MultimodalVA

Verbal autopsy interviews record both structured symptom questions and a free-text narrative of the circumstances before a death. MultimodalVA brings the two sources together in one reproducible workflow, so researchers can train and compare narrative-only, structured-data, and multimodal cause-of-death classifiers under consistent data splits and evaluation metrics.

GitHub · Zenodo release · Preprint (medRxiv) · Dissertation preprint

What it enables

• Prepare narrative and structured verbal-autopsy records for reproducible modeling
• Train narrative-only, indicator-only, and multimodal cause-of-death classifiers
• Compare data fusion, feature fusion, soft voting, and stacking under shared data splits and metrics
• Run hyperparameter optimization, evaluation, and visualization workflows
• Use prompt-based pipelines for open-weight chat models through Hugging Face or local serving with Ollama