Prediction of Overall Patient Characteristics that Incorporate Multiple Outcomes in Acute Stroke: Latent Class Analysis

Background Previous prediction models have predicted a single outcome (e.g. gait) from several patient characteristics at one point (e.g. on admission). However, in clinical practice, it is important to predict an overall patient characteristic by incorporating multiple outcomes. This study aimed to develop a prediction model of overall patient characteristics in acute stroke patients using latent class analysis. Methods This retrospective observational study analyzed stroke patients admitted to acute care hospitals (37 hospitals, N=10,270) between January 2005 and March 2016 from the Japan Association of Rehabilitation Database. Overall, 6,881 patients were classified into latent classes based on their outcomes. The prediction model was developed based on patient characteristics and functional ability at admission. We selected the following outcome variables at discharge for classification using latent class analysis: Functional Independence Measure (functional abilities and cognitive functions), subscales of the National Institutes of Health Stroke Scale (upper extremity function), length of hospital stay, and discharge destination. The predictor variables were age, Functional Independence Measure (functional abilities and comprehension), subscales of the National Institutes of Health Stroke Scale (upper extremity function), stroke type, and amount of rehabilitation (physical, occupational, and speech therapies) per day during hospitalization. Results Patients (N=6,881) were classified into nine classes based on latent class analysis regarding patient characteristics at discharge (class size: 4–29%). Class 1 was the mildest (shorter stay and highest possibility of home discharge), and Class 2 was the most severe (longer stay and the highest possibility of transfers including deaths). Different gradations characterized Classes 3–9; these patient characteristics were clinically acceptable. Predictor variables at admission that predicted class membership were significant (odds ratio: 0.0– 107.9, P <.001). Conclusions Based on these findings, the model developed in this study could predict an overall patient characteristic combining multiple outcomes, helping determine the appropriate rehabilitation intensity. In actual clinical practice, internal and external validation is required.