Predicting premature treatment termination in inpatient psychotherapy: A machine learning approach.
The study aimed to quantify types of premature treatment termination in a psychosomatic hospital and to investigate if patient characteristic at the beginning of inpatient or day-clinic treatment can predict a subsequent premature termination.
Routine clinical data were analysed retrospectively for N = 2017 patients. Based on treatment length and medical discharge reports, patients were categorized as premature treatment termination related to therapeutic reasons (PTT-T) or not (NPTT-T) by two independent raters. Next, two prediction models were built using random forest algorithms. Model 1 included general clinical information, model 2 additional data from self-report measures at treatment admission. Finally, a SHAP feature importance plot was calculated for each model.
In this study, 12.1% of the patients ended treatment prematurely related to therapeutic reasons. Model 1 predicted 29% of PTT-Ts and 95% of NPTT-Ts in the holdout sample correctly. Model 2 was able to identify 59% of PTT-Ts and 80% of NPTT-Ts. The absence of self-report measures at admission was the predictor variable with the greatest influence.
PTT-T in inpatient treatment can substantially be predicted by routine clinical intake data, which can therefore provide a valuable source for identifying patients at risk.
Routine clinical data were analysed retrospectively for N = 2017 patients. Based on treatment length and medical discharge reports, patients were categorized as premature treatment termination related to therapeutic reasons (PTT-T) or not (NPTT-T) by two independent raters. Next, two prediction models were built using random forest algorithms. Model 1 included general clinical information, model 2 additional data from self-report measures at treatment admission. Finally, a SHAP feature importance plot was calculated for each model.
In this study, 12.1% of the patients ended treatment prematurely related to therapeutic reasons. Model 1 predicted 29% of PTT-Ts and 95% of NPTT-Ts in the holdout sample correctly. Model 2 was able to identify 59% of PTT-Ts and 80% of NPTT-Ts. The absence of self-report measures at admission was the predictor variable with the greatest influence.
PTT-T in inpatient treatment can substantially be predicted by routine clinical intake data, which can therefore provide a valuable source for identifying patients at risk.
Authors
Engstler Engstler, Lutz Lutz, Schwartz Schwartz, Jennissen Jennissen, Friederich Friederich, Dinger Dinger
View on Pubmed