An Optimized Model for Cardiovascular Risk Prediction in Wearable Devices

Cardiovascular disease prediction has evolved significantly with advances in wearable sensing, Internet of Things (IoT) systems, and machine-learning methodologies. Early research focused on traditional statistical and machine-learning models applied to structured clinical datasets. Recent work has leveraged deep-learning architectures capable of extracting complex temporal and physiological patterns from continuous sensor data. Parallel developments in TinyML and edge computing have further enabled resource-efficient deployment of intelligent health-monitoring systems on wearable and low-power devices. This literature review synthesizes prior contributions across these domains, highlighting key methodologies and limitations.

Bhatt et al. [3] highlighted the continued relevance of classical machine-learning methods such as logistic regression, decision trees, and Support Vector Machines (SVM) for heart-disease prediction due to their interpretability and suitability for structured clinical data. Their approach incorporated essential preprocessing steps, including feature selection and discretization (e.g., binning age, blood pressure, and cholesterol), which helped capture nonlinear relationships and improved classifier performance. They also employed k-modes clustering to identify latent patterns in categorical medical data, further enhancing predictive accuracy.

Traditional ML approaches have additionally been applied to wearable-sensor data. Siirtola et al. [4] and Martin-Gonzalez et al. [5] used Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA) to predict cardiac-related events from sleep-sensor signals. While initial user-independent models suffered from variability in individual sleep patterns, personalized models significantly improved performance by tailoring decision boundaries to each user’s physiological characteristics.

Wearable and IoT technologies are rapidly reshaping cardiovascular monitoring by enabling continuous, non-clinical data collection and analytics. Singhal and Cowie [6] reviewed how wearables support heart-failure management, highlighting their promise and current limitations such as data validity and clinical integration. Perez-Pozuelo et al. [7] demonstrated wearable devices could detect sleep outside the clinic, supporting scalable early anomaly detection.

Alday et al. [8] presented the 2020 PhysioNet challenge on 12-lead ECG classification, underlining the integration of sensor data into predictive workflows. Lin et al. [9] provided a detailed survey of wearable sensors and devices for real-time cardiovascular disease monitoring, covering hardware, signals — ECG, Photoplethysmography (PPG) — and deployment platforms. De Zambotti et al. [10] showed how a commercially available wristband could capture sleep and cardiac function in adolescents, confirming the feasibility of consumer wearables in cardiovascular research. Sarmah [11] reported an IoT-based deep-learning system for heart-disease prediction, combining device data with modified neural networks, but warned of system-level risks like latency and data quality.

Ali et al. [12] proposed an ensemble deep-learning and feature-fusion approach using wearable and IoT-derived variables for heart-disease prediction, offering strong accuracy yet raising issues of interpretability. Kundrick et al. [13] applied machine learning to wearable fitness tracker data to predict hospitalizations and cardiovascular events. This demonstrated improved risk stratification using continuous real-world physiological signals. However, the study was limited by device-specific data and cohort dependence, which might affect generalizability.

Deperlioglu et al. [14] applied an autoencoder–Deep Neural Network (DNN) in a secure Internet of Health Things (IoHT) framework for disease diagnosis, trading transparency for performance. Pakhomov et al. [15] integrated electronic medical records with sensor data for heart-failure identification, improving prediction depth but facing interoperability and privacy hurdles.

Recent research on deep-learning approaches for heart-disease prediction has demonstrated significant advances in representation learning, feature augmentation, and hybrid network design. García-Ordás et al. [16] showed that deep neural architectures combined with feature augmentation substantially improved prediction accuracy on structured clinical datasets. Alqurashi et al. [17] integrated Predator Crow Optimization with deep neural networks to automate hyperparameter tuning and feature selection, achieving strong performance but at the cost of increased computational complexity.

Mohammad and Al-Ahmadi [18] proposed a hybrid Wavelet Transform–Convolutional Neural Network (WT-CNN) model that extracted wavelet-based ECG features before applying convolutional layers, reporting accuracy levels near 97%. This work highlighted the value of signal-driven deep learning. Shankar et al. [19] demonstrated that even relatively simple CNN architectures could outperform classical machine-learning models on heart-disease datasets when trained with proper regularization.

Ram Kumar et al. [20] introduced a hybrid CNN–DNN architecture that combined convolutional feature extractors with dense layers for final classification, achieving improved predictive performance but also noted potential overfitting risks on limited datasets. Collectively, these studies confirmed that deep-learning models can capture nonlinear feature interactions more effectively than traditional methods, while also emphasizing ongoing challenges related to dataset size, model interpretability, and robust cross-dataset validation.

Subashini and Kanaka Raju [21] presented an IoT-based heart-disease diagnosis framework that integrated physiological data collected via smart sensors with a hybrid learning pipeline combining gradient boosting for tabular features and a deep convolutional neural network for improved classification accuracy; however, the study was limited by evaluation on a restricted dataset and the lack of external clinical validation, which might affect generalizability. Abutalip et al. [22] proposed a machine-learning–driven heart-disease detection system using data acquired from wearable devices, where conventional ML classifiers were employed to identify cardiovascular risk patterns, but the work was constrained by limited feature diversity and short-term wearable data, reducing robustness across populations. Al Reshan et al. [23] developed a robust heart-disease prediction approach using hybrid deep neural networks, including CNN, LSTM, and a combined CNN–LSTM architecture, achieving high predictive performance on multiple benchmark datasets. Nevertheless, the model’s reliance on curated public datasets and increased computational complexity posed challenges for real-time deployment and clinical interpretability.

Xia et al. [24] proposed an intelligent cardiovascular disease diagnosis framework that integrated Ant Colony Optimization for feature selection with a deep-learning–enhanced neural network, further optimized using Bayesian hyperparameter tuning to improve classification accuracy. Although the approach achieved strong performance, the combined use of metaheuristic optimization and deep learning increased computational complexity and training cost, which might limit its suitability for real-time or resource-constrained healthcare applications.

To make these models practical for wearables, researchers have applied TinyML techniques. Arooj et al. [25] explored structured pruning to reduce model complexity, while Neri et al. [26] showed that 8-bit quantization significantly reduced memory use and power consumption without major accuracy loss. Qureshi and Krishnan [27] demonstrated that TinyML-compatible CNN–LSTM models could deliver accurate, real-time predictions on microcontrollers, balancing performance with energy efficiency.

Recent work by Sun et al. [28] introduced a TinyML methodology for continuous, cuff-less blood-pressure estimation using only PPG signals. Their approach shrunk conventional CNN architectures (AlexNet, LeNet, SqueezeNet, ResNet, MobileNet) via pruning and quantization, and deployed them on constrained edge platforms. Their evaluation used thousands of ICU patient records and showed performance comparable to server-based systems while meeting the Association for the Advancement of Medical Instrumentation and the British Hypertension Society (AAMI/BHS) standards. The memory footprint was reduced to <1 MB and inference latency to around 10 ms on a Cortex‑M microcontroller, although generalization across ambulatory settings remained a concern. Their work demonstrated the viability of TinyML for cardiovascular monitoring but highlighted the trade-off between model size and inter-subject robustness.

Mahardika et al. [29] proposed a CNN–LSTM architecture trained on the MIMIC‑III arterial-blood-pressure (ABP) and PPG dataset. Their optimized configuration (5 convolutional layers + 1 LSTM + 2 dense layers) achieved a mean absolute error (MAE) of 7.89 ± 3.79 mmHg for systolic BP and 5.34 ± 2.89 mmHg for diastolic BP — meeting the AAMI and BHS limits. Deployment considerations included window segmentation (4 s, 500 points) and Principal Component Analysis (PCA) feature reduction. While the accuracy was promising, the architecture was still too heavy (around 10 M parameters) for ultra‑low‑power microcontroller deployment without further compression or dedicated hardware acceleration. This work underscored the value of hybrid CNN–LSTM models for cardiovascular regression tasks but also emphasized the need for TinyML‑aware optimization.

Arthi and Krishnaveni [30] proposed a fog-enabled TinyML with explainable‑AI pipeline for healthcare decision support. Their system achieved an F1‑score of 0.93 for abnormal‑health‑event detection while employing Modified Lempel‑Ziv‑Welch (mLZW) data compression and Lightweight Shapley Additive explanations (SHAP) on edge/fog nodes. The paper reported memory usage of around 800 kB and latency of 30 ms for anomaly inference. However, the study used general health‑sensor features rather than specialized cardiovascular signals and lacked prospective wearable deployment. It provided a valuable proof‑of‑concept for combining compression, interpretability, and TinyML in health monitoring.

Elhanashi et al. [31] provided a comprehensive survey of TinyML in embedded and IoT‑based healthcare applications, covering more than 150 papers up to 2024. The review reported that the median parameter count of deployed TinyML models in health was around 35 k parameters, average memory footprint 256 kB, and average latency 20 ms. It also cited key challenges: lack of standardized evaluation benchmarks (e.g., sensitivity, calibration), limited cross‑device generalization, and weak clinical validation. Their conclusions emphasized that, although TinyML is technically feasible and deployment‑ready, clinically validated TinyML systems for cardiovascular applications remain scarce. Our work builds on these advances by deploying a compressed CNN–LSTM model for multi‑class cardiovascular risk prediction, reporting clinically relevant metrics (sensitivity, specificity, calibration) and validating on edge hardware under realistic constraints.

For fair comparison, the standalone CNN and LSTM baselines were implemented using standardized and lightweight architectures aligned with prior deep-learning studies on physiological signal analysis. The CNN model consisted of three 1D convolutional layers (Conv1D: 64 filters, kernel size 3; Conv1D: 128 filters, kernel size 3; Conv1D: 256 filters, kernel size 3), each followed by ReLU activation and max-pooling, and a final dense layer for classification.

The LSTM model comprised two stacked LSTM layers with 128 and 64 units, respectively, followed by a fully connected classification layer. Both models were trained using the same preprocessing pipeline, Adam optimizer, learning-rate schedule, batch size, and early-stopping strategy as the proposed CNN-LSTM. All models were evaluated using stratified 5-fold cross-validation, and performance metrics were reported as mean ± standard deviation across folds. The detailed results of this comparison are provided in Table 3 and illustrated in Figure 6.

Table 3. Performance metrics comparison with state-of-the art models

Figure 6. Performance metrics comparison with state-of-the art models

To examine the feasibility of deploying the optimized CNN-LSTM model in a real-world edge environment, we conducted performance testing on a Raspberry Pi 4 Model B (4GB RAM), a widely used embedded platform suitable for simulating wearable-device workloads. The Raspberry Pi 4 features a 1.5 GHz quad-core ARM Cortex-A72 CPU, LPDDR4 memory, and support for Python-based edge frameworks such as TensorFlow Lite, making it a suitable intermediate platform for validating lightweight machine-learning models before migration to ultra-low-power microcontroller units (MCUs).

The pruned and quantized model (190MB) was deployed on the Raspberry Pi using TensorFlow Lite, where we measured inference latency, memory consumption, and CPU usage during continuous streaming of wearable-like physiological data. The device operated within acceptable thermal and power limits, and no throttling was observed during continuous execution, demonstrating that the optimized model is capable of real-time cardiovascular risk inference in a mobile-edge environment. Although the Raspberry Pi exceeds the resource constraints of commercial wrist-worn devices, it provides a realistic and controlled environment for evaluating embedded performance, identifying bottlenecks, and guiding further model compression or distillation for future deployment on low-power MCUs.

In addition to computational performance, we addressed key engineering and regulatory considerations essential for wearable medical systems. Since inference is performed entirely on-device, user physiological data does not need to be transmitted to cloud servers, reducing privacy and security risks.

The performance of the proposed CNN-LSTM model was assessed not only in terms of classification accuracy but also with respect to its suitability for deployment on resource-constrained wearable devices.

Figure 7. Memory usage

Figure 8. Model’s accuracy

Figure 7 shows the memory usage across the baseline, pruned, and quantized versions of the model. The unoptimized baseline required approximately 610MB of memory, which is far beyond the capacity of typical edge devices. With pruning, memory consumption dropped to 370MB, and further quantization reduced it to just 190MB-representing a cumulative reduction of nearly 70%. This dramatic decrease in memory footprint demonstrates the effectiveness of TinyML techniques in preparing deep learning models for embedded environments where hardware limitations pose a significant concern.

Figure 8 highlights the classification accuracy of each model version. The baseline model achieved an accuracy of 95.27%, which decreased to 73.7% after pruning, then the fine-tuning process restored the accuracy back to 95.23% and the fine-tuning after quantization resulted in an accuracy of 95.14%. The results indicate that the core predictive capability of the CNN-LSTM architecture is largely preserved, even after aggressive model compression.

Figure 9. Power consumption and inference time

Figure 9 compares power consumption and inference time for the three configurations. The baseline model exhibited the highest power usage at 16.74 watts and the longest inference time at 71 seconds, which is impractical for continuous monitoring on battery-powered devices. The pruned model showed substantial improvement, reducing power draw to 10.2 watts and processing time to 50 seconds. The quantized version performed even better, consuming only 7.9 watts and completing inference in 48 seconds. These gains in speed and energy efficiency underscore the viability of TinyML-optimized deep learning for wearable deployment, enabling fast, on-device inference while conserving battery life.

Table 4. Performance metrics comparison (V1: Baseline CNN-LSTM, V2: Pruned CNN-LSTM (After fine-tuning), V3: Quantized CNN-LSTM (After fine-tuning))

Table 5. Performance with statistically grounded metrics

Table 4 presents a comparison of clinically relevant diagnostic metrics for the baseline, pruned, and quantized CNN-LSTM models. The baseline model demonstrates the strongest overall performance, with high sensitivity and specificity, indicating excellent ability to correctly identify both at-risk and normal individuals. The pruned model retains performance very close to the baseline, showing that model compression does not significantly affect its diagnostic reliability. The quantized model exhibits a slight reduction in sensitivity and specificity, but still maintains strong AUC values, confirming that it continues to provide clinically meaningful discrimination between risk levels.

Table 5 summarizes the performance of the three model variants with statistically grounded metrics. The baseline model achieved the highest accuracy (95.27%±0.28, 95% CI: 94.81-95.73) and AUC (0.963±0.004, CI: 0.956-0.970), while the pruned and quantized models showed only marginal reductions across precision, sensitivity, specificity, and AUC. The narrow confidence intervals and low standard deviations across all metrics indicate strong stability of the results and confirm that pruning and quantization introduced only minimal performance degradation. These statistical measures reinforce the reliability and robustness of the proposed TinyML-optimized model versions.

To ensure clinical relevance and dataset independence, the proposed model was further validated on the MIMIC-IV Waveform Database (PhysioNet), which contains clinically recorded ECG, PPG, arterial blood pressure (ABP), respiration, and SpO₂ signals.

All validation data underwent comprehensive preprocessing. This included noise filtering, baseline correction, and resampling. Following preprocessing, feature extraction was performed to compute heart rate, heart rate variability (HRV), blood pressure indices, oxygen saturation, and respiration rate. These features were consistent with the feature set used in the Kaggle Wearables dataset.

These features were normalized using the same scaling parameters to maintain uniformity across datasets. The model achieved an average accuracy of 93.8%±0.3, precision of 91.6%±0.3, recall of 93.1%±0.3, F1-score of 92.3%±0.3, and AUC of 0.953±0.002, confirming its ability to generalize effectively from wearable to clinical data. These results validate that the proposed CNN-LSTM model remains stable and accurate across heterogeneous signal sources when appropriate preprocessing and feature alignment are applied.

Table 6. Cross-validation results of the proposed CNN-LSTM model on MIMIC-IV dataset

Table 6 explores the impact of varying learning rates on the quantized model’s accuracy, precision, recall, and F1-score. The five-fold cross-validation results demonstrate stable and consistent model performance across all evaluation metrics. The model achieved a mean accuracy of 93.8% ± 0.3%, with precision, recall, and F1-score of 91.6% ± 0.3%, 93.1% ± 0.3%, and 92.3% ± 0.3%, respectively, indicating a balanced classification performance. Additionally, the high mean AUC of 0.953 ± 0.002 reflects strong discriminative capability and robustness of the model across different validation folds.

Figure 10. Calibration curves for model variants

The calibration curve in Figure 10 shows that all three models produce well-calibrated probability estimates, with their curves closely following the diagonal reference line. At low predicted probabilities (0.0-0.3), all models slightly underestimate risk, which is expected in datasets dominated by normal cases. In the mid-range (0.5-0.7), the observed frequencies rise sharply, indicating that predicted probabilities accurately correspond to actual risk levels. At higher probabilities (0.8-1.0), the baseline CNN-LSTM aligns most closely with the ideal line, while the pruned and quantized models show only minor deviation. These results confirm that compression minimally affects calibration quality and that all three models provide reliable probability estimates for cardiovascular risk prediction.

The confusion matrices in Figures 11(a)-(c) reflect classification performance across the four risk classes. Each model demonstrates high diagonal concentrations, with correctly identified samples. Off-diagonal entries remain comparatively low, typically within the range of 1 to 15 samples, indicating limited dispersion of predictions across neighboring categories.

Table 7 highlights how memory usage, inference time, and power consumption vary across the three model stages. The baseline CNN-LSTM model is the most resource-intensive, requiring 610MB of memory and consuming over 16 watts of power. Pruning substantially reduces these requirements, cutting memory needs to 370MB and decreasing power usage by almost 40%. Quantization further optimizes the model, reducing memory to 190MB and lowering power consumption to below 8 watts. These optimizations make the model feasible for deployment on energy-limited wearable devices.

The CNN-LSTM model was trained using the Adam optimizer with a learning rate of 0.001, batch size of 32, and default momentum parameters (β₁=0.9, β₂=0.999) with ε=1e-8. Training was performed for a maximum of 100 epochs using the categorical cross-entropy loss function. To prevent overfitting and ensure stable convergence, an early-stopping strategy was applied with a patience of 10 epochs and a minimum required improvement of 0.001 in validation loss. Additionally, model checkpointing was enabled to automatically save the best-performing model based on validation accuracy during training, ensuring that the final deployed model corresponded to the optimal epoch.

(a)

(b)

(c)

Figure 11. (a) Confusion matrix- CNN-LSTM baseline model (b) Confusion matrix- CNN-LSTM pruned model (c) Confusion matrix- CNN-LSTM pruned and quantized model model

Table 7. Resource efficiency comparison

Table 8. Effect of batch size on performance (Pruned model)

Figure 12. Accuracy vs batch size

Table 8 and the corresponding plot in Figure 12 together analyze the impact of batch size on model performance and efficiency for the pruned model. A batch size of 32 achieves the highest accuracy (95.27%) and F1-score, indicating an optimal balance between stable gradient updates, memory utilization, and learning effectiveness. Increasing the batch size to 64 and 128 leads to a slight reduction in accuracy while improving inference speed, highlighting the typical trade-off between computational efficiency and learning stability in deep learning models. Overall, the results show that moderate batch sizes provide the best performance, whereas very large batch sizes, despite reducing training noise, result in marginally lower accuracy due to less frequent weight updates.

Table 9. Effect of learning rate on accuracy (Quantized model)

Figure 13. Accuracy vs learning rate (Quantized model)

Table 9 and Figure 13 together illustrate the effect of learning rate on model performance. A learning rate of 0.001 achieves the best accuracy, providing an optimal balance between stable convergence and effective learning. In contrast, a high learning rate (0.1) results in unstable training and reduced performance, while a very low learning rate (0.0001) slows convergence without meaningful accuracy improvements. These results highlight the importance of careful learning-rate tuning, even after model compression.

The results presented in Table 10 and Figure 14 demonstrate the impact of dropout on the baseline CNN-LSTM model’s accuracy and generalization. A dropout rate of 0.2 provides the best balance, achieving 95.27% accuracy while effectively reducing overfitting. In contrast, the absence of dropout leads to slightly higher accuracy but increases overfitting risk, whereas excessive dropout (0.6) causes underfitting and degrades predictive performance. These results highlight the importance of moderate regularization for robust wearable healthcare models.

Table 10. Effect of dropout rate

Figure 14. Accuracy vs dropout rate

Table 11 and Figure 15 illustrate the relationship between pruning percentages and the resulting accuracy and model size. As pruning levels increase from 30% to 90%, model size shrinks dramatically from 610MB to just 170MB—while accuracy gradually declines. Up to 50% pruning, the accuracy remains relatively stable (above 95%), demonstrating that substantial compression is possible without severely affecting model performance. Beyond 70% pruning, the model’s predictive ability drops more noticeably.

Table 11. Pruning ratio vs accuracy and model size

Table 12 compares different quantization techniques: post-training int8 quantization, dynamic range quantization, and Float16 quantization. Post-training int8 quantization delivers the highest memory and energy savings, reducing model size to 190MB with only a minor decrease in accuracy. Float16 maintains slightly higher accuracy but requires more memory, indicating a trade-off between precision retention and deployment feasibility on low-power hardware.

Figure 15. Pruning vs accuracy and model size

Table 12. Effect of quantizationtype

Table 13. Number of LSTM units vs accuracy

Insights from Table 13 and the corresponding trends in Figure 16 demonstrate the effect of increasing LSTM units on accuracy, memory consumption, and inference time. Adding more units improves accuracy, with peak performance (~95.7%) at 256 units. However, memory usage rises sharply with more units, from 290MB at 32 units to 880MB at 256 units. This trade-off is critical: 128 LSTM units offer a sweet spot, balancing high accuracy and reasonable memory requirements for wearable applications.

Table 14 presents a detailed evaluation of how combined pruning and quantization affect model performance, power consumption, inference time, and model size. Moderate pruning (30%-50%) coupled with quantization yields the best balance, sustaining accuracy above 94% while reducing power consumption and model size substantially. Heavier pruning (70%-90%) significantly cuts memory and energy costs but at the expense of noticeable drops in predictive performance, making it less ideal for critical healthcare applications where reliability is crucial.

The heatmap in Figure 17 illustrates how different levels of pruning combined with int8 quantization impact model performance metrics. As pruning levels increase from 30% to 90%, both model size and power consumption decrease sharply, making the models more suitable for edge devices. However, this compression comes at the cost of reduced accuracy, especially beyond 70% pruning. The heatmap highlights the importance of carefully choosing the pruning threshold to maintain acceptable accuracy while maximizing resource efficiency.

Figure 16. Accuracy and memory vs number of LSTM units

Table 14. Combined pruning + quantization levels impact

Figure 17. Optimization levels vs performance

The CNN-LSTM baseline model delivered the highest classification performance with an accuracy of 95. 27%, a precision of 92. 8% and an F1 score of 94.8%. However, it required 610MB of memory and consumed 16.74 watts of power, rendering it infeasible for deployment on low-power wearable devices. To overcome these limitations, TinyML optimization techniques were applied.

These results clearly demonstrate that the primary contribution of pruning is memory and power reduction, not accuracy improvement. The slight decrease in accuracy is a trade-off for significant computational gains, which is often acceptable in wearable applications where hardware constraints are critical. The interpretation of the work is that TinyML techniques can preserve diagnostic reliability while transforming otherwise heavy deep learning models into lightweight versions suitable for edge-level inference. In other words, the model does not become more accurate after optimization; it becomes more efficient while retaining sufficient accuracy for practical use in the real world.

Ultimately, the actual result and central insight of this research is the demonstration that a carefully optimized CNN-LSTM model can serve as a reliable, low-power, real-time diagnostic engine for wearable health monitoring devices.