Master's thesis: multimodal machine learning framework for clinical trial outcome prediction with explainability and structured clinical knowledge extraction
Used clinical natural-language processing on unstructured oncology/clinical trial reports to extract structured signals for downstream prediction tasks. The work involved converting free-text protocols and clinical features into model-ready representations while handling missingness and ensuring robust, leakage-controlled evaluation. Explainability was integrated to attribute predictions to clinically relevant features across phases. • Extracted structured signal from free-text clinical trial reports • Managed missing data and uncertainty in real-world text • Designed leakage-controlled pipeline with out-of-distribution evaluation • Added SHAP-based phase-level interpretability and feature stability checks