet count data for a patient; processing, by the computing system, the time-series of platelet count data using a pre-trained autoencoder model to generate a reconstructed time-series ofplatelet count data, wherein the autoencoder model was trained on a training dataset of platelet count time-series from a cohort of patients without HIT; calculating, by the computing system, a reconstruction error based on a difference between the timeseries of platelet count data for the patient and the reconstructed time-series of platelet count data; and generating, by the computing system, a HIT alert in response to determining that the reconstruction error exceeds a predefined threshold.
[0018] In some cases, the pre-trained autoencoder model is a Long Short-Term Memory (LSTM) autoencoder model.
[0019] In some cases, the method further comprising generating the time-series of platelet count data by selecting a minimum daily platelet count value from raw platelet count data for each day within a post-operative monitoring period.
[0020] In some cases, the method further comprising preprocessing the timeseries of platelet count data by imputing one or more missing values using interpolation.
[0021] In some cases, the patient is a patient who has undergone cardiac surgery, and wherein the time-series of platelet count data is monitored over a predefined period post-surgery.
[0022] In some cases, generating the HIT alert comprises transmitting the HIT alert to a display device associated with a clinician or integrating the HIT alert into an electronic health record (EHR) interface.
[0023] In some cases, the method further comprising determining the predefined threshold by executing an iterative process to identify a threshold value that optimizes an F1 -score based on a test dataset.
[0024] In another broad aspect, a computing system is provided for detecting heparin-induced thrombocytopenia (HIT) in patients, comprising:a dataset creation module configured to retrieve surgery patient records from a patient database, including platelet count trends, HIT labels, and hospitalization durations;a data engineering module configured to handle missing values, perform platelet count preprocessing, and standardize input sequences for predictive modeling;a statistical modeling module configured to detect HIT risk using percentilebased HIT detection (PHD) and AutoRegressive Integrated Moving Average (ARIMA) techniques;a machine learning module utilizing long short-term memory autoencoders (LSTM-AE) to learn temporal dependencies and identify anomalies within platelet count trends;a model evaluation module configured to assess detection accuracy using F1-score, precision, and recall metrics;wherein the system generates an HIT alert when statistical and machinelearning models indicate a significant deviation in platelet trends consistent with HIT onset.
[0025] In some cases, the data engineering module retains the minimum daily platelet count per patient to emphasize the clinical significance of lower platelet values in HIT assessment.
[0026] In some cases, missing values within platelet count sequences are imputed using interpolation to enhance predictive reliability.
[0027] In some cases, platelet count trajectories exceeding dynamically determined ARIMA residual thresholds are flagged as potential HIT cases.
[0028] In another broad aspect, a method is provided for detecting heparin-induced thrombocytopenia (HIT) in patients, comprising:collecting historical and real-time platelet count data from the patients; preprocessing data to standardize sequence lengths and impute missing values for consistency;applying percentile-based HIT detection (PHD) to flag platelet counts below a statistical threshold derived from non-HIT training data;utilizing AutoRegressive Integrated Moving Average (ARIMA) to model time-series platelet trends and identify anomalous deviations;training a long short-term memory autoencoder (LSTM-AE) to extract temporal features and recognize abnormal platelet drop patterns;validating model performance using F1-score, precision, and recall for optimal HIT detection accuracy;outputting an HIT risk indicator when platelet trends exhibit statistical and learned deviations consistent with HIT progression.
[0029] In some cases, an augmented Dickey-Fuller (ADF) test determines stationarity in platelet count sequences, optimizing ARIMA model order selection for anomaly detection.
[0030] In some cases, the method includes processing electronic health record (EHR) data to refine patient monitoring and risk assessment.
[0031] In some cases, platelet count trends are dynamically monitored over a n-day period post-surgery to optimize early HIT detection, where n is a natural number.
[0032] In some cases, an early warning HIT alert is generated and communicated to clinicians upon identifying patients meeting predefined risk thresholds.
[0033] Other features and advantages of the present disclosure will become apparent from the following detailed description. It should be understood, however, that the detailed description and the specific examples, while indicating embodiments of the disclosure, are given by way of illustration only and the scope of the claims should not be limited by these embodiments, but should be given the broadest interpretation consistent with the description as a whole.DRAWINGS
[0034] Certain embodiments of the disclosure will now be described in greater detail with reference to the attached drawings in which:
[0035] FIG. 1A is a block diagram of a computer system for automatic HIT detection and alert communication in accordance with at least some embodiments.
[0036] FIG. 1 B is a block diagram of a configuration of a HIT detection module shown in FIG. 1A, in accordance with at least some embodiments.
[0037] FIG. 2 is a block diagram of a computer in accordance with at least some embodiments.
[0038] FIG. 3 shows the evolution of median platelet count over time following cardiac surgery in accordance with at least some embodiments. The data represented includes patient data collected from 2002 to 2019.
[0039] FIG. 4 shows the evolution of median Absolute Monocytes over time following cardiac surgery in accordance with at least some embodiments. The data represented includes patient data collected from 2002 to 2019.
[0040] FIG. 5 shows the evolution of median Absolute Neutrophils count over time following cardiac surgery in accordance with at least some embodiments. The data represented includes patient data collected from 2002 to 2019.
[0041] FIG. 6 shows the daily likelihood of HIT on the first 10 post-surgery days in accordance with at least some embodiments.
[0042] FIG. 7 shows the overview of the proposed pipeline for the early identification of HIT event detection in accordance with at least some embodiments.
[0043] FIG. 8 shows the basic structure of an encoder in accordance with at least some embodiments.
[0044] FIG. 9 shows the general workflow of LSTM autoencoder for HIT event identification in accordance with at least some embodiments.
[0045] FIG. 10 shows detailed LSTM autoencoder layers for HIT event identification in accordance with at least some embodiments.
[0046] FIG. 11 shows the recall, precision and F1 -Score for each experiment in ARIMA model in accordance with at least some embodiments.
[0047] FIG. 12 shows LSTM-AE Training and Validation Loss in Data in accordance with at least some embodiments.
[0048] FIG. 13 shows LSTM-AE Training and Validation Loss in Data B in accordance with at least some embodiments.
[0049] FIG. 14 shows LSTM-AE Training and Validation Loss in Data C in accordance with at least some embodiments.
[0050] FIG. 15 shows recall, precision and Fl-Score for each experiment in accordance with at least some embodiments.
[0051] FIG. 16 shows sensitivity distribution over 100 test splits of non-HIT platelet count data in accordance with at least some embodiments.
[0052] FIG. 17 shows specificity distribution over 100 test splits of non-HIT platelet count data in accordance with at least some embodiments.
[0053] FIG. 18 shows 2D Space of latent representation of our LSTM-AE architecture using K-mean clustering with t-SNE in accordance with at least some embodiments.DETAILED DESCRIPTION
[0054] Unless otherwise indicated, the definitions and embodiments described in this and other sections are intended to be applicable to all embodiments and aspects of the present disclosure herein described for which they are suitable as would be understood by a person skilled in the art. It is also to be understood that the terminology used herein is for the purpose of describing particular aspects only and is not intended to be limiting.
[0055] In understanding the scope of the present disclosure, the term “comprising” and its derivatives, as used herein, are intended to be open ended terms that specify the presence of the stated features, elements, components, groups, integers, and / or steps, but do not exclude the presence of other unstated features, elements, components, groups, integers and / or steps. The foregoing also applies to words having similar meanings such as the terms, “including”, “having” and their derivatives. The term “consisting” and its derivatives, as used herein, are intended to be closed terms that specify the presence of the stated features, elements, components, groups, integers, and / or steps, but exclude the presence of other unstated features, elements, components, groups, integers and / or steps. The term “consisting essentially of’, as used herein, is intended to specify the presence of the stated features, elements, components,groups, integers, and / or steps as well as those that do not materially affect the basic and novel characteristic(s) of features, elements, components, groups, integers, and / or steps.
[0056] Terms of degree such as “substantially”, “about” and “approximately” as used herein mean a reasonable amount of deviation of the modified term such that the end result is not significantly changed. These terms of degree should be construed as including a deviation of at least ±5% of the modified term if this deviation would not negate the meaning of the word it modifies. In addition, all ranges given herein include the end of the ranges and also any intermediate range points, whether explicitly stated or not.
[0057] As used in this disclosure, the singular forms “a”, “an” and “the” include plural references unless the content clearly dictates otherwise.
[0058] In embodiments comprising an “additional” or “second” component, the second component as used herein is chemically different from the other components or first component. A “third” component is different from the other, first, and second components, and further enumerated or “additional” components are similarly different.
[0059] The term “and / or” as used herein means that the listed items are present, or used, individually or in combination. In effect, this term means that “at least one of” or “one or more” of the listed items is used or present.
[0060] The abbreviation, “e.g.” is derived from the Latin exempli gratia and is used herein to indicate a non-limiting example. Thus, the abbreviation “e.g.” is synonymous with the term “for example.” The word “or” is intended to include “and” unless the context clearly indicates otherwise.
[0061] It will be understood that any component defined herein as being included may be explicitly excluded by way of proviso or negative limitation, such as any specific compounds or method steps, whether implicitly or explicitly defined herein.
[0062] System For Detecting HIT and Generating Alerts
[0063] In some cases, a computing system is provided for automated HIT event identification and alerting, implemented across one or more networked computingdevices. The system is configured to ingest patient clinical data, transform the data into machine-learnable time-series representations, model expected post-operative platelet recovery behavior, and automatically generate alerts when a patient’s platelet trajectory deviates in a manner consistent with HIT.
[0064] In some cases, the computing system includes one or more data sources providing patient laboratory data, including platelet counts associated with individual patients over time. In some cases, the computing system further includes a data ingestion and normalization subsystem that retrieves, maps, and preprocesses heterogeneous clinical records. In some cases, the computing system further includes a HIT detection subsystem comprising statistical and machine-learning models configured to evaluate platelet count time series. In some cases, an alert generation subsystem is configured to issue HIT alerts when an abnormal platelet trajectory is detected.
[0065] In some cases, the computing system includes one or more computing processors, including specialized processors (e.g., graphical processing units or tensor processing units, etc.) that are configured for neural network computations, memory resources, and communication interfaces supporting execution of the system in a clinical or a cloud-based environment.
[0066] In some cases, the computing system may be implemented using one or more servers, cloud computing platforms, or hybrid architectures, and may interface with electronic health record (EHR) systems.
[0067] In some cases, the computing system executes data transformation and time-series modeling. For example, in operation, the computing system receives platelet count measurements for a given patient over a defined post-operative period. The system converts raw laboratory records into a fixed-length temporal sequence, for example by selecting a minimum platelet count per day and normalizing the sequence to a uniform time window. This transformation enables reliable computational analysis despite missing values, variable sampling frequency, and short hospital stays. In some cases, the transformed platelet time series is processed by the HIT detection subsystem, which is configured to learn normal post-surgical platelet recovery patterns from non-HIT patient data. Rather than relying on predefined clinicalrules, the system models expected physiological behavior using statistical forecasting and machine-learning-based sequence reconstruction.
[0068] In some cases, the computing system executes machine-learning based anomaly detection. In some cases, the HIT detection subsystem includes a recurrent neural network autoencoder, such as a long short-term memory (LSTM) autoencoder, trained using platelet count sequences from patients without HIT. The autoencoder learns a latent representation that captures typical post-operative platelet dynamics.
[0069] In some cases, for a given patient, the trained autoencoder attempts to reconstruct the patient’s platelet time series. The system computes a reconstruction error representing the degree to which the observed platelet behavior deviates from learned normal patterns. Elevated reconstruction error indicates an anomalous platelet trajectory that may correspond to HIT.
[0070] This anomaly-based modeling approach allows the system to identify HIT as a deviation from physiological norms, rather than as a direct classification problem, thereby improving computational robustness in sparse, noisy, and imbalanced clinical datasets.
[0071] In some cases, the computing system executes statistical corroboration and alert generation. In addition to machine-learning models, the system may apply complementary statistical techniques, such as percentile-based thresholds or short-horizon time-series forecasting models, to further characterize abnormal platelet behavior. Outputs from one or more analytical models may be combined or evaluated independently. In some cases, when the system determines that the data of a patient’s platelet trajectory satisfies an abnormality criterion indicative of HIT, the alert generation subsystem automatically generates a HIT alert. The alert may be transmitted to clinicians, displayed within an EHR interface, or stored for audit and quality assurance purposes.
[0072] In some cases, the computing system provides automated modeling of patient-specific temporal behavior rather than reliance on static clinical rules. In some cases, the computing system provides improved detection accuracy in short, high-variance platelet sequences common after cardiac surgery. In some cases, the computing system enables scalable deployment in distributed or cloud computingenvironments. In some cases, the computing system provides early and reliable identification of rare adverse events, enabling timely clinical intervention. In some cases, by transforming clinical time-series data into anomaly scores using trained sequence models, the system improves the operation of computing systems used for clinical monitoring and decision support.
[0073] Referring to FIG. 1A, a computing system 100 is shown for automatic HIT detection and alert communication. In particular, the computing system identifies HIT events and generates automated alerts based on patient clinical data. The computing system 100 includes one or more client computer systems 102a, 102b, 102c (collectively referred to as client computer systems 102). Each client computer system 102 may correspond to a user device, clinical workstation, electronic health record (EHR) terminal, or other computing devices operated by healthcare personnel or integrated healthcare software.
[0074] The client computer systems 102 are communicatively coupled to a network 104, which may include one or more local area networks, wide area networks, wireless networks, or combinations thereof. The network 104 may be wired or wireless, or a combination of both wired and wireless technologies. The network 104 facilitates bi-directional communication between the client computer systems 102 and other components of the computing system 100.
[0075] The computing system 100 further includes a data lake 106 configured to store clinical and analytical data associated with HIT analysis. In the illustrated embodiment, the data lake 106 includes a patient database 108, a HIT database 110, and a HIT artifacts repository 112.
[0076] In some cases, clinical data (e.g., blood test data, patient data, etc.) is inputted from a client computer system 102, and the clinical data is transmitted over the network for processing by the server system 120. For example, a data mapping module 122 and a conflict resolution and detection module 124 process the clinical data, and afterward the data is stored and indexed in the patient database 108.
[0077] The data management system 121 includes the data mapping module 122 and the conflict resolution and detection module 124, to preprocess and store the clinical data in the patient database 108. In some cases, clinical data about a givenpatient could come from different client computer systems. The data mapping module 122 maps the incoming clinical data about the given patient to a data record specific to the patient. The conflict resolution and detection module 124 determines if there is an inconsistency from of two or more separate instances of data about the given patient from two or more different client computer system 102. If there is an inconsistency, then the conflict resolution and detecting module 124 reconciles the data inconsistency using data management, before storing the data in the patient database 108.
[0078] In some cases, the conflict resolution and detection module 124 is configured to identify and resolve conflicting, duplicate, or inconsistent data values within patient records, such as multiple laboratory measurements associated with overlapping time intervals.
[0079] The patient database 108 stores patient-specific clinical information, including laboratory measurements such as platelet counts and related metadata. The HIT database 110 stores HIT-related data, including labeled HIT cases, reference cohorts, or derived HIT indicators. The HIT artifacts repository 112 stores analytical outputs, intermediate data products, trained models, feature representations, or other artifacts generated during HIT detection and evaluation.
[0080] The data lake 106 is communicatively coupled to a server system 120, for example via one or more secure communication links. In some embodiments, the server system 120 may be implemented as one or more physical servers, virtual machines, or cloud-based computing instances. In some cases, the data lake 106 resides on the server system 120.
[0081] The server system 120 also includes a HIT subsystem 125. The HIT subsystem 125 includes a HIT data ingestor 126, which is configured to obtain (e.g., retrieve) patient data from the data lake 106, including data stored in the patient database 108 and the HIT database 110, and to supply the retrieved data to analytical components of the system.
[0082] The HIT data ingestor 126 provides data to a HIT detection module 128, which is configured to analyze patient platelet count time-series data to identify abnormal patterns indicative of HIT. The HIT detection module 128 may implementone or more statistical models, machine learning models, or combinations thereof, as described in further detail with respect to FIG. 1B.
[0083] Outputs generated by the HIT detection module 128 are provided to an alert module 130, which is configured to generate one or more HIT alerts when a potential HIT event is detected. The alert module 130 may transmit alerts to one or more client computer systems 102, store alert information within the data lake 106, or integrate alerts into external clinical workflows, or a combination thereof.
[0084] FIG. 1B shows an internal functional architecture of the HIT detection module 128 shown in FIG. 1A, according to an embodiment.
[0085] As shown, the HIT detection module 128 includes a dataset creation module 140, which is configured to construct patient-level datasets from raw clinical data. The dataset creation module 140 may select relevant time windows, label datasets, and assemble platelet count time-series sequences associated with individual patients.
[0086] The HIT detection module 128 further includes a data engineering module 142, which is configured to preprocess the datasets generated by the dataset creation module 140. In some embodiments, the data engineering module 142 performs operations such as selecting a minimum daily platelet count, handling missing values, normalizing time-series lengths, or generating fixed-length sequences suitable for computational modeling.
[0087] The HIT detection module 128 also includes a statistical modeling module 144, which is configured to apply one or more statistical analysis techniques to the engineered data. Such techniques may include percentile-based anomaly detection, time-series forecasting models, or other statistical methods for identifying abnormal platelet behavior.
[0088] A machine learning module 146 is further provided within the HIT detection module 128. The machine learning module 146 is configured to analyze platelet count time-series data using one or more trained machine learning models, such as recurrent neural network autoencoders. In particular embodiments, themachine learning module 146 learns representations of normal post-operative platelet dynamics and identifies deviations corresponding to potential HIT events.
[0089] The HIT detection module 128 additionally includes a model evaluation module 148, which is configured to evaluate the outputs of the statistical modeling module 144 and the machine learning module 146. The model evaluation module 148 may compute performance metrics, determine anomaly scores, compare outputs to detection thresholds, and generate evaluation results used by the alert module 130.
[0090] Together, the modules illustrated in FIG. 1B enable the HIT detection module 128 to transform raw patient laboratory data into analytical determinations indicative of HIT, thereby supporting automated detection and alert communication as illustrated in FIG. 1A.
[0091] Referring now to FIG. 2, there is illustrated a simplified block diagram of a computer 200 in accordance with at least some embodiments. The computer 200 is also herein interchangeably called a computing system. Computer 200 is an example implementation of a computer such as a client computer system 102, a server system 120, and / or a data lake 106. Computer 200 has at least one processor 210 operatively coupled to at least one memory 220, at least one communications interface 230 (also herein called a network interface), and at least one input / output device 240.
[0092] The at least one memory 220 includes a volatile memory that stores instructions executed or executable by processor 210, and input and output data used or generated during execution of the instructions. Memory 220 may also include nonvolatile memory used to store input and / or output data - e.g., within a database -along with program code containing executable instructions.
[0093] Processor 210 may transmit or receive data via communications interface 230, and may also transmit or receive data via any additional input / output device 240 as appropriate.
[0094] In some cases, the processor 210 includes a system of central processing units (CPUs) 212. In some other cases, the processor includes a system of one or more CPUs and one or more Graphical Processing Units (GPUs) 214 that are coupled together. For example, a large language model executes neural networkcomputations on CPU and GPU hardware, such as the system ofCPUs 212 and GPUs 214. In some cases, the GPU hardware is used to execute machine learning or artificial intelligence computations. In some cases, the GPU hardware is used in machine learning or artificial intelligence computations executed for training a machine learning model, or for executing a machine learning model, or for running operations by an agentic Al agent, or for running operations by a machine learning data pipeline, or a combination thereof.EXAMPLES
[0095] The following non-limiting examples are illustrative of the present disclosure:
[0096] Example 1: Retrospective analysis of HIT in post-major cardiac surgery patients.
[0097] Methods
[0098] A retrospective analysis of all patients who underwent major cardiac surgery with cardiopulmonary bypass from 2002 to 2019 from hospitals in Hamilton, Ontario, Canada was completed. Major cardiac surgery was defined as any thoracoabdominal or intrathoracic operation on the large intrathoracic vessels or heart requiring cardiopulmonary bypass (CPB). Anonymized data from the Transfusion Registry for Utilization, Surveillance, and Tracking (TRUST) research database housed at McMaster University was used. Demographic data and daily platelet count values were extracted from baseline (pre-operatively) until hospital discharge or post-operative day 21. HIT-positive patients were identified as those with a positive enzyme-linked immunosorbent assay (ELISA) for anti-platelet factor 4 (PF4) antibodies and a positive serotonin release assay (SRA) confirming platelet activation with therapeutic doses of heparin. Based on the data, platelet count recovery was defined as the first postoperative day when the platelet count increased by at least 30% from the nadir value and continued to rise for at least 2 subsequent days.
[0099] The following mathematical equation was used to define platelet count recovery post cardiac surgery:fP (j-recovery) — Pnadir "I" ^.3 X Pnadir[P(t + 1) > P(t) for subsequent days following trecovery
[0100] Where: P(f) = platelet count on postoperative day t, Pnadir = the lowest platelet count observed postoperatively, and trecovery = the day of platelet count recovery.
[0101] Results and Discussion
[0102] Of 25,040 adult patients who underwent cardiac surgery, 92 (0.4%) patients developed HIT post-operatively. The number of patients with HIT per year did not change significantly over the follow up period. Median days of heparin exposure was 8 [interquartile range (IQR) 7-10] for HIT patients and 6 (IQR 5-7) for non-HIT patients, (p < 0.001) (see Table 1).April 2002 -December 2019Index admissionsNon-HIT Cases [N=24948] HIT Cases [N=92] Female 6695 [26.8%] 39(42.4%]Age at admission [years] 69 [IQR 60,76] 71 (IQR 65,78]Female 71 [IQR 63,77] 73 (IQR 65,80]CPB time [minutes] 275.00 [IQR: 243.00-315.00] 285.50 (IQR: 267.50-346.00] Median Baseline platelet count156 [IQR 104-226] 187 (IQR 106-253] (x 109 / L]Nadir platelet count [x 109 / L] 102 [IQR 79-127] 62.5 (IQR 34-83] Median Nadir Day [day] l(IQR0-3] 10(IQR 5-12] Percent platelet drop from31[IQR 0-52.6] 68.3(IQR 45.3-79.4] baseline [median, IQR]Time to platelet count4 (IQR 2-6] 15 (IQR 12-17] recovery post-surgery [days]Length of stay [days] 7 (IQR 6,11] 22 (IQR 16,31] In-hospital mortality 817(3.3%] 12(13.0%] Number of platelet transfusion 1 (IQR 1,2] 2 (IQR 1,3]Thrombosis 237(0.95%] 8(8.7%]Day on Heparin 6 (IQR 5, 7] 8(IRQ 7, 10]
[0103] Where CPB = cardiopulmonary bypass, and IQR = interquartile range.
[0104] Patients in the HIT group were more commonly female (42.4% vs. 26.8%, p <0.001) and had longer CPB times CPB (285.5 minutes, IQR 267.5-346; vs. 275 minutes, IQR 243-315, p=0.013). Patient age (median age 73 vs. 71) was similar between groups. Patients with HIT had longer hospital stays (22 vs. 7 days, p < 0.001 )and higher in-hospital mortality (13.0% vs. 3.3%, p < 0.001 ). Time to HIT diagnosis was 9 days (IQR 8-12). The percentage of thrombosis in HIT and Non-HIT patients was 8.7% and 0.95%, respectively (p < 0.001).
[0105] In all post cardiac surgery patients, the platelet count declined rapidly. In patients without HIT, the median nadir platelet count was 102 x 109 / L (IQR 79-127), representing a 31% drop from baseline, and it occurred on POD 1 (median, IQR 0 - 3; FIG. 3). Platelet counts increased from POD 3 onward, reflecting normal platelet recovery and criteria for platelet count recovery was met by POD 4 (median; IQR 2-6). In contrast, in patients with HIT, the thrombocytopenia was more severe, with a median nadir platelet count of 62.5 x 109 / L (IQR 34-83) occurring on POD 10 (IQR 5-12), representing a 68.3% drop from baseline. In HIT patients, the platelet count did not recover even by POD 6. Absolute Monocytes were slightly higher in HIT patients until POD 10, then decreased (FIG. 4), and Absolute Neutrophils peaked on POD 2 and remained higher in HIT patients than in non-HIT patients throughout (FIG. 5).
[0106] T o compute the HIT likelihood per day, the daily platelet count trends were tracked first for each patient, identifying significant drops that align with HIT patterns. An estimate of the probability of HIT occurrence was completed by comparing the platelet count variations of HIT and non-HIT patients using statistical modeling. Finally, the daily likelihood was calculated as the proportion of patients meeting HIT criteria on each postsurgery day, visualized through a time series plot to highlight peak risk days (FIG. 6).
[0107] The data disclosed in this application provides key indicators after major cardiac surgery that should prompt clinicians to think of HIT: 1) a drop in platelet count by at least 68% from baseline; and 2) lack of platelet recovery by POD 6. While these features are sensitive (they may indicate HIT), they are not specific and other causes of post-operative thrombocytopenia should be considered (e.g., sepsis, dissemflinated intravascular coagulation and drug-induced thrombocytopenia). In some cases, the computing system processes time series data to characterize temporal patterns of platelet count values over time immediately, or soon after, after cardiac surgery, which can generate a prediction interval that can be established for the platelet count values on every post-operative day. For example, there is a median sensitivity of 86.4% (95% Cl: 76.6%, 93.5%) and a median specificity of 84.7% (95% Cl: 78.4%, 90.5%). Insome cases, the computing process of this trend analysis can inform daily prediction models to improve the timeliness of HIT diagnosis.
[0108] Example 2: Model to rapidly identify patients at risk of HIT in a postcardiac surgery
[0109] Methods
[0110] Dataset Creation. Data were collected from a patient database 108. In some cases, the patient database 108 is called a TRUST (Transfusion Registry for Utilization, Surveillance, and Tracking) database, which includes records from McMaster University-affiliated hospitals in Hamilton, Ontario, Canada. See also FIG. 7. The computing process includes automatically obtaining raw data files in CSV format for cardiac surgery patients from 2002 to 2019. These files contained patients' health records, laboratory test results, pharmaceutical administration data, transfusion details, hospital length of stay, surgery dates, and information on HIT (Heparin-Induced Thrombocytopenia) events during hospitalization. Data was recorded daily throughout each patient's hospital stay. From the raw files, the computing system generates a comprehensive dataset containing information on 25,040 patients. For this study, the files include data from the first day of hospitalization as baseline data and extracted platelet counts, HIT labels, and hospitalization days for all patients to create a subset. This subset, used in the analysis, includes 92 patients (0.4%) with HIT labels and 24,948 patients (99.6%) without, highlighting the small proportion of HIT cases.
[0111] Data Engineering. In the data engineering stage, interpolation and padding were applied to handle missing values and performed pre-processing to ensure that the data were suitable for modeling. Since platelet count data were collected daily, some patients had multiple entries per day. To address this, the computing system retained the minimum platelet count for each day, based on discussions with clinicians emphasizing the clinical importance of lower counts. In some cases, the computing system identifies patients with fewer than 21 days of data, and, after this identification, for the records of these patients, the computing system automatically pads the sequence using the last recorded platelet count, ensuring uniform input lengths for the model. Conversely, for patients with more than 21 days of data, the computing system automatically limits the dataset to the first 21 days, aligning with the focus on earlydetection within this timeframe. Missing values within daily platelet counts were imputed using interpolation, facilitating the computing system to address minor data gaps without compromising predictive accuracy. The dataset was divided into a training set, which included non-HIT data only, and a test set which included all HIT data and 1% of non-HIT data.
[0112] Statistical Models. In this stage, two statistical models are presented: one based on a simple approach (percentiles) and the other on a complex one (Auto Regressive Integrated Moving Average).
[0113] Percentile-based HIT Detection (PHD). To identify patients with HIT using this method, fifth percentile of platelet counts were used from the training data as the threshold for abnormal values. The threshold represents the lower limit of the normal platelet count distribution among patients without HIT. Mathematically, the threshold is computed byth5= min{%6 Platelet Counts : D x) > 0.05}
[0114] With D(x) the cumulative distribution function of platelet counts in the training set. In this method, platelet counts below ths signal a significant deviation indicative of HIT, aligning with the clinical characteristic of a substantial platelet drop. During the testing phase, each patient’s daily platelet count is compared against the 5th-percentile thresholds derived from the training data. The patient is flagged as a potential HIT case if a platelet count on any day falls below th5.
[0115] Auto Regressive Integrated Moving Average (ARIMA). This approach is a statistical method used to analyze and forecast time series data. Its utility in anomaly detection lies in its ability to model temporal patterns and identify deviations that suggest irregularities. In some cases, the computing system use ARIMA instead of other methods like Double Exponential Smoothing (LES), Logarithmic Regression model, or Box-Jenkins models, due to the short length of the time series (21 days). This very short length of each patient's platelet counts trajectory limits the effectiveness of more complex time series forecasting approaches, which typically require longer sequences to capture meaningful temporal dependencies. In some cases, ARIMA's ability to model short-term trends and residual patterns makes it more suitable for our anomaly detection task, where the primary goal is to identify deviations rather than long-term forecasting.ARIMA uses a sliding window approach to focus on platelet count trends over the early post-surgery days. Let xtrepresent the platelet count on day t. Here, ARIMA predicts future values based on past observations and residuals. The mathematical formulation of ARIMA is given by:xt= fiXt-t + 02^t-2 + - + Q + + faet-2
[0116] Where < >£, and fa, are model parameters, and e£is the error term. The residual is computed as Rt= xt- xt, where ytis the predicted platelet count for day t. Anomalous sequences are flagged when their residuals exceed a dynamically determined threshold, maximizing the F1-Score.
[0117] For this ARIMA model, to determine the most optimized in order (p, d, q), it is based on a systematic method of relying on the principles of time series analysis. Thus, it is stated by evaluating the stationarity of the series using the Augmented Dickey-Fuller (ADF) test, which tests the existence of a unit root in the series. For a non-stationary series (p-value > 0.05), differencing is applied iteratively until stationarity is achieved, thereby defining the differencing order d. The order of autoregression p is determined by the partial autocorrelation function (PACF), whereby large indications by lags shows a direct dependency on their previous values. To determine the average moving order q, the computing system executes an autocorrelation function (ACF), which checks for dependencies on previous error terms. At the end of this method, a the computing system performs a grid search over the values of p and q, and ARIMA models are fitted for each combination. In some cases, the computing system automatically selects the optimal order of p and q based on the lowest value of the Akaike Information Criterion (AIC), which balances complexity and goodness of fir for a model. In some cases, this selection approach guarantees the best ARIMA configuration for identifying HIT anomalies in the platelet count time series.
[0118] Representation Learning Method. Representation learning aims to automatically learn meaningful feature representations from raw data, reducing the need for manual feature engineering. LSTM-based Autoencoders (LSTM-AE) are particularly suited for sequential medical data like platelet count trends, as they capture temporal dependencies through LSTM while reducing dimensionality and identifying anomalies via reconstruction error. Unlike Generative Adversarial Networks (GANs),which are more suitable for data augmentation and synthetic data generation, LSTM-AE is directly optimized for anomaly detection and robust feature learning, making it more interpretable for clinical applications than alternatives like Variational Autoencoders (VAE) or Transformer-based embeddings.
[0119] In some cases, the Long Short-Term Memory (LSTM) architecture includes a set of recurrently connected sub-networks, called memory blocks. The idea behind this memory block is to preserve its state over time and control the information flow thought nonlinear gating units. The architecture of a LSTM block includes the input signal, the gates, the activation functions, the output, and peephole connections. All the gates and the block input are recurrently connected to the block's output. In this project, LSTM is used to capture the temporal nature of platelet count fluctuations over time.
[0120] In some cases, FIG. 7 illustrates an example embodiment of an end-to-end HIT detection workflow 700 implemented by the computing system (e.g., the server system 120) described herein, including data preparation, model training, anomaly detection, and alert generation.
[0121] In some cases, the workflow 700 begins with raw platelet count data 702 associated with a plurality of patients. The raw platelet count data 702 may include multiple platelet measurements per patient per day (or some other time interval) collected during a post-operative hospitalization period.
[0122] The raw platelet count data 702 is provided to a daily platelet aggregation process 704, which is configured to generate a single representative platelet value per day for each patient. In some cases, the daily platelet aggregation process 704 selects a minimum daily platelet count for each day to capture clinically relevant platelet nadirs.
[0123] The aggregated daily platelet data is then supplied to a time-series normalization and sequencing block 706, which is configured to construct a fixed-length platelet count sequence for each patient. In some embodiments, the sequence length corresponds to a defined post-operative window (e.g., twenty-one days), and shorter sequences may be padded or truncated to achieve a uniform temporal length.
[0124] In some cases, the operations associated with 702, 704, and 706 are herein referred to as data input / ingestion and pre-processing.
[0125] After the input data is pre-processed, it is automatically partitioned by the computing system. The normalized platelet time-series data is provided to a dataset partitioning block 708, which separates the data into at least a training dataset 710 and a testing dataset 712.
[0126] In the illustrated embodiment, the training dataset 710 comprises platelet count sequences from non-HIT patients, while the testing dataset 712 comprises platelet count sequences that may include HIT-positive patients. This partitioning enables unsupervised or semi-supervised learning of normal platelet behavior.
[0127] After partitioning of the data, the computing system executes model training. The training dataset 710 is supplied to one or more LSTM autoencoder models 714., which is configured to train a sequence reconstruction model. In some cases, the machine learning model training block 714 trains a long short-term memory (LSTM) autoencoder using only non-HIT platelet sequences.
[0128] During training, training data are also processed by statistical models 716.
[0129] The testing dataset 712 is provided directly to a model evaluation block 718. The model evaluation block 718 also receives outputs from both the statistical models 716 and the LSTM autoencoder models 714.
[0130] Accordingly, the model evaluation block 718 is configured to evaluate model performance by applying the statistical models 716 and the LSTM autoencoder models 714 to the test data 712. The model evaluation block 718 compares outputs generated by the statistical models 716 and the LSTM autoencoder models 714 using the test data 712 to determine detection performance.
[0131] The model evaluation block 718 outputs one or more metrics 720, which quantify performance of the evaluated models. As shown in FIG. 7, the metrics 720 include F1 score, precision, and recall. These metrics 720 provide quantitative measures of HIT detection effectiveness for the statistical models 716 and the LSTM autoencoder models 714 based on the test data 712.
[0132] An autoencoder (AE) is a type of unsupervised learning that learns the underlying properties of input data by encoding it into a new representation and then decoding it back to the original input. In other words, AE techniques are intended toconvert the input data into an internal latent representation, which is subsequently used to generate an output similar to the input data. Autoencoders usually include encoding and decoding processes (FIG. 8). It is used to learn how to reconstruct normal platelet count trends (from non-HIT patients). The combination of LSTM and autoencoder (AE) architectures in our project is essential for identifying HIT events based on platelet count data collected daily on cardiac surgery patients. Indeed, the LSTM network is good at learning the temporal dependencies and trends in the time series platelet count data, which are key to HIT progression. In contrast, AE involves data compression in high dimensions down to latent variables, which can detect anomalies by reconstructing normal patterns and identifying deviations. The proposed hybrid model (FIG. 9) takes advantage of a temporal model and anomaly detection to generate a reliable identification profile for variations in trajectories that indicate HIT events.
[0133] In particular, FIG. 9 illustrates an example embodiment of a hybrid long short-term memory (LSTM) autoencoder system 900 configured to identify anomalous platelet count trajectories indicative of heparin-induced thrombocytopenia (HIT). As shown in FIG. 9, the system 900 includes an LSTM encoder 902 that receives input platelet count data 903. The input platelet count data 903 comprises a time-ordered sequence of platelet count values, illustrated as values x1 ,x2,..., , corresponding to platelet measurements collected over time.
[0134] The LSTM encoder 902 includes one or more stacked LSTM layers, which process the temporal platelet count sequence to learn sequential dependencies and transform the input platelet count data 903 into a compressed representation.
[0135] The output of the LSTM encoder 902 is provided to a latent space 906, which includes one or more latent vectors z. The latent space 906 represents a reduced-dimensional encoding of the input platelet count data 903 and captures underlying temporal features learned by the LSTM encoder 902.
[0136] The latent vectors generated in the latent space 906 are provided to an LSTM decoder 904. The LSTM decoder 904 includes one or more stacked LSTM layers configured to reconstruct the input platelet count sequence from the latent representation.
[0137] The LSTM decoder 904 produces output data 905, illustrated as reconstructed platelet count values xA1,xA2,... , corresponding to the original input platelet count data 903.
[0138] The input platelet count data 903 and the output data 905 are provided to a reconstruction loss block 908, which computes a reconstruction loss representing a difference between the original platelet count sequence and the reconstructed platelet count sequence. The reconstruction loss quantifies how accurately the LSTM encoder 902 and LSTM decoder 904 reconstruct the input platelet count data 903.
[0139] The reconstruction loss generated by the reconstruction loss block 908 is provided to an anomaly score calculation block 910, which computes an anomaly score based on the reconstruction loss. The anomaly score represents the degree to which the input platelet count data deviates from platelet count patterns learned by the hybrid LSTM autoencoder system 900.
[0140] The anomaly score is evaluated by a decision block 912, which compares the anomaly score to a predefined threshold, shown as SCORE > T. The decision block 912 determines whether the anomaly score exceeds the threshold.
[0141] If the anomaly score exceeds the threshold, the decision block 912 outputs a signal indicating an anomaly, which results in an anomalous condition output 914, corresponding to transmission of a HIT alert. This triggers the server system 120 to automatically communicate a HIT alert for the given patient to the one or more client computing systems 102 that are associated with the given patient. For example, the one or more client computing systems 102 that are associated with the patient are the same one or more client computing systems 102 that sent clinical data about the patient data to the server system 120 and / or data lake 106.
[0142] If the anomaly score does not exceed the threshold, the decision block 912 outputs a signal indicating normal behavior, which results in a normal condition output 916, corresponding to no action or communication of an update or nominal status to the one or more client computing systems 102 that are associated with the given patient.
[0143] As illustrated in FIG. 9, the hybrid LSTM autoencoder system 900 combines temporal sequence modeling via the LSTM encoder 902 and LSTM decoder 904 with latent representation learning in the latent space 906, enabling reconstruction-based anomaly detection for platelet count trajectories. This configuration supports identification of abnormal temporal patterns indicative of HIT by detecting deviations from reconstructed normal platelet count behavior.
[0144] FIG. 10 details a process 1000 that includes the network layers used in FIG. 9. For example, the patient’s platelet count data 903’ is input into the LSTM encoder 902’, which generates vector values in the latent space 906’, which are decoded by the LSTM decoder 908’, to output a reconstructed output sequence 905’.
[0145] Regarding FIG. 10, let XeIRrxdrepresent the input sequence for a patient, where T is the timesteps or the sequence length (days post- surgery), and is the number of features (e.g., platelet count 1002). Xis defined by X - [%1,x2,x3, -,xr]. In some other cases, additional features, in addition to platelet count are considered by the computing system.
[0146] The LSTM Encoder consists of LSTM layers that encode the sequence X into a latent representation given by:ht= LSTM xtht^Pe)
[0147] Where are the encoder LSTM weights, and htis the hidden state at time step t. The final hidden state hTrepresents the entire sequence in a compressed form.
[0148] The latent space is defined as zelR" with z repeated across all time steps to reconstruct the sequence Z:Z = [z, z, z, ... , z], 6JRTX
[0149] The LSTM Decoder reconstructs the sequence from Z using another LSTM:xt= LSTM^tXt^a)
[0150] Where xtis the reconstructed patient’s platelet count at timestep t, and / 3dare the decoder LSTM weights. The loss function, as given in the equation below,Ti y^reconstruction= / J**t=l
[0151] measures the mean absolute error between the original sequence xtand the reconstructed sequence xt. The reconstruction error for each timestep is defined as 2et= ||xt- xt||2
[0152] The hybrid architecture classifies a patient as HIT if the overall reconstruction error exceeds a thresholds ththat is obtained by optimizing the F1 -score.( T \^et> th\t=i /
[0153] In some cases, the LSTM-AE detailed architecture includes an encoder, a latent space, and a decoder as shown in FIG. 10. The encoder consists of a series of LSTM layers with LeakyReLU (Leaky Rectified Linear Unit) activation and Dropout. Through the use of a LeakyReLU activation, the vanishing gradient problem is addressed and permits better leaning of the complex patterns present in our input sequence. Dropout prevents overfitting by randomly dropping out network units during training, thereby creating a more robust model with getter generalization. The encoder takes the input platelet count sequence and compresses it into a latent representation. Latent representation is then repeated and fed onto the decoder, which is the reverse of the encoder’s setup that reconstructs the input sequence.
[0154] In some cases, the reconstruction error (et) is computed as the mean absolute difference between the actual and the predicted values across all time steps. A higher et indicates a platelet count those deviates from the model’s learned normal platelet count, potentially indicating an HIT event. In some cases, a desirable threshold minimize dales positives (FP) and false negatives (FN). In some cases, the reconstruction error serves as an anomaly score, and determines the threshold for classifying patients as HIT or non-HIT. Thus, this is based on an iterative computing approach to identify the best threshold that minimizes both FP and FN. The computing approach uses three main steps: Set Threshold Range: a range of possible thresholds between the minimum and the maximum reconstruction errors are defined in the testset and divide this range into evenly spaced intervals for evaluation. Evaluate Each Threshold: for each threshold, a sample is classified as HIT if its reconstruction error (e) exceeds the thresholds (th). The predictions xpredare compared with the true label to compute performance metrics such as precision, Recall, F1 -score, and Confusion Matric (used in this disclosure to evaluate the methods).> rl if et> thxpred [Q otherwise
[0155] To select the optimal threshold, this disclosure uses the F1 -Score as the primary optimization metric, which balances precision (avoiding false positives) and recall (avoiding false negatives). The f1 -Score is defined by the equation2 ■ (Precision ■ Recall)Fl — Score = — - — - - - n —Precision 4- Recall
[0156] To assess the architecture, metrics evaluation computations includes a confusion metric, a F1-Score, a Recall, and a precision. To assess the robustness of the model’s sensitivity and specificity to different test sets, a sensitivity and specificity analysis are processed by partitioning the non-HIT data into 100 fixed subsets. Each subset is used as the test set once, alongside the HIT data, while the remaining 9 subsets are used fortraining. This procedure is repeated 100 times. The sensitivity and specificity of the model are recorded for each iteration. The median sensitivity and specificity and their 95% confidence interval (Cl) are computed, and the distribution of sensitivities and specificities are visualized using a histogram. This analysis evaluates whether the model’s sensitivity and specificity remain stable despite variation in the test data.
[0157] In some cases, the time-series data processed by the HIT detection module 128 is a multi-variate sequence that includes additional laboratory parameters beyond platelet counts. For example, the input data for each time step may include values for absolute monocytes and absolute neutrophils. As shown in FIG. 4 and FIG.5, the temporal patterns of monocyte and / or neutrophil levels in the post-operative period can also exhibit different trends between HIT patients and non-HIT patients. By training the autoencoder model on a multi-variate time-series that includes platelet counts, monocyte levels, or neutrophil levels, or a combination thereof, the system canlearn a more complex and nuanced representation of a patient's normal physiological state. The inclusion of these additional, relevant variables can enhance the model's ability to distinguish subtle, anomalous deviations indicative of HIT, thereby improving the overall precision and accuracy of the detection system.
[0158] In some cases, the alert module 130 is configured to generate graded alerts to provide clinicians with more nuanced risk information. The system may analyze not only the absolute magnitude of the reconstruction error but also its trend over time. For instance, the system may calculate the rate of change, or velocity, of the reconstruction error over a recent time window (e.g., the last 24-48 hours). This velocity can be computed as the slope or first derivative of the reconstruction error time-series.
[0159] Based on this analysis, the alert module 130 may issue different levels of alerts. For example, if the reconstruction error exceeds the primary predefined threshold, the system generates a critical "red alert," indicating a high probability of a HIT event requiring immediate attention. However, if the reconstruction error is currently below the primary threshold but the velocity of the reconstruction error exceeds a predefined slope threshold, the system may generate a less critical "yellow alert." This yellow alert serves as an early warning, notifying clinicians that while the patient's platelet trajectory is not yet critically anomalous, it is deviating from the norm at a rapid rate, suggesting a developing issue that warrants closer monitoring. This two-tiered alert system enables earlier and more proactive clinical intervention.
[0160] In some cases, the dataset creation module 140 is configured to incorporate a patient's baseline platelet count, obtained from blood testing performed prior to surgery, as part of the input time-series data. This pre-operative baseline value serves as the initial data point (e.g., at time t=0) for the patient's platelet count sequence. The inclusion of the baseline count allows the computing system to contextualize the post-operative platelet count trend relative to the patient's own starting physiological state. When the autoencoder model processes the time-series, the magnitude of any subsequent drop in platelet count is implicitly compared against this baseline. This provides a more personalized and accurate basis for anomaly detection, as a significant percentage drop from a patient's specific baseline can be a stronger indicator of HITthan an absolute platelet value considered in isolation. This comparison helps the model to more accurately inform the detection of HIT.
[0161] Example Results and Discussion
[0162] In this dataset of 24,948 non-HIT platelet count observations, 99% are allocated of the data for training, enabling the model to learn the baseline patterns of normal platelet behavior. The remaining 1 % was combined with the 92 HIT observations to form the test set. The length of a platelet count sequence is 21. In some cases, there is an evaluation whether, after being trained on normal platelet count patterns, the model could successfully identify HIT data as anomalies. This section begins by presenting results obtained from statistical models, specifically the Percentile-Based HIT Detection (PHD) and Auto Regressive Integrated Moving Average (ARIMA) models, followed by results from the representation learning method.
[0163] Table 2: ARIMA Confusion Matrix with Data ANON-HIT HITNON-HIT 189 59HIT 38 59
[0164] Table 3: ARIMA Confusion Matric with Data BNON-HIT HITNON-HIT 32 216HIT 13 79
[0165] Percentile-Based HIT Detection (PHD). The confusion matrix shows that applying the Percentile-Based HIT Detection (PHD) method to identify anomalies resulted in a 100% false negative rate, with the method classifying all test data as anomalies (sensitivity: 0%, specificity: 27%). This indicates that PHD is not effective for our specific objective and is therefore not suitable for this project.
[0166] Auto Regressive Integrated Moving Average (ARIMA) model. Three experiments were conducted using the dataset to evaluate the performance of differentapproaches with the ARIMA model. In Data A, the ARIMA model was trained using normal platelet count data from our training set. In Data B, new data was generated based on the daily difference in platelet counts. In Data C, both datasets were combined (normal platelet counts and daily differences) to create two training features for the input data. These experiments aimed to determine which approach yields the best results. The confusion matrices for each experiment are presented in Tables 2, 3, and 4 while Table 5 summarizes the precision, recall, and Fl-Score for each input sequences data. Additionally, FIG. 11 provides a graphical representation of these results. Based on these evaluations, Data A demonstrated the best overall performance between HIT identification (64%) and non-HIT identification (76%). However, despite its relatively better results, the detection rates for both HIT and non-HIT cases remain suboptimal.
[0167] Table 4: ARIMA Confusion Matrix with Data CNON-HIT HITNON-HIT 32 216HIT 13 79
[0168] Table 5: ARIMA, Recall, Precision and F1-ScorePrecision (%) Recall (%) Fl-Score (%)Data A 66 67 66Data B 49 49 31Data C 55 52 30
[0169] Representation Learning Method. The next step is to evaluate the performance of the representation learning model for anomaly detection. To maintain the same experimental setup as described in the previous section, using Data A, Data B, and Data C. For each Data (A, B and C), train and test our LSTM-AE architecture to identify which approach yields the best results. The input sequence length remains 21 days. Once the optimal input sequence data is identified using the 21- day input sequence, the model's performance is further assessed in detecting anomalies specifically from day 5 to day 21. Since deviations in platelet count data can begin as early as day 5, it is crucial to evaluate the model's sensitivity and responsiveness indetecting HIT cases at the earliest possible stage. A detailed discussion of this evaluation will follow, including an assessment of the explainability of the model to ensure transparency and interpretability of the results. The confusion matrices for each experiment are presented in Tables 6, 7, and 8, while FIGs. 12, 13, and 14 illustrate the training and validation loss curves for each input sequences data. The architecture employs an early stopping mechanism during training to prevent overfitting and ensure optimal performance. Additionally, FIG. 15 provides a comprehensive comparison of recall, precision, and Fl-Score across the three input sequences data. The percentage of HIT cases correctly identified versus non-HIT cases identified is as follows: 83% vs.88% in Data A, 88% vs. 78% in Data B, and 86% vs. 85% in Data C.
[0170] Based on Table 9 and FIG. 16, Data A demonstrates the best overall performance with respect to recall, precision, and Fl-Score. However, Data B shows superior sensitivity for detecting HIT anomalies, achieving 88.04% when the sequence length equals 21. To further evaluate the differences between each experiment in detecting HIT and non-HIT patients, Table 9 presents the daily evaluation of each method using our LSTM-AE architecture. From day 6 to day 12, Data A achieves good detection performance and accuracy. This period corresponds clinically to the time when HIT is most likely to occur. However, if the primary objective is to prioritize the early identification of HIT events, Data A emerges as the preferred approach for the input sequence. It is worth noting that Data A utilizes normal platelet counts as its primary feature.
[0171] Table 6: LSTM-AE Confusion matrix with Data ANON-HIT HITNON-HIT 217 31HIT 16 76
[0172] Table 7: LSTM-AE Confusion matrix with Data BNON-HIT HITNON-HIT 194 54HIT 11 81
[0173] Table 8: LSTM-AE Confusion Matricwith Data CNON-HIT HITNON-HIT 210 38HIT 17 75
[0174] Table 9: Representation Learning Method: Recall, Precision and F1- ScorePrecision (%) Recall (%) Fl-Score (%)Data A 82 85 83Data B 77 83 79Data C 79 83 81
[0175] Table 10: Summary of metrics evaluation by sequence length according to the input sequences data used.Days5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 HIT Cid 71 89 82 84 60 65 67 83 60 72 75 77 87 81 78 84 83 Non-HIT Cid 42 39 41 63 78 79 92 73 90 88 84 86 82 88 92 89 88 Precision 55 56 54 57 67 70 79 42 78 79 77 79 79 81 85 81 82 Data ARecall 56 54 54 57 69 72 74 78 75 80 79 81 84 84 85 87 85 Fl-Score 49 37 41 44 67 71 76 73 76 79 78 80 80 82 85 83 83 HIT Cid 100 100 100 70 74 82 78 73 88 86 83 92 89 94 78 95 88 Non-HIT Cid 0 0 0 63 61 69 77 78 67 69 77 65 69 75 83 71 78 Precision 64 64 64 59 64 70 73 72 74 74 77 74 74 77 77 76 77 Data BRecall 50 50 50 62 68 75 77 78 77 78 80 81 80 84 80 83 83 Fl-Score 22 22 22 59 62 70 74 69 71 72 76 73 74 78 78 76 79 HIT Cid 99 99 100 69 61 69 79 85 82 87 87 86 78 86 76 87 82 Non-HIT Cid 0 0 0 62 74 77 65 73 74 75 79 77 84 82 88 81 85Precision 13 13 64 62 65 70 69 73 76 78 79 80 78 79 80 78 79 Data CRecall 49 49 50 65 68 73 73 79 78 81 81 81 81 85 82 84 83Fl-Score 21 21 22 61 66 71 68 74 73 76 76 76 77 79 81 80 81
[0176] Where Cid = Correctly Identify, All the data in this table are in percentage.
[0177] FIGs. 16 and 17 present histograms (distributions) of the sensitivity and specificity scores obtained from the 100 test fragments of patients without HIT. There is a median sensitivity 86.4% with a 95% confidence interval (Cl) of (76.6, 93.5) and a median specificity 84.7% with a 95% Cl of (78.4, 90.5). The results demonstrate that the sensitivity and specificity remain consistent between different test partitions of non-HIT patients to detect HIT data as anomaly. These outcomes confirm the robustness of the model in detecting HIT cases irrespective of the specific non-HIT samples included in the test set.
[0178] Discussion and Explainability of the LSTM-AE model. The results obtained with the LSTM-AE architecture surpass those achieved with the ARIMA time series analysis, particularly in detecting HIT events at an early stage through daily analysis. This architecture demonstrates robust performance in identifying HIT anomalies, achieving 82% precision on day 10 after the operation for input data. This indicates that the computing system effectively characterizes treat HIT events as anomalies and detects the HIT events as early as possible from the five postoperative days.
[0179] In some cases, to understand the computing system’s identification precision, K-means clustering was used combined with t-distribution Stochastic Neighbor Embedding (t-SNE) to visualize the latent space of out LSTM-AE architecture (FIG. 18). K-means clustering is an unsupervised machine learning algorithm that groups data points into clusters based on their similarity. When paired with d-SNE, a dimensionality reduction technique, the high-dimensional latent features of the model are projected into a 2D space, allowing for a clearer visualization of patterns and subgroup structures in the platelet count data. Since the LSTM-AE model was trained on non-HIT data, clustering enables the observation of whether distinct subgroups or outliers exist, which might not be evident in the original high-dimensional feature space.
[0180] In FIG. 18, the two classes are visually distinguishable, represented by distinct clusters. However, certain data points overlap between these two clusters, which help to characterize features of the model identification accuracy. This overlap suggests inherent similarities in platelet count patterns between some HIT and non-HIT cases, posing a challenge for perfect classification.
[0181] Any recitation of numerical ranges by endpoints herein includes all numbers and fractions subsumed within that range (e.g., 1 to 5 includes 1, 1.5, 2, 2.75, 3, 3.90, 4, and 5). It is also to be understood that all numbers and fractions thereof are presumed to be modified by the term "about" which means a variation of up to a certain amount of the number to which reference is being made if the result is not significantly changed.
[0182] Some elements herein may be identified by a part number, which is composed of a base number followed by an alphabetical or subscript-numerical suffix (e.g., 102a, or 102b). All elements with a common base number may be referred to collectively or generically using the base number without a suffix (e.g., 102).
[0183] The systems and methods described herein may be implemented as a combination of hardware or software. In some cases, the systems and methods described herein may be implemented, at least in part, by using one or more computer programs, executing on one or more programmable devices including at least one processing element, and a data storage element (including volatile and non-volatile memory and / or storage elements). These systems may also have at least one input device (e.g., a pushbutton keyboard, mouse, a touchscreen, and the like), and at least one output device (e.g. a display screen, a printer, a wireless radio, and the like) depending on the nature of the device. Further, in some examples, one or more of the systems and methods described herein may be implemented in or as part of a distributed or cloud-based computing system having multiple computing components distributed across a computing network. For example, the distributed or cloud-based computing system may correspond to a private distributed or cloud-based computing cluster that is associated with an organization. Additionally, or alternatively, the distributed or cloud-based computing system be a publicly accessible, distributed or cloud-based computing cluster, such as a computing cluster maintained by MicrosoftAzure™, Amazon Web Services™, Google Cloud™, or another third-party provider. In some instances, the distributed computing components of the distributed or cloudbased computing system may be configured to implement one or more parallelized, fault-tolerant distributed computing and analytical processes, such as processes provisioned by an Apache Spark™ distributed, cluster-computing framework or a Databricks™ analytical platform. Further, and in addition to the CPUs described herein, the distributed computing components may also include one or more graphics processing units (GPUs) capable of processing thousands of operations (e.g., vector operations) in a single clock cycle, and additionally, or alternatively, one or more tensor processing units (TPUs) capable of processing hundreds of thousands of operations (e.g., matrix operations) in a single clock cycle.
[0184] Some elements that are used to implement at least part of the systems, methods, and devices described herein may be implemented via software that is written in a high-level procedural language such as object-oriented programming language. Accordingly, the program code may be written in any suitable programming language such as Python or Java, for example. Alternatively, or in addition thereto, some of these elements implemented via software may be written in assembly language, machine language or firmware as needed. In either case, the language may be a compiled or interpreted language.
[0185] At least some of these software programs may be stored on a storage media (e.g., a computer readable medium such as, but not limited to, read-only memory, magnetic disk, optical disc) or a device that is readable by a general or special purpose programmable device. The software program code, when read by the programmable device, configures the programmable device to operate in a new, specific, and predefined manner to perform at least one of the methods described herein.
[0186] Furthermore, at least some of the programs associated with the systems and methods described herein may be capable of being distributed in a computer program product including a computer readable medium that bears computer usable instructions for one or more processors. The medium may be provided in various forms, including non-transitory forms such as, but not limited to, one or more diskettes,compact disks, tapes, chips, and magnetic and electronic storage. Alternatively, the medium may be transitory in nature such as, but not limited to, wire-line transmissions, satellite transmissions, internet transmissions (e.g., downloads), media, digital and analog signals, and the like. The computer usable instructions may also be in various formats, including compiled and non-compiled code.
[0187] While the above description provides examples of one or more processes or systems, it will be appreciated that other processes or systems may be within the scope of the accompanying claims.
[0188] To the extent any amendments, characterizations, or other assertions previously made (in this or in any related patent applications or patents, including any parent, sibling, or child) with respect to any art, prior or otherwise, could be construed as a disclaimer of any subject matter supported by the present disclosure of this application, Applicant hereby rescinds and retracts such disclaimer. Applicant also respectfully submits that any prior art previously considered in any related patent applications or patents, including any parent, sibling, or child, may need to be revisited
[0189] While the present disclosure has been described with reference to examples, it is to be understood that the scope of the claims should not be limited by the embodiments set forth in the examples, but should be given the broadest interpretation consistent with the description as a whole.
[0190] All publications, patents and patent applications are herein incorporated by reference in their entirety to the same extent as if each individual publication, patent or patent application was specifically and individually indicated to be incorporated by reference in its entirety. Where a term in the present disclosure is found to be defined differently in a document incorporated herein by reference, the definition provided herein is to serve as the definition for the term.
Claims
What is claimed is:
1. A computing system for generating an alert for heparin-induced thrombocytopenia (HIT), the system comprising: a memory storing instructions; and one or more processors operatively coupled to the memory, wherein the instructions, when executed by the one or more processors, cause the system to: receive a timeseries of platelet count data for a patient; process the time-series of platelet count data using a pre-trained autoencoder model to generate a reconstructed time-series of platelet count data, wherein the autoencoder model was trained on a training dataset of platelet count time-series from a cohort of patients without HIT; calculate a reconstruction error based on a difference between the time-series of platelet count data for the patient and the reconstructed time-series of platelet count data; and generate a HIT alert in response to determining that the reconstruction error exceeds a predefined threshold.
2. The system of claim 1, wherein the pre-trained autoencoder model is a Long Short-Term Memory (LSTM) autoencoder model.
3. The system of claim 2, wherein the LSTM autoencoder model comprises an LSTM encoder configured to generate a latent representation of the time-series of platelet count data, and an LSTM decoder configured to generate the reconstructed time-series of platelet count data from the latent representation.
4. The system of claim 1, wherein the instructions further cause the system to generate the time-series of platelet count data by selecting a minimum daily platelet count value for each day within a post-operative monitoring period.
5. The system of claim 1, wherein the instructions further cause the system to preprocess the time-series of platelet count data by imputing one or more missing values using interpolation.
6. The system of claim 1, wherein the instructions further cause the system to preprocess the time-series of platelet count data by standardizing a length of the timeseries to a uniform temporal length by padding or truncating the time-series.
7. The system of claim 1, wherein the patient is a patient who has undergone cardiac surgery, and the time-series of platelet count data corresponds to a predefined post-surgery period.
8. The system of claim 1 , wherein generating the HIT alert comprises transmitting the HIT alert to a client computer system associated with a clinician or to an electronic health record (EHR) system.
9. The system of claim 1, wherein the predefined threshold is determined by optimizing an F1-score that balances precision and recall metrics for HIT detection.
10. A computer-implemented method for generating an alert for heparin-induced thrombocytopenia (HIT), the method comprising: receiving, at a computing system, a time-series of platelet count data for a patient; processing, by the computing system, the time-series of platelet count data using a pre-trained autoencoder model to generate a reconstructed time-series of platelet count data, wherein the autoencoder model was trained on a training dataset of platelet count time-series from a cohort of patients without HIT; calculating, by the computing system, a reconstruction error based on a difference between the time-series of platelet count data for the patient and the reconstructed time-series of platelet count data; and generating, by the computing system, a HIT alert in response to determining that the reconstruction error exceeds a predefined threshold.
11. The method of claim 10, wherein the pre-trained autoencoder model is a Long Short-Term Memory (LSTM) autoencoder model.
12. The method of claim 10, further comprising generating the time-series of platelet count data by selecting a minimum daily platelet count value from raw platelet count data for each day within a post-operative monitoring period.
13. The method of claim 10, further comprising preprocessing the time-series of platelet count data by imputing one or more missing values using interpolation.
14. The method of claim 10, wherein the patient is a patient who has undergone cardiac surgery, and wherein the time-series of platelet count data is monitored over a predefined period post-surgery.
15. The method of claim 10, wherein generating the HIT alert comprises transmitting the HIT alert to a display device associated with a clinician or integrating the HIT alert into an electronic health record (EHR) interface.
16. The method of claim 10, further comprising determining the predefined threshold by executing an iterative process to identify a threshold value that optimizes an F1 -score based on a test dataset.
17. A computing system for detecting heparin-induced thrombocytopenia (HIT) in patients, comprising:a dataset creation module configured to retrieve surgery patient records from a patient database, including platelet count trends, HIT labels, and hospitalization durations;a data engineering module configured to handle missing values, perform platelet count preprocessing, and standardize input sequences for predictive modeling;a statistical modeling module configured to detect HIT risk using percentilebased HIT detection (PHD) and AutoRegressive Integrated Moving Average (ARIMA) techniques;a machine learning module utilizing long short-term memory autoencoders (LSTM-AE) to learn temporal dependencies and identify anomalies within platelet count trends;a model evaluation module configured to assess detection accuracy using F1-score, precision, and recall metrics;wherein the system generates an HIT alert when statistical and machinelearning models indicate a significant deviation in platelet trends consistent with HIT onset.
18. The computing system of claim 17, wherein the data engineering module retains the minimum daily platelet count per patient to emphasize a clinical significance of lower platelet values in HIT assessment.
19. The computing system of claim 17, wherein missing values within platelet count sequences are imputed using interpolation to enhance predictive reliability.
20. The computing system of claim 17, wherein platelet count trajectories exceeding dynamically determined ARIMA residual thresholds are flagged as potential HIT cases.
21. A method for detecting heparin-induced thrombocytopenia (HIT) in patients, comprising:a) collecting historical and real-time platelet count data from the patients;b) preprocessing data to standardize sequence lengths and impute missing values for consistency;c) applying percentile-based HIT detection (PHD) to flag platelet counts below a statistical threshold derived from non-HIT training data;d) utilizing AutoRegressive Integrated Moving Average (ARIMA) to model time-series platelet trends and identify anomalous deviations;e) training a long short-term memory autoencoder (LSTM-AE) to extract temporal features and recognize abnormal platelet drop patterns;f) validating model performance using F1 -score, precision, and recall for optimal HIT detection accuracy;g) outputting an HIT risk indicator when platelet trends exhibit statistical and learned deviations consistent with HIT progression.
