Comparing interpretable machine learning models for fall risk in middle-aged and older adults with and without pain

Abstract

Pain is common in middle-aged and older adults, has also been identified as a fall risk factor, whereas the mechanism of falls in pain is unclear. This study included 13,074 middle-aged and older adults from the China health and retirement longitudinal study (wave 2011-2015) to separately develop four-year fall risk prediction models for older adults with and without pain, using five machine learning algorithms with 145 input variables as candidate features. Shapley Additive exPlanations (SHAP) was used for the prediction model explanations. Adjusted logistic regression (LR) models showed that pain (OR 1.40 [1.29, 1.53]) was associated with a higher fall risk. Among pain characteristics, lower limb pain had the highest risk (OR 1.71 [1.22, 2.18]), followed by severe pain (OR 1.53 [1.36, 1.73]) and multisite pain (OR 1.43 [1.28, 1.55]). Among the fall prediction models for pain and non-pain, the LR model performed best with AUC-ROC values of 0.732 and 0.692, respectively. Common important features included fall history and height. Unique features for the pain model were functional limitation, SPPB, WBC, chronic disease score, life satisfaction, platelets, cooking fuel, and pain quantity, while marital status, age, depressive symptoms, cognitive function, hearing, rainy days, tidiness, and sleep duration were exclusive to the non-pain model. Pain characteristics are associated with falls among middle-aged and older adults. Prediction model can help identify people at high risk of falls with pain. Important features of falls differ between pain and non-pain populations, and prevention strategies should target specific populations for fall risk prediction.

Keywords: Falls; Machine learning; Older adults; Pain; Risk factors.

Fig. 1

Flow chart of participant selection.

Fig. 2

Receiver operating characteristic curve performance of five models on the test set. (A) Pain populations. (B) Non-pain populations. ML machine learning, LR logistic regression, NB naive Bayesian, RF random forest, XGBoost extreme gradient boosting, ANN artificial neural network, AUC-ROC area under the receiver operating characteristic curve.

Fig. 3

Feature Importance Ranking (top 10) with the SHAP summary plot for the logistic regression models. (A) Pain population model, (B) Non-pain population model. SPPB short physical performance battery, WBC white blood cell.