Optimized Heart Rate Forecasting in Healthcare using Hybrid LSTM-XGBoost with Reinforcement Learning and Attention Mechanism

Abstract

Forecasting time series data is very useful in healthcare, mainly to understand heart rate variability (HRV), which provides insights into cardiac health and stress response. Conventional statistical and machine learning techniques have been extensively used, but their capacity to accurately model the complex, non-linear nature of HRV remains limited. Deep learning techniques such as Long Short-Term Memory (LSTM) networks and Temporal Convolutional Networks (TCN) provide stronger predictive power, with TCN handling long-term patterns and Prophet capturing short-term variations. But, when used in isolation, these models are not able to deliver reliable forecasts.In this work, we have explored many hybrid models to improve prediction accuracy. Static and adaptive weighting schemes combining LSTM and XGBoost showed incremental gains, but their performance was not sufficient. Therefore, we have designed a novel hybrid framework that integrates LSTM and XGBoost with an attention mechanism, supported by reinforcement learning for dynamic weighting. To guarantee robustness, the framework uses a comprehensive preprocessing pipeline, including wavelet-based noise reduction, min–max normalization and interpolation to handle the missing data. Experimental evaluation on HRV datasets validates that the proposed framework captures both short-term fluctuations and long-term dynamics more effectively than baseline models, there by achieving the best overall accuracy. It makes it highly suitable for real-time HRV monitoring in wearable computing devices and IoT-enabled healthcare systems. The model's computational efficiency and dynamic adaptation mechanism enable deployment in resource-constrained environments for continuous cardiovascular health monitoring.