Investigating feature-engineered predictors for systolic blood pressure changes in an mHealth-based disease...

Mobile health (mHealth)-based disease management programs enable continuous monitoring of blood pressure (BP) and related health behaviors. Feature engineering may help to extract informative predictors from longitudinal data, potentially improving BP change prediction. This study aimed to evaluate whether feature-engineered predictors can improve the prediction of systolic BP (SBP) changes using an mHealth-based disease management program. We analyzed data from participants with hypertension, dyslipidemia, or diabetes mellitus who completed the 24-week Mystar program, which combined phone-based coaching, remote monitoring, and app-based logging of BP and behavioral data. The primary outcome was the change in morning SBP from baseline to the end of the program. Prediction models for SBP changes were developed using ElasticNet regression at weeks 4, 8, 12, and 22 by comparing models with and without feature-engineered variables generated by feature tools. In total, 2318 participants were included in the analysis. At week 4, the top feature after feature engineering showed a stronger correlation with SBP change (r = 0.561) than the best original predictor (r = 0.455), although the model-level performance was similar (r = 0.561 vs. 0.559). By week 22, both models achieved a high correlation of approximately 0.85 with no substantial difference in performance. Feature engineering increased the correlation between individual predictors and SBP change in the early phase; however, the overall prediction performance of the ElasticNet model remained largely unchanged. Further studies are required to confirm these findings and examine their applicability in broader clinical and implementation contexts. Data from a 24-week mHealth-based program were analyzed to predict systolic blood pressure SBP changes using feature-engineered variables and ElasticNet regression. In the early phase, feature-engineered predictors ranked highest in importance, although overall model performance remained similar with and without feature engineering. Prediction accuracy improved over time, with correlations reaching ~0.85 by week 22.