Lifestyle Data Did Not Boost Heart Risk Forecasts
Health experts looked at a large health data set. The set is known as NHANES. The time frame was 2017 through March 2020. They asked if simple health details could help forecast high blood pressure. They tested five computer models. The models were logistic regression, random forest, gradient boosting, bagged decision trees, and XGBoost. Each model used local fold steps. They also used bootstrap confidence intervals.
The team added lifestyle facts. These facts covered activity, drinking, smoking, and sleep. The aim was better results. The models already had age, sex, and body measures. The team checked held out results. The paired ROC AUC scores had zero inside all intervals. So no clear gain showed up. A stricter blood pressure test gave the same result. The team also tried without synthetic data.
The team then checked model agreement. Mean cross model SHAP rank agreement went from 0.873 to 0.806. That was for all coded columns. The drop was mostly from shared baseline columns. There were 19 of them. The scores were 0.873 and 0.863. The same pattern held under both checks. Different sized feature sets may need a fair comparison space.
In a small look at adults who did not get treatment, the models found tiny positive shifts. All five models had small gains. Two had unadjusted intervals that did not include zero. The team called for more confirmation. The final result stayed small. The lifestyle block did not add much. The baseline was already useful.