Kumar et al. (35) analyzed the classification of heart diseases prediction, which involved five stages such as heart disease detection and diagnostics, machine learning models and algorithms used for healthcare, feature engineering and optimization techniques, evolving and advanced techniques in healthcare, and different applications of AI across various diseases and health conditions. This study analyzed the deep learning models for early diagnosis of heart disease prediction with evaluation techniques like sensitivity, specificity, and area under the curve (AUC). It also discusses ethical issues, dataset challenges, and transparency of the model. This paper clearly pointed out the advantages and challenges of modern equipment and advanced prediction systems. Rajkumar et al. (36) proposed IoT-based framework and advanced and enhanced deep learning framework. A Hungarian heart disease dataset is used, which is preprocessed by a median studentized residual approach for reducing error values and missing data. The Harris Hawk Optimization (HHO) approach is applied to select the features during preprocessing, which are classified using Modified Deep Long Short-Term Memory (MDLSTM). This output is updated by the Improved Spotted Hyena Optimization (ISHO) algorithm and achieved 98.01% accuracy in implementation. Hammadi et al. (37) presented an updated framework for cardiovascular disease prediction with the hybrid ensemble learning method. A soft voting method is introduced. Class imbalance is given preference here and the model achieves 97.4% accuracy in score 1, 83.6% accuracy in score 2, and 93% accuracy in score 3. This advanced technique applies ensemble learning for early detection of heart diseases. The Department of Computer Science & Engineering BRAC University in Dhaka, Bangladesh. Rokoni et al. (38) focused on model interpretability and applied one-dimensional Convolutional Neural Networks (1D CNNs) and logistic regression for classifying diseases. It achieves 80% overall accuracy, and Local Interpretable Model-Agnostic Explanations (LIME) provides transparency by finding the influenced features like glucose, blood pressure, and troponin.

Wang and Song (39) presented an edge-assisted IoMT framework for monitoring aged people having chronic diseases. The IoMT-based smart home monitoring model is utilized to access medical data and diagnose diseases for aged persons by continuously monitoring and communicating hastily using edge computing. Martinek et al. (40) used federated learning and blockchain technology to ensure privacy for healthcare monitoring. The IoMT technology used in this work employs sensors for health monitoring, and the sensed data are stored and managed using a fog-cloud-assisted network. The federated learning and fraud detection mechanism–enabled blockchain framework is used to process application workloads and validate the quality of service. Kumar et al. (41) proposed a novel IoMT-based healthcare monitoring system using rooted elliptic curve cryptography with Vigenere cipher (RECC-VC) for securing the environment. RECC-VCC enhances security, and the exponential K anonymity model is used for privacy protection. The Improved Extension Neural Network (IENN) framework is used to analyze the level of sensitive data, and the Gaussian mutated chimp optimizer is used to update the weight. Blockchain technology is used to store and manage transactions on the cloud server. Kishor and Chakraborty (42) presented an IoT-based health monitoring system by predicting various diseases such as heart disease, diabetes, breast cancer, dermatology, thyroid, liver disease, and surgery data using machine learning approaches like Decision Tree, Naïve Bayes, Random Forest, Support Vector Machine (SVM), Adaptive Boosting, Artificial Neural Network, and K-nearest neighborhood (KNN). Shafiq et al. (43) presented a deep learning framework using CNN for detecting heart disease symptoms by analyzing the biosensor input for detecting heart disorders. The PASCAL dataset is used to train the CNN model, and the real-time data sensed by IoT sensors are stored in the cloud. The sound of the heart is given as input for classifying whether patients are affected by heart disease. Kumar and Gandhi (44) introduced a three-tier IoT architecture for detecting heart disease. Wearable sensors are used to observe the patient’s health condition, and Apache HBase stores the patient’s monitored data in cloud computing. Apache Mahout is utilized to implement a logistic regression framework for heart diseases. Panja et al. (45) utilized IoMT architecture to monitor and assess the health issues of infected patients and trigger an alert message to their clinicians and relatives. The real-time data collected from patients are transmitted to the cloud via edge devices for processing. A severity analysis of the infection is carried out using fuzzy logic to detect the risk status of COVID-19 patients effectively. Jain et al. (46) proposed point-of-care testing to rapidly detect infectious diseases and give spot results for taking early action.

IoMT devices are also used to capture the patient’s vitals for early detection of diseases such as malaria, influenza A, Ebola virus, Zika virus, COVID-19 virus, and dengue fever. Rezaee et al. (47) suggested a meta-heuristic fuzzy inference system for emotion recognition using the IoMT. Patients are tested by playing music videos to detect their emotional states. Electroencephalography (EEG) signals are captured before and after meditation. Using an optimized, innovative Gunner algorithm, a fuzzy inference-based classification approach is used to classify emotions. Krupa et al. (48) presented an IoMT-based deep learning framework for automatically detecting fetal QRS. The framework used two methods: one for detecting fetal QRS complex using a deep neural network and the second for classifying the results by acclimatizing transfer learning to improve accuracy. The method uses a time-frequency image as input for an IoT-based deep neural network in the abdominal ECG without removing the maternal components. Lu et al. (49) developed a novel IoMT-based fetal monitoring model incorporating an automatic fetal heart rate (FHR) rating method to evaluate fetal health conditions inside the uterus using digital cardiotocographic signals. The monitoring system uses Kreb’s Fischer, improved Fischer, and American College of Obstetricians and Gynecologists (ACOG) classifiers to detect and classify fetal conditions as good and bad for comparison. Rahmani et al. (50) used fog computing–based e-Health gateways by offering higher-level services for storage provisioning and data processing to form a geo-distributed middle layer between the cloud and IoT sensors. The framework uses an early warning score for monitoring health to facilitate energy efficiency, interoperability, reliability, mobility, and performance. Nandy et al. (51) proposed an IoMT-based intelligent agent mechanism to detect brain response using an electroencephalography signal. A bag of neural network categorizes the complex brain signals captured by the IoMT sensors and detects the brain responses. The IBoNN framework is compared with standard machine learning algorithms. Yadav et al. (52) presented biomarker-based electrochemical immuno sensors for diagnosing COVID-19 using the IoMT and artificial intelligence. The smart sensing technique is used with bioinformatics approaches for monitoring non-invasive SARS-COV2.

Verma et al. (53) summarized nano-integrated wearable biosensors and the use of 5G in the Internet of Things for healthcare applications. Fouad et al. (54) presented a numerical approach using the Gautschi model for vertebral tumor prediction. The IoMT technology is used for predicting tumors employing heuristic hock transformation for evaluating possible perpetual incapacity caused by tumors on Haar-like characteristics (HLC), logistics models (LM), conservative therapy method (CTM), and carbon fiber including reinforced materials (CFRM) approaches. Figure 3 describes the percentage of AI and machine learning approaches used in the healthcare industry. In recent years, machine learning models have revolutionized the diagnosis of various diseases and the assessment of the risk factors involved in making accurate decisions. According to a background study, supervised learning approaches like 35% of logistic regression, 26% of decision tree, and 24% of neural networks, as well as 5% of unsupervised learning methods like clustering and anomaly detection, have been used by the healthcare industry to assess the risks.

Ribeiro et al. (55) presented a novel agnostic approach called LIME, which is used to provision the comprehensiveness of the decisions made by the banking staff by determining them with simple and interpretable models. The LIME model helps improve accuracy, transparency, and trustworthiness. This model can be used with complex machine learning algorithms without any knowledge of their working mechanisms. Lundberg and Lee (56) proposed a new technique, SHAPELY Additive exPlanations (SHAP), to interpret the existing complex machine learning models. The SHAP algorithm provides global and local interpretations to help borrowers easily understand the predictions made by black box techniques. SHAP can be used in various machine learning approaches and deep neural networks and effectively work with real-world datasets.

Muddamsetty et al. (57) presented an evaluation for understanding the outcomes of machine learning models. Thus, it is evident that the XAI model helps to present outcomes with transparency and provides solutions to black box models. Explanations for the clinical prediction outcome entail the justification of reliability and trustworthiness that can be achieved using XAI models (58). Onan (59) presented a hierarchal graph-based model for text classification of dynamic fusion using BERT. The framework uses seven stages for graph-based text classification and analysis with various benchmark functions. Onan (60) proposed a genetic technique combined with graph-based neural networks for generating augmented text having high dimensional feature space. The objective function is based on perplexity when evaluating the quality of generating augmented text data. Onan (61) proposed a Semantic Role Labeling algorithm with an Ant colony optimization approach for generating training data to improve the performance of the natural language processing (NLP) framework. The semantic roles are identified using semantic role labeling (SRL) for text augmentation to enhance the quality of training data. Onan (62) suggested a bidirectional convolutional recurrent neural network framework for semantic analysis using gated recurrent unit (GRU) and LSTM layers. Feature extraction is carried out by the bidirectional layers to reduce dimensionality and extract high-quality features. Onan et al. (63) presented a two-stage topic extract model using a word embedding approach and cluster analysis. The word vectors are extracted by Word2Vec, POS2vec, LDA2vec, and word position2vec schemes. A comparison of Naïve Bayes, SVM, Random Forest, and Logistic regression with ensemble methods is used for evaluating the statistical key extraction model (64). Onan et al. (65) presented a consensus cluster mechanism using an undersampling model with five supervised learning algorithms and three ensemble learners for imbalanced learning. Onan and Korukoğlu (66) utilized an ensemble model for feature selection with a genetic-optimized algorithm for sentimental analysis. Sentiment classification based on a hybrid ensemble pruning model with consensus clustering is utilized for text classification (67). Sentimental analysis for product reviews (68), online course evaluation (69), and mining opinions for instructors (70) is done using deep neural networks. Onan (71) presented a comparative analysis of feature engineering models using five base learners for text genre classification and language function analysis. Onan and Toçoğlu (72) suggested inverse gravity moment utilizing bidirectional LSTM for representing text documents. The LSTM framework is evaluated on the basis of the sarcasm identification corpus. The deep learning model is utilized to identify sarcasm for predicting the performance of sentiment analysis. Vakharia et al. (73) proposed three deep learning frameworks with optimized explainable artificial intelligence for predicting the discharge capacity of the battery. The jellyfish optimization algorithm is used with the XAI model to improve the predictive performance. Ali et al. (74) presented an SVM model based on four Ant Bee Colony (ABC) algorithms, Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Whale Optimization Algorithm (WAO) for tuning hyperparameters. A teaching learning–based optimization algorithm with a heat transfer searching model is used to select features and identify faults. Suthar et al. (75) and Vakharia et al. (76) highlighted a comparative study of feature ranking approaches for fault identification by using the Fisher score, ReliefF, Gain ratio, Wilcoxon rank, and Memetic feature selection model. The literature survey shows that the XAI framework improves prediction accuracy with interpretability and explainability in healthcare applications because of the crucial nature of the decision-making process and public health safety. Tables 1, 2 depict a comparison of various black box models used for healthcare applications.

This study aims to answer all the following questions: 1.

How can this IoMT sensor device help predict the heart disease effectively?

The IoMT integrates wearable devices with sensor technologies to monitor health parameters, track physical activity, and enable remote patient monitoring. The literature survey focuses on two specific sensor categories: vital signs and motion sensors. For vital signs sensors, Rao et al. (104) present a non-invasive wearable device that accurately monitors BP without requiring invasive catheterization. The device utilizes capacitive wrist and/or foot sensors to acquire pulse waveform data, which are then processed using artificial neural networks to determine systolic, diastolic, and mean arterial pressures. A comparison with invasive arterial line data confirmed the device’s accuracy, making it a viable alternative for continuous BP monitoring in critically ill infants. In motion sensors, Jakob et al. (105) evaluate the effectiveness of wearable sensors in analyzing motion patterns in individuals with Parkinson’s disease. The study assesses the accuracy and reliability of the sensor system in detecting and quantifying motor symptoms associated with Parkinson’s disease, such as bradykinesia and shuffling gait. Wearable sensors distinguish Parkinson’s patients from healthy controls, showing their potential for clinically relevant gait assessments in flexible environments. These research papers are examples of studies conducted on sensors used in IoMT projects. The surveyed literature demonstrates the significance of vital signs sensors in non-invasive blood pressure monitoring and the potential of motion sensors in analyzing motor symptoms in Parkinson’s disease. The following parameters mentioned in Figure 4 are taken into consideration while designing existing and ongoing IoMT systems that are positively helping to transform the IoMT domain through cutting-edge technology. Physiological parameters, biochemical parameters, electrical activity, respiratory parameters, motion and activity, sleep patterns, environmental factors, and medical adherence are the IoT devices applied for treatments. Some of the IoMT projects have incorporated the above-mentioned parameters and have made the availability of diagnostics more accessible, for example, Ovularing (106), VitalPatch (107), SmartPill (108), GlucoWear (109), MindMotion Pro (110), BioStampRC (111), Biotricity (112), SmartMat (113), PillCam (114), WAND (Wireless Artifact-free Neuromodulation Device) (115), Abilify MyCite (116), Embrace (117), and Insulet Omnipod (118).

The following case studies outline why XAI will prove to be a revolutionary change required in the IoMT.

CVD takes the lives of around 18 million people every year and is the primary cause of death. The rate of accountability of death reports due to CVD is around 31%. A total of 80% of deaths associated with CVD are mainly due to heart attack and stroke. These attacks are observed in groups of people who are less than 70 years old. With this in mind, a dataset (126) has been prepared as an amalgamation of observations recorded from Cleveland (303), Hungarian (294), Switzerland (123), Long Beach, VA (200), and Stalog Dataset (270). After removing duplicates, the final dataset contains 918 instances with 11 important features for analyzing CVD diseases. The dependent target class is Heart Failure. The other independent features are Age, Sex, Chest Pain Type, ST_Slope, Cholesterol, Resting BP, Blood Sugar, Resting ECG, Exercise Angina, Old Peak, and Maximum Heart Rate (MaxHR). Some of the features are numeric, and some of the features are non-numeric. Table 4 provides the list of features converted to numeric data. The string data are transformed using the Label Encoder preprocessing technique with Min-Max scalar transformation. The dataset (126) has no missing values or class imbalance. Heart Rate, Variable Heart Rate, Blood Glucose Level, MaxHR, Blood Pressure, and ECG (Polar H10 Sensor) are measured by IoMT sensors. The other readings are observed in the oscilloscopes and treadmills (ST_Slope, Restring Angina), and some data are collected directly from patients and their attenders (Name, Age, Sex, etc.).

Figure 6 provides an overview of interfacing the ML algorithms discussed in this section with XAI. In terms of monitoring and managing cardiovascular health, IoMT devices play an important role in the use of advanced transformation techniques. These devices increase power connectivity, analyze data, and monitor remotely, providing advanced care for cardiac patients and improving patient health. There are many IoMT devices for monitoring the heart behavior of patients, such as remote ECG monitors, wearable heart rate monitors, pacemakers, BP monitors, temperature monitors, and medication dispensers. These devices help healthcare professionals to monitor patients continuously. They can personalize treatment plans, and they can easily predict previous symptoms and take immediate action.

Machine learning techniques play a substantial role in identifying heart diseases with the help of IoMT devices. Data are collected from IoMT sensor devices, and ML algorithms understand the data, detect anomalies, and produce solutions for accurate heart diagnosis. Many ML algorithms can be applied to train the model to analyze data and recognize patterns. A large volume of data can be processed by ML algorithms from IoMT sensor devices, such as blood pressure measurements, ECG reading results, and heart rate information. ML algorithms like decision trees, Random Forest, SVM, and Logistic Regression are applied here with IoMT devices.

AI algorithms suggest transparent and interpretable explanations for making decisions or predictions. Traditional AI algorithms work as black boxes and produce results with less transparency. When we use explainable AI, it produces an understanding of the reasoning behind its results. Data are collected from IoMT devices and sent for preprocessing, followed by model selection and training. After training the data, feature analysis is done, and during local interpretability, explainable AI uses LIME to understand how specific features contribute to identifying the heart disease. As global interpretability, SHAPLEY helps explainable AI analyze overall behavior and features and their relationship to the decision-making process. Model-agnostic interpretability independently understands the prediction process and aims to apply it to any algorithm. The results can finally be visualized so that appropriate decisions can be made.

This section describes two algorithms, one for the heart risk evaluation through Algorithm 1 and the other for explaining heart failure through the Algorithm 2. These two algorithms comprehensively analyze and explain the risk of Heart Failure as a complete solution. In Algorithm 1, the performance metrics such as accuracy, precision, recall, sensitivity, specificity, and F1-score are evaluated. During the testing phase, when xtest becomes 1, the heart failure alarm will be activated. Otherwise, the result indicates that the function of the heart is normal. In Algorithm 2, the model with local surrogates explains the appropriate decision after heart failure when the prediction is local. In case the probability of the prediction is global, explainability is achieved in global surrogates.

The dataset under consideration must be protected and authenticated. Hence, rigorous data access control permissions must be set in the cloud to access it properly. A secure environment-based attribute access control system is required in this context to protect unauthorized access to the data in the cloud. The model is divided into two categories: static and dynamic. Users with the lowest role, such as those looking for recommendations, access information in a static environment. This audience will only be permitted to obtain legal information; no other transactions will be permitted. In a dynamic state, different parameters are measured and recorded at various instances of time. Thus, many data acquisition and update cycles are a series of transactions carried out in the cloud in big time. These states only allow special users such as clinicians and administrators.

The development of a digital identity is the first step. The key used in the digital identity protects and guarantees a transmission between the server and the client, and the key is pk. The user shares its digital account identity and the symmetric key ek, and the corresponding data information can be obtained by decrypting the key pk. Various functions used in the algorithm, such as IssueRole, revokeIssueRole, and partialExtension, help the framework achieve a secured space to function. After the digital identity is authenticated and a role is identified, the model can access the framework accordingly. Each of the entity’s transactions is considered along with its authorization. Therefore, a secure environment for the fuzzy framework is achieved.

The 11 parameters that determine the failure of the heart are acquired from various sources across various countries and used in this work. These parameters have a strong influence on determining heart failure in real time. Most of these parameters are embedded with IoMT sensors, which can be integrated through information fusion in cloud platforms. Later, these data are classified by cloud machine learning models and transformed into a valid dataset. One such dataset is used in this work for experimental analysis. Because the problem is binary, the experimentation is done with machine learning models such as SVM, Logistic Regression, Decision Tree, and Random Forest. The explanation of this dataset is provided by LIME and SHAPELY values. The classification probability of the random forest model is evaluated due to its high classification accuracy with various explanations for clarity.

This work has the following challenges (not limited to), which are required to be addressed in the future. The sensors may go out of order and hence can provide false alarms to the cloud and database. The electronic faults may induce false alarms regarding heart failure. The medical data are subjected to be private. Explaining may compromise the privacy and integrity of the individual medical data. Medical data stored in the cloud are vulnerable to attacks if no security mechanisms are provided. If the medical record is stored in a blockchain model, it is extremely difficult to access and explain the same with the XAI model. The reliability of the explanation and privacy need to be enhanced by Federated Learning. Training and demonstration are required for medical practitioners to handle data from wearable sensors and the cloud.

The essential contributions of this paper helps identify the complete purpose of this research. This paper provides a complete illustration of all the sections of IoMT-enabled XAI infrastructure. It also discusses various IoMT applications and case studies related to heart failure in detail. This paper works with a dataset with all the vital parameters required for heart failure prediction. It provides solutions for the explanation of heart failure through local and global surrogates with the explanations of LIME and SHAPELY. This study discusses various state-of-the-art IoMT sensors with practical applicability in medical applications with a discussion of advantages and disadvantages.

Improvements can be made to this study by applying many advanced techniques. The application of 6G may improve the connectivity and network-related issues associated with wearable sensors. Application of Federated Learning would improve the privacy, reliability, and safety of medical data. Meta-verse applications can enhance IoMT sensor support and provide real-time solutions to heart problems. Industry 5.0 can enhance the quality of service of the proposed system with a human-centric man–machine interface. Web 3.0 standards can provide better semantics, security, and reliability in cloud service.