Background and purpose: We developed a machine learning model to allow early functional outcome prediction for patients presenting with posterior circulation (pc)-stroke based on CT-imaging and clinical data at admission. The proposed algorithm utilizes quantitative information from automated multidimensional assessments of posterior circulation Acute Stroke Prognosis Early CT-Score (pc-ASPECTS) regions. Discriminatory power was compared to predictions based on conventional pc-ASPECTS ratings.
Methods: We retrospectively analyzed non-contrast CTs and clinical data of 172 pc-stroke patients. 90 days outcome was dichotomized into good and poor using modified Rankin Scale (mRS) cut-offs. Predictive performance was assessed for outcome differentiation at mRS 2, 3, 4 and survival prediction (mRS ≤ 5) using random forest algorithms. Results were compared to conventional pc-ASPECTS and clinical parameters. Models were evaluated in a nested fivefold cross-validation approach.
Results: Receiver operating characteristic areas under the curves (ROC-AUCs) of the test sets using conventionally rated pc-ASPECTS reached 0.63 for mRS ≤ 4 to 0.68 for mRS ≤ 5 and 0.73 for mRS ≤ 5 to 0.85 for mRS ≤ 2 if clinical data were considered. Pure imaging-based machine learning classifier ROC-AUCs were lowest for mRS ≤ 4 (0.81) and highest for mRS ≤ 5 (0.87). The combined clinical data and machine learning-based model had the highest predictive performance with ROC-AUCs reaching 0.90 for mRS ≤ 2.
Conclusion: Machine learning-based evaluation of pc-ASPECTS regions predicts functional outcome of pc-stroke patients with higher accuracy than conventional assessments. This could optimize triage for additional diagnostics and allocation of best possible medical care and might allow required arrangements of the social environment at an early point of time.
Keywords: Machine learning; Outcome prediction; Pc-ASPECTS; Posterior circulation stroke.
© 2022. The Author(s).