Efficient Multimodal Clinical Question Answering for Pulmonary Embolism Risk Assessment

The study presents a benchmark for multimodal large language models (MLLMs) applied to pulmonary embolism (PE) risk assessment, utilizing the INSPECT dataset comprising 23,248 CTPA studies. The research evaluates models like Gemma4 E4B and Gemma4 E2B across various input modalities (CTPA only, EHR only, and combined) using zero-shot and few-shot prompting, revealing better performance in diagnostic tasks compared to prognostic ones. This work highlights the potential of compact multimodal models in enhancing early-stage PE risk detection and clinical decision-making.