Using machine learning to develop a clinical prediction model for SSRI-associated bleeding: a feasibility study.

Area of Science:

• Pharmacovigilance and Machine Learning
• Clinical Informatics
• Computational Health

Background:

• Adverse drug events (ADEs) lead to poor patient outcomes and increased healthcare costs.
• Predictive tools for ADEs can mitigate risks and improve patient safety.
• Selective serotonin reuptake inhibitors (SSRIs) are commonly prescribed, but associated bleeding events require monitoring.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting bleeding events associated with SSRI use.
• To identify key clinical and demographic features that contribute to SSRI-associated bleeding.
• To assess the feasibility of using large-scale electronic health record (EHR) data for ADE prediction.

Main Methods:

• Utilized the National Institute of Health All of Us (AoU) database, including EHR data from over 10,000 participants exposed to SSRIs.
• Selected 88 features encompassing sociodemographics, lifestyle, comorbidities, and medication use.
• Applied logistic regression, decision tree, random forest, and extreme gradient boost models to predict bleeding events, assessing performance via AUC.

Main Results:

• 9.6% of 10,362 SSRI-exposed participants experienced a bleeding event.
• Machine learning models achieved AUCs ranging from 0.632 to 0.698.
• Clinically significant predictors included health literacy (for escitalopram) and bleeding history and socioeconomic status (for all SSRIs).

Conclusions:

• Demonstrated the feasibility of using ML for predicting SSRI-associated bleeding events.
• Identified specific patient characteristics that are significant risk factors for bleeding.
• Suggests potential for improved ADE prediction by incorporating genomic data and deep learning.