MIMIC-\RNum{4}-Ext-22MCTS: A 22 Millions-Event Temporal Clinical Time-Series Dataset with Relative Timestamp for Risk...

Area of Science:

• Medical Informatics
• Natural Language Processing
• Machine Learning

Background:

• High-quality clinical time series data are essential for accurate machine learning-based risk prediction in healthcare.
• Existing datasets like MIMIC-IV-Note contain unstructured discharge summaries, posing challenges due to length and missing explicit timestamps for clinical events.

Purpose of the Study:

• To create a novel, large-scale dataset of clinical time series events (MIMIC-4-Ext-22MCTS) from unstructured discharge summaries.
• To develop a robust framework for extracting clinical events and their temporal information from lengthy medical texts.
• To improve the performance of machine learning models in healthcare applications through the use of this new dataset.

Main Methods:

• Developed a framework to process lengthy discharge summaries by segmenting them into smaller chunks.
• Utilized contextual BM25 and semantic search to identify relevant text chunks containing clinical events.
• Employed prompt engineering with the Llama-3.1-8B model to identify and infer timestamps for clinical events.

Main Results:

• Created the MIMIC-4-Ext-22MCTS dataset, comprising 22,588,586 clinical time series events.
• Fine-tuning standard models on this dataset led to significant performance improvements in healthcare tasks.
• BERT models achieved a 10% accuracy increase in medical question answering and a 3% increase in clinical trial matching.

Conclusions:

• The MIMIC-4-Ext-22MCTS dataset provides informative and transparent clinical time series data.
• This dataset effectively enhances the performance of machine learning models for critical healthcare applications.
• The proposed framework offers a scalable solution for extracting temporal clinical event data from unstructured medical records.