NLM DIR Seminar Schedule

Abstract:

Early detection and timely treatment are critical in medicine. For example, surgical excision of skin lesions can cure early-stage skin cancer, but once metastasis occurs, even the most advanced therapies often fail. In this work, we introduce MIMIC-EXT-TE, a large-scale dataset provides a structured timeline of over a million clinical events from MIMIC-IV-Note. It is the first dataset with temporal information of events in patient level. To achieve the dataset, we propose to integrate retrieval-augmented generation with large language models to capture the temporal trajectories of patient events. To evaluate the dataset, we introduce TimeLife, a temporal-aware medical question answering system by fine-tuning the Qwen3-4B-Base language model on our dataset. TimeLife achieves an 18% overall accuracy boost on MedMCQA dataset compared with the base model. By fine-tuning TimeLife with downstream tasks such as PubMedQA and MedMCQA, TimeLife achieves the superiority most of the time compared with fine-tuning only on the base model without our dataset.