Real-Time Forecasting from Wearable-Monitored Heart Rate Data Through Autoregressive Models

Heart rate (HR) analysis is of paramount importance in healthcare, particularly for monitoring cardiovascular health, a global concern. The advent of wearable sensors has enabled continuous HR monitoring, with researchers attempting to develop early detection systems by forecasting HR in a univariate fashion. This study analyzes real-world HR time series gathered during participants daily routines to critically assess the predictive power of past HR data in short-term, univariate forecasting. The literature emphasizes a minute-by-minute, univariate forecasting approach, where state-of-the-art predictive models predominantly employ autoregressive integrated moving average (ARIMA). Yet, its superiority has been proved without studying its optimized hyper-parameters, which could not only improve forecast accuracy but also provide valuable insights. By leveraging the interpretability of ARIMA, we tune its hyper-parameters within a minute-by-minute forecasting structure to address the central research question: how does historical HR data contribute to generate accurate short-term HR forecasts? Our analysis finds that the random walk model, a special case of ARIMA, consistently performs comparably to, or even better than, more complex ARIMA specifications. This indicates that HR values alone offer limited predictive power for short-term forecasting, casting doubt on the value of further refinement in univariate models for alarm system development. These findings highlight the limitations of univariate HR forecasting in real-time health monitoring. Rather than increasing model complexity, future research might benefit from exploring alternative approaches to improve early warning system capabilities in real-world settings.