Interpretable machine learning framework for frailty risk prediction using NHANES 2007-2018: A cross-sectional study

Frailty is a multidimensional clinical syndrome associated with adverse health outcomes, yet existing prediction models often sacrifice interpretability for accuracy. This study aimed to develop and validate an interpretable machine learning framework for frailty risk prediction using nationally representative survey data. We conducted a cross-sectional analysis of 3817 adults aged 20 years and older from the National Health and Nutrition Examination Survey (NHANES, 2007-2018). Frailty was defined using a deficit accumulation Frailty Index (FI ≥ 0.21). The prediction model incorporated 48 features spanning metabolic biomarkers (n = 29), inflammatory markers (n = 11), and demographic covariates (n = 8). We developed a 3-tier interpretable framework using Automatic Piecewise Linear Regression (APLR) and compared its performance to extreme gradient boosting (XGBoost) and logistic regression (LR) via 10-fold cross-validation. Model interpretability was assessed using Anchor explanations. A sensitivity analysis was performed in the subgroup aged 60 years and older (n = 2629). APLR achieved the highest Area Under the Receiver Operating Characteristic Curve (AUC) of 0.801 ± 0.018, significantly outperforming XGBoost (AUC = 0.776 ± 0.015, P < .05) and LR (AUC = 0.778 ± 0.024, P < .05). Body mass index (BMI), poverty-income ratio (PIR), and age were the top predictors in the full sample. Anchor explanations achieved a mean precision of 98.2% ± 1.7%. In the 60+ sensitivity analysis, APLR retained the highest AUC (0.799 ± 0.022), remaining statistically significantly superior to both XGBoost (P = .003) and LR (P = .012). Blood urea nitrogen (BUN) emerged as the second most important predictor in the older subgroup, reflecting the increased relevance of renal function in geriatric frailty. The APLR-based interpretable framework offers a promising approach to frailty risk stratification that balances accuracy with clinical transparency, particularly in older adults. The identified features should be interpreted as cross-sectional predictive markers rather than causal determinants of frailty. External and longitudinal validation, including formal assessment of calibration and clinical utility, is warranted before clinical deployment.