Multi-dimensional MRI representation and privileged learning approaches to functional outcome prediction for ischemic stroke patients.
Authors
Affiliations (7)
Affiliations (7)
- Gilbert S. Omenn Computational Medicine and Bioinformatics, University of MIchigan, Ann Arbor, MI, USA. [email protected].
- Department of Neurosurgery, Virginia Commonwealth University, Richmond, VA, USA.
- Department of Neurological Surgery, MIchigan Medicine, Ann Arbor, MI, USA.
- Gilbert S. Omenn Computational Medicine and Bioinformatics, University of MIchigan, Ann Arbor, MI, USA.
- Department of Emergency Medicine, MIchigan Medicine, Ann Arbor, MI, USA.
- MIchigan Institute for Data and AI in Society (MIDAS), University of MIchigan, Ann Arbor, MI, USA.
- Center for Data-Driven Drug Development and Treatment Assessment (DATA), University of MIchigan, Ann Arbor, MI, USA.
Abstract
Accurate prediction of long-term functional outcomes for stroke patients remains a clinical challenge, despite advances in diagnostics and treatments. Most machine learning (ML) and artificial intelligence (AI) outcome prediction models lack robust strategies to incorporate high-dimensional clinical imaging data or to account for data availability at the point of care. In this study, we present an enhanced representation learning pipeline that fuses 2.5D magnetic resonance images (MRI), clinical data, and imaging biomarkers to predict 90-day modified Rankin Scale (mRS) for ischemic stroke patients. Using autoencoder-generated MRI embeddings, we systematically evaluate multiple ML/AI methods for both classification (90-day mRS > 2) and ordinal regression. Critically, we incorporate a privileged information paradigm, leveraging features available during training to enhance model generalizability without requiring them during inference. We developed and validated our models on a large public dataset (N = 974) and an external validation dataset (N = 738), demonstrating performance comparable to state-of-the-art convolutional neural network-based approaches with the added benefit of modularity (Test AUC 0.801; F1 0.699; MAE 1.179). Our results highlight the promise of representation learning and privileged information paradigms for bridging the gap between research and bedside prognosis in computational neurology.