Development of an interpretable machine learning model for frailty risk prediction in older adult care institutions: a mixed-methods, cross-sectional study in China.

OBJECTIVE

To develop and validate an interpretable machine learning (ML)-based frailty risk prediction model that combines real-time health data with validated scale assessments for enhanced decision-making and targeted health management in integrated medical and older adult care institutions (IMOACIs) in central China.

DESIGN

Mixed-methods, cross-sectional study.

SETTING

13 IMOACIs across seven cities in Hunan province, central China, from 8 to 16 July 2022.

PARTICIPANTS

Five healthcare experts and two data scientists participated in the requirements analysis stage. A total of 586 older adults were included in the assessment data collection stage, and 15 participants (10 healthcare professionals and five data scientists) were involved in the model evaluation stage.

METHODS

A collaborative requirements analysis involving healthcare professionals and data scientists guided the design of an interpretable frailty risk prediction model. Five machine learning models were developed and evaluated: logistic regression, support vector machines (SVM), random forest, extreme gradient boosting (XGBoost) and a multimodel ensemble approach. Hyperparameter optimisation was performed using stratified fivefold cross-validation with grid search, incorporating class-weighted loss functions to address class imbalance and model-specific regularisation techniques to maximise performance while preventing overfitting. To enhance interpretability, the model incorporated Shapley Additive Explanations. The final model was integrated into a user-facing platform and validated using cross-sectional standardised assessment data collected from 13 IMOACIs. A mixed-methods evaluation approach combined quantitative performance metrics with qualitative user experience assessments.

RESULTS

The dataset (n=586) was randomly split into training (n=468) and validation (n=118) sets (4:1 ratio). Among models, XGBoost demonstrated superior performance, achieving an accuracy of 0.89 and an area under the receiver operating characteristic curve (AUC) of 0.89 on the training set. On the validation set, the XGBoost model achieved a precision of 0.76, recall of 0.72, F1 score of 0.74, accuracy of 0.83 and AUC of 0.80, outperforming other models. User experience surveys yielded high mean ratings for satisfaction (4.20/5), perceived accuracy (4.20/5), interpretability (4.30/5) and application value (4.10/5). Qualitative analysis of user feedback identified six key themes: practical and application value, performance and data analysis, interpretability and comprehensibility, impact and integration into practice, limitations and areas for improvement, and future development and innovation prospects, highlighting the model's strong potential for practical implementation.

CONCLUSIONS

This novel, interpretable ML-based frailty risk prediction model can enhance decision-making in the care of older adults by providing transparent predictions and identifying crucial factors associated with frailty. It establishes a foundation for targeted management and broader ML applications in healthcare systems, such as IMOACIs, particularly in developing regions.