Preparing for the bedside-optimizing a postpartum depression risk prediction model for clinical implementation in a...

Area of Science:

• * Computational medicine and health informatics.
• * Clinical informatics and machine learning applications in healthcare.

Background:

• * Postpartum depression (PPD) poses a significant public health challenge.
• * Electronic Health Records (EHRs) offer a rich data source for developing predictive models.
• * Integrating predictive models into clinical workflows requires rigorous evaluation.

Purpose of the Study:

• * To develop and externally validate a machine-learning model for PPD prediction using EHR data.
• * To pre-implement evaluate the model for performance, fairness, and clinical appropriateness.
• * To refine the model using debiasing techniques to ensure equitable performance.

Main Methods:

• * Utilized EHR data from an academic medical center and a clinical research network (2014-2020).
• * Evaluated predictive performance using area under the curve and sensitivity; assessed fairness with disparate impact, equal opportunity, and predictive parity metrics.
• * Employed decision curve analysis and expert review for clinical appropriateness; compared five debiasing approaches.

Main Results:

• * The baseline PPD model showed some unfairness in academic medical center data but fair performance in clinical research network data.
• * A debiasing approach, 'fairness through blindness,' was applied, yielding improved overall performance and fairness.
• * Expert reviewers confirmed the clinical appropriateness of the revised model.

Conclusions:

• * Thorough pre-implementation evaluation of predictive models is crucial, encompassing performance, fairness, and clinical appropriateness.
• * Debiasing techniques can mitigate fairness issues in machine learning models for healthcare.
• * The refined PPD prediction model shows promise for clinical decision support in EHR systems.