Predicting patient outcomes in psychiatric hospitals with routine data: a machine learning approach

Area of Science:

• Psychiatric hospital care organization
• Machine learning applications in healthcare
• Predictive modeling in mental health

Background:

• Data availability at the point of decision-making is a challenge in machine learning.
• Psychiatric hospital care organization can benefit from predictive insights.
• Routine admission data offers a potential source for such predictions.

Purpose of the Study:

• To predict aspects of psychiatric hospital care organization using readily available admission data.
• To compare the predictive performance of machine learning (gradient boosting) against traditional logistic regression and a naive baseline.
• To assess the utility of administrative data for clinical decision support.

Main Methods:

• Utilized data from 45,388 inpatient episodes across nine psychiatric hospitals in Hesse, Germany (2017-2018).
• Compared stochastic gradient boosting (GBM) with multiple logistic regression and a naive baseline classifier.
• Evaluated model performance using the area under the Receiver Operating Characteristic curve (AUC) on unseen patient data.

Main Results:

• Model performance varied by outcome: high for coercive treatment (AUC: 0.83) and 1:1 observations (0.80), lower for short length of stay (0.69) and non-response to treatment (0.65).
• Gradient boosting slightly outperformed logistic regression.
• Both machine learning and logistic regression were significantly better than the naive baseline classifier.

Conclusions:

• Administrative routine data is valuable for predicting key aspects of psychiatric hospital care organization.
• Machine learning approaches demonstrate potential for enhancing clinical practice and hospital management.
• Further research is needed to determine the predictive performance required for effective clinical assistance.