Leveraging deep neural network and language models for predicting long-term hospitalization risk in schizophrenia

Area of Science:

• Psychiatry and Mental Health
• Artificial Intelligence in Healthcare
• Biomedical Informatics

Background:

• Early identification of long-term hospitalization risk in schizophrenia (SCZ) is vital for resource management and tailored patient care.
• Current prediction methods often lack the comprehensive data integration needed for high accuracy.

Purpose of the Study:

• To develop and validate a deep learning model for forecasting extended hospital stays in SCZ patients at admission.
• To assess the impact of integrating demographic, behavioral, and blood test data on prediction accuracy.

Main Methods:

• A retrospective cohort study utilizing a deep learning model trained on multimodal data (demographic, behavioral, blood tests).
• Natural language processing (NLP) techniques were employed to extract 95% of unstructured electronic health records (EHRs) data.
• Model performance was evaluated using classification accuracy and Area Under the Curve (AUC), with interpretability and ablation studies for validation.

Main Results:

• The deep learning model achieved a classification accuracy of 0.81 and an AUC of 0.9.
• Key predictors for long-term hospitalization included advanced age, longer disease duration, elevated neutrophil-to-lymphocyte ratio, lower lymphocyte percentage, and reduced albumin levels.
• Integrating multimodal data significantly improved prediction accuracy compared to demographic data alone (0.81 vs. 0.73).

Conclusions:

• Deep learning models integrating multimodal data offer superior performance in predicting long-term SCZ hospitalizations compared to traditional methods.
• This approach provides a robust framework for early intervention and personalized treatment strategies in schizophrenia management.
• The model demonstrates potential for reducing discrimination and erroneous dependencies in clinical predictions.