AI and mental health: evaluating supervised machine learning models trained on diagnostic classifications.

Area of Science:

• Psychiatry
• Artificial Intelligence
• Computational Neuroscience

Background:

• Machine learning (ML) is increasingly applied in psychiatry, showing initial promise for diagnostics.
• Current ML models often rely on the Diagnostic and Statistical Manual of Mental Disorders (DSM) categories.
• DSM categories have known limitations, including heterogeneity and low predictive validity, impacting psychiatric diagnoses.

Purpose of the Study:

• To critically evaluate the limitations of supervised ML models in mimicking clinician-based psychiatric judgments.
• To argue that current ML applications in psychiatry, focused on DSM classifications, offer limited added value to patients.
• To propose a shift in ML focus towards improving prognosis, treatment selection, and prevention in mental health.

Main Methods:

• Critical analysis of supervised ML techniques applied to psychiatric data.
• Evaluation of the impact of training data validity (DSM categories) on ML model performance.
• Conceptual framework for reorienting ML applications in psychiatry.

Main Results:

• Supervised ML models trained on DSM classifications inherit the validity issues of these categories.
• High accuracy in ML models predicting DSM classifications is misleading and does not validate the classification itself.
• Current ML approaches offer little demonstrable added value to patient outcomes due to inherent diagnostic limitations.

Conclusions:

• ML models mimicking clinician judgments based on DSM categories have limited utility in improving psychiatric patient outcomes.
• A paradigm shift is needed, focusing ML on transdiagnostic goals like prognosis, treatment selection, and prevention.
• Reorienting ML towards these goals can enhance personalized treatment strategies and better support clinicians in mental healthcare.