Creating Clinically Labeled MIMIC-III PPG Datasets for CAD and CVD Research: A Reproducible Workflow

Technical Summary: Creating Clinically Labeled MIMIC-III PPG Datasets for CAD and CVD Research

Problem Statement
Despite the widespread availability of Photoplethysmography (PPG) signals in public datasets, there is a significant scarcity of resources that combine patient-level PPG waveforms with clinically meaningful labels for Coronary Artery Disease (CAD) and broader Cardiovascular Disease (CVD). Existing public datasets are primarily designed for heart-rate estimation, stress recognition, or cuff-less blood pressure prediction, lacking the structured diagnosis codes and patient-level aggregation necessary for disease cohort construction. Furthermore, the lack of transparent documentation regarding preprocessing, diagnosis-code logic, and patient-level separation in existing resources limits reproducibility and risks information leakage when waveform windows from the same subject are used across different analysis partitions.

Methodology
The authors present a five-step reproducible workflow to construct clinically labeled PPG datasets from the matched waveform and clinical components of the MIMIC-III database. The workflow is implemented in Python and is designed to be rerun locally by authorized users.

1. PPG Waveform Extraction:

   • Waveforms were extracted from the MIMIC-III matched waveform database (mimic3wdb-matched/1.0) using WFDB APIs.
   • The workflow identified PPG channels by searching for "PLETH" or "PPG" in channel names.
   • Records were processed to retain the first eight minutes of finite samples. The first two minutes were discarded, and minutes 3–8 were retained as a six-minute segment.
   • Quality Control: Segments were excluded if the standard deviation was below 0.01 or the peak-to-peak amplitude was below 0.05.
   • Accepted six-minute segments were split into six one-minute windows (ppg_min_3 through ppg_min_8).

2. Demographic Linkage:

   • Waveform rows were linked to the PATIENTS and ADMISSIONS tables using SUBJECT_ID.
   • Gender was merged directly from PATIENTS.
   • Age was calculated based on the earliest admission time and the date of birth. Ages >300 (de-identified) were mapped to 90, and ages >89 were grouped as 90 for analysis.

3. Diagnosis Events and Labeling:

   • Labels were derived from the DIAGNOSES_ICD table linked with ADMISSIONS.
   • Strict CAD Label: Assigned if the patient had an ICD-9 code starting with 4140 (Coronary atherosclerosis).
   • Broad CVD Label: Assigned if the patient had ICD-9 codes starting with 410–414 (Ischemic heart disease), 420–429 (Other heart disease), or 440–448 (Arterial diseases).
   • Comorbidities: Flags for Diabetes (250), High Cholesterol (272), and Obesity (278) were generated at the subject level.
   • Label Assignment Logic: Since extracted PPG rows lacked explicit timestamps or admission IDs to match specific diagnosis events, labels were assigned at the SUBJECT_ID level. A subject was labeled positive if any relevant diagnosis evidence existed in their record.

4. Patient-Level Aggregation:

   • To prevent information leakage, the workflow reduced multiple waveform records per subject to a single selected record per SUBJECT_ID.
   • A deterministic safeguard sorted by record_id to ensure reproducibility if duplicate rows occurred.
   • The final output consists of one row per patient containing demographics, binary disease/comorbidity labels, and six one-minute PPG windows.

5. EDA and Validation:

   • The final datasets underwent validation for patient-level uniqueness, binary label integrity, case-control consistency, and waveform window completeness.

Key Contributions

• Reproducible Workflow: A documented, end-to-end pipeline for extracting, filtering, and labeling PPG data from MIMIC-III, available via a GitHub repository.
• Two Distinct Cohorts:
  • A Strict CAD Dataset focusing on specificity (ICD-9 4140), excluding unresolved cardiovascular cases from the control group.
  • A Broad CVD Dataset focusing on coverage (ICD-9 410–448), encompassing ischemic heart disease, other heart conditions, and arterial diseases.
• Leakage-Aware Design: Patient-level aggregation ensures that waveform windows from the same subject do not appear as independent samples in training/testing splits.
• Rich Metadata: The released CSV files include age, sex, comorbidity flags, waveform metadata (sampling frequency, source record), and the six PPG windows.

Results
The study produced two final patient-level datasets:

• Strict CAD Dataset: Contains 3,465 unique patients (1,625 CAD-positive, 1,840 controls).
  • CAD-positive subjects were significantly older (70.8 ± 12.6 years) than controls (55.3 ± 17.1 years), with a large effect size (Cohen's d = 1.02).
  • CAD-positive subjects showed higher prevalence of diabetes (44.6% vs. 20.1%), high cholesterol (63.1% vs. 24.2%), and obesity.
• Implemented CVD Dataset: Contains 5,456 unique patients (3,616 CVD-positive, 1,840 controls).
  • The CVD-positive group was also significantly older (68.4 ± 14.8 years) than controls (55.3 ± 17.1 years), with Cohen's d = 0.84.
  • This cohort represents 66.3% CVD-positive cases.
• Signal Characteristics: All retained records contain six one-minute windows sampled at 125 Hz.

Significance and Claims
The authors position this work not as a predictive model or a benchmark for algorithm performance, but as a dataset-contribution resource. The primary significance lies in lowering the entry barrier for researchers interested in clinically linked PPG analysis. By providing a transparent, documented workflow that handles the non-trivial tasks of waveform extraction, quality filtering, clinical table linkage, and ICD-9 label construction, the study enables the regeneration of these specific cohorts.

The authors emphasize that the datasets are intended to support future biomedical engineering studies using MIMIC-derived physiological waveforms. They explicitly note that the resource favors "transparent cohort construction and leakage-aware patient organization" over exhaustive temporal phenotyping. The paper acknowledges limitations, including the retrospective nature of the data, the use of ICD-9 codes which may introduce coding uncertainty, and the inability to verify temporal alignment between the specific PPG waveform and the diagnosis admission due to missing timestamps in the extracted rows. Consequently, the authors advise users to account for age and sex confounding and to avoid interpreting model performance as disease-specific signal evidence without appropriate adjustment.