Development of a computerized adaptive diagnostic screening tool for psychosis

Area of Science:

• Psychiatry
• Machine Learning
• Computational Neuroscience

Background:

• Accurate diagnosis of psychotic disorders is crucial for effective treatment.
• Distinguishing between schizophrenia, schizoaffective disorder, and mood disorders with psychosis presents clinical challenges.

Purpose of the Study:

• To develop and validate a machine learning-based diagnostic classification system for psychotic disorders.
• To differentiate schizophrenia and schizoaffective disorder from depression and bipolar disorder with psychosis (Stage 1).
• To differentiate schizophrenia from schizoaffective disorder (Stage 2).

Main Methods:

• Development of a two-stage diagnostic system using an extremely randomized trees algorithm.
• Utilized a bank of 73 clinician-rated items from inpatient and outpatient samples.
• Performance evaluated using Area Under the Receiver Operator Characteristic Curve (AUC) for out-of-sample classification.

Main Results:

• Stage 1 achieved outstanding classification accuracy (AUC = 0.93, 95% CI = 0.89, 0.94) differentiating schizophrenia/schizoaffective disorder from depression/bipolar disorder with psychosis.
• Stage 2 demonstrated excellent classification accuracy (AUC = 0.86, 95% CI = 0.83, 0.88) differentiating schizophrenia from schizoaffective disorder.
• The system requires an average of 5 items for Stage 1 and 6 items for Stage 2.

Conclusions:

• The developed two-stage machine learning system shows high accuracy in classifying psychotic disorders.
• This approach offers a promising, item-efficient method for aiding in the differential diagnosis of complex psychiatric conditions.
• The system's performance suggests potential for clinical application in diagnostic support tools.