Multi-dimensional MRI representation and privileged learning approaches to functional outcome prediction for ischemic...

Area of Science:

• Computational Neurology
• Medical Imaging Analysis
• Machine Learning in Healthcare

Background:

• Predicting long-term functional outcomes for stroke patients is a significant clinical challenge.
• Existing machine learning (ML) and artificial intelligence (AI) models struggle to integrate high-dimensional imaging data and handle point-of-care data limitations.

Purpose of the Study:

• To develop an enhanced representation learning pipeline for predicting 90-day modified Rankin Scale (mRS) in ischemic stroke patients.
• To fuse 2.5D magnetic resonance images (MRI), clinical data, and imaging biomarkers for improved outcome prediction.

Main Methods:

• Utilized autoencoder-generated MRI embeddings to create feature representations.
• Evaluated multiple ML/AI methods for classification (mRS > 2) and ordinal regression.
• Implemented a privileged information paradigm, using training-available features to boost model generalizability for inference.

Main Results:

• Models were developed and validated on large public (N=974) and external (N=738) datasets.
• Achieved performance comparable to state-of-the-art convolutional neural network approaches (Test AUC 0.801, F1 0.699, MAE 1.179).
• Demonstrated modularity and enhanced generalizability through the privileged information paradigm.

Conclusions:

• Representation learning and privileged information paradigms show significant promise for improving stroke outcome prediction.
• The developed pipeline offers a modular approach to bridge the gap between research and bedside prognosis in computational neurology.