Towards phenotyping stroke: Leveraging data from a large-scale epidemiological study to detect stroke diagnosis

Area of Science:

• Medical informatics
• Epidemiology
• Machine learning in healthcare

Background:

• Stroke diagnosis relies on accurate phenotyping for effective treatment and research.
• Traditional methods like ICD-9 codes have limitations in precision.
• Large-scale epidemiological studies generate extensive data valuable for refining diagnostic tools.

Purpose of the Study:

• To develop and assess machine learning (ML) algorithms for detecting stroke cases and subtypes using hospitalization data.
• To evaluate ML algorithm performance against established methods (ICD-9 codes, study nurses) using real-world epidemiological data.
• To identify key predictors for developing high-precision stroke phenotypic signatures.

Main Methods:

• Utilized 8,131 hospitalization events from the Greater Cincinnati/Northern Kentucky Stroke Study (2005, 2010).
• Extracted demographic and clinical variables from patient medical records.
• Trained ML algorithms to predict stroke cases and subtypes, validated against physician-adjudicated labels.

Main Results:

• The best ML algorithm achieved 88.57% accuracy for stroke case detection and 87.39% for subtype detection.
• ML algorithms significantly outperformed ICD-9 code classifications (P<0.001) across all measures.
• ML performance was comparable to study nurses, offering a better precision-recall tradeoff.

Conclusions:

• Machine learning demonstrates significant promise for improving stroke diagnosis accuracy from hospitalization data.
• This approach can enhance statistical power for subsequent genetic and genomic studies.
• Identified predictive variables can guide the development of precise stroke phenotyping algorithms.