[Paper Review] An Extensive Data Processing Pipeline for MIMIC-IV

Mehak Gupta, Brennan Gallamoza+4 more|arXiv (Cornell University)|Apr 29, 2022

Machine Learning in Healthcare8 citations

TL;DR

This paper presents a customizable, open-source data processing pipeline for MIMIC-IV that standardizes EHR data cleaning, feature engineering, and time-series construction for clinical prediction tasks. It enables reproducible, user-defined cohort creation and integrates end-to-end modeling and evaluation for readmission, length of stay, mortality, and phenotype prediction, with demonstrated performance across multiple deep learning models including LSTM and TCN.

ABSTRACT

An increasing amount of research is being devoted to applying machine learning methods to electronic health record (EHR) data for various clinical purposes. This growing area of research has exposed the challenges of the accessibility of EHRs. MIMIC is a popular, public, and free EHR dataset in a raw format that has been used in numerous studies. The absence of standardized pre-processing steps can be, however, a significant barrier to the wider adoption of this rare resource. Additionally, this absence can reduce the reproducibility of the developed tools and limit the ability to compare the results among similar studies. In this work, we provide a greatly customizable pipeline to extract, clean, and pre-process the data available in the fourth version of the MIMIC dataset (MIMIC-IV). The pipeline also presents an end-to-end wizard-like package supporting predictive model creations and evaluations. The pipeline covers a range of clinical prediction tasks which can be broadly classified into four categories - readmission, length of stay, mortality, and phenotype prediction. The tool is publicly available at https://github.com/healthylaife/MIMIC-IV-Data-Pipeline.

Motivation & Objective

• To address the lack of standardized, reproducible preprocessing for MIMIC-IV, which hinders comparability and reliability in EHR-based ML research.
• To lower the barrier to entry for researchers by providing a user-friendly, wizard-style pipeline that supports customizable cohort definition and data preprocessing.
• To enhance reproducibility by recording and sharing all preprocessing and modeling choices, including feature selection, imputation, and model hyperparameters.
• To support a broad range of clinical prediction tasks—readmission, length of stay, mortality, and phenotype prediction—through a unified, extensible framework.
• To integrate end-to-end modeling and evaluation, including fairness assessment and model interpretability, to support robust clinical ML development.

Proposed method

• The pipeline is implemented as a Jupyter Notebook with sequential, wizard-style code blocks guiding users through data extraction, cleaning, and pre-processing steps.
• It supports outlier removal, missing value imputation, and clinical feature grouping using standardized coding systems (e.g., ICD-10, LOINC) for dimensionality reduction.
• Time-series data are constructed by binning clinical observations into user-defined intervals (e.g., 2-hour windows), with options to filter sequence length.
• The pipeline enables customizable cohort selection based on user-defined criteria, such as specific diagnoses or time windows (e.g., first 24 or 48 hours post-admission).
• It includes built-in support for training and evaluating multiple ML and deep learning models (e.g., LSTM, TCN, GRU) across four clinical prediction tasks.
• Evaluation includes standard metrics (AUROC, AUPRC), fairness assessment (e.g., demographic parity), and model interpretability techniques, all configurable via a parameter file.

Experimental results

Research questions

• RQ1 How can a standardized, customizable pipeline improve reproducibility and usability of MIMIC-IV for clinical prediction research?
• RQ2 To what extent can a unified pipeline support diverse clinical prediction tasks, including readmission, mortality, length of stay, and phenotype prediction?
• RQ3 Can the integration of modeling and evaluation components within the pipeline reduce the technical burden for non-expert users in EHR-based ML research?
• RQ4 How does the pipeline’s configurability impact cohort definition and feature selection for downstream prediction tasks?
• RQ5 What is the performance of state-of-the-art models (e.g., LSTM, TCN) when trained using the pipeline’s standardized preprocessing and time-series construction?

Key findings

• The pipeline achieved AUROC scores of 0.86 and AUPRC of 0.49 for 30-day readmission prediction using an LSTM model on a chronic kidney disease cohort.
• For mortality prediction, the LSTM model reached an AUROC of 0.87 and AUPRC of 0.49, demonstrating strong performance on a clinically relevant task.
• The TCN model achieved an AUROC of 0.85 and AUPRC of 0.47 for 30-day readmission, showing competitive performance with other architectures.
• The pipeline successfully enabled fairness evaluation, with one model (LSTM for mortality) showing a demographic parity difference of 0.041, indicating moderate fairness performance.
• The LSTM model achieved an AUROC of 0.85 and AUPRC of 0.46 for predicting length of stay >3 days, demonstrating robustness across different outcome types.
• The pipeline’s modular design allows for full reproducibility, with all preprocessing and modeling choices recorded and shareable, enhancing transparency and comparability.