MIMIC-III-Ext-PPG: A PPG Benchmark Dataset for Cardiorespiratory Analysis

@article{PhysioNet-mimic-iii-ext-ppg-1.1.0,
  author = {Moulaeifard, Mohammad and Charlton, Peter H and Strodthoff, Nils},
  title = {{MIMIC-III-Ext-PPG: A  PPG Benchmark Dataset for Cardiorespiratory Analysis}},
  journal = {{PhysioNet}},
  year = {2026},
  month = mar,
  note = {Version 1.1.0},
  doi = {10.13026/r6k1-xt76},
  url = {https://doi.org/10.13026/r6k1-xt76}
}

Abstract

MIMIC-III-Ext-PPG is a large-scale, annotated, and quality-assessed dataset of photoplethysmogram (PPG) segments extracted from the MIMIC-III Waveform Database Matched Subset. It contains around 6.3 million non-overlapping 30-second segments from 6,189 critically-ill patients and serves as the largest public resource for heart rhythm classification. The dataset contains signals recorded during several heart rhythms: sinus rhythm, atrial fibrillation, atrial flutter, heart block, pacing, and several less common rhythms. PPG segments were extracted within the 15-minute period prior to a patient's heart rhythm being recorded in the medical record, providing reference rhythm annotations in line with best practice from the literature. In addition, simultaneous electrocardiogram (ECG, lead II), arterial blood pressure (ABP), and respiratory (RESP) signals are provided where available. Additional annotations were derived from these signals, including blood pressure, heart rate, and respiratory rate. Signal quality metrics are provided for all signals. The MIMIC-III-Ext-PPG dataset is a valuable resource for a variety of downstream prediction tasks, enabled by its size, range of reference annotations, and signal quality metrics.

Background

Physiological signals such as PPG, ECG, ABP, and RESP are widely used to monitor health. In the Intensive Care Unit (ICU), these signals are often monitored continuously, whilst wearables such as smartwatches use PPG and ECG for health monitoring. Much research focuses on extracting physiological information from these signals, with tasks including arrhythmia detection [1], blood pressure estimation [2], and signal quality assessment. However, public datasets are often limited in scale or lack support for comprehensive classification and regression tasks (e.g., DeepBeat [3], PulseDB [4], MIMIC-BP [5]).

To fill this gap, we introduce MIMIC-III-Ext-PPG, a large-scale, curated dataset with around 6.3 million non-overlapping 30-second PPG segments. Each segment is annotated with a heart rhythm obtained from a rhythm documentation in the medical notes, at most 15 minutes after the waveform data. In addition, simultaneous ECG (lead II), ABP, and RESP signals and derived parameters from them are also included where available.

Methods

Data Processing Pipeline

• Waveform Selection:
  We used records from the MIMIC-III Waveform Database Matched Subset [6], which contains digitized Critical Care waveforms (PPG, ECG, ABP, RESP) sampled at 125 Hz.
  • Inclusion criteria:
    A waveform record was included if it met all of the following:
    1. Contains a PPG/PLETH channel with valid sampling metadata.
    2. Its time interval overlaps with at least one rhythm chart event.
    3. Can be successfully parsed in WFDB format.
    4. Contains ≥30 seconds of non-flat, non-empty PPG signal before quality filtering.
    Exclusion criteria:
    A waveform record was excluded if:
    1. It did not contain a PPG signal.
    2. No rhythm chart events occurred within its waveform time span.
    3. The WFDB file was corrupted or unreadable.
    4. The PPG channel was entirely flat, constant, NaN, or missing.
    5. Remaining usable PPG was <30 seconds after extraction.
• Clinical Metadata Extraction:
  Chart events, including heart rhythm annotations and demographic measurements, were extracted from the MIMIC-III clinical database using both CareVue and MetaVision systems.
  • Chart event inclusion criteria:
    1. The event timestamp falls within the waveform recording.
    2. Rhythm label text is non-empty and parsable.
    Chart event exclusion criteria:
    1. Non-rhythm chart events.
    2. Rhythm labels that were blank, ambiguous, or malformed.
• Event Matching:
  For each waveform, we identified rhythm events that occurred within its time span and retrieved:
  (i) the current rhythm annotation,
  (ii) the previous rhythm event, and
  (iii) the nearest height and weight records.
• Annotation Harmonization:
  Rhythm labels from CareVue and MetaVision were unified into a single standardized vocabulary (see Table: heart-rhythms_before_filter).
• Diagnosis Mapping:
  ICD-9 discharge diagnoses were mapped to ICD-10 codes [7], expanded hierarchically, and truncated to 3-digit levels to support generalizable analysis.
• Demographics:
  Patient age, sex, ethnicity, height (in cm), and weight (in kg) were extracted from structured fields in the clinical data.
• Data Stratification:
  We divided the dataset into 10 non-overlapping folds using unique patient identifiers. Stratification was balanced by rhythm class, age strata, and ICD-10 Chapter IX cardiovascular diagnoses to minimize data leakage.
• Signal Segment Extraction:
  For each matched event, we extracted waveform data ending at the annotation time and starting from the previous rhythm annotation or a point up to 15 minutes earlier, whichever was more recent. Segments were then split into 30-second windows, and leftover fragments under 30 seconds were discarded. All segments contained PPG (PLETH) signals. In addition, where available, simultaneous ECG (lead II), ABP, and RESP signals were also extracted.
• Signal Storage:
  All extracted 30-second waveform segments were stored in WFDB format (.hea and .dat) following PhysioNet conventions.
• Timestamp Handling:
  The event_time (timestamp of the rhythm chart event) and start_segment (start time of each 30-second waveform window) values directly inherit MIMIC-III’s date-shifted timestamps.

Enrichment with Additional Annotations and Signal Quality Metrics

Each 30-second segment was enriched with additional annotations and signal quality metrics based on the available waveforms:

• RESP was analyzed over the full 30-second segment.
• ABP, PPG, and ECG were evaluated in three consecutive 10-second windows, producing a 3-value vector representing the intervals 0–10 seconds, 10–20 seconds, and 20–30 seconds. For these signals, both the vector and aggregated statistics (median with IQR) were reported.

The processes for extracting additional annotations and signal quality metrics are now described.

Enrichment with Additional Annotations: SBP, DBP, HR, and RR

• Respiratory Rate (RR, from RESP, if available):
  RR was derived over the full 30-second segment using the algorithm in [8]. An instantaneous RR was calculated from each valid breath duration, and RRs were summarized for each segment as the median and IQR.
• Heart Rate (HR, from ECG, if available):
  HR was calculated from beat-to-beat intervals within each 10-second window using the Neurokit QRS detector [9], with the final HR obtained from the median RR interval [10]. For each segment, a 3-element vector (one per window) was extracted, along with the segment-level median and IQR.
• Systolic and Diastolic Blood Pressure (SBP and DBP from ABP, if available):
  SBP and DBP were computed using the slope sum function (SSF) beat detection algorithm [11] within each 10-second window. Similarly to HR, a 3-element vector and segment-level median and IQR were reported.

Enrichment with Signal Quality Metrics

We implemented a signal quality assessment pipeline for RESP, PPG, ECG, and ABP signals. Quality was quantified using an integer-based signal quality index (SQI) code.

SQI codes:

Data Description

Dataset Summary

The following table summarizes the data available for each task (heart rhythm classification, RR, HR, and BP estimation) in the MIMIC-III-Ext-PPG dataset. It provides the number of subjects, segments, and demographic statistics available.

We also summarized the dataset in terms of heart rhythm labels (obtained from the 'event_rhythm' column of the 'metadata.csv' file).

Waveforms

This dataset is structured in WFDB format, making it compatible with a wide range of signal processing tools, including PhysioNet’s WFDB software, Python (wfdb, biosppy, neurokit2), MATLAB, and Julia packages. Each record contains waveform data (PPG channels and, when available, ECG, ABP, RESP channels) segmented into 30-second waveform segments, along with a rich set of metadata provided in accompanying .csv files.

It is worth noting that in our internal analyses (e.g., HR, RR, BP, and SQI calculations) preprocessing was performed within dedicated signal-specific pipelines, and only the resulting derived features were stored in the metadata. However, the provided WFDB files contain segmented raw waveform signals extracted from the original MIMIC-III matched subset. No preprocessing (e.g., filtering, denoising, or normalization) has been applied to the released signals. Therefore, users are responsible for applying appropriate preprocessing steps (e.g., bandpass filtering) before conducting signal analysis or training AI models.

Metadata

This dataset includes a rich set of segment-level metadata for each signal sample, provided in a structured metadata.csv file.

Signal Information

Heart Rhythm

BP

HR

RR

PPG, ECG, ABP; 10-second Window SQIs

RESP; 30-second Window SQIs

Subject Demographics

Diagnostic Information

Stratification

Usage Notes

Folder structure

We organized the data in a folder structure analogous to the MIMIC-III Waveform Matched Subset:

MIMIC-III-Ext-PPG
p00
  ├── p000052
  │  ├── 3238451_0005_0_1.hea
  │  ├── 3238451_0005_0_1.dat
  │  ├── ...
  ├── p000107
  │  ├── 3805787_0011_0_4.hea
  │  ├── 3805787_0011_0_4.dat
  │  ├── ...
  ├── ...
p01
  ├── p010045
  │  ├── 3120713_0001_0_0.hea
  │  ├── 3120713_0001_0_0.dat
  │  ├── ...
  ├── p010049
  │  ├── 3456766_0016_0_1.hea
  │  ├── 3456766_0016_0_1.dat
  │  ├── ...
  ├── ...
...
p09
  ├── p090012
  │  ├── 3522957_0007_0_0.hea
  │  ├── 3522957_0007_0_0.dat
  │  ├── ...
  ├── ...

Access Logic

Locating WFDB Files and Corresponding Metadata

I) From Metadata → WFDB Signal → Metadata Lookup

1. Use the path in the folder_path column of the metadata.
2. Use the following logic based on subject_id and signal_file_name:

• Files are organized as pXX/pXXXXXX/, where:
• pXX is the first two digits of subject_id (zero-padded).
• pXXXXXX is subject_id zero-padded to 6 digits.
• Inside each pXXXXXX folder, the WFDB files are named using signal_file_name (signal_file_name.hea and signal_file_name.dat).
• Minimal Code Example:

import wfdb
import pandas as pd
import os

# --------------------------------------------------------
# 1. Load metadata
# --------------------------------------------------------
df = pd.read_csv("metadata.csv")

# Select a specific segment (e.g., first row)
meta_row = df.iloc[X]

# --------------------------------------------------------
# 2. Build the WFDB path from metadata fields
# --------------------------------------------------------
folder = meta_row["folder_path"]
signal_file = meta_row["signal_file_name"]

wfdb_path = os.path.join(folder, signal_file)

# --------------------------------------------------------
# 3. Load the waveform segment using WFDB
# --------------------------------------------------------
record = wfdb.rdrecord(wfdb_path)

II) From WFDB File → Find Metadata Entry

To retrieve metadata for a given .hea or .dat file in the physionet/ folder:

• Extract the 6-digit subject ID from the folder name pXXXXXX; this gives you subject_id.
• Extract the file name without the extension; this gives you signal_file_name.
• Use these two values to look up the corresponding row in metadata.csv.
• Minimal Code Example:

import wfdb
import pandas as pd

# Load metadata
df = pd.read_csv("metadata.csv")

# Suppose we manually pick a WFDB file path:
example_path = "p03/p0300123/3238451_0005_0_1"  # without .hea/.dat
signal_file = example_path.split("/")[-1]     # "3238451_0005_0_1"

# Look up metadata row
row = df[df["signal_file_name"] == signal_file]

Intended Uses and Limitations

The dataset supports various research and machine learning tasks, including:

• Arrhythmia detection and classification
• Cuffless blood pressure estimation
• Heart rate and respiratory rate estimation
• Signal quality assessment

Additionally, demographic and diagnostic metadata enable stratified and subgroup analyses. This dataset is well-suited for benchmarking deep learning models.

Release Notes

The current version of the database is v1.1.0.

Revision history:

MIMIC-III-Ext-PPG v1.0.0

Initial public release.

MIMIC-III-Ext-PPG v1.1.0

In v1.0.0, we excluded numerous segments with apparently mislabeled channel assignments, which we identified based on conservative filtering criteria. Subsequent investigation revealed that these issues stemmed from corrupted downloads rather than systematic problems in the original MIMIC-III waveform files themselves. For v1.1.0, we revised our download procedure and verified file integrity using the provided checksums. Re-validation of the signals confirmed that the previously excluded segments were not inherently mislabeled. Consequently, we released v1.1.0 as an extended version of the dataset, retaining 99.3% of v1.0 waveforms while adding 1,513,372 previously excluded 30-second PPG segments. In the metadata of v1.1.0, we introduced a small number of updates compared to v1.0.0. Specifically, we removed a few columns (e.g., vector_10s_mean_rpeak_deriv_pleth and vector_10s_mean_peak_deriv_pleth). These columns were originally introduced as conservative checks to detect potentially corrupted segments. Furthermore, we updated a few metadata values for signals shared between v1.1.0 and v1.0.0 (e.g., weight, height, event_time, and event_rhythm) to improve consistency and accuracy.

Ethics

This dataset is derived entirely from the MIMIC-III database, a publicly available critical care resource generated at the Beth Israel Deaconess Medical Center.

MIMIC-III was released under the oversight of the Institutional Review Boards (IRBs) of both Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology (MIT), with all data collection performed under the same IRB approvals and HIPAA-compliant de-identification procedures.

All patient data in MIMIC-III were included under the IRB-approved waiver of informed consent, due to the retrospective nature of the data and the use of rigorous de-identification protocols.

The MIMIC-III-Ext-PPG dataset does not introduce any new patient data; it consists solely of reprocessed and curated outputs derived from the existing MIMIC-III waveform and clinical records.

Accordingly, it inherits the same IRB approvals, consent waivers, and ethical framework as MIMIC-III. No additional data collection, patient contact, or identifiable information was involved in the creation of this dataset.

Acknowledgements

The project (22HLT01 QUMPHY) has received funding from the European Partnership on Metrology, co-financed from the European Union’s Horizon Europe Research and Innovation Programme and by the Participating States. Funding for the University of Cambridge was provided by Innovate UK under the Horizon Europe Guarantee Extension, grant number 10091955. PHC acknowledges funding from the British Heart Foundation (BHF) grant [FS/20/20/34626].

Conflicts of Interest

The authors declare no conflicts of interest.

References

1. Moulaeifard M, Coquelin L, Rinkevičius M, Sološenko A, Pfeffer O, Bench C, Hegemann N, Vardanega S, Nandi M, Alastruey J, Heiss C. Machine-learning for photoplethysmography analysis: Benchmarking feature, image, and signal-based approaches. arXiv preprint arXiv:2502.19949. 2025 Feb 27.
2. Moulaeifard M, Charlton PH, Strodthoff N. Generalizable deep learning for photoplethysmography-based blood pressure estimation--A Benchmarking Study. arXiv preprint arXiv:2502.19167. 2025 Feb 26.
3. Torres-Soto J, Ashley EA. Multi-task deep learning for cardiac rhythm detection in wearable devices. NPJ digital medicine. 2020 Sep 9;3(1):116.
4. Wang W, Mohseni P, Kilgore KL, Najafizadeh L. PulseDB: A large, cleaned dataset based on MIMIC-III and VitalDB for benchmarking cuff-less blood pressure estimation methods. Frontiers in Digital Health. 2023 Feb 8;4:1090854.
5. Sanches I, Gomes VV, Caetano C, Cabrera LS, Cene VH, Beltrame T, Lee W, Baek S, Penatti OA. MIMIC-BP: A curated dataset for blood pressure estimation. Scientific Data. 2024 Nov 15;11(1):1233.
6. Moody B, Moody G, Villarroel M, Clifford G, Silva I. MIMIC-III waveform database matched subset. MIMIC-III Waveform Database Matched Subset v1. 0. 2020 Apr.
7. World Health Organization. International statistical classification of diseases and related health problems: 10th revision (ICD-10). https://icd.who.int/browse10/2019/en. 1992.
8. Charlton PH, Bonnici T, Tarassenko L, Clifton DA, Beale R, Watkinson PJ, Alastruey J. An impedance pneumography signal quality index: Design, assessment and application to respiratory rate monitoring. Biomedical signal processing and control. 2021 Mar 1;65:102339.
9. Makowski D, Pham T, Lau ZJ, Brammer JC, Lespinasse F, Pham H, Schölzel C, Chen SA. NeuroKit2: A Python toolbox for neurophysiological signal processing. Behavior research methods. 2021 Aug;53(4):1689-96.
10. Orphanidou C, Bonnici T, Charlton P, Clifton D, Vallance D, Tarassenko L. Signal-quality indices for the electrocardiogram and photoplethysmogram: Derivation and applications to wireless monitoring. IEEE journal of biomedical and health informatics. 2014 Jul 23;19(3):832-8.
11. Sun JX. Cardiac output estimation using arterial blood pressure waveforms (Doctoral dissertation, Massachusetts Institute of Technology).