1. Introduction

The Internet of Things (IoT) is a major catalyst for advancements in information and communication technology, steering various industries toward greater automation and decentralized intelligence. The IoT continuously evolves and influences all aspects of our lives, much like a dynamic entity [1]. Although they have developed independently, cloud computing and IoT features are mutually supportive and enhance each other. In recent years, these two technologies have converged to form what is now known as the cloud–IoT paradigm. This convergence opens up vast opportunities for developing innovative services and applications [2]. Embedded systems based on IoT are evolving into interconnected smart devices equipped with sensors. The primary issue with smart devices was their inadequate storage capacity and processing power, which cloud computing has addressed by offering extensive processing and storage capabilities [3]. To enhance smart healthcare systems, integrating IoT sensors with cloud computing is essential. In smart healthcare systems, real-time patient monitoring is a key feature. An intelligent framework that leverages deep learning (DL) and artificial intelligence (AI) is crucial for delivering high-quality, cost-effective healthcare quickly [4]. Smart healthcare systems are increasingly utilizing affordable and advanced IoT sensors. These include wearable devices that track various health metrics such as blood pressure (BP), heart rate, respiration rate, electroencephalogram (EEG), electrocardiogram (ECG), electromyogram (EMG), blood sugar levels, and cholesterol levels [5]. The real-time data generated by IoT sensors is intricate and demands effective handling methods, including big data analytics, cloud computing, AI, and DL. Figure 1 shows types of heart diseases (HDs).

A significant challenge in smart healthcare systems, where numerous IoT sensors generate vast amounts of data, is the development of a framework that delivers cost-effective and efficient solutions for all stakeholders [6]. Advances in AI have significantly enhanced its effectiveness in healthcare. AI has simplified the diagnosis of symptoms, enabling earlier detection of diseases [7]. DL models identify diseases by analyzing patterns in data from previous patients. Additionally, these models assist healthcare professionals in identifying common causes of diseases and can help patients adjust their habits to prevent illness [8]. A significant amount of data related to heart patients is generated and stored in public databases. Therefore, there is a need to create an innovative solution that combines cloud computing and DL to accurately predict HDs. Analytics involves both quantitative and qualitative assessments of relevant data to support effective decision-making, while predictive analytics uses advanced methods to forecast future events based on existing data [9–12]. Figure 2 shows the risk factors involved in HD.

2. Literature Survey

In this section, the researchers have used different methods to develop an IoT-based HD prediction system. The existing heterogeneity problem faces the most significant challenge, and this issue arises from the various range of data source and format, such as heterogenous sensor data, different patient demographics, and various healthcare standards, which are essential for integrating to form a cohesive predictive model. From the below related works, we have analyzed the existing problems in the IoT-based HD prediction system using wearable devices. Here, Nancy et al. [21] introduced a cutting-edge smart healthcare monitoring system that merges IoT technology with cloud computing to forecast HD using DL methodologies. The system captures real-time data from wearable sensors and IoT devices to monitor cardiovascular health continuously. Umar et al. [22] presented an IoT-based smart monitoring system tailored for individuals with acute heart failure. By employing a network of IoT-enabled medical devices, the system collects and transmits essential health metrics such as heart rate, BP, and oxygen saturation to a central monitoring unit. Raju et al. [23] have proposed a smart HD prediction system that integrates IoT technology with fog computing and utilizes a cascaded DL approach. The system uses edge computing (EC) resources alongside cloud analytics to process and interpret heart health data from IoT devices efficiently. Islam et al. [24] innovated a DL-based IoT system designed for remote health monitoring and early detection of potential health issues in real time. The system utilizes IoT sensors to gather continuous physiological data from patients, which is then analyzed using advanced DL algorithms. This real-time analysis allows for early identification of health problems, enabling timely medical interventions and improving overall patient care. The author in [25] presented a Health Cloud which is an advanced system for tracking the health status of heart patients through the integration of ML and cloud computing technologies. The system aggregates health data from various sources and processes it using sophisticated ML techniques in the cloud. Zhang et al. [26] created a novel approach that combines physics-based modeling with DL techniques for the functional assessment of cardiovascular diseases (CVDs) within an IoT-enabled smart health system. By integrating domain-specific physical principles with advanced DL algorithms, the system enhances the accuracy and reliability of cardiovascular assessments. Yuan et al. [27] have introduced a robust AI-based model designed for predicting HD within the framework of the Internet of Medical Things (IoMT). The model supports both binary and multiclass classification to identify various types of heart conditions. By employing stable and scalable AI techniques, the system ensures accurate predictions and effective risk stratification, thereby enhancing early detection and management of HD. Islam et al. [28] innovated a system named as Predicts that combines IoT technology with ML to assess and predict risk levels of CVD. The system collects health data from IoT-enabled devices and applies ML algorithms to analyze this information and estimate individual risk levels. Absar et al.’s [29] research evaluates the effectiveness of a ML-enhanced smart system for predicting HD. Furthermore, the system integrates various ML techniques to analyze health data and generate predictive insights. Mishra et al. [30] proposed an IoT-enabled system for HD prediction that utilizes a three-layer DL model combined with a metaheuristic optimization approach. Moreover, the system analyzes ECG data collected from IoT devices to detect signs of HD with high precision. Dami et al. [31] explored a DL-based methodology for predicting cardiovascular events within the framework of the IoT. By applying IoT devices to collect extensive health data, the proposed approach employs advanced DL algorithms to analyze and predict potential cardiovascular issues. Reddy et al. [32] present a smart HD monitoring system that integrates IoT technology with a metaheuristic-based fuzzy-LSTM (long short-term memory) model. Additionally, the system utilizes IoT devices to gather real-time health data, which is then processed by the fuzzy-LSTM model optimized through metaheuristic techniques. This approach is aimed at enhancing the accuracy and reliability of HD monitoring and prediction, offering a sophisticated tool for proactive cardiovascular health management. Gupta et al. [33] have presented DEEP-CARDIO as a recommendation system designed for predicting CVD through an IoT network. The system collects and analyzes health data from IoT-enabled devices using DL algorithms to provide accurate predictions and recommendations for cardiovascular health. By integrating real-time data processing with predictive analytics, DEEP-CARDIO is aimed at facilitating early intervention and personalized treatment strategies for HD. Nancy et al. [34] have introduces a fog-based smart system for predicting CVD, utilizing a modified gated recurrent unit (GRU) network. The system uses fog computing to process data from IoT devices closer to the edge, reducing latency and improving response times. The modified GRU model enhances predictive accuracy by effectively capturing temporal dependencies in cardiovascular data, aimed at supporting timely and informed decision-making in cardiovascular care. Iqbal et al. [35] developed a real-time big data analytics model for healthcare, focusing on improving the quality of service (QoS) in cardiac disease prediction using IoT devices. The model processes vast amounts of health data generated by IoT sensors to enhance prediction accuracy and operational efficiency. By integrating advanced analytics with real-time data processing, the system is aimed at optimizing cardiac disease prediction and supporting high-quality healthcare delivery. Furthermore, the author in paper [36] innovated FETCH which is an advanced system that integrates DL with fog computing and IoT technology for healthcare monitoring and diagnosis. The system uses fog computing to process data closer to the source, reducing latency and enhancing real-time analysis. Roy et al. [37] have presented Conv-random forest–based IoT, a DL model that integrates convolutional neural networks (CNNs) with random forest algorithms for the classification and analysis of valvular HDs. The model utilizes IoT devices to collect detailed cardiovascular data, which is processed through a combined CNN and random forest framework. The author in [38] has introduced an IoT-based remote monitoring system designed specifically for tracking patients with heart failure. The system employs a network of IoT devices to continuously collect and transmit vital health metrics, such as heart rate and BP, to a central monitoring platform. Zhuhai et al. [39] developed a wearable ECG monitor that utilizes DL techniques for real-time detection of CVD. Moreover, the wearable device continuously records ECG signals, which are then analyzed using advanced DL algorithms to identify potential cardiovascular issues. Sam et al. [40] have presented an edge-based device for HD prediction that uses IoT technology. The device processes health data from IoT sensors directly at the edge of the network, minimizing latency and enhancing prediction accuracy.

As medical datasets in the healthcare sector continue to expand in size and complexity, there is an increasing need for automated systems powered by DL technologies to help medical professionals make accurate and efficient decisions. Despite this progress, a significant challenge for DL methods is maintaining accuracy when managing extensive datasets. To solve this issue, LIFE-CARE handles the heterogeneity of data sources by providing a flexible edge-based and multicloud architecture. This system effectively manages data variability and ensures that predictions remain accurate and relevant across different healthcare settings and device types. Overall, LIFE-CARE represents a significant advancement in creating a robust, adaptive, and collaborative HD prediction system. Moreover, the model employs a CTL-ST, which addresses the need for using knowledge from multiple models to improve predictive performance. Finally, this collaborative approach allows for continuous model updates and improvements through knowledge distillation, ensuring the model evolves with new data and insights. Table 1 compares the existing and proposed research gaps.

3. Preliminaries

In fault analysis application of unsupervised domain adaptation, diverse source of training data with known labels, along with the unlabeled testing data, can be gathered from various equipment operating conditions. Precisely, the source domain data y^so from $$ are drawn from different but related distributions $$, where $$ represents the dataset for the $$th source domain dataset and $$ indicates the number of samples in the source domain. The label $$ corresponds to the health status, with ℵ denoting the total number of categories in the health status. The domain of the target dataset $$ is drawn from the distributed target $$, where $$ denotes the number of samples in the target domain, $$. Unlike the source domains, which have known status labels, the target domain lacks these labels but shares the same label space, a scenario referred to as closed-set domain adaptation. The goal of cross-domain fault diagnosis is to develop an intelligent network model $$ that can learn domain-invariant and distinguishable features by minimizing the discrepancy between distributions. This approach is aimed at reducing the risk of incorrect predictions on the target domain, $$ with unsupervised domain adaptation. In HD prediction, IoT devices such as wearable heart monitors and health trackers are instrumental in gathering continuous health data. Integrating these devices with transfer learning techniques can significantly enhance predictive models for CVD conditions. Transfer learning involves adapting a model that has been trained on extensive datasets to a specific, smaller dataset related to heart health. This method is especially useful in medical applications where acquiring large amounts of labeled data can be challenging. For example, a model trained on general health information can be fine-tuned using data from IoT devices focused on cardiovascular metrics. This approach not only improves the model’s ability to identify early signs of HD but also minimizes the need for extensive computational resources and training time. Transfer learning can also aid in detecting rare heart conditions by applying insights from models trained on related health issues. By combining IoT data with transfer learning, healthcare systems can provide more accurate and timely HD predictions, allowing for early intervention and personalized care. This innovative integration of IoT and transfer learning is advancing the field of preventative healthcare, making it more responsive and tailored to individual patient needs. Figure 3 shows the overall architecture of the proposed LIFE-CARE model.

4. LIFE-CARE Model

This section explains the IoT–cloud-enabled heart prediction framework for HD patients. This research is intended to simplify the workload of medical professionals using IoT, EC, and cloud computing technologies, respectively. In order to monitor the patient’s heart status, we have utilized sensors such as blood pressure sensor (BPS), temperature sensor (TS), blood glucose sensor (BGS), heart rate sensor (HRS), and ECG signal. Technically, for every patient pa^j = {pa¹, pa², ⋯, paⁿ}, we acquire patient-specific heart status analysis which is forwarded to the distributed Gw. The gathered medical IoT data from the GW^j are forwarded to DES^j for data preprocessing. The sensed medical data of the patients by GW^j are represented as Y_d, d = 1, 2, ⋯, D. Y_d is provided to DES^j which performs preprocessing in dual folds such as data filtering and data normalization, respectively. The preprocessed data is then forwarded to MCS^j for HD prediction.

4.3. Collaborative HD Prediction Layer

The preprocessed heart rate data points s_j are then provided to the CTL-ST for automated HD prediction. Initially, we transform every s_j into 2D frequency–time matrix for generating image representations. For that, we calculate short-time Fourier transform (STFT) as

$$ ()

From the above equation, the window function of the signal is denoted as wi[m], the discrete input signal is denoted by s_j[m], the size of hop is denoted as G, and M_s denotes the length of STFT. After that, we calculate the squared magnitude of spectrogram as STFT and mapping the mel scale frequency.

The designed CTL-ST is composed of dual sequential transformer modules, domain adaption, and knowledge blocks, respectively. The dual sequential transformer blocks accept varied dimension tokens. The tokens can be separated based on the time axis which is illustrated in Figure 4. The architecture depth can be increased F times by the proposed CTL-ST. The processed token is then provided to the multilayer perceptron (MLP) for final HD prediction result.

4.3.1. Linear Projection and Tokenization

The input to the initial stage is spectrogram with $$ frequency bins and $$ time slots. The proposed research divides the spectrogram input into k · m token with size of v × v. The size of patch considered in this work is 1 × 1 for processing patches in the finest level with better visualization. The token patches are further provided to the linear projection block which performs token projection into dim-dimensional vectors. The linear projection tensor output can be denoted as T ∈ ℝ^m×k×dim in which T_j,i denotes the projected token and latent dimension is denoted by dim ∈ M.

4.3.2. Upright Transformer

The projected tokens are separated into subsamples in time domain and can be denoted as T_:,i = [T_1,i, T_2,i, ⋯, T_k,i] ∈ ℝ^k×dim which thereby obtain m data samples. Note that every data sample contains k tokens. Furthermore, the token class T_[CLS] can be replicated to perform positional embedding on every token. The additional m individual data samples are further processed by the upright transformer.

4.3.3. Horizontal Transformer

The data sample output by the upright transformer is denoted by $$. The attained data samples are amalgamated into tensor T^ver ∈ ℝ^m×k×dim. Furthermore, the class token can be decoupled and then average pooling applied to the class token as $$. The tensor T^ver can be separated into frequency domain of data subsamples as $$. The subsamples can be generated to form m tokens. Similar to the upright transformer, the token class can be replicated $$ in k times to every data sample $$. We additionally performed position embedding on every token in which the obtained samples are provided as an input to the horizontal transformer.

4.3.4. Transformer Module

The operations involved on the upright and horizontal transformer are similar and varied by input data sample format. The sequence of n tokens (i.e., $$ or T_:,i) can be denoted as γ ∈ ℝ^n×dim in which n ∈ {k, m}, and dim is the dimension embedding of every token. The multihead attention module is denoted by MHA, normalization layer is denoted as Norm, MLP is denoted by MLP, and the auxiliary tensors are denoted as ϱ, ϖ, respectively. The formulation is provided as below:

$$ ()

$$ ()

The main aim of the transformer module is to capture the long range dependencies and contextual information from the heart rate data. The MHA performs well in attaining contextual information which is composed of triple learnable weight matrices as $$ for attaining query, key, and values, respectively, from γ. Otherwise, γ is initially projected to the weighted matrices to obtain Qu = γ · wei^Qu, ke = γ · wei^Ke, Va = γ · wei^Va. The output by the self-attention can be formulated as

$$ ()

where R denotes the self-attention output, with $$ as the transpose of ke.

4.3.5. Domain Adaption Layer

In HD prediction context-based domain adaption, we utilize maximum mean discrepancy (MMD) distance measure for computing the feature distribution among varied domains of similar classes. Mostly, the MMD methods are utilized for global feature distribution alignment and can be formulated as

$$ ()

From the above equation, λ denotes the reproducing kernel Hilbert space (RKHS) which is embedded with the k-th kernel in which ker(y^so, y^ta) = 〈θ(y^so), θ(y^ta)〉 in which 〈..〉 denotes the inner vector products. The feature mapping function is denoted as θ(.) which denotes the feature mapping function in the source domain to the RKHS. y^so and y^ta denotes the instances of source and target domains, respectively.

For the case of MMD, aligning similar feature distributions is often complex and leads to sparse feature distribution. To overcome that, LMMD is adopted which firmly aligns the feature of similar subdomain in local perspective. The formulation is provided as below:

$$ ()

From the above equation, $$ and $$ denote the distribution of k-th class in Γ_so and Γ_ta, respectively. The above equation is considered as a loss function of the propounded CTL-ST model. From the optimization of loss function in iterative fashion, we attain that the similar class distribution of every related subdomain is close to each other. The above equation can be recomputed as

$$ ()

From the above equation, $$ and $$ denote the weights of probabilities which belong to similar classes. Furthermore, $$ and $$ are equal to 1, and weighted sum of class k is denoted by $$. More clearly, for a sample y_j, $$ can be formulated as

$$ ()

From the above equation, for class k, the probability sample y_j is denoted as $$. For instance, y_j comes from Do_so; $$ can take the one hot vector label of the k-th term, whereas if y_j comes from the target domain Do_ta, it had not the one hot vector label. By adding domain adaption layer in the CTL-ST, the loss among the source and target domains is the same. But for varied domains, the formulation is defined as below:

$$ ()

From the above equation, Fe(.) denotes the function of feature extraction. By diminishing the domain loss in the CTL-ST, the similar distribution classes are firmly aligned better to form better predictions.

5. Experimental Results

5.4. Comparative Analysis

The proposed work is evaluated according to the prediction accuracy with cutting-edge techniques which attach the dataset of HD. Table 3 lists the analysis comparison study of the proposed LIFE-CARE model accuracy with existing models as per maximizing accuracy. Figure 5 displays the comparative result of the proposed model with state of the art. The result of comparison with correlated existing HD diagnosis systems shows that the performance of the proposed system exceeds that of prior works.

Table 3. Detailed comparison with existing works.

Ref. no.	Method/algorithm	Accuracy (%)
[21]	DL	85.5
[22]	Real-time data collection and analysis using IoT sensors	86.2
[23]	ML with DL	80.15
[24]	DL to analyze real-time data from IoT devices	90.6
[25]	ML algorithms process health data in the cloud	95.8
[26]	Integrates physical models with DL for data analysis	93.5
[27]	AI models for classification of HD types in IoMT environments	92.4
[28]	ML algorithms analyze IoT sensor data to assess risk levels	90.6
[29]	ML algorithms analyze patient data for predictive accuracy	93.6
[30]	Three-layer DL model combined with metaheuristic optimization	95.6
[31]	DL	94.99
[32]	Combines fuzzy logic with LSTM networks optimized by metaheuristic methods	96.5
[33]	IoT network data analyzed with DL for prediction and recommendations	95.6
[34]	Modified GRU processes data at the fog layer	97
[35]	Big data analytics model processes real-time data from IoT devices.	97.5
[36]	Mask-guided attention U-Net	98
[37]	CNN with random forest	98.5
[38]	Remote monitoring of patient health through IoT sensors	96.5
[39]	Wearable ECG device combined with DL for real-time detection	97.5
[40]	Edge computing device processes IoT data for HD prediction	98.2
Proposed	CTL-ST	98.92

Large-scale IoT-driven tasks in healthcare systems demand rapid processing due to their context-sensitive and time-critical nature. The rapid growth in IoT devices and the data they generate have led to a significant increase in data flow, which places considerable strain on service and bandwidth resources. The centralized cloud computing model struggles with connectivity and latency issues, as it cannot efficiently manage these challenges on its own due to inherent limitations in handling large-scale data and bandwidth demands.

To address these issues, decentralized EC models have been developed, which enable computation and storage to occur closer to the data source. Moreover, this model adapts MCS with collaborative transfer learning capabilities. This edge–cloud hierarchy significantly reduces delays and mitigates latency, providing a more efficient framework for managing IoT data.

5.5. Result and Discussion

The experimentation is conducted to estimate the proposed model with changing number of samples reaching from 10% to 100% on generic learning approach including ML, DL, and the proposed model CTL-ST method. Table 4 illustrates the performance analysis in basis of evaluation metrics of accuracy, precision, recall/sensitivity, specificity, and F1 score of ML, DL, and CTL-ST models. Besides, Figures 6a, 6b, 6c, 6d, and 6e signify the impact of accuracy, precision, recall, specificity, and F1 score attained by ML, DL, and CTL-ST. The records are maximized from 10% to 100% for the experiment analysis of considered three learning approaches.

Table 4. Comparison of CTL-ST models on conventional ML and DL models.

Data (%)	Accuracy			Precision			Recall
	ML	DL	CTL-ST	ML	DL	CTL-ST	ML	DL	CTL-ST
10	90.41	92.26	94.89	90.62	92.33	94.85	90.57	92.42	94.99
20	90.58	92.68	94.92	90.81	92.78	94.96	90.92	92.73	95.05
30	90.76	92.45	95.64	90.99	92.49	95.66	90.68	92.88	95.68
40	90.93	92.88	95.36	90.74	92.54	95.24	90.75	92.53	95.34
50	90.89	93.74	95.78	91.02	93.29	95.45	91.08	93.95	95.19
60	91.05	93.56	96.23	91.25	93.86	96.23	91.36	94.09	96.13
70	91.55	93.82	96.86	91.60	93.68	96.47	91.14	94.34	96.76
80	91.14	93.93	97.33	91.44	93.98	97.02	91.68	94.21	97.44
90	91.78	94.58	97.16	91.82	94.42	97.68	91.74	94.63	97.09
100	92.08	94.96	98.72	92.13	94.73	98.25	92.35	94.85	98.79
Data (%)	Specificity					F1 score
	ML	DL		CTL-ST		ML	DL	CTL-ST
10	90.01	92.28		94.42		90.11	92.35	94.58
20	90.68	92.78		94.67		90.57	92.63	94.74
30	90.53	92.43		95.95		90.74	92.82	95.69
40	90.38	92.57		95.29		90.42	92.49	95.33
50	90.73	92.89		95.68		90.85	92.92	95.98
60	91.25	93.14		96.38		91.29	93.56	96.47
70	91.62	93.88		96.85		91.67	93.22	97.43
80	91.81	93.30		97.52		91.58	93.94	96.92
90	91.54	94.38		97.23		91.82	94.42	97.68
100	91.96	94.69		98.55		91.99	94.72	98.78

The result of analyzing the performance metrics of the proposed CTL-SL, ML, and DL approaches exposes that the proposed model outperforms the performance of other two learning approaches. Overall performances of the proposed CTL-ST model, ML, and DL methods are compared in Table 5 and Figure 7. Considering multiple performance standards, it is judicious to suppose that the proposed work outperforms other learning methods. In terms of predicting cardiac disease, ML and DL techniques have made considerable strides. The strong performance metrics of traditional ML techniques, with their fewer complex algorithms, are demonstrated by their 95.17% accuracy, 95.19% precision, 95.28% recall, 95.58% specificity, and 95.69% F1 score. These techniques, however, are constrained by the fact that they frequently necessitate manual feature engineering and are unable to automatically extract complicated features from raw data. Furthermore, different kinds of medical data, such those from different sensors utilized in IoT-enabled environments, can be difficult for ML models to integrate. When fresh data becomes available, these constraints may hinder their capacity to adjust and get better over time.

Table 5. Performance comparison of proposed learning approach.

Performance metrics	ML	DL	CTL-ST
Accuracy (%)	95.17	97.23	98.72
Precision (%)	95.19	97.30	98.76
Recall (%)	95.28	97.43	98.84
Specificity (%)	95.58	97.66	98.86
F1 score (%)	95.69	97.82	98.92

DL techniques, on the other hand, provide better results on all metrics: 97.23% accuracy, 97.30% precision, 97.30% recall, 96.66% specificity, and 97.82% F1 score. The power of DL is found in its capacity to automatically extract hierarchical characteristics from large, complicated datasets, hence improving its prediction power. Despite these benefits, DL models can be difficult to comprehend and frequently require large amounts of data and computer power. The suggested model, “LIFE-CARE,” outperforms both ML and DL techniques with remarkable accuracy of 98.72%, precision of 98.76%, recall of 98.84%, specificity of 98.86%, and an F1 score of 98.92%. It does this by utilizing CTL-ST. The strength of this model lies in its collaborative knowledge distillation technique, which offers more accurate and robust predictions by continuously improving the prediction capabilities through shared learning across models. This iterative improvement combines the advantages of edge-based and cloud-based processing to constantly increase performance, thereby addressing the shortcomings of both ML and DL.

5.6. Real-Time Performance Analysis of the LIFE-CARE Model

In order to examine the applicability of LIFE-CARE model on real-world settings, we create a hypothetical scenario with respect to number of patients and IoT devices on varied combinations. To be clearer, we also show how the designed model performed on varied number of patient volume and IoT devices, respectively.

5.6.1. Scenario Arrangement

Hypothetically, we took about 1000 patients in which every patient is equipped with about 1–8 IoT devices that include cholesterol level, blood sugar, respiration rate, EMF, EEG, heart rate, and BPSs, respectively. Furthermore, we also integrate patient corresponding EHRs from the hospital database. We utilize real-time metrics for assessing the model performance such as the following:

• •

  Scalability—the capability of the designed model to effectively withstand with amount of IoT devices and patients without compromising the model performance

• •

  Model update frequency—the amount of frequency at which the designed model updates/transfers knowledge from another model

• •

  Processing time—the amount of time taken to examine and generate model predictions

• •

  Prediction accuracy—the percentage of precise HD prediction results

5.6.2. Hypothetical Result Analysis

In this supplementary section, we built four hypothetical scenarios with varied number of patients and IoT devices. Here is a detailed breakdown of the scenario analysis.

5.6.2.1. Scenario I: 100 Patients With 2 IoT Devices

In this scenario, patients utilized heart rate and BPS, respectively. As the patient number and IoT devices are limited, the scalability of the model is relatively higher with 0.4 s as the processing time. The frequency of model update is about every 8 s with prediction accuracy of 90%.

5.6.2.2. Scenario II: 500 Patients With 4 IoT Devices

In this scenario, patients utilized cholesterol level, ECG, heart rate, and BPS, respectively, for HD status acquisition. As the devices and patients get slight increases than Scenario I, we attain only moderate model performance with processing time of 1.5 s per patient, frequency of model update about 15 s, and prediction accuracy of 93%. From the analysis, adding more IoT devices can significantly enhance the prediction accuracy with higher processing time and model update frequency than Scenario I.

5.6.2.3. Scenario III: 1000 Patients With 6 IoT Devices

In this scenario, patients utilized cholesterol level sensor, respiration rate sensor, EEG sensor, ECG sensor, and BPS for HD data acquisition. As the patients attain their maximum limit of 1000 and devices increase, we obtain model update frequency of 28 s with 1.7 s as processing time per patients and prediction accuracy of 97%. By examining this scenario, the prediction accuracy is high than the other two scenarios with higher processing time and model update frequency. The scalability is better; however, higher patients and devices adversely affect the model update frequency.

5.6.2.4. Scenario IV: 1000 Patients With 8 IoT Devices

In this scenario, we adopt cholesterol level sensor, blood sugar sensor, respiration rate sensor, EMG sensor, EEG sensor, ECG sensor, HRS, and BPS, respectively. As the patients and IoT devices reach the scenario maximum limit, there is higher processing time and model update frequency of 3 s and 50 s, respectively, whereas the accuracy attained is 99%. Firmly, there is increment in prediction accuracy and decrement in model scalability in terms of processing time and update frequency, respectively, due to larger number of devices.

In summary, we obtain higher prediction even though the number of IoT devices and patients gets increased. The reason for such higher prediction accuracy is the availability of comprehensive data. In terms of processing time and model update frequency, the performance gets decreased due to increment in patient and IoT devices. The scalability of model shows better on certain limit (i.e., 500 patients with 4 IoT devices) and lags with beyond threshold limit. Table 6 and Figures 8a, 8b, and 8c show the real-time analysis of LIFE-CARE model.

Table 6. Applicability analysis of the proposed model.

# of patients	# of IoT devices	Model update frequency (s)	Processing time (s)	Prediction accuracy (%)
100	2	8 s	0.4 s	90%
500	4	15 s	1.5 s	93%
1000	6	28 s	1.7 s	97%
1000	8	50 s	3 s	99%

6. Conclusion

In this research, we adopt an IoT–multicloud-enabled heart prediction system for HD prediction using real-time sensor data. This research adopts medical IoT sensors and EHRs for acquisition of patient health status. The acquired data are then passed to the DES via Gw for preprocessing. The preprocessed data is then provided to the MCC for HD prediction using CTL-ST algorithm. Our model resolves the problem of heterogeneity by accepting the varied input of different patient and acquiring knowledge from the other models through TL and domain adaption techniques. By examining the hearty failure factors of the patients, the survival rate of the patients can be improved through timely prediction.

The effectiveness of the designed CTL-ST is trained on heart failure UCI dataset with various performance metrics. The results show that the proposed work outperforms the existing works.

Disclosure

No third-party individuals or external services were involved in the research or preparation of the manuscript.

Conflicts of Interest

The authors declare no conflicts of interest.

Author Contributions

We confirm that all authors have contributed significantly to the manuscript preparation and research activities. E. Michael Priya: conceptualization, methodology, experimentation, validation, and final approval. K. Sundara Krishnan: data analysis, writing (review and editing), and visualization.

Funding

No funding was received for this research.