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

TL;DR

This work releases MIMIC-IV-Ext-22MCTS, a large-scale temporal clinical time-series dataset containing $22{,}588{,}586$ events with relative timestamps extracted from $267{,}284$ discharge summaries, addressing the challenge of long, timestamp-sparse notes. It presents an end-to-end retrieval-augmented annotation framework that first gathers candidate event chunks via contextual BM25 and semantic search (with $0.75$ similarity threshold) and then uses Llama-3.1-8B prompts to infer temporal information from the chunks. The authors show that fine-tuning Temporal BERT on this data improves medical question answering by up to $10\%$ and clinical trial matching by $3\%$ over baselines, and that GPT-2 fine-tuned on the dataset yields more clinically reliable generations. The dataset and code are publicly available, enabling improved risk prediction, causal reasoning, and decision-support applications in healthcare, with future work on fairness and per-note dependencies.

Abstract

A crucial component for clinical risk prediction is developing a reliable prediction model is collecting high-quality time series clinical events. In this work, we release such a dataset that consists of 22,588,586 Clinical Time Series events, which we term MIMIC-\RNum{4}-Ext-22MCTS. Our source data are discharge summaries selected from the well-known yet unstructured MIMIC-IV-Note \cite{Johnson2023-pg}. The general-purpose MIMIC-IV-Note pose specific challenges for our work: it turns out that the discharge summaries are too lengthy for typical natural language models to process, and the clinical events of interest often are not accompanied with explicit timestamps. Therefore, we propose a new framework that works as follows: 1) we break each discharge summary into manageably small text chunks; 2) we apply contextual BM25 and contextual semantic search to retrieve chunks that have a high potential of containing clinical events; and 3) we carefully design prompts to teach the recently released Llama-3.1-8B \cite{touvron2023llama} model to identify or infer temporal information of the chunks. The obtained dataset is informative and transparent that standard models fine-tuned on the dataset achieves significant improvements in healthcare applications. In particular, the BERT model fine-tuned based on our dataset achieves 10\% improvement in accuracy on medical question answering task, and 3\% improvement in clinical trial matching task compared with the classic BERT. The dataset is available at https://physionet.org/content/mimic-iv-ext-22mcts/1.0.0. The codebase is released at https://github.com/JingWang-RU/MIMIC-IV-Ext-22MCTS-Temporal-Clinical-Time-Series-Dataset.

Figures (5)

• Figure 1: The pipeline of end-to-end annotation framework.
• Figure 2: An example of the original and contextualized chunk demonstrating clinical findings related to lung examination.
• Figure 3: Comparison of chunk selection distribution and histogram of chunk frequency at each step.
• Figure 4: A synthetic patient topic in TREC 2022.
• Figure 5: Example of a clinical trial "NCI-G96-1000" with title, summary, and criteria.

Theorems & Definitions (2)

• Definition 1
• Example 1