Using Machine Learning to Study Factors Affecting Discharge Destination in Recovery Units

Background: In recent years, machine learning has been developed in the medical community to construct multidimensional datasets consisting of many variables and perform simultaneous factor analysis.

Objective: This study aimed to construct a multidimensional dataset of 50 items by incorporating supervised machine learning in a random forest algorithm to predict whether patients will be discharged home or to a facility after a stroke.

Methods: Thirty patients hospitalized with cerebrovascular diseases who were subsequently discharged were considered as the study subjects. The dataset used for analysis consisted of attributes such as characteristics (three items), physical and cognitive functions (seven items), functional independence measure (FIM) (18 items), blood data (16 items), and social characteristics (six items). The discharge destination variable was either a home or a facility. Machine learning was used to extract factors important for this classification. The accuracy of the random forest was calculated by five-fold cross-validation. The mean decrease Gini, a measure of importance in classification, was calculated for each fold.

Results: The results indicated that FIM, a measure of activities of daily living (ADL), and cognitive function, including memory, which strongly influenced the prediction equation, were important factors in the proposed algorithm. The results of the analysis revealed that the algorithm predicted home discharge or institutionalization with 87.1% accuracy.

Conclusion: Through this study, ADL and cognitive function were identified as important factors in predicting home discharge for patients with cerebrovascular disease.