Multivariate prediction of functional outcome using lesion topography characterized by acute diffusion tensor imaging

Area of Science:

• Neuroimaging
• Stroke Medicine
• Machine Learning

Background:

• Stroke outcome prediction traditionally relies on manual lesion segmentation, which may oversimplify complex stroke topography.
• Continuous imaging parameters reflecting ischemia severity could offer more nuanced insights into long-term functional recovery.
• Diffusion Tensor Imaging (DTI) provides quantitative measures of acute ischemic injury.

Purpose of the Study:

• To investigate if advanced imaging parameters from DTI, analyzed with machine learning, can improve prediction of long-term functional outcome after stroke.
• To identify critical brain regions contributing to functional recovery by analyzing model contributions.
• To compare the predictive accuracy of DTI-derived parameters against standard lesion segmentation.

Main Methods:

• Eighty-seven thrombolyzed stroke patients underwent DTI at 24 hours post-stroke.
• Functional outcome was assessed at 3 months using the modified Rankin Scale (mRS), dichotomized into good (mRS ≤ 2) and poor (mRS > 2).
• Support Vector Machines (SVM) were employed to build predictive models using DTI parameters (FA, MD, AD, RD asymmetry maps), lesion segmentation, and clinical variables (age, recanalization, thrombectomy).

Main Results:

• SVM classifiers using axial diffusivity (AD) maps achieved the highest predictive accuracy (median 82.8%) compared to lesion segmentations (76.7%).
• Clinical variables, particularly lesion volume, thrombectomy treatment, and recanalization status, significantly contributed to outcome prediction.
• Highest model weights identified deep white matter regions at the crossroads of major tracts as critical for recovery.

Conclusions:

• Axial diffusivity is a more sensitive imaging marker for characterizing stroke topography in predicting long-term functional outcome than binary lesion segmentation.
• Machine learning models integrating advanced DTI parameters and clinical data enhance stroke outcome prediction accuracy.
• This approach provides valuable insights into neuroanatomical substrates critical for stroke recovery.