MIMIC-IV Dataset Overview

Updated 18 January 2026

MIMIC-IV is a comprehensive public EHR resource capturing over 10 years of critical care records at BIDMC, enabling robust clinical and epidemiologic studies.
Its modular structure organizes data into core, hospital, ICU, note, and specialty modules, supporting detailed analyses such as time series, NLP, and phenotyping.
Standardized preprocessing pipelines and benchmarks facilitate reproducible modeling for tasks like risk prediction, mortality assessment, and multimodal analysis.

The Medical Information Mart for Intensive Care IV (MIMIC-IV) dataset is a large-scale, publicly accessible electronic health record (EHR) corpus encompassing over a decade of adult inpatient and ICU care at Beth Israel Deaconess Medical Center. It is designed as a research-standard resource for clinical machine learning, benchmarking, epidemiologic surveillance, process mining, and fairness auditing. MIMIC-IV’s modular architecture supports detailed analyses of time series, clinical notes, multimodal imaging, and event logs and serves as the foundation for sophisticated derived datasets and benchmarks spanning risk prediction, phenotyping, and NLP.

1. Structural Composition and Submodules

MIMIC-IV is organized into several schemas reflecting key clinical domains:

Core tables: Demographics (patients), admissions, and ICU stays permit cohort extraction by age, diagnosis, or care setting.
Hospital tables: Diagnoses (ICD-9/10), laboratory events, prescriptions, procedures, and in-hospital outcome flags.
ICU modules: High-frequency time series from charted vitals, interventions, and outputs.
Note modules: De-identified discharge summaries, radiology reports, and other clinical documentation.
Specialty extensions: Emergency Department (MIMIC-IV-ED), ECG signals (MIMIC-IV-ECG), and derived phenotype datasets, such as MIMIC-IV-Ext-PE and MEETI.

This architecture supports extraction of static and longitudinal features, linkage of notes to events, and reproducible splits by patient or hospital encounter (Nguyen et al., 2023, Bui et al., 2024).

2. Data Processing, Preprocessing, and Benchmark Pipelines

Several open-source pipelines have emerged for cleaning and featurizing raw MIMIC-IV data:

Feature harmonization involves unit conversion, ICD mapping (including ICD-9→ICD-10), and cohort definition by age, diagnosis, or condition (Gupta et al., 2022).
Outlier removal resorts to percentile clipping or winsorization; for example, numeric fields clipped at their 1st and 99th percentiles.
Time-series construction is typically performed by aggregation into fixed intervals (e.g., 1–2 h bins), forward/backward imputation, and concatenation with static features.
Missingness management includes binary indicators for imputed values and statistical summaries for feature inclusion.
Train/validation/test splits are executed at the patient or admission level to prevent leakage, using stratified or randomized cross-validation (Liao et al., 2023, Nguyen et al., 2023).

For benchmarking, standardized feature sets are developed for tasks such as mortality, length of stay (LOS), readmission, and phenotype prediction. These are distributed as reproducible CSV/Parquet, with code repositories guiding researchers through extraction, preprocessing, and evaluation (Gupta et al., 2022, Liao et al., 2023).

3. Key Benchmark Tasks and Derived Datasets

3.1 Extreme multilabel ICD coding

The Mimic-IV-ICD benchmark links discharge summaries to all ICD-9/10 codes (diagnoses and procedures), organizing over 330,000 admissions for multilabel text classification. Unique code counts reach 11,311 (ICD-9) and 26,096 (ICD-10), with label cardinality per sample at 13.4–16.1. Evaluation uses micro/macro F1 and Precision/Recall@k. Top-50 code subsets facilitate standardization across studies (Nguyen et al., 2023).

3.2 Event log and process mining

The MIMICEL event log enables analytic process mining of ED throughput—extracting activity traces (enter ED, triage, vital sign check, medication dispensation, discharge) with precise timestamps and case attributes. Length-of-stay ΔLoS, path durations, transition frequencies, and crowding indicators are formalized to facilitate conformance checking and flow optimization (Wei et al., 26 May 2025).

3.3 NLP and phenotype extraction

MIMIC-IV-Ext-PE extracts nearly 20,000 PE phenotype labels through manual adjudication and transformer-based NLP (VTE-BERT) of CTPA radiology reports. Reports are classified into acute/chronic/equivocal/negative PE. VTE-BERT achieves sensitivity 92.4%, PPV 87.8%, and specificity 98.9% on external MIMIC-IV validation, outperforming ICD code–based billing data for precision and coverage (Lam et al., 2024).

MIMIC-IV-Ext-22MCTS corpus parses 267,284 discharge summaries into 22.6M short “event” spans with relative timestamps (hours from admission), derived via chunking, BM25/semantic retrieval, and Llama-3–assisted prompt engineering. This enables fine-tuned BERT and GPT-2 models to achieve substantial downstream gains on clinical QA and trial matching tasks (Wang et al., 1 May 2025).

4. Modeling Strategies, Fairness, and Interpretability

Time series models: XGBoost delivers the strongest baseline on tabular/flattened features (AUROC up to 0.87 for in-ICU mortality), often outperforming deep sequence models (LSTM, TCN) in irregular, sparse EHRs. Advanced methods (BoXHED 2.0) leverage nonparametric hazards and time-varying covariate trees for real-time risk monitoring, exceeding classical Cox and recurrent neural network performance in held-out ICU data (AUC-ROC 0.83) (Nowroozilarki et al., 2021, Bui et al., 2024).
Interpretability: ROAR-based feature ablation, Integrated Gradients, DeepLIFT, and attention scores reveal critical variables (labs, respiratory parameters, demographics, SAPS-II components) for mortality tasks. ArchDetect achieves the best empirical feature ranking fidelity. Glassbox attention aids IMV-LSTM–based interpretability (Meng et al., 2021).
Fairness audits: Minimum and macro-AUC across subgroups quantify model fairness; disparities are observed in mechanical ventilation utilization, LOS prediction sensitivity, and dependency on ethnicity/insurance attributes. Post-hoc subgroup metrics and group-feature importance parity guide mitigation strategies; IMV-LSTM offers both high accuracy (AUROC 0.955) and subgroup fairness (Meng et al., 2021, Kakadiaris, 2023).

5. Advanced Modalities and Multimodal Learning

MEETI extends MIMIC-IV-ECG by incorporating four aligned modalities per ECG record: raw waveform, paper-style image, beat-level parameters (RR, PR, QT, QRS intervals, amplitudes), and GPT-4o–generated interpretation text. This comprehensive multimodal structure (indexed by study ID) supports transformer fusion, vision-language architectures, and explainability studies for detailed cardiac phenotyping and AI-aided diagnosis (Zhang et al., 21 Jul 2025).

6. Accessibility, Reproducibility, and Usage Considerations

MIMIC-IV access is regulated via PhysioNet credentialing and HIPAA-compliant data use agreements. Nearly all derived benchmarks, event logs, and processing pipelines are distributed as open-source, with explicit documentation, configuration controls, and provenance tracking for reproducibility. Python, SQL, R, and Jupyter-based workflows are standard, enabling transparent cohort definition, feature selection, model training, and fairness audits (Liao et al., 2023, Gupta et al., 2022, Nguyen et al., 2023, Wei et al., 26 May 2025).

Researchers are advised to initiate modeling efforts with validated modular pipelines, critically assess population bias, imputation, and outlier handling, and document all analytic steps for cross-study comparability. Benchmark splits and code lists are shared to foster reproducible care trajectory analyses and facilitate external validation (Nguyen et al., 2023, Meimeti et al., 18 Mar 2025).

7. Impact, Future Directions, and Methodological Extensions

MIMIC-IV and its extensions have accelerated research in risk stratification, clinical process mining, and multimodal disease phenotyping. By easing the extraction bottleneck, standardizing featurization, and supporting multimodal fusion, they lower barriers to clinically meaningful AI. Methodological trajectories include adaptation to continuous-time model architectures (ODE-RNN, GRU-D), advanced attention–based interpolation for sparse series, transfer learning across institutions, and integration of complex event logs for process-aware prediction (Bui et al., 2024, Liao et al., 2023, Wang et al., 1 May 2025, Zhang et al., 21 Jul 2025).

Expanding on techniques pioneered in RegEx-based report identification, multimodal temporal annotation, and transformer-driven phenotype extraction, future research is poised to incorporate additional phenotypes, richer imaging and waveform data, and causality-aware modeling frameworks. The ongoing external validation, rigorous benchmarking, and open sharing of protocols reinforce MIMIC-IV’s role as the foundational dataset for critical care ML and data-driven translational research.