Machine Learning Models Classify Frailty in Adults With Mild Cognitive Impairment

A cross-sectional investigation examining 3,124 adults aged 45 years and older with mild cognitive impairment in China found that machine learning algorithms effectively classified prevalent frailty using a compact set of routinely collected patient characteristics. Frailty was identified in 13.0 percent of the cohort, representing 406 individuals among the total study population drawn from the 2015 China Health and Retirement Longitudinal Study. Because frailty is linked with adverse health consequences in people living with mild cognitive impairment, accurate identification of the condition remains an important clinical objective.

To evaluate screening approaches, investigators split the sample into training and test datasets using a seven-to-three ratio. The analytical workflow restricted missing-data imputation, feature selection, and algorithm development entirely to the training partition to avoid data leakage. Candidate predictors were filtered through least absolute shrinkage and selection operator regression, ultimately isolating six key functional and patient-reported attributes: activities of daily living, instrumental activities of daily living, self-reported health, chronological age, pain, and sleep disturbance.

The researchers evaluated eight machine learning models across several statistical performance dimensions, including the area under the receiver operating characteristic curve, precision–recall area under the curve, sensitivity, specificity, F1-score, Brier score, calibration analysis, and decision curve analysis. Model stability was further examined through repeated cross-validation within the training cohort alongside bootstrap resampling performed on the held-out test set.

Standard logistic regression and the gradient-boosted decision tree algorithm XGBoost demonstrated comparable discriminative capability, though their strengths varied across distinct evaluation metrics. Logistic regression attained the highest test-set area under the receiver operating characteristic curve at 0.887 and the leading precision–recall area under the curve at 0.537. Meanwhile, XGBoost achieved a very similar area under the receiver operating characteristic curve of 0.885, a precision–recall area under the curve of 0.522, and the highest overall F1-score at 0.535.

Resampling procedures, including bootstrap iterations and repeated cross-validation, confirmed that neither logistic regression nor XGBoost established a definitive discrimination advantage over the other. To provide complementary model interpretation alongside logistic regression coefficients and odds ratios, researchers conducted SHapley Additive exPlanations analysis on the XGBoost model. This interpretability framework verified that activities of daily living, instrumental activities of daily living, self-reported health, age, pain, and sleep disturbance all contributed directly to the classification decisions generated by XGBoost.

The study authors emphasized that routinely accessible functional indicators and patient-reported measures can effectively support multidimensional frailty classification in adults diagnosed with mild cognitive impairment. However, because the study relied strictly on a cross-sectional design, the resulting algorithms reflect existing frailty status rather than predicting the future onset of frailty over time. The authors noted that prospective external validation across independent cohorts remains necessary prior to deploying these computational tools in clinical care settings.

The methodological framework developed by the team highlighted how missing-data imputation and candidate feature selection were strictly confined to the training subset to preserve unbiased evaluation during validation. By evaluating eight distinct machine-learning models side by side, the authors sought to clarify whether complex algorithms offered meaningful advantages over conventional statistical techniques when identifying prevalent frailty from routine observations. In addition to area under the receiver operating characteristic curve and precision–recall metrics, comprehensive evaluation included Brier score, specificity, sensitivity, calibration analysis, and decision curve analysis to appraise clinical utility across varying decision thresholds.

Complementary interpretability analyses further established how each selected parameter guided classification. While logistic regression supplied baseline coefficients and odds ratios for the retained features, SHapley Additive exPlanations analysis mapped the contributions of individual factors within the tree-based XGBoost model, confirming that activities of daily living, instrumental activities of daily living, self-assessed health status, chronological age, bodily pain, and sleep disturbance consistently drove model outputs. Nevertheless, the authors cautioned that because the data reflected a cross-sectional snapshot from the 2015 national survey, prospective external validation in diverse clinical environments is essential before these models can be translated into routine diagnostic workflows.

Subject of Research: Machine learning classification of frailty in adults with mild cognitive impairment

Article Title: Machine learning-based classification of frailty among adults with mild cognitive impairment in China: a cross-sectional study

Article References: Xie, J., Aldirawi, A., Zhang, J., Song, Y.-N., Luo, L., Tao, M.-Y., & Su, W.-Y. (2026). Machine learning-based classification of frailty among adults with mild cognitive impairment in China: a cross-sectional study. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03988-y

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03988-y

Keywords: machine learning, frailty, mild cognitive impairment, China Health and Retirement Longitudinal Study, logistic regression, XGBoost, LASSO regression, activities of daily living, instrumental activities of daily living, self-reported health, sleep disturbance, cross-sectional study, SHAP analysis, geriatric health, precision-recall AUC, calibration analysis