MIMIC-IV-Ext-22MCTS: A 22 Million-Event Temporal Clinical Time-Series Dataset for Risk Prediction.

Area of Science:

• Medical Informatics
• Machine Learning
• Natural Language Processing

Background:

• High-quality time series clinical events are essential for reliable machine learning-based healthcare risk prediction.
• Existing datasets like MIMIC-IV-Note contain unstructured discharge summaries, posing challenges for event extraction and temporal information retrieval.

Purpose of the Study:

• To create a comprehensive dataset of clinical time series events (MIMIC-IV-Ext-22MCTS) with extracted temporal information.
• To develop a novel framework for processing lengthy discharge summaries and inferring timestamps for clinical events.
• To improve the performance of machine learning models in healthcare applications through fine-tuning on the new dataset.

Main Methods:

• Developed a framework to break down long discharge summaries into manageable text chunks.
• Utilized contextual BM25 and semantic search to identify relevant text chunks containing clinical events.
• Employed carefully designed prompts with the Llama-3.1-8B model to extract or infer event timestamps.
• Extracted 22,588,586 clinical events and their associated temporal information from MIMIC-IV-Note.

Main Results:

• The MIMIC-IV-Ext-22MCTS dataset provides informative and transparent clinical time series data.
• Fine-tuning BERT on this dataset resulted in a 10% accuracy improvement in medical question answering and a 3% improvement in clinical trial matching.
• GPT-2 models fine-tuned on the dataset demonstrated more clinically reliable results for clinical questions.

Conclusions:

• The proposed framework effectively extracts clinical events and temporal information from unstructured discharge summaries.
• The MIMIC-IV-Ext-22MCTS dataset significantly enhances the performance of machine learning models in various healthcare tasks.
• This work facilitates the development of more accurate and reliable clinical risk prediction tools.