AN AI-ENABLED MULTIFUNCTIONAL DEVICE FOR HEART DISEASE PREDICTION, DETECTION, AND ANALYSIS

Abstract

The present invention relates to anAI-enabled multifunctional device for heart disease prediction, detection, and analysis with long-term synthesis of multimodal data archives on a cloud. The proposed device includes an edge device integrated with an intelligent explainable AI model, plurality of sensor for collecting ECG data, and other health parameters, and sensor for measuring Troponin-I level in blood, which is a significant indicator of heart disease. The proposed device can be used by both common man, and medical practitioner, for real-time detection and prediction of ischemic heart disease. The proposed device integrated with intelligent AI model further aims to generate multimodal data archive using real-time data, wherein said collected data is stored onto a cloud repository, and this data is used for continuous learning. The proposed device is integrated with explainable AI model, aims to find out the actual factors that are possible underlying cause of onset heart disease.

Specification

Description:FIELD OF THE INVENTION
The present disclosure relates to a multifunctional device for heart disease prediction, detection, and analysis. In more particular manner, the present invention relates to an Explainable AI enabled multifunctional device for ischemic heart disease prediction, detection and analysis with long term synthesis of multimodal data archive on cloud.
BACKGROUND OF THE INVENTION
Ischemic Heart Disease (IHD) is one of the most significant cardiac diseases, wherein in this inadequate oxygen supply of the myocardial cells is observed. It has become a prevalent disease, affecting number of people in a year, besides being a significant contributor to the mortality rate. The increase in the number of people affected by this disease has urged the requirement of an intelligent model that could be used by both the common man and medical practitioners for the early detection of IHD.
Application of Artificial Intelligence (AI) has increase in the healthcare. One prior art proposes a supervised model for classifying non-obstructive versus obstructive CAD using XGBoost integrated with SHAP for explainability to identify significant classification features. Another uses XGBoost and explainable AI for heart failure survival prediction, employing feature selection pre-processing and achieving explainability through feature importance computation.
A different prior art exploits federated learning for ECG-based heart care monitoring with explainable AI and CNNs, categorizing arrhythmias using a CNN-based autoencoder. One approach selects the best ensemble tree algorithm using a data pipeline and feature selection technique, employing post-hoc techniques for explainability analysis in heart failure survival detection.
Another prior art presents a classification technique using Genetic Algorithm and Adaptive Neural Fuzzy Inference System (ANFIS) for heart attack detection, addressing explainability through graphs and exploring symptom significance via an evaluation function. One prior art deploys a Poisson Binomial based Comorbidity discovery (PBC) with explainable AI to analyze Electronic Health Records for comorbid diagnoses, procedures, and medications.
For multi-disease prediction, a model uses deep learning with hybrid meta-heuristic algorithms (Lion Algorithm and Butterfly Optimization Algorithm) for optimal feature selection. Other inventions focus on cardiovascular disease prediction from symptoms using AI algorithms on medical history data.
One approach aims to provide insights on XAI interpretability for medical practitioners, performing experiments on heart disease datasets to improve trust in healthcare. Another invention analyzes household air pollution's impact on IHD mortality rates, comparing trends between India and China using data from the Global Burden of Diseases Study 2019.
Some other related technologies are also utilized for performing heart disease predictions, however, none of the existing technology facilitates the home based detection of the heart disease by a common man at an early onset with an easily accessible dataset.
The technologies utilized in the existing state-of-the arts also carried pros and cons, wherein ECG has less sensitivity, and with it only post event confirmation is possible. Echocardiogram offer high sensitivity but it requires specialist intervention and specialized machinery, similar to angiogram, which also requires specialist and expensive machinery, and also angiogram is tedious and cannot always be relied on by a common man.
In present, the prediction of heart disease takes lot of time, and despite large number of deep learning models for prediction of heart disease, yet existing state-of-the-art solutions poses concerns over the reliability or interpretability of the results, creating a need for an intelligent model that is accurate and can be utilized by both common man and medical practitioners, and perform early detection of IHD. Also the aforementioned discussion of state-of-the-art solutions makes it evident that, a single source for prediction of heart disease will not provide the necessary precision. Therefore, it becomes necessary to implement a multimodal dataset for heart health analysis, which is absent in discussed state-of-the-art solutions.
In the view of the forgoing discussion, it is clearly portrayed that there is a need for a device that overcome the aforementioned issues by utilizing data for prediction in a manner that increases the utility of device for a common man. For accurate prediction of IHD, said device focuses on the user of blood levels, especially Troponin levels as an early indicator for detecting IHD, and has a design similar to a hand-held meter with strips, ensuring early detection of IHD at the user end. The present invention provides a handheld multifunctional device that, generates a real-time multimodal dataset along with explainable AI, and utilize a troponin blood level indicators for the detection of the presence of ongoing cardiac event in real-time.
SUMMARY OF THE INVENTION
The present disclosure relates to a multifunctional device for heart disease prediction, detection, and analysis. The proposed handheld device performs early detection, and analysis of IHD, along with long-term real-time synthesis of multimodal data archive. The proposed device mainly utilizes a trained explainable AI model, and a test kit for testing the troponin blood levels, for the real-time detection of IHD. The proposed device can be used by common man, and any medical practitioner, wherein said device offers end-user the facility of testing their chances of contracting IHD at the convenience of their residence, thereby limiting the demand for consulting a doctor for light symptoms. The accuracy of reading based on highly sensitive strips detects each component of blood accurately. The device also provides accurate reading of Troponin levels which additionally integrated with the trained model enabling key decision making about medical condition of the user. Additionally, the proposed device gives an alert warning through notification, when a sudden onset of slightly elevated troponin levels is detected, wherein device recommends the userfor medical attention. The proposed device can also be used by medical practitioners, wherein the device will help cardiologists in screening people for IHD, in real-time, and thereby facilitates an early detection of IHD. The explainable AI model integrated in the proposed device also performs identification of significant features that may have contributed to the disease.
The present disclosure seeks to provide anAI-enabled multifunctional device for heart disease prediction, detection, and analysis with long-term synthesis of multimodal data archive on a cloud. The device comprises: an edge device equipped with one or more sensors configured to collect real-time ECG data and integrate multimodal data; a pre-processing unit coupled to said edge device configured to resize and normalize said collected ECG data into images of a user-defined dimension and range; a feature extraction unit coupled to said pre-processing unit configured to extract a set of features from pre-processed ECG data, said set of features selected from both morphological and frequency domain features, including durations of a P-wave, QRS complex, and T-wave, R-R intervals, ST-segment elevation, and frequency bands using Fast Fourier Transform (FFT) and Wavelet Decomposition; a vision transformer-based artificial intelligence (AI) model connected to said pre-processing unit configured to classify heart conditions from ECG data upon detecting heart disease patterns from ECG signals; a plurality of sensors integrated with said edge device configured to acquire real-time physiological data such as blood pressure, resting heart rate, and exercise-induced heart rate; a medical history input module, configured to gather user-specific historical data, such as age, gender, and family history, through questionnaires or forms; a multimodal data archive stored on a cloud-based system, comprising real-time ECG data, physiological parameters, and medical history, facilitating long-term data analysis; a troponin-I sensor integrated with said edge device configured to detect presence of heart failure indicators in real-time upon analyzing a troponin-I concentration and comparing against predefined clinical thresholds to detect patterns indicative of heart failure; an explainable AI (XAI) unit, providing transparency and insights into predictions made by said AI model using local and global interpretability techniques selected from LIME and Grad-CAM; and a cloud-based continuous learning system that retrains said AI model periodically with new data from a multimodal archive, ensuring up-to-date predictions, wherein said continuous learning system updates said AI model through periodic retraining based on newly collected multimodal data, ensuring said model adapts to evolving patient health patterns.
In an embodiment, said explainable AI (XAI) unit comprises: a data input unit configured to receive and preprocess input data from external sources, including image, time-series, or tabular data, said data input unit having a feature scaling module for normalizing or encoding input features to ensure compatibility with the AI model and a data preprocessing pipeline for handling missing values, data augmentation, or other necessary preprocessing steps, wherein said AI model configured to perform predictions based on preprocessed input data, said AI model contains a model inference component for feeding the preprocessed data into said AI model and generating prediction outputs, including probabilities and class labels; a global interpretability module configured to provide global explanations using Grad-CAM, said global interpretability module comprises a gradient calculation unit for computing a gradient of said AI model's output with respect to feature maps of a final convolutional layer, an activation mapping unit for multiplying computed gradients with activations to identify influential regions in said input data, and a heatmap visualization unit for overlaying activation heatmaps on said input data, highlighting one or more regions contributing most to model's predictions; a local interpretability module configured to provide local explanations using LIME, said local interpretability module comprises a perturbation generator for creating a perturbed dataset by randomly altering input features while keeping a target instance fixed, a local approximation model for training a simpler interpretable model on perturbed dataset to approximate said AI model's behavior locally, and a feature importance analyzer for evaluating and displaying the impact of each feature on said AI model's prediction for a specific instance; an explanation interface configured to present the global and local explanations to the user, said explanation interface includes: a global explanation display for showing said Grad-CAM heatmaps on said input data, and a local explanation display for presenting LIME-generated feature importance charts or bar plots.
In an embodiment, explainable AI (XAI) unit further comprises: a threshold control unit configured to adjust a detail level of said explanations, said threshold control unit includes granularity settings for controlling level of detail for global and local explanations, allowing users to specify the desired level of interpretability; and a real-time explanation generator configured to compute and display explanations alongside each prediction in real-time, said real-time explanation generator includes a parallel processing module for optimizing computation and ensuring timely generation of explanations even with complex models and large datasets.
In an embodiment, said multimodal sensor data collection includes real-time synchronization of data inputs from ECG, pressure, heart rate, and temperature sensors using a unified clock on the Jetson Nano and archiving data in a structured database on said device for local analysis and future cloud transmission.
An objective of the present disclosure is to provide a multifunctional device for heart disease prediction, detection, and analysis
Another objective of the present disclosure is to develop an explainable AI model for IHD prediction that makes use of ECG data trained on vision transformers.
Another objective of the present disclosure is to prepare a real-time multimodal data archive using the trained model deployed on edge device and connected sensors for ECG and other medical history data.
Another objective of the present disclosure is to integrate high-sensitive blood level tests kit onto the edge device for detection of ongoing cardiac events with the most accurate indicator Troponin-I.
Another objective of the present disclosure is to integrate explainable AI model for identifying the feature contributing toward IHD, by analyzing multimodal data archive.
Yet, another object of the present disclosure is to provide an intelligent hand-held device for home-based IHD detection, wherein said device utilizes intelligent explainable AI model, and high sensitive troponin-I test kit, for the detection of IHD, wherein said device can be used by both common man, and medical practitioners.
To further clarify advantages and features of the present disclosure, a more particular description of the invention will be rendered by reference to specific embodiments thereof, which is illustrated in the appended drawings. It is appreciated that these drawings depict only typical embodiments of the invention and are therefore not to be considered limiting of its scope. The invention will be described and explained with additional specificity and detail with the accompanying drawings.

BRIEF DESCRIPTION OF FIGURES
These and other features, aspects, and advantages of the present disclosure will become better understood when the following detailed description is read with reference to the accompanying drawings in which like characters represent like parts throughout the drawings, wherein:
Figure 1 illustrates a block diagram of anAI-enabled multifunctional device for heart disease prediction, detection, and analysis with long-term synthesis of multimodal data archive on a cloud, in accordance with an embodiment of the present disclosure;
Figure 2 illustrates a diagram representing the design of the proposed device, in accordance with an embodiment of the present disclosure;
Figure 3 illustrates a diagram representing the workflow of the various phases towards the proposed device, in accordance with an embodiment of the present disclosure;
Figure 4 illustrates a diagram representing the proposed ECG based initial classification phase, in accordance with an embodiment of the present disclosure;
Figure 5 illustrates a diagram representing the vision transformer architecture in accordance with an embodiment of the present disclosure; and
Figure 6 illustrates a diagram representing the workflow of the multimodal data archive collection.
Further, skilled artisans will appreciate that elements in the drawings are illustrated for simplicity and may not have been necessarily been drawn to scale. For example, the flow charts illustrate the method in terms of the most prominent steps involved to help to improve understanding of aspects of the present disclosure. Furthermore, in terms of the construction of the device, one or more components of the device may have been represented in the drawings by conventional symbols, and the drawings may show only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the drawings with details that will be readily apparent to those of ordinary skill in the art having benefit of the description herein.

DETAILED DESCRIPTION:
For the purpose of promoting an understanding of the principles of the invention, reference will now be made to the embodiment illustrated in the drawings and specific language will be used to describe the same. It will nevertheless be understood that no limitation of the scope of the invention is thereby intended, such alterations and further modifications in the illustrated system, and such further applications of the principles of the invention as illustrated therein being contemplated as would normally occur to one skilled in the art to which the invention relates.It will be understood by those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not intended to be restrictive thereof.
Reference throughout this specification to "an aspect", "another aspect" or similar language means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Thus, appearances of the phrase "in an embodiment", "in another embodiment" and similar language throughout this specification may, but do not necessarily, all refer to the same embodiment.The terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process or method that comprises a list of steps does not include only those steps but may include other steps not expressly listed or inherent to such process or method. Similarly, one or more devices or sub-systems or elements or structures or components proceeded by "comprises...a" does not, without more constraints, preclude the existence of other devices or other sub-systems or other elements or other structures or other components or additional devices or additional sub-systems or additional elements or additional structures or additional components.
Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The system, methods, and examples provided herein are illustrative only and not intended to be limiting.Embodiments of the present disclosure will be described below in detail with reference to the accompanying drawings.
The functional units described in this specification have been labeled as devices. A device may be implemented in programmable hardware devices such as processors, digital signal processors, central processing units, field programmable gate arrays, programmable array logic, programmable logic devices, cloud processing systems, or the like. The devices may also be implemented in software for execution by various types of processors. An identified device may include executable code and may, for instance, comprise one or more physical or logical blocks of computer instructions, which may, for instance, be organized as an object, procedure, function, or other construct. Nevertheless, the executable of an identified device need not be physically located together, but may comprise disparate instructions stored in different locations which, when joined logically together, comprise the device and achieve the stated purpose of the device.
Indeed, an executable code of a device or module could be a single instruction, or many instructions, and may even be distributed over several different code segments, among different applications, and across several memory devices. Similarly, operational data may be identified and illustrated herein within the device, and may be embodied in any suitable form and organized within any suitable type of data structure. The operational data may be collected as a single data set, or may be distributed over different locations including over different storage devices, and may exist, at least partially, as electronic signals on a system or network.
Reference throughout this specification to "a select embodiment," "one embodiment," or "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the disclosed subject matter. Thus, appearances of the phrases "a select embodiment," "in one embodiment," or "in an embodiment" in various places throughout this specification are not necessarily referring to the same embodiment.
Furthermore, the described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. In the following description, numerous specific details are provided, to provide a thorough understanding of embodiments of the disclosed subject matter. One skilled in the relevant art will recognize, however, that the disclosed subject matter can be practiced without one or more of the specific details, or with other methods, components, materials, etc. In other instances, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring aspects of the disclosed subject matter.
In accordance with the exemplary embodiments, the disclosed computer programs or modules can be executed in many exemplary ways, such as an application that is resident in the memory of a device or as a hosted application that is being executed on a server and communicating with the device application or browser via a number of standard protocols, such as TCP/IP, HTTP, XML, SOAP, REST, JSON and other sufficient protocols. The disclosed computer programs can be written in exemplary programming languages that execute from memory on the device or from a hosted server, such as BASIC, COBOL, C, C++, Java, Pascal, or scripting languages such as JavaScript, Python, Ruby, PHP, Perl or other sufficient programming languages.Some of the disclosed embodiments include or otherwise involve data transfer over a network, such as communicating various inputs or files over the network. The network may include, for example, one or more of the Internet, Wide Area Networks (WANs), Local Area Networks (LANs), analog or digital wired and wireless telephone networks (e.g., a PSTN, Integrated Services Digital Network (ISDN), a cellular network, and Digital Subscriber Line (xDSL)), radio, television, cable, satellite, and/or any other delivery or tunneling mechanism for carrying data. The network may include multiple networks or sub networks, each of which may include, for example, a wired or wireless data pathway. The network may include a circuit-switched voice network, a packet-switched data network, or any other network able to carry electronic communications. For example, the network may include networks based on the Internet protocol (IP) or asynchronous transfer mode (ATM), and may support voice using, for example, VoIP, Voice-over-ATM, or other comparable protocols used for voice data communications. In one implementation, the network includes a cellular telephone network configured to enable exchange of text or SMS messages.
Examples of the network include, but are not limited to, a personal area network (PAN), a storage area network (SAN), a home area network (HAN), a campus area network (CAN), a local area network (LAN), a wide area network (WAN), a metropolitan area network (MAN), a virtual private network (VPN), an enterprise private network (EPN), Internet, a global area network (GAN), and so forth.
Figure 1 illustrates a block diagram of anAI-enabled multifunctional device for heart disease prediction, detection, and analysis with long-term synthesis of multimodal data archive on a cloud, in accordance with an embodiment of the present disclosure.
Referring to Figure 1, the multifunctional device (100) includes an edge device (102) equipped with one or more sensors configured to collect real-time ECG data and integrate multimodal data.
In an embodiment, a pre-processing unit (104) is coupled to said edge device (102) configured to resize and normalize said collected ECG data into images of a user-defined dimension and range.
In an embodiment, a feature extraction unit (106) is coupled to said pre-processing unit (104) and is configured to extract a set of features from pre-processed ECG data, said set of features selected from both morphological and frequency domain features, including durations of a P-wave, QRS complex, and T-wave, R-R intervals, ST-segment elevation, and frequency bands using Fast Fourier Transform (FFT) and Wavelet Decomposition.
In an embodiment, a vision transformer-based artificial intelligence (AI) model (108) is connected to said pre-processing unit (104) configured toclassify heart conditions from ECG data upon detecting heart disease patterns from ECG signals.
In an embodiment, a plurality of sensors (110) is integrated with said edge device (102) configured to acquire real-time physiological data such as blood pressure, resting heart rate, and exercise-induced heart rate.
In an embodiment, a medical history input module (112) is configured to gather user-specific historical data, such as age, gender, and family history, through questionnaires or forms.
In an embodiment, a multimodal data archive (114) is stored on a cloud-based system (116), comprising real-time ECG data, physiological parameters, and medical history, facilitating long-term data analysis.
In an embodiment, a troponin-I sensor (118) is integrated with said edge device (102) configured to detect presence of heart failure indicators in real-time upon analyzing a troponin-I concentration and comparing against predefined clinical thresholds to detect patterns indicative of heart failure.
In an embodiment, an explainable AI (XAI) unit (120) provides transparency and insights into predictions made by said AI model (108) using local and global interpretability techniques selected from LIME and Grad-CAM.
In an embodiment, a cloud-based continuous learning system (122) retrains said AI model (108) periodically with new data from a multimodal archive, ensuring up-to-date predictions, wherein said continuous learning system (122) updates said AI model (108) through periodic retraining based on newly collected multimodal data, ensuring said model adapts to evolving patient health patterns.
In an embodiment, said edge device (102) is equipped with a progressive vision transformer pipeline, converting ECG data into image patches for classification into multiple heart conditions into four typed selected from (i) Normal, Left/Right bundle branch block, Atrial escape and Nodal escape, (ii) Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) Premature ventricular contraction and Ventricular escape, and finally (iv) Paced and Fusion of paced and normal.
In an embodiment, said multimodal data archive (114) facilitates long-term synthesis and evolutionary analysis of heart-related data to track patient health over time and refine heart disease prediction models.
In an embodiment, said XAI unit (120) provides local explanations for individual predictions through LIME by perturbing said input ECG data and determining optimum influential regions affecting decision, wherein said XAI unit (120) uses Grad-CAM to generate visual heatmaps of critical ECG regions contributing to model predictions, offering insights into a model's decision-making process.
In an embodiment, said troponin-I sensor (118) provides real-time blood level monitoring for early detection of heart failure, triggering immediate alerts if abnormal levels are detected, wherein said troponin-I biosensor contains a specific antibody-antigen interaction mechanism to detect said concentration of cardiac troponin-I (cTnI), a key biomarker of heart failure and myocardial infarction.
In an embodiment, said heart disease patterns are selected from arrhythmia patterns, including atrial fibrillation through irregular R-R intervals and absent P-waves, ventricular tachycardia through wide QRS complexes, and bradycardia through slow heart rates and ischemic patterns identified through ST-segment elevation or depression and prolonged QT intervals.
In an embodiment, the device (100) further comprises:a diagnostic reporting module (124) that generates a comprehensive report including identified heart conditions, time-stamped episodes of arrhythmias, heart rate variability (HRV) metrics, and visual plots of said ECG signal annotated with abnormalities; andan alert and notification unit (126) configured to provide real-time alerts to users and healthcare providers in case of life-threatening conditions, wherein said alert and notification unit (126) sends notifications through mobile applications, SMS, or email.
In an embodiment, said explainable AI (XAI) unit (120) comprises:a data input unit (120a) configured to receive and preprocess input data from external sources, including image, time-series, or tabular data, said data input unit having a feature scaling module for normalizing or encoding input features to ensure compatibility with the AI model (108) and a data preprocessing pipeline for handling missing values, data augmentation, or other necessary preprocessing steps, wherein said AI model (108) configured to perform predictions based on preprocessed input data, said AI model (108) contains a model inference component for feeding the preprocessed data into said AI model (108) and generating prediction outputs, including probabilities and class labels. The explainable AI (XAI) unit (120) further comprises a global interpretability module (120b) configured to provide global explanations using Grad-CAM, said global interpretability module (120b) comprises: a gradient calculation unit for computing a gradient of said AI model's output with respect to feature maps of a final convolutional layer, an activation mapping unit for multiplying computed gradients with activations to identify influential regions in said input data, and a heatmap visualization unit for overlaying activation heatmaps on said input data, highlighting one or more regions contributing most to model's predictions. The explainable AI (XAI) unit (120) further comprises: a local interpretability module (120c) configured to provide local explanations using LIME, said local interpretability module (120c) comprises:a perturbation generator for creating a perturbed dataset by randomly altering input features while keeping a target instance fixed;a local approximation model for training a simpler interpretable model on perturbed dataset to approximate said AI model's behavior locally; and a feature importance analyzer for evaluating and displaying the impact of each feature on said AI model's prediction for a specific instance. The explainable AI unit (120) further comprises an explanation interface (120d) configured to present the global and local explanations to the user, said explanation interface includes:a global explanation display for showing said Grad-CAM heatmaps on said input data, anda local explanation display for presenting LIME-generated feature importance charts or bar plots.
In an embodiment, said explainable AI (XAI) unit (120) further comprises:a threshold control unit (120e) configured to adjust a detail level of said explanations, said threshold control unit (120e) includes granularity settings for controlling level of detail for global and local explanations, allowing users to specify the desired level of interpretability; and a real-time explanation generator (120f) configured to compute and display explanations alongside each prediction in real-time, said real-time explanation generator (120f) includes a parallel processing module for optimizing computation and ensuring timely generation of explanations even with complex models and large datasets.
In an embodiment, said multimodal sensor data collection includes real-time synchronization of data inputs from ECG, pressure, heart rate, and temperature sensors using a unified clock on the Jetson Nano and archiving data in a structured database on said device for local analysis and future cloud transmission.
The present invention relates to a multifunctional handheld device that performs heart disease prediction and can be used by both common man and medical practitioner. In more particular manner, the present invention relates to an intelligent hand-held multifunctional device for the detection and prediction of IHD. The device integrates intelligence derived from the trained model and high-sensitive troponin-I tests into a single edge device. The device automates the task of IHD prediction and detection with an explainable AI model, wherein the proposed device also provides information relating to the possible factor causing the underlying disease. Additionally, the design of a hand-held home-based IHD monitoring device helps the common man in the early detection of IHD at home, facilitating early treatment options. The device with disease prediction and interpretation capabilities will provides a feature of alerting the medical practitioners to respond immediately. Furthermore, the interpretability provided through the explainable AI model act as a continuous learning paradigm for the device, which further makes it more accurate. Most importantly, the proposed device helps in generation of a multimodal dataset comprising data that can be used for accurate prediction of heart disease, which is considered as the most significant gap in the intelligent heartcare domain is the key highlight of the device.
In an embodiment, the edge device is equipped with a progressive vision transformer pipeline that segments the ECG data into structured image patches, each patch representing discrete segments of ECG waveform for classification, wherein the classification is further performed using a multilayer self-attention mechanism that iteratively refines feature representation through transformer encoder layers, thereby categorizing heart conditions into predefined types, including (i) Normal, Left/Right bundle branch block, Atrial escape and Nodal escape, (ii) Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) Premature ventricular contraction and Ventricular escape, and (iv) Paced and Fusion of paced and normal, based on specific morphological and temporal attributes.
In this embodiment, the edge device processes ECG data by transforming it into structured image patches, each representing specific time segments of the ECG waveform. For example, an ECG waveform recorded over ten seconds is divided into segments of one second each, resulting in ten discrete image patches. Each patch captures morphological elements of the waveform-such as the P-wave, QRS complex, and T-wave-and these features are preserved as visual data.
Once segmented, the data undergoes classification through a progressive vision transformer pipeline. This pipeline uses a series of transformer encoder layers, which apply a multilayer self-attention mechanism. In each encoder layer, the self-attention mechanism identifies relationships between patches by examining the contribution of various segments to specific heart conditions. For example, if an abnormal R-R interval is detected in one patch, it may suggest an arrhythmia, and the attention mechanism will emphasize this patch over others.
As the data moves through these layers, the transformer refines the feature representations of each patch. This refinement enables the system to categorize heart conditions into predefined classes, including Normal rhythms, Bundle branch blocks, and specific arrhythmias. For instance, in cases where wide QRS complexes are prominent across several patches, the system may classify the condition as a Left or Right bundle branch block. Similarly, if a premature beat is detected without a preceding P-wave, the device can categorize it as a Ventricular premature contraction.
The final classification stage organizes the output into categories, such as (i) Normal, Left/Right bundle branch block, Atrial escape, and Nodal escape, (ii) Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) Premature ventricular contraction and Ventricular escape, and (iv) Paced rhythms. Each category reflects specific morphological and temporal attributes, allowing for targeted diagnostic insights. For example, conditions under category (iii) will focus on ventricular patterns, while category (ii) captures atrial-origin arrhythmias, thus enhancing the precision and granularity of the heart condition classification process.
In an embodiment, the XAI unit provides local explanations by segmenting input ECG data into multiple overlapping temporal windows, each representing a portion of the signal, and applies LIME by generating perturbed versions of each segment, measuring the impact of each temporal window on the prediction outcome, while simultaneously generating Grad-CAM heatmaps by mapping convolutional layer activations back to these temporal windows, thereby visualizing critical ECG regions associated with the prediction, and wherein said troponin-I sensor includes a microfluidic analysis chamber designed to isolate blood plasma and quantify cardiac troponin-I concentrations using electrochemical immunoassay principles, wherein said concentration is correlated with predefined clinical thresholds indicative of heart failure or myocardial infarction, and triggers automated alerts if troponin-I levels exceed these thresholds, providing immediate feedback to healthcare providers.

In this embodiment, the AI model identifies arrhythmia patterns by examining specific ECG waveform characteristics and integrating supplementary physiological data for a comprehensive analysis. For example, to detect atrial fibrillation, the model calculates the variability of R-R intervals, which typically lacks regularity in this condition. By identifying irregular R-R intervals over a period, such as variations exceeding 20% between successive intervals, the model flags a potential atrial fibrillation pattern. For ventricular tachycardia, the model focuses on the width of the QRS complex. If the QRS duration surpasses 0.12 seconds consistently, it suggests a ventricular origin of rapid heartbeats. Similarly, the model analyzes the ST-segment for ischemic conditions, identifying deviations such as elevations or depressions of 0.1 mV or more, indicative of myocardial ischemia.
This detection process is augmented by cross-referencing ECG findings with real-time physiological data, such as blood pressure and heart rate. For instance, during an arrhythmia episode characterized by an irregular R-R interval, the model correlates this with a simultaneous spike in blood pressure or heart rate to confirm the anomaly's relevance to the patient's overall cardiovascular state. By integrating these additional data points, the model increases the accuracy of its pattern recognition, minimizing the risk of false positives or negatives in diagnosing specific heart conditions.
The Explainable AI (XAI) unit enhances transparency by including both global and local interpretability modules. The global interpretability module utilizes Grad-CAM to provide insights into the vision transformer model's focus areas. For instance, as the transformer model classifies ECG image patches, Grad-CAM calculates the gradient relative to feature map activations in the final encoder layers. This gradient information is then mapped back onto the ECG image patches, generating heatmaps that highlight significant waveform features, such as regions within the QRS complex or ST-segment, which influenced the model's classification. For example, during ventricular tachycardia detection, the heatmap may emphasize wide QRS complexes, helping clinicians understand which features the model considered most indicative of the condition.
The local interpretability module applies LIME to gain further insights by creating perturbed datasets at the pixel level within each ECG image patch. In this approach, slight variations are introduced to specific pixels-representing minute changes in ECG waveform morphology-to evaluate their effect on the classification outcome. If the model's prediction changes significantly with minor pixel alterations in the ST-segment region, it implies a high level of importance for this feature in the context of ischemic pattern detection. By approximating feature importance and local dependencies, the local interpretability module offers a granular understanding of how each part of the ECG contributes to the model's decision-making process, thereby supporting clinicians in validating AI-based classifications with confidence.
In this embodiment, the Explainable AI (XAI) unit enhances model transparency by breaking down ECG data into overlapping temporal windows. Each window captures a discrete portion of the ECG signal, such as a 50-millisecond segment that may contain a single heartbeat's QRS complex. These windows overlap, ensuring that critical waveform features, like the ST-segment or P-wave, are captured across multiple segments for a more detailed analysis.
The system applies Local Interpretable Model-agnostic Explanations (LIME) to each window by creating perturbed versions. For example, within a specific 50-millisecond segment, minor adjustments are made to the amplitude of the QRS complex, simulating variations that could occur due to physiological changes. The impact of these perturbations on the model's prediction is measured, revealing how sensitive the model is to changes in this segment. By analyzing these responses, LIME helps identify which parts of the ECG most significantly influence the prediction, such as an elevated ST-segment indicating potential ischemia.
Simultaneously, the system generates Grad-CAM heatmaps, which provide visual insights into the model's decision-making. Grad-CAM operates by mapping activations from the convolutional layers back onto the original ECG windows, highlighting critical regions. For instance, if the model detects an abnormality associated with a high-risk heart condition, such as prolonged QT intervals, Grad-CAM will overlay heatmaps on the specific windows containing these intervals. This allows clinicians to visually assess which portions of the ECG signal most influenced the model's decision, such as red-colored regions highlighting the ST-segment during a myocardial infarction prediction.
In parallel, the troponin-I sensor operates using a microfluidic analysis chamber to isolate blood plasma for cardiac biomarker detection. This sensor employs electrochemical immunoassay principles where troponin-I molecules bind to specific antibodies within the chamber, causing an electrical signal that varies based on the concentration of troponin-I present. For instance, if a troponin-I level of 0.05 ng/mL is detected-above the typical threshold of 0.04 ng/mL associated with myocardial infarction-the device triggers an automated alert. This alert is immediately transmitted to healthcare providers, ensuring they receive real-time data on the patient's cardiac status. This prompt feedback enables swift clinical decisions, potentially averting severe outcomes associated with elevated cardiac troponin-I levels.
In an embodiment, the AI model is configured to detect arrhythmia patterns based on R-R interval variability for atrial fibrillation, identification of wide QRS complexes for ventricular tachycardia, and the analysis of ST-segment deviations for ischemic conditions, said detection incorporates temporal analysis and cross-references with the physiological data, such as blood pressure and heart rate, to corroborate ECG findings, enhancing the accuracy of pattern recognition for each heart disease type, and wherein said explainable AI (XAI) unit further includes a global interpretability module which applies Grad-CAM to the vision transformer model by calculating gradients relative to feature map activations within the final transformer encoder layers, wherein activation heatmaps are overlaid onto ECG image patches, identifying significant waveform features contributing to the model's decision, while a local interpretability module implements LIME by generating perturbed datasets at the pixel level within each ECG image patch, allowing the model to approximate feature importance and local dependencies affecting specific classification outcomes.
In an embodiment, the XAI unit's threshold control unit enables granular adjustments of explanation detail by configuring the number of perturbed samples in LIME and modifying activation thresholds within Grad-CAM, thereby facilitating a tailored interpretability experience, and said real-time explanation generator operates in a parallel processing environment to ensure concurrent computation and real-time display of explanations for individual predictions, optimizing response times even with computationally intensive models and extensive datasets.
In this embodiment, the Explainable AI (XAI) unit improves model transparency by segmenting the ECG signal into overlapping temporal windows. Each window spans a specific duration-such as 50 milliseconds-encompassing discrete portions of the ECG waveform, including features like the P-wave, QRS complex, and T-wave. By overlapping these windows, the XAI unit ensures that each critical waveform feature is captured across multiple segments, allowing for a more granular and continuous view of the ECG signal.
For instance, if the ECG signal duration is 500 milliseconds, the XAI unit may generate 10 overlapping windows, each offset by 25 milliseconds. This overlap enables the analysis of key features-like the QRS complex-within several windows, capturing its variations across different parts of the signal. Such a method enhances the detail available for analysis, as it ensures that any significant feature, such as a prolonged ST-segment indicative of ischemia, is represented within multiple windows.
This segmentation approach allows the XAI unit to employ techniques like LIME (Local Interpretable Model-Agnostic Explanations) to generate explanations for the AI model's decisions. By perturbing the data within each window, LIME assesses how slight variations impact the model's prediction, identifying which portions of the signal are most influential. For example, if a 50-millisecond window containing a wide QRS complex results in a change to the model's classification when perturbed, the XAI unit can attribute significant importance to this region, suggesting that it was crucial in identifying an arrhythmia.
Additionally, Grad-CAM (Gradient-weighted Class Activation Mapping) further augments the transparency by mapping the influence of activations within the model back onto these temporal windows, visually highlighting areas where the model focused most intently. For example, Grad-CAM may generate a heatmap over the windows containing ST-segment elevations, visually confirming to a clinician that this feature was key to the AI's ischemia diagnosis. By providing these overlapping and highlighted temporal windows, the XAI unit offers a clear, interpretable visualization of how the model evaluated the ECG data, facilitating a better understanding of the underlying decision-making process.
Figure 2 illustrates a diagram representing the design of the proposed device, in accordance with an embodiment of the present disclosure.
Referring to Figure 2, the design of the proposed device includes plurality of sensor for the data collection, wherein ECG data and troponin blood level data are obtained. The proposed device comprises a trained explainable AI model, that based on the obtained data performs the early detection of IHD. The device integrates intelligence derived from the trained model and high-sensitive troponin-I tests into a single edge device, as shown in Figure 2.
The proposed device for performing accurate detection of heart disease uses various data including combination of ECG, patient medical history along with Troponin-I levels in blood, which is the most accurate indicator of possible heart disease, wherein all these data makes multimodal dataset that is used analysis and detection if ischemic heart disease. The data is used for training an explainable AI model, wherein the training begins with training an ECG based heart disease prediction model by using vision transformers, deployed on the edge device such as jetson nano. This device is later used for collecting the real-time multimodal data archive which consists of a combination of ECG data and medical history from actual end users. The collected data is proposed to be stored onto a cloud repository due to the limitations in storage on an edge device. Over a period of time, the combination of real-time ECG and medical history is used for continuous learning and explainable AI aims to find out the actual factors influencing the onset of heart disease from and across both the modalities. The dataset that is used by the device are described as under, in detail.
Electrocardiogram (ECG): The ECG data is initially utilized for the training of AI model for heart disease prediction, wherein said ECG data for training purpose is based on publicly available datasets.In addition, the model training is performed along with explainable AI adding trustworthiness into the model, to make the device or AI model more reliable and trustworthy. During the time of testing and for multimodal dataset generation, the ECG data is collected using ECG sensors that are connected to the edge device.
Patient medical history: This data is important for knowing any premedical condition of the patient. The proposed device performs a synthesis of a real-time data archive, wherein said archive combines the most important and accessible data to common user that is ECG and medical history data.
Troponin-I levels: This data is most important, as it is a significant indicator of underlying heart disease, because troponin is a protein found in heart for the repair of heart muscles, wherein a high level of troponin indicates the bad health of cardiac muscle health. The level of Troponin-I can be obtained by a blood test, therefore such as sensor is integrated with the edge device that can measure the level of troponin-I in blood in real time.
Figure 3 illustrates a diagram representing the workflow of the various phases towards the proposed device, in accordance with an embodiment of the present disclosure.
Referring to Figure 3, the proposed device mainly comprises three main modules or phases, wherein the first phase is to train the AI model using ECG data, wherein training includes data collection, model design, and training, and testing along with explainability.The second phase is to deploy the trained AI model on the edge device, wherein the deployment of the trained model is done along with the deployment of other sensors such as ECG and other health parameter sensor, data collected from which helps in generation of multimodal data archive. The third phase is to incorporate troponin based blood level indicator on the edge device for the detection of heart disease in real time. Together combination of these three phases fabricates the proposed multifunctional hand-help device for heart disease prediction.
The training of the AI model for ECG based heart disease classification is carried out by using vision transformers and then the model is deployed on the edge device, after complete training. The prediction from this model later act as a ground truth to the proposed synthesis of a real time multimodal data archive, wherein the collective real-time data across two modalities are collected from end users wherein the ECG data can be collected from sensors connected to the edge. Medical history data can be classified as observed and collected data. Observed data such as real-time data like blood pressure, resting heartrate, exercise induced heartrate can be collected from additional sensors connected to the edge device whereas actual data such as age, gender, and other historical information may be collected via questionnaire and forms. The chance that a person may get a heart disease can be acquired from the trained AI model and a combined data repository can be formulated. This formulated multimodal data archive that is combined data repository of important classification data, can be used over a period of time, in medical field for analyzing the effects of medical history data on the variations in heart rate and ECG data. The dataset may also be used for learning the evolutionary dependencies from heart related multimodal data. The combined dataset can also be used to train a multimodal deep learning architecture and can be utilized to predict an improved heart disease prediction system. The device also consists of blood level indicator sensors that aids in detecting quickly the ongoing presence of an undesirable event from highly sensitive Troponin-I which is the most accurate indicator of heart failure.
A detailed explanation of each phase namely, training phase, deploying of trained model, with other sensors, and deploying sensor for troponin-I level, is provided below along with appropriate diagrams.
Figure 4 illustrates a diagram representing the proposed ECG based initial classification phase, in accordance with an embodiment of the present disclosure.
Figure 5 illustrates a diagram representing the vision transformer architecture in accordance with an embodiment of the present disclosure.
The training phase mainly includes vision transformer training, and explainable AI based adaptive model training, wherein a workflow of the proposed training framework is illustrated in Figure 4. The model begins training on a few 200 patient records that are manually labeled. As opposed to the conventional deep learning models that are trained offline, herein continual learning is incorporated that enables the model to learn during its operation as well, because of the ever-changing patient data, it becomes crucial to work on a model that relies on continuous learning. Continuous training seeks to automatically and continuously retrain the model to adapt to changes that might occur in the data. The continual learning algorithm that best suits the proposed application is periodic learning that retrains the model in periodic intervals with the arrival of new patient information. The model aims at classifying ECG signals into four major classes. Since ECGs can be visualized as image data, the recently popular vision transformers have been adopted for the same, wherein the architecture of which is illustrated in Figure 5.
For the data collection and pre-processing, the data for training is collected from customized collection of Physio net's MIT-BIH Arrhythmia Dataset in the form of images. The images are initially resized in the dimension of 100*100 and normalized by dividing with 255.0, the highest value possible in order to bring down the values into the range 0 and 1. Four classes of images have been considered and trained. The four categories of ECG signals along with their properties considered are (i) N with properties as Normal, Left/Right bundle branch block, Atrial escape and Nodal escape, (ii) S with properties as Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) V with Premature ventricular contraction and Ventricular escape, and finally (iv) Q with properties Paced and Fusion of paced and normal. The dataset has been split in the ratio 80:20 for training and testing respectively. The training has been performed for 100 epochs. Further, in order to save the best model, early stopping has been employed and a patience factor of 10 for saving the best model. Evaluation has been performed based on quantitative metrics and the best model was chosen as the optimal one.
Referring to Figure 5, an overview of the vision transformer structure is given that is utilized for training the classification model. On contrary to standard transformers, the image is reshaped into x ∈ RHxWXC into a sequence of flattened 2D patches xp∈ RNX(P2.C), where (H, W) is the resolution of the original image, C is the number of channels, (P, P) is the resolution of each image patch, and N = HW/P2 is the resulting number of patches, which also serves as the effective input sequence length for the Transformer. A constant vector size D is used in the transformer through all of its layers. Patch embeddings are generated by flattening the patched and mapping to D called linear projections.
A learnable embedding to the sequence of embedded patches (z0 = Xclass) is prepended, whose output of the Transformer encoder (z) is the depiction of image y. Both during pre-training and fine-tuning, a classification head is attached to z0L. During pre-training, an MLP with a single hidden layer and during fine-tuning, a single linear layer implement the classification head. To preserve positional information, position embeddings are appended to the patch embeddings, wherein conventional learnable ID position embeddings are employed because of no appreciable performance improvements with more sophisticated 2D-aware position embeddings. The encoder receives this sequence of embedding vectors as input.
The MLP contains two layers with a GELU non-linearity and the equations that govern the training are as in equations (1), (2), (3) and (4).
Z0 = [Xclass; Xp1E: Xp2E;… Xpn E]+ Epos E∈R(P2.C)XD, Epos ∈ R(N+1)×D (1)
z'l = MSA(LN(Zl-1)) + Zl-1, l = 1…..L (2)
zl = MLP(LN(z'l)) + z'l, l = 1…..L (3)
y = LN(z0L) (4)
The proposed AI model is integrated with explainable AI, wherein this is done for interpreting the decision of the trained model. Explainability algorithms can be divided into local and global ones. The proposed model has experimented with both and has reported results for the same. Two explainability algorithms LIME, and Grad-CAM are implemented herein, which are described below along with their working.
LIME: LIME (Local Interpretable Model-Agnostic Explanations) is an Explainability technique that helps interpret the predictions made by machine learning models, including those applied to image classification tasks. LIME provides local explanations, meaning it focuses on explaining individual predictions rather than the model as a whole. When applying LIME to image classification, the basic idea is to understand which parts of an image were most influential in the model's decision-making process. LIME achieves this by perturbing the input image and observing how the predictions change. The major steps in LIME Explainability are:
1. Generate perturbed instances for chosen image: Create slightly modified versions of the original image by perturbing its pixels. This can be done by applying random transformations or by systematically altering small patches of the image.
2. Select representative instances from model predictions: Choose a subset of the perturbed images and their corresponding predictions to create a local neighborhood. The selection can be based on proximity to the original image or using a sampling method.
3. Generate interpretable features: Convert the selected perturbed instances into a format that can be easily interpreted by human observers. For example, you might convert the images into superpixels (contiguous patches of pixels) and compute their average pixel values or intensities.
4. Train an interpretable model: Build an interpretable model, such as a linear model or decision tree, using the interpretable features obtained in the previous step and their corresponding predicted probabilities or scores. This model approximates the behavior of the underlying complex model within the local neighborhood.
5. Compute feature importance: Assess the importance of each interpretable feature (e.g., super pixel) by examining the learned weights or feature importance values of the interpretable model. The higher the weight/importance, the more influential the corresponding feature is for the model's prediction.
6. Generate explanation: Highlight the important features identified in the previous step on the original image. This can be done by overlaying the identified super pixels or using other visualization techniques.

Grad-CAM: Grad-CAM (Gradient-weighted Class Activation Mapping) is a technique used for visualizing and understanding the decision-making process of deep learning models especially CNNs, when applied to image classification tasks. It provides insights into which regions of an input image are relevant for the network's prediction. The main goal of Grad-CAM is to project the salient areas of an image that contribute to a specific class prediction made by the neural network. It generates a heatmap that indicates the importance of each pixel in the input image for the final prediction. The important steps in Grad-CAM calculation are:
1. Gradient Calculation: The gradients of the member class corresponding to the activations of the final convolutional layer are computed. These gradients represent the importance of each feature map in the final prediction.
2. Global Average Pooling: The gradients are spatially pooled by taking their global average, resulting in a vector representing the importance of each feature map.
3. Weighted Sum: The feature maps are weighted by their corresponding importance values obtained from the global average pooling step.
4. Heatmap Generation: The weighted sum of the feature maps is passed through a ReLU activation function, and this activation map is up sampled to the dimension of the image, creating a heatmap.
5. Visualization: The heatmap is overlaid on the input image to highlight the regions that contributed most to the prediction. The intense regions in the heatmap indicate the areas that the network attended to when making the classification decision.

By using Grad-CAM, insights can be gained regarding how the decisions are made in the CNN and identify visual cues that influenced the predictions.
The trained intelligent AI model with explainability is deployed onto the proposed device as shown in Figure 1. Troponin-I level is a major indicator of IHD. A proposed device is integrated with strips for the detection of IHD, wherein the integration of the trained model will further enhance the possibility of an early prediction of heart disease, and the strips being designed are highly sensitive ones which can detect slight variations in the different levels in the blood sample.
The combined intelligence deployed on device can further be utilized for synthesizing the multimodal dataset. ECG sensors can be utilized for collecting the ECG data from user end and the predictions from the ECG based heart disease prediction model on edge can be used as the ground truth for the new medical history data collected. Explainability integrated in the proposed device, makes this output reliable and accountable.
Figure 6 illustrates a diagram representing the workflow of the multimodal data archive collection.
Data attributes corresponding to observed real time ones such as blood pressure, heart rate, glucose levels, etc. can be collected from corresponding sensors connected to the device and other static data such as age, history, gender, etc. can be collected from the users based on a questionnaire. The workflow for the generation of multimodal data archive collection is depicted in Figure 6. The combined real-time multimodal data is one of its kind which can be used for early prediction, and analysis of the various features that contribute to heart disease from a multimodal perspective. The data collected from the end users are stored on cloud after removing the identity details. Over a period of time once the data matures enough, this can be used for training a novel deep learning model that is capable of reading the multiple modalities of data and learn the complex features from them which can be insightful to the cardiovascular domain. Explainable AI further ensures the trustworthiness of the dataset.
In an embodiment, a deep convolutional model for heart disease prediction based on the ECG data with explainable AI is developed, wherein said model leverages convolutional neural networks (CNNs) to extract features from ECG data and experiments with different explainable AI techniques to provide insights into the decision-making process. The proposed approach aims to enhance diagnostic accuracy while ensuring transparency and interpretability of the model's predictions for clinical applications.
The present invention holds several possible industrial applications, for end users, cardiologists, and research and development. The design of intelligent hand-held home based intelligent IHD detection device will offer end-user the facility of testing their chances of contracting IHD at the convenience of their residence, thereby limiting the demand for consulting a doctor for light symptoms. The accuracy of reading based on highly sensitive strips detects each component of blood accurately gives you an accurate reading of ones Troponin levels which additionally integrated with the trained model enabling key decision making about an ongoing cardiac event. Additionally, a sudden onset of slightly elevated troponin levels is a warning sign that one needs medical attention. The proposed model will definitely help cardiologists screen people for IHD (ischemic heart disease) at its onset and early intervention by a specialist for a full diagnosis will get those patients into earlier and more effective treatments. The multimodal dataset can be utilized by doctors and researchers for gaining insights into the actual features that contribute to heart disease. Additionally, the inclusion of troponin-based test module makes the device more reliant by aiding in rapid detection of an ongoing event resulting in reduced mortality rate. Identification of the significant features from a multimodal perspective contributes to the disease by the use of explainable AI. Eventually, the unique dataset can be utilized for training a multimodal deep learning architecture for enhanced predictions, forecasts and analysis and correlation of the features across modalities. The proposed device is a multifunctional device that can be used by both common man and medical practitioners, for the detection, and analysis of heart disease and synthesis of multimodal data archive. The proposed device: predicts the early onset of heart disease using AI model based on vision transformers trained on ECG data; analyze the correlation between multiple modalities of data and identification of new insights leading to early intervention and preventive care; and detects the presence of an ongoing cardiac event based on the most accurate indicator Troponin I for timely detection of cardiac event. The proposed model is capable of generating a long term multimodal real-time data archive by the combination of ECG and medical history. In an embodiment, the proposed device can further be integrated with additional sensors for retrieving accurate data, capability of detecting other heart ailments such as sudden cardiac arrest, coronary heart diseases, etc.
The drawings and the forgoing description give examples of embodiments. Those skilled in the art will appreciate that one or more of the described elements may well be combined into a single functional element. Alternatively, certain elements may be split into multiple functional elements. Elements from one embodiment may be added to another embodiment. For example, orders of processes described herein may be changed and are not limited to the manner described herein. Moreover, the actions of any flow diagram need not be implemented in the order shown; nor do all of the acts necessarily need to be performed. Also, those acts that are not dependent on other acts may be performed in parallel with the other acts. The scope of embodiments is by no means limited by these specific examples. Numerous variations, whether explicitly given in the specification or not, such as differences in structure, dimension, and use of material, are possible. The scope of embodiments is at least as broad as given by the following claims.
, Claims:
1. An AI-enabled multifunctional device for heart disease prediction, detection, and analysis with long-term synthesis of multimodal data archive on a cloud, comprising:
an edge device equipped with one or more sensors configured to collect real-time ECG data and integrate multimodal data;
a pre-processing unit coupled to said edge device configured to resize and normalize said collected ECG data into images of a user-defined dimension and range;
a feature extraction unit coupled to said pre-processing unit configured to extract a set of features from pre-processed ECG data, said set of features selected from both morphological and frequency domain features, including durations of a P-wave, QRS complex, and T-wave, R-R intervals, ST-segment elevation, and frequency bands using Fast Fourier Transform (FFT) and Wavelet Decomposition;
a vision transformer-based artificial intelligence (AI) model connected to said pre-processing unit configured toclassify heart conditions from ECG data upon detecting heart disease patterns from ECG signals;
a plurality of sensors integrated with said edge device configured to acquire real-time physiological data such as blood pressure, resting heart rate, and exercise-induced heart rate;
a medical history input module, configured to gather user-specific historical data, such as age, gender, and family history, through questionnaires or forms;
a multimodal data archive stored on a cloud-based system, comprising real-time ECG data, physiological parameters, and medical history, facilitating long-term data analysis;
a troponin-I sensor integrated with said edge device configured to detect presence of heart failure indicators in real-time upon analyzing a troponin-I concentration and comparing against predefined clinical thresholds to detect patterns indicative of heart failure;
an explainable AI (XAI) unit, providing transparency and insights into predictions made by said AI model using local and global interpretability techniques selected from LIME and Grad-CAM; and
a cloud-based continuous learning system that retrains said AI model periodically with new data from a multimodal archive, ensuring up-to-date predictions, wherein said continuous learning system updates said AI model through periodic retraining based on newly collected multimodal data, ensuring said model adapts to evolving patient health patterns.

2. The device as claimed in claim 1, wherein said edge device is equipped with a progressive vision transformer pipeline, converting ECG data into image patches for classification into multiple heart conditions into four typed selected from (i) Normal, Left/Right bundle branch block, Atrial escape and Nodal escape, (ii) Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) Premature ventricular contraction and Ventricular escape, and finally (iv) Paced and Fusion of paced and normal, and wherein said multimodal data archive facilitates long-term synthesis and evolutionary analysis of heart-related data to track patient health over time and refine heart disease prediction models.

3. The device as claimed in claim 1, wherein said XAI unit provides local explanations for individual predictions through LIME by perturbing said input ECG data and determining optimum influential regions affecting decision, wherein said XAI unit uses Grad-CAM to generate visual heatmaps of critical ECG regions contributing to model predictions, offering insights into a model's decision-making process, and wherein said troponin-I sensor provides real-time blood level monitoring for early detection of heart failure, triggering immediate alerts if abnormal levels are detected, wherein said troponin-I biosensor contains a specific antibody-antigen interaction mechanism to detect said concentration of cardiac troponin-I (cTnI), a key biomarker of heart failure and myocardial infarction, and wherein said heart disease patterns are selected from arrhythmia patterns, including atrial fibrillation through irregular R-R intervals and absent P-waves, ventricular tachycardia through wide QRS complexes, and bradycardia through slow heart rates and ischemic patterns identified through ST-segment elevation or depression and prolonged QT intervals.

4. The device as claimed in claim 1, further comprises:
a diagnostic reporting module that generates a comprehensive report including identified heart conditions, time-stamped episodes of arrhythmias, heart rate variability (HRV) metrics, and visual plots of said ECG signal annotated with abnormalities; and
an alert and notification unit configured to provide real-time alerts to users and healthcare providers in case of life-threatening conditions, wherein said alert and notification unit sends notifications through mobile applications, SMS, or email.

5. The device as claimed in claim 1, wherein said explainable AI (XAI) unit comprises:
a data input unit configured to receive and preprocess input data from external sources, including image, time-series, or tabular data, said data input unit having a feature scaling module for normalizing or encoding input features to ensure compatibility with the AI model and a data preprocessing pipeline for handling missing values, data augmentation, or other necessary preprocessing steps;
said AI model configured to perform predictions based on preprocessed input data, said AI model contains a model inference component for feeding the preprocessed data into said AI model and generating prediction outputs, including probabilities and class labels;
a global interpretability module configured to provide global explanations using Grad-CAM, said global interpretability module comprises:
a gradient calculation unit for computing a gradient of said AI model's output with respect to feature maps of a final convolutional layer,
an activation mapping unit for multiplying computed gradients with activations to identify influential regions in said input data,
a heatmap visualization unit for overlaying activation heatmaps on said input data, highlighting one or more regions contributing most to model's predictions;
a local interpretability module configured to provide local explanations using LIME, said local interpretability module comprises:
a perturbation generator for creating a perturbed dataset by randomly altering input features while keeping a target instance fixed;
a local approximation model for training a simpler interpretable model on perturbed dataset to approximate said AI model's behavior locally;
a feature importance analyzer for evaluating and displaying the impact of each feature on said AI model's prediction for a specific instance;
an explanation interface configured to present the global and local explanations to the user, said explanation interface includes:
a global explanation display for showing said Grad-CAM heatmaps on said input data, and
a local explanation display for presenting LIME-generated feature importance charts or bar plots.

6. The device as claimed in claim 1, wherein said explainable AI (XAI) unit further comprises:
a threshold control unit configured to adjust a detail level of said explanations, said threshold control unit includes granularity settings for controlling level of detail for global and local explanations, allowing users to specify the desired level of interpretability;
a real-time explanation generator configured to compute and display explanations alongside each prediction in real-time, said real-time explanation generator includes a parallel processing module for optimizing computation and ensuring timely generation of explanations even with complex models and large datasets, and wherein said multimodal sensor data collection includes real-time synchronization of data inputs from ECG, pressure, heart rate, and temperature sensors using a unified clock on the Jetson Nano and archiving data in a structured database on said device for local analysis and future cloud transmission.

7. The device as claimed in claim 1, wherein said edge device is equipped with a progressive vision transformer pipeline that segments the ECG data into structured image patches, each patch representing discrete segments of ECG waveform for classification, wherein the classification is further performed using a multilayer self-attention mechanism that iteratively refines feature representation through transformer encoder layers, thereby categorizing heart conditions into predefined types, including (i) Normal, Left/Right bundle branch block, Atrial escape and Nodal escape, (ii) Atrial premature, Aberrant atrial premature, Nodal premature, Supra-ventricular premature, (iii) Premature ventricular contraction and Ventricular escape, and (iv) Paced and Fusion of paced and normal, based on specific morphological and temporal attributes.

8. The device as claimed in claim 1, wherein said XAI unit provides local explanations by segmenting input ECG data into multiple overlapping temporal windows, each representing a portion of the signal, and applies LIME by generating perturbed versions of each segment, measuring the impact of each temporal window on the prediction outcome, while simultaneously generating Grad-CAM heatmaps by mapping convolutional layer activations back to these temporal windows, thereby visualizing critical ECG regions associated with the prediction, and wherein said troponin-I sensor includes a microfluidic analysis chamber designed to isolate blood plasma and quantify cardiac troponin-I concentrations using electrochemical immunoassay principles, wherein said concentration is correlated with predefined clinical thresholds indicative of heart failure or myocardial infarction, and triggers automated alerts if troponin-I levels exceed these thresholds, providing immediate feedback to healthcare providers.

9. The device as claimed in claim 1, wherein said AI model is configured to detect arrhythmia patterns based on R-R interval variability for atrial fibrillation, identification of wide QRS complexes for ventricular tachycardia, and the analysis of ST-segment deviations for ischemic conditions, said detection incorporates temporal analysis and cross-references with the physiological data, such as blood pressure and heart rate, to corroborate ECG findings, enhancing the accuracy of pattern recognition for each heart disease type, and wherein said explainable AI (XAI) unit further includes a global interpretability module which applies Grad-CAM to the vision transformer model by calculating gradients relative to feature map activations within the final transformer encoder layers, wherein activation heatmaps are overlaid onto ECG image patches, identifying significant waveform features contributing to the model's decision, while a local interpretability module implements LIME by generating perturbed datasets at the pixel level within each ECG image patch, allowing the model to approximate feature importance and local dependencies affecting specific classification outcomes.

10. The device as claimed in claim 9, wherein said XAI unit's threshold control unit enables granular adjustments of explanation detail by configuring the number of perturbed samples in LIME and modifying activation thresholds within Grad-CAM, thereby facilitating a tailored interpretability experience, and said real-time explanation generator operates in a parallel processing environment to ensure concurrent computation and real-time display of explanations for individual predictions, optimizing response times even with computationally intensive models and extensive datasets.

Documents