MACHINE-LEARNING MODELS FOR PREDICTING BLOOD PRESSURE AND CARDIAC FUNCTION USING WEARABLE SENSOR DATA

Abstract

Cuffless monitoring can degrade because motion noise, inconsistent wear, and context alter wearable signals, weakening prediction of blood pressure (BP) and cardiac function. This study tested an integrated wearable plus Likert survey framework for predicting systolic BP (SBP), diastolic BP (DBP), and RMSSD in a quantitative, cross-sectional, case-study-based design. The case sample included 180 observations with data completeness 0.86; outcomes averaged SBP 128.4 mmHg and DBP 79.6 mmHg, with RMSSD 32.5 ms. Key variables comprised wearable predictors (pulse arrival time (PAT), PPG upstroke time, pulse width, activity intensity) and five Likert constructs (stress, sleep quality, activity routine, adherence, usability), with acceptable reliability (Cronbach alpha 0.78 to 0.88). The analysis plan applied descriptive statistics, Pearson correlations, multivariable regression (wearable-only then wearable plus Likert), and 5-fold cross-validated machine-learning benchmarking using MAE, RMSE, and R². Headline findings showed PAT correlated with SBP (r = -0.41, p < .001) and DBP (r = -0.29, p < .001), while stress correlated positively with SBP (r = 0.30, p < .001) and negatively with RMSSD (r = -0.33, p < .001). Wearable-only regression explained R² = 0.46 (SBP), 0.33 (DBP), and 0.38 (RMSSD), improving to 0.55, 0.40, and 0.47 after adding Likert predictors. In benchmarking, XGBoost delivered the best SBP model (MAE 5.0 mmHg, RMSE 7.3 mmHg, R² 0.70), improving on wearable-only performance (MAE 5.6; R² 0.66), with gains for DBP (MAE 3.7; R² 0.57) and RMSSD (MAE 4.8; R² 0.58). These results support enterprise-style monitoring pipelines that combine sensor features with survey measures and quality screening, while causal claims require longitudinal validation.