[Predicting the need for long-term care: development and comparative evaluation of advanced machine learning models].

Abstract

Background and aim of the article: In recent years, the Bavarian Medical Service (Medizinischer Dienst Bayern, MD Bavaria) has recorded an increasing number of initial assessments without determination of a care level (Pflegegrad) within the assessment process for determining long-term care needs. These initial assessments without a care level determination create considerable administrative burdens as well as avoidable effort for the MD Bavaria and caregiving households. Against this backdrop, this study investigates whether machine learning methods can be used to predict the success of initial applications for long-term care needs based on selected features that are not part of the standardized assessment instrument according to the German Social Code XI.

Data foundation and methodology: The dataset consists of a complete collection of data from the MD Bavaria that includes all initial assessments to determine long-term care needs from 2019. Using mutual information, we first identified relevant input variables and then checked for possible multicollinearity. Based on the selected characteristics, we trained various machine learning procedures and evaluated them in terms of their predictive performance.

Results: Particularly relevant input variables for predicting long-term care needs include the age of the person being assessed and the ability to prepare simple meals. The analysis shows that the random forest classifier achieves the highest prediction performance with an accuracy of 0.88 and an AUC value of 0.95.