Acute-Phase Machine Learning Prediction of 12-Month Aphasia and Discourse Recovery

Area of Science:

• Neuroscience
• Computational Linguistics
• Medical Imaging

Background:

• Aphasia affects 30-40% of stroke patients at 12 months.
• Early forecasting of language recovery is crucial for guiding rehabilitation and clinical trial enrichment.
• Current machine learning (ML) prediction models often rely on chronic-phase data, which is unavailable at the acute decision point.

Purpose of the Study:

• To investigate whether acute-phase features can predict 12-month language recovery outcomes after ischemic stroke.
• To determine if global aphasia severity and connected-speech recovery share common substrates within an ML framework.
• To develop and validate ML models for forecasting aphasia resolution and discourse normalization using acute clinical and imaging data.

Main Methods:

• Studied 73 patients with acute left-hemisphere ischemic stroke and aphasia.
• Defined 12-month outcomes as aphasia resolution (Western Aphasia Battery-Revised Aphasia Quotient [WAB-AQ] ≥93.8) and discourse normalization (Modern Cookie Theft content units ≥22.1).
• Trained four ML algorithms on hierarchical feature sets (clinical, volumetric, anatomical, network-disconnection) using nested cross-validation and SHapley Additive exPlanations (SHAP) stability analysis.

Main Results:

• Acute WAB-AQ was the dominant predictor for aphasia resolution (mean |SHAP| = 13.60).
• Random forest model achieved high accuracy for aphasia resolution (F1 = 0.874), with clinical features alone yielding F1 = 0.851.
• Support vector regression model achieved moderate accuracy for discourse normalization (F1 = 0.725), with shared predictors including acute WAB-AQ, lesion volume, and left pars triangularis.

Conclusions:

• Acute-phase ML models can accurately forecast 12-month aphasia resolution and modestly forecast discourse normalization.
• Clinical features account for the majority of predictive variance in language recovery.
• Acute imaging data reveals shared and outcome-specific neural substrates, supporting early patient stratification for rehabilitation and clinical trials.