Machine learning for early dynamic prediction of functional outcome after stroke.

Background: Prediction of outcome after stroke is critical for treatment planning and resource allocation but is complicated by fluctuations during the first days after onset. We propose a machine learning model that can provide hourly predictions based on the integration of continuous variables acquired within 72 h of hospital admission.

Methods: We analyzed 2492 admissions for ischemic stroke in the Geneva University Hospital from 01.01.2018 to 31.12.2021, amounting to 2'131'752 unique data points. We developed a transformer model that continuously included clinical, physiological, imaging, and biological data recorded within 72 h of admission. This model was trained to generate hourly predictions of mortality and morbidity. Shapley additive explanations were used to identify the most relevant predictors to explain outcomes for each patient. The MIMIC-III database was used for external validation.

Results: Our transformer model predicts mortality, with an area under the receiver operating characteristic curve of 0.830 (95% CI 0.763-0.885) on admission, reaching 0.893 (95% CI 0.839-0.933) 72 h later for a 3-month outcome. Validated in an independent cohort, it outperforms all static models. Based on their mean explanatory weights, the top predictors included continuous clinical evaluation, baseline patient characteristics, timing from admission to acute treatment, and markers of inflammation and organ dysfunction.

Conclusions: The performance of our transformer model demonstrates the potential of machine learning models integrating clinical, physiological, imaging, and biological variables over time after stroke. The clinical applicability of our model is further strengthened by access to hourly updated predictions along with accompanying explanations.