Advanced Multimodal AI for Predicting Long-Term Functional Outcomes After Ischemic Stroke Using Only Admission Data

Prognostication after acute ischemic stroke often relies on limited variables and simple risk scores, despite richer information being available at admission. We developed a multimodal AI model using admission data to predict modified Rankin Scale (mRS) outcomes and compared it to established tools.

In a retrospective study of ischemic stroke/TIA patients, we trained three modality-specific models on admission non-contrast head CT, history and physical notes, and structured clinical variables, and combined them in a weighted-average ensemble. We predicted binary (mRS 0-2 versus 3-6) and ordinal mRS (0-6) outcomes at discharge and 90 days. Performance on an external test cohort was compared with THRIVE and SPAN-100 scores using AUROC, AUPRC, Brier score, mean absolute error (MAE), and quadratic weighted kappa (QWK).

A total of 6,915 patients were split into training, validation and testing cohorts in a 3:1:1 ratio. For discharge binary mRS (n = 1596), the multimodal ensemble achieved significantly better discrimination (AUROC 0.859, AUPRC 0.858) with 25-61% lower Brier scores than THRIVE or SPAN-100 (all p < 0.001). For 90-day binary mRS (n = 207), the model also outperformed both THRIVE and SPAN-100 (AUROC 0.838, AUPRC 0.805, with 3-38% lower Brier scores). Ordinal mRS prediction showed similarly strong performance with significantly better QWK at discharge and numerically lower MAE. The multimodal ensemble model reassigned about one-third of patients to different risk categories versus THRIVE and was closer to the true discharge outcome in ~ 74% of discordant cases.

We developed a well-calibrated multimodal AI model for prediction of discharge and 90-day post-stroke functional outcomes using only data present at the time of admission. This model outperforms existing prognostic tools and can support early clinical decision-making.