README.md

---
license: mit
task_categories:
- time-series-forecasting
language:
- en
tags:
- eNose
- multivariate time series
- disease screening
- medical dataset
- sensor drift
- breath analysis
pretty_name: S-OH
size_categories:
- 1K<n<10K
---
# Scent of Health (S-O-H) Dataset

## Overview

The **Scent of Health (S-O-H)** dataset is the largest public clinical electronic nose (eNose) collection for non-invasive disease screening via exhaled breath analysis. It comprises **1,234 patients** across **eight diagnostic groups** (healthy controls and seven diseases), each providing a **17-channel multivariate time series** of breath measurements.

| Property | Value |
|----------|-------|
| **Patients** | 1,234 |
| **Diagnostic groups** | 9 (healthy + 8 diseases) |
| **Time series channels** | 17 (eNose sensors) + auxiliary sensors |
| **Sampling rate** | 0.4 Hz |
| **Duration per sample** | 895 seconds (~15 minutes) |
| **Collection period** | 13 consecutive weeks |
| **Clinical sites** | 2 |

## Repository Structure

```
S-OH/
├── README.md                       # This file
├── LICENSE.txt                     # MIT License
├── metadata.csv                    # Patient metadata (demographics, diagnosis, site, week)
├── data/
│ ├── manifest.json                 # Index of all patient files
│ ├── Z00/                          # Healthy controls
│ │ ├── patient_1.json
│ │ ├── patient_2.json
│ │ └── ...
│ ├── B18/                          # Hepatitis B/C
│ │ ├── patient_10.json
│ │ └── ...
│ ├── K29/                          # Gastritis and duodenitis
│ ├── K76/                          # Non-alcoholic fatty liver disease
│ ├── E11/                          # Diabetes mellitus type II
│ ├── N18/                          # Chronic renal failure
│ ├── J44/                          # COPD
│ ├── C34/                          # Lung cancer
│ └── A15/                          # Respiratory tuberculosis
└── scripts/
  ├── quick_start.py                # Load metadata and patient JSONs
  ├── validate_metadata.py          # Metadata check
  ├── validate_temporal_splits.py   # Temporal splits check
  ├── baseline_lstm_lung_cancer.py  # Example: ML/AI Use Case (LSTM baseline for C34)
  ├── baseline_cnn_z00.py           # Temporal splits check
  └── baseline_resnet18_z00.py      # Example: ML/AI Use Case (LSTM baseline for C34)
```

## Dataset Structure

### `metadata.csv`

CSV file containing patient metadata with the following columns:

| Column | Description |
|--------|-------------|
| `Patient_id` | Unique patient identifier |
| `Patient_age` | Age in years |
| `Patient_gender` | Gender (0 = female, 1 = male) |
| `Diagnosis` | ICD-10 diagnosis code |
| `D_class` | Disease class (0–7) |
| `D_bin_class` | Binary class for one-vs-rest classification |
| `Datetime` | Collection timestamp |
| `Week` | Collection week (1–13) |
| `Site` | Clinical site (SiteA or CiteB) |

### `data/` - Per-Patient JSON Files

Each patient is stored as a separate JSON file in a subdirectory named after their ICD-10 diagnosis code. The file name format is `patient_{Patient_id}.json`. Example path: `data/Z00/patient_1.json`.

Each patient JSON file has the following structure:

```json
{
    "patient_id": 1,
    "patient_diag_class": 0,
    "startDateTime": "2025-09-01T08:05:47.676148Z",
    "startTimeGases": 20,
    "endTimeGases": 450,
    "durationSec": 895,
    "sensors": [
        {
            "id": "enose",
            "sampleRate": 0.4,
            "channels": [
                {"id": "R1", "samples": [float, ...]},
                {"id": "R2", "samples": [float, ...]},
                ...
                {"id": "R17", "samples": [float, ...]},
                {"id": "humidity", "samples": [float, ...]},
                {"id": "temperature", "samples": [float, ...]}
            ]
        },
        {
            "id": "ze03",
            "sampleRate": 0.4,
            "channels": [{"id": "0", "samples": [float, ...]}]
        },
        {
            "id": "mhz14",
            "sampleRate": 0.4,
            "channels": [{"id": "0", "samples": [float, ...]}]
        },
        {
            "id": "ze08",
            "sampleRate": 0.4,
            "channels": [{"id": "0", "samples": [float, ...]}]
        },
        {
            "id": "bme280",
            "sampleRate": 0.4,
            "channels": [
                {"id": "pressure", "samples": [float, ...]},
                {"id": "temperature", "samples": [float, ...]},
                {"id": "humidity", "samples": [float, ...]}
            ]
        }
    ]
}
```

### `data/manifest.json`

Index file mapping patient IDs to their JSON file paths:

```json
{
    "total_patients": 1234,
    "files": [
        {"patient_id": 1, "diagnosis": "Z00", "file": "Z00/patient_1.json"},
        {"patient_id": 2, "diagnosis": "Z00", "file": "Z00/patient_2.json"},
        ...
    ]
}
```

### `scripts/`

Utility scripts for loading and processing the dataset.

## eNose Channels (17 channels)

The eNose sensor array consists of 17 channels printed on a single chip:

| Channel ID | Material |
|-------------|-----------|
| R1–R17 | ZnO and metal-doped ZnO (In-ZnO, Ag-ZnO, Ce-ZnO, Ni-ZnO) |

## Auxiliary Sensors

| Sensor ID | Measurements |
|-------------|-----------|
| ze03  | Ozone (O₃) |
| mhz14 | Carbon dioxide (CO₂) |
| ze08  | Carbon monoxide (CO) |
| bme280  | Pressure, temperature, humidity |

## Quick Start

``` python
import json
import pandas as pd

# 1. Load metadata
metadata = pd.read_csv('../metadata.csv')

# 2. Load patient data via manifest
with open('../data/manifest.json', 'r') as f:
    manifest = json.load(f)

# 3. Load a specific patient
patient_id, icd = '1', 'Z00'
entry = next(e for e in manifest['files'] if e['patient_id'] == patient_id)
with open(f"../data/{icd}/{entry['file']}", 'r') as f:
    patient_data = json.load(f)

# 4. Extract eNose signals
for sensor in patient_data['sensors']:
    if sensor['id'] == 'enose':
        for channel in sensor['channels']:
            print(f"{channel['id']}: {len(channel['samples'])} samples")
```

## Temporal Train/Test Splits

The dataset includes explicit temporal splits to enable drift-aware evaluation. For each disease, test weeks were selected to be temporally separated from training weeks, simulating real-world deployment conditions.

## Ethics

- The study protocol was approved by the Local Ethics Committee at Anonymized Clinical Institute (SiteA) and Anonymized Research Institute (SiteB).

- All participants provided written informed consent.

## Citation

If you use this dataset in your research, please cite:
``` bibtex
@article{soh2026,
    title = {Scent of Health (S-OH): Olfactory Multivariate Time Series Dataset for Non-Invasive Disease Screening},
    author = {Anonymized and ...},
    journal = {MICCAI Open Data},
    year = {2026}
}
```

## License

This dataset is released under the **MIT License**. See `LICENSE.txt` for full terms. You are free to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the dataset, subject to the condition that the copyright notice and permission notice are included in all copies or substantial portions.

## Contact

For questions or issues, please open an issue on this repository or contact the corresponding author (see paper for details).