Public-oriented personalized long-term care service recommendation using machine learning

Long-term care prediction models often rely on large multidomain assessments, which can limit interpretability and reuse across multiple service outcomes.

To determine whether a common reduced predictor set could support prediction of 20 later recorded long-term care service-use outcomes while quantifying performance loss relative to the full set.

This retrospective cohort study linked initial Care Management Evaluation Form assessments to administrative service records for 9,297 adults from southern Taiwan. The development cohort included 8,367 participants, and the fixed-holdout re-evaluation cohort included 930. We trained logistic regression, decision tree, random forest, extreme gradient boosting, and a shallow artificial neural network. Development-only recursive feature elimination ranked 378 processed predictors and produced nested sets of 31 and 16 predictors. Performance was assessed using the area under the receiver operating characteristic curve, area under the precision-recall curve, Brier score, and log loss.

With all 378 predictors, macro area under the receiver operating characteristic curve ranged from 0.711 to 0.772. The corresponding ranges were 0.706 to 0.745 with 31 predictors and 0.685 to 0.702 with 16 predictors. The 31-predictor set incurred smaller losses than the 16-predictor set. Performance varied by algorithm and service. The artificial neural network did not consistently outperform logistic regression, random forest, or extreme gradient boosting.

The shared 31-predictor set provided the strongest balance between parsimony and discrimination and emerged as the preferred reduced configuration for subsequent geographic, temporal, and prospective evaluation.