Machine learning based reanalysis of clinical scores for distinguishing between ischemic and hemorrhagic stroke in low resource setting.

Background: Identifying ischemic or hemorrhagic strokes clinically may help in situations where neuroimaging is unavailable to provide primary-care prior to referring to stroke-ready facility. Stroke classification-based solely on clinical scores faces two unresolved issues. One pertains to overestimation of score performance, while other is biased performance due to class-imbalance inherent in stroke datasets. After correcting the issues using Machine Learning theory, we quantitatively compared existing scores to study the capabilities of clinical attributes for stroke classification.

Methods: We systematically searched PubMed, ERIC, ScienceDirect, and IEEE-Xplore from 2001 to 2021 for studies that validated the Siriraj, Guys Hospital/Allen, Greek, and Besson scores for stroke classification. From included studies we extracted the reported cross-tabulation to identify and correct the above listed issues for an accurate comparative analysis of the performance of clinical scores.

Results: A total of 21 studies were included. Comparative analysis demonstrates Siriraj Score outperforms others. For Siriraj Score the reported sensitivity range (Ischemic Stroke-diagnosis) 43-97% (Median = 78% [IQR 65-88%]) is significantly higher than our calculated range 40-90% (Median = 70% [IQR 57-73%]), also the reported sensitivity range (Hemorrhagic Stroke-diagnosis) 50-95% (Median = 71% [IQR 64-82%]) is higher than our calculated range 34-86% (Median = 59% [IQR 50-79%]) which indicates overestimation of performance by the included studies. Guys Hospital/Allen and Greek Scores show similar trends. Recommended weighted-accuracy metric provides better estimate of the performance.

Conclusion: We demonstrate that clinical attributes have a potential for stroke classification, however the performance of all scores varies across demographics, indicating the need to fine-tune scores for different demographics. To improve this variability, we suggest creating global data pool with statistically significant attributes. Machine Learning classifiers trained over such dataset may perform better and generalise at scale.