Predicting Conversion From Unipolar Depression to Bipolar Disorder and Schizophrenia: A 10-Year Retrospective Cohort...

Area of Science:

• Psychiatry and Mental Health
• Computational Psychiatry
• Machine Learning in Healthcare

Background:

• Initial depressive symptoms in bipolar disorder (BD) and schizophrenia (SCZ) can lead to misdiagnosis as unipolar depression (UD).
• Lack of consensus on individualized, time-varying predictors for conversion from UD to BD and SCZ.
• Need for predictive models to stratify risk for BD/SCZ conversion in UD patients.

Purpose of the Study:

• To examine the conversion rate from UD to BD and SCZ.
• To develop predictive models for short-, medium-, and long-term risk stratification of BD/SCZ conversion.
• To identify key features contributing to disease progression using explainable AI.

Main Methods:

• Retrospective analysis of 12,182 depressive inpatients' electronic medical records (EMRs) from 2009-2020.
• Application of four machine-learning algorithms utilizing sociodemographic, clinical, laboratory, vital signs, symptoms, and treatment data.
• Utilized SHapley Additive exPlanations (SHAP) and Break Down for feature contribution analysis.

Main Results:

• Conversion rates: 2.82% to BD and 0.53% to SCZ.
• Risk factors for BD: female sex, severe depression, mood stabilizers, beta-blockers, antipsychotics.
• Risk factors for SCZ: male sex, psychotic symptoms, antipsychotics, antiside effect drugs, psychotherapy.
• Family history of mental illness significantly increased susceptibility to both BD and SCZ conversion.
• Refractory and psychotic UD patients showed elevated transition risk.

Conclusions:

• Predictive model performance decreased over time (AUC for BD: 0.771 at 1 year to 0.733 at 7 years; for SCZ: 0.866 at 1 year to 0.752 at 7 years).
• Social-demographic factors, lifestyle, vital signs, and blood markers became significant risk factors over follow-up.
• Validated models could offer clinicians dynamic risk information for disease conversion.