Predicting recovery following stroke: Deep learning, multimodal data and feature selection using explainable AI.

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Machine learning (ML) holds promise for predicting post-stroke symptoms and rehabilitation response.
• High-dimensional neuroimaging data, small datasets, and integrating diverse data types pose significant challenges.
• Accurate prediction is crucial for personalized stroke recovery strategies.

Purpose of the Study:

• To evaluate strategies for combining neuroimaging and tabular data for post-stroke aphasia prediction.
• To introduce and assess a novel convolutional neural network (CNN) approach integrating regions of interest (ROIs) with tabular data.
• To predict spoken picture description ability (aphasic vs. non-aphasic) in stroke survivors.

Main Methods:

• Compared 2D image summaries of MRI scans and feature selection strategies.
• Developed a CNN trained on combined MRI-derived ROIs and symbolic tabular data representations.
• Evaluated 2D and 3D CNN architectures on MRI and tabular data from 758 stroke survivors.
• Utilized a five-group split for training, validation, and a held-out test set (lock-box).

Main Results:

• Baseline logistic regression achieved 0.678 accuracy (lesion size alone).
• Accuracy increased to 0.757 (adding symptom severity) and 0.813 (adding recovery time).
• The best performance (0.854 accuracy, 0.899 AUC, 0.901 F1) was achieved with a 2D Residual Neural Network (ResNet) using 8 ROIs and tabular data.
• This model also performed best on a challenging subset of 286 participants with moderate to severe initial aphasia (AUC = 0.865).

Conclusions:

• Combining neuroimaging (ROIs) and tabular data via CNNs significantly enhances post-stroke aphasia classification accuracy.
• This approach demonstrates effectiveness even with limited dataset sizes typical in ML.
• Future work can further improve accuracy by incorporating images directly from hospital scanners.