Predicting new-onset post-stroke depression from real-world data using machine learning algorithm.

Area of Science:

• Neurology
• Psychiatry
• Medical Informatics

Background:

• Post-stroke depression (PSD) is a significant complication following ischemic stroke, necessitating early detection for effective clinical management.
• Real-world data offers a valuable resource for developing predictive models for new-onset PSD.

Purpose of the Study:

• To develop and validate machine learning models for predicting the occurrence of new-onset PSD after ischemic stroke.
• To identify key clinical features associated with PSD development at various time points post-stroke.

Main Methods:

• Utilized a large dataset of 61,460 ischemic stroke patients from Taiwanese medical institutions (2001-2019) for model development.
• Validated models on an independent cohort of 15,366 patients, assessing specificity and sensitivity at 30, 90, 180, and 365 days post-stroke.
• Employed machine learning to identify and rank important clinical predictors of PSD.

Main Results:

• The prevalence of PSD in the study cohort was 1.3%.
• Machine learning models achieved average specificities between 0.83-0.91 and sensitivities between 0.30-0.48.
• Ten critical features, including advanced age, new-onset hypertension, and post-stroke sleep/anxiety disorders, were identified as significant PSD predictors.

Conclusions:

• Machine learning models show promise as predictive tools for identifying patients at risk of PSD.
• Identifying key clinical factors enables clinicians to proactively screen and manage depression in high-risk stroke survivors.