Comparing Deep Learning and Conventional Machine Learning Models for Predicting Mental Illness from History of...

Area of Science:

• Computational psychiatry
• Artificial intelligence in healthcare

Background:

• Mental illness poses a global health challenge, necessitating advanced diagnostic tools.
• Accurate mental illness diagnosis is complex due to overlapping symptoms and predispositions.

Purpose of the Study:

• To systematically compare deep learning and conventional machine learning models for mental illness prediction.
• To evaluate model performance using free-text patient history data.

Main Methods:

• Seven deep learning models and two conventional machine learning models were assessed.
• The CB-MH architecture and an attention model were specifically highlighted.
• Integrated Gradients interpretability method was employed to identify influential features.

Main Results:

• The CB-MH model achieved the highest F1 score (0.62).
• An attention model demonstrated the best F2 score (0.71).
• Analysis revealed key features in meaningful contexts for true positives, while false negatives presented challenges with unclear contexts.

Conclusions:

• Computational models show potential for mental illness prediction from clinical notes.
• Model interpretability is crucial for understanding diagnostic predictions and improving accuracy.
• Further validation and refinement of models are needed for clinical application.