Applying density-based outlier identifications using multiple datasets for validation of stroke clinical outcomes.

Area of Science:

• Neurology
• Data Science
• Biostatistics

Background:

• The modified Rankin Scale (mRS) and Barthel Index (BI) are crucial for assessing stroke patient outcomes.
• Accurate measurement is vital for developing reliable machine learning models for stroke prediction.
• Data quality directly impacts the validity and performance of predictive models.

Purpose of the Study:

• To evaluate density-based outlier detection methods for identifying measurement errors in stroke outcome data.
• To assess the efficacy of these methods across multiple large-scale stroke datasets.
• To enhance the quality of data used in stroke outcome prediction models.

Main Methods:

• Applied three density-based outlier detection algorithms: DBSCAN, HDBSCAN, and LOF.
• Utilized a large dataset from a Taiwanese prospective stroke registry for initial development.
• Validated the methods on four independent NINDS-funded stroke datasets.

Main Results:

• DBSCAN demonstrated high accuracy across various mRS scores, with the highest average accuracy at mRS 4 (99.2%).
• LOF showed comparable performance to DBSCAN.
• HDBSCAN required parameter tuning for optimal performance.

Conclusions:

• Density-based outlier detection is a promising approach for validating stroke outcome measures.
• An algorithm developed on a large registry dataset proved effective on external datasets.
• This tool can improve data quality for real-world stroke outcome assessments.