Machine Learning Based Reanalysis of Clinical Scores for Distinguishing Between Ischemic and Hemorrhagic Stroke in Low Resource Setting

BACKGROUND Identification of stroke and classifying them as ischemic and hemorrhagic type using clinical scores alone faces two unaddressed issues. One pertains to over-estimation of performance of scores and the other involves class imbalance nature of stroke data leading to biased accuracy. We conducted a quantitative comparison of existing scores, after correcting them for the above-stated issues. We explored the utility of Machine Learning theory to address overestimation of performance and class imbalance inherent in these clinical scores. METHODS We included validation studies of Siriraj (SS), Guys Hospital/Allen (GHS/AS), Greek (GS), and Besson (BS) Scores for stroke classification, from 2001-2021, identified from systematic search on PubMed, ERIC, ScienceDirect, and IEEE-Xplore. From included studies we extracted the reported cross tabulation to identify the listed issues. Further, we mitigated them while recalculating all the performance metrics for a comparative analysis of the performance of SS, GHS/AS, GS, and BS. RESULTS A total of 21 studies were included. Our calculated sensitivity range (IS-diagnosis) for SS is 40-90% (median 70%[IQR:57-73%], aggregate 71%[SD:15%]) as against reported 43-97% (78%[IQR:65-88%]), for GHS/AS 35-93% (64%[IQR:53-71%], 64%[SD:17%]) against 35-94% (73%[IQR:62-88%]), and for GS 60-74% (64%[IQR:62-69%], 69%[SD:7%]) against 74-94% (89%[IQR:81-92%]). Calculated sensitivity (HS-diagnosis), for SS, GHS/AS, and GS respectively, are 34-86% (59%[IQR:50-79%], 61%[SD:17%]), 20-73% (46%[IQR:34-64%], 44%[SD:17%]), and 11-80% (43%[IQR:27-62%], 51%[SD:35%]) against reported 50-95% (71%[IQR:64-82%]), 33-93% (63%[IQR:39-73%]), and 41-80% (78%[IQR:59-79%]). Calculated accuracy ranges, are 37-86% (67%[IQR:56-75%], 68%[SD:13%]), 40-87% (58%[IQR:47-61%], 59%[SD:14%]), and 38-76% (51%[IQR:45-63%], 61%[SD:19%]) while the weighted accuracy ranges are 37-85% (64%[IQR:54-73%], 66%[SD:12%]), 43-80% (53%[IQR:47-62%], 54%[SD:13%]), and 38-77% (51%[IQR:44-64%], 60%[SD:20%]). Only one study evaluated BS. CONCLUSION Quantitative comparison of existing scores indicated significantly lower ranges of performance metrics as compared to the ones reported by the studies. We conclude that published clinical scores for stroke classification over-estimate performance. We recommend inclusion of equivocal predictions while calculating performance metrics for such analysis. Further, the high variability in performance of clinical scores in stroke identification and classification could be improved upon by creating a global data-pool with statistically important attributes. Scores based on Machine Learning from such globally pooled data may perform better and generalise at scale.