TL;DR
Adding lifestyle data—like sleep, smoking, and exercise habits—to machine learning models does not significantly improve hypertension prediction over a simple baseline of demographics and body measurements. While these factors are clinically vital for health, this study suggests they provide redundant predictive information once age and BMI are already accounted for.
What Problem Does This Paper Address?
Hypertension is a leading contributor to cardiovascular disease, and clinical guidelines heavily emphasize modifiable lifestyle factors for prevention. Naturally, many researchers have attempted to build machine learning models that use this lifestyle data to identify at-risk individuals.
However, the author identifies a critical gap: many previous studies fail to use "leakage-controlled" evaluations. In ML, "data leakage" occurs when information from the test set (the data the model shouldn't see yet) accidentally influences the training process—for example, by calculating the average BMI of the entire dataset before splitting it into training and testing groups. Furthermore, it wasn't clear if lifestyle data provides incremental value. If a model already knows a patient's age and BMI, does knowing their sleep schedule actually make the prediction more accurate?
Methods
The researcher compared two "feature arms":
- Baseline: Demographics (age, sex, race, education, income) and anthropometrics (weight, height, BMI, waist circumference).
- Lifestyle-Augmented: The baseline plus 11 derived lifestyle features (physical activity, alcohol use, smoking status, and sleep duration/quality).
To ensure the results were robust, the study employed:
- Five Model Families: Logistic Regression, Random Forest, Gradient Boosting, Bagged Decision Trees, and XGBoost.
- Leakage Control: All preprocessing (imputation, scaling, and SMOTE for class imbalance) was performed strictly within cross-validation folds.
- Paired Bootstrapping: This allowed the author to calculate confidence intervals for the difference in performance between the two arms, rather than just comparing two static scores.
- SHAP Interpretability: Used to analyze which features the models prioritized and how consistently different models agreed on those priorities.
Datasets
The study utilized the NHANES (National Health and Nutrition Examination Survey) 2017–March 2020 cycle.
- Cohort Size: 8,024 adults.
- Primary Outcome: "Phenotype B," defined as a measured blood pressure $\ge$ 130/80 mmHg or currently taking antihypertensive medication.
- Secondary Outcome: "Phenotype A," which relied solely on measured blood pressure (ignoring medication status).
Key Insights and Ideas
- The Redundancy Effect: The "aha moment" here isn't that lifestyle doesn't matter, but that it's redundant for prediction. For instance, vigorous activity is moderately correlated with age. Once a model has age, the incremental "signal" from activity levels is too weak to move the needle on accuracy.
- The "Matched Comparison Space": A major methodological contribution involves how we measure model agreement. When you add more features, the "agreement" between different models (e.g., Random Forest vs. XGBoost) naturally appears to drop because there are more variables to rank. The author demonstrates that to fairly assess model consistency, you must compare how they rank the shared features (the baseline) before and after adding the new ones.
- Subgroup Nuance: In an exploratory analysis of adults not taking medication, the models actually showed a small, positive improvement when lifestyle data was added. This suggests that lifestyle data might be more predictive in specific populations where medication isn't masking the underlying blood pressure trends.
Key Results
- No Significant Gain: Across all five model families, the change in ROC-AUC (a measure of model discrimination) was negligible. Point estimates for the improvement ranged from -0.0028 to +0.0002, and every 95% confidence interval included zero.
- Strong Baseline: The baseline models were already quite effective, achieving ROC-AUCs around 0.81. Age and BMI were the dominant predictors.
- SHAP Agreement: When looking at the 19 shared baseline features, model agreement remained high (Spearman $\rho$ of 0.873 vs. 0.863). This proves that adding lifestyle data didn't "confuse" the models or fundamentally change how they viewed the core demographic risks.
Limitations and Future Work
- Scope of "Lifestyle": The study focused on activity, alcohol, smoking, and sleep. It did not include detailed nutritional/dietary data or laboratory biomarkers (like cholesterol), which might offer more predictive power.
- Single Dataset: The findings are based on one NHANES cycle. The author notes that external validation on different cohorts is necessary to see if this "redundancy" holds true globally.
- Medication Masking: Because many participants were already on blood pressure medication, their measured BP might not reflect their "natural" risk, potentially dampening the predictive signal of lifestyle factors.
- Future Directions: The author suggests future work should investigate blood pressure control (how well meds are working) rather than just diagnosis, and explore more complex modifiable domains like diet.
Summary updated