Attention-Kan for Proactive Cardiovascular Care Using A Hybrid Iot–Ml Framework For Early Detection And Real-Time Monitoring

Abstract

Cardiovascular diseases (CVDs) are the foremost cause of global mortality, and contemporary clinical workflows frequently detect pathology only after symptoms appear. Continuous physiological sensing from wearables paired with robust predictive models can enable earlier intervention and preventive care. Existing systems commonly rely on single modalities (ECG, PPG, accelerometry, or EHR), struggle with multimodal fusion, and perform poorly under motion artifacts, noisy recordings, severe class imbalance, and demographic heterogeneity. Wearable deployments further face strict power, latency, and connectivity constraints that impede reliable on-device inference and personalized feedback. We propose Attention-Kolmogorov–Arnold Network (Attention-KAN), a hybrid architecture combining a lightweight attention mechanism for cross-modal temporal–spatial aggregation with the KAN for high-precision function approximation. The pipeline ingests synchronized multimodal inputs: raw ECG waveforms, PPG traces, triaxial accelerometer streams, and engineered EHR-derived features, then applies sensor-specific preprocessing: bandpass filtering, adaptive baseline wander removal, motion-artifact rejection using accelerometer-guided gating, and sample-wise calibration. The attention layer computes cross-modal context vectors and dynamic modality weights; these embeddings feed stacked KAN blocks that approximate complex nonlinear interactions through ensembles of univariate subnetworks and an outer mixing network. Training uses a composite loss (cross-entropy + focal loss for imbalance + calibration loss), domain-aware augmentation, and edge-aware knowledge distillation to produce an efficient wearable-ready variant. From the public datasets, Attention-KAN achieves higher early detection sensitivity and AUC than gradient-boosted trees, Bi-LSTM, and conventional CNN baselines, while reducing false alarm rates. Extensive experimental results demonstrate that Attention-KAN achieves 96.8% accuracy, 95.7% F1-score, and 95.3% sensitivity, significantly outperforming existing methods such as Bi LSTM, CNN+NB, and Swarm-ANN. Furthermore, the framework reduced inference latency by 23%, model size by 18%, and energy consumption by 21%, proving its suitability for low-power, wearable deployments.