Add COVID Red Dataset

[jpolsgrove]

Add COVID-RED Dataset, Detection/Prediction Tasks, and Example

Summary

This PR adds support for the COVID-RED (Remote Early Detection of SARS-CoV-2 infections) dataset to PyHealth, including:

• A wearable device dataset loader (COVIDREDDataset)
• Classification task functions (covidred_detection_fn, covidred_prediction_fn)
• A runnable usage example (covidred_example.py)

This provides a clinically relevant wearable device dataset for PyHealth users and supports reproducible research in early infectious disease detection using consumer wearables.

Feature

1. COVIDREDDataset

• Loads wearable device data (heart rate, steps, sleep) from the COVID-RED study
• Returns unified time series format consistent with PyHealth signal datasets
• Supports multiple data splits: split="train" | "test" | "all"
• Configurable sliding window approach with window_days parameter
• Two task modes:
  • Detection: Classify COVID-19 positive vs negative during illness period
  • Prediction: Early detection - predict COVID-19 onset before symptom appearance (1-14 days pre-symptomatic)
• Automatic train/test split with reproducible random seed
• Feature extraction from multivariate time series:
  • Resting heart rate statistics (mean, std, min, max)
  • Activity metrics (total steps, mean hourly steps)
  • Sleep metrics (duration, efficiency)

2. Task Functions

covidred_detection_fn Maps dataset samples into PyHealth task format for COVID-19 detection:

{
    "patient_id": str,
    "visit_id": str,
    "signal": Tensor(n_features × window_days),
    "label": int(0 or 1),
    "metadata": dict
}

covidred_prediction_fn Maps dataset samples for early COVID-19 prediction (pre-symptomatic detection):

• Identifies patterns 1-14 days before symptom onset
• Critical for early intervention and transmission reduction
• Same output format as detection task

covidred_multiclass_fn (optional extension) Extends to multiclass severity classification:

• 0: COVID-19 negative
• 1: Mild (recovered at home, no assistance)
• 2: Moderate (recovered at home with assistance)
• 3: Severe (hospitalized)

3. Example Script

• Demonstrates complete pipeline: loading → task definition → LSTM classifier training
• Implements bidirectional LSTM with attention to temporal patterns
• Includes proper evaluation metrics (accuracy, precision, recall, F1, AUC)
• Handles class imbalance with weighted loss
• Saves best model based on F1-score
• Serves as a minimal reproducible example for users

Dataset Details

Dataset: COVID-RED - Remote Early Detection of SARS-CoV-2 infections
Source: Utrecht University, Netherlands
DOI: 10.34894/FW9PO7
URL: https://dataverse.nl/dataset.xhtml?persistentId=doi:10.34894/FW9PO7

Data characteristics:

• Wearable device measurements (Fitbit, Apple Watch, Garmin)
• Daily aggregated metrics: heart rate, steps, sleep
• COVID-19 test results and symptom onset dates
• Longitudinal data across pandemic period
• Focus on pre-symptomatic and asymptomatic detection

Clinical significance:

• Early detection 1-14 days before symptom onset
• Enables early intervention and isolation
• Reduces community transmission
• Demonstrates utility of consumer wearables for public health surveillance

Tests

Basic verification performed:

• Dataset loads correctly from CSV files
• Train/test split works as expected (70/30 split)
• Both detection and prediction task functions output PyHealth-compliant dictionaries
• Example script runs end-to-end (CPU/GPU tested)
• Feature extraction handles missing values appropriately
• Label distribution matches expected class imbalance
• LSTM model architecture validated with sample data

Note on Dataset Download

The COVID-RED dataset must be manually downloaded from DataverseNL.

Users must:

1. Visit: https://dataverse.nl/dataset.xhtml?persistentId=doi:10.34894/FW9PO7
2. Download required files:
   • heart_rate.csv - Daily resting heart rate measurements
   • steps.csv - Daily step counts
   • sleep.csv - Daily sleep duration and efficiency
   • labels.csv - COVID-19 test results and symptom dates
3. Place files in a directory (e.g., /data/covidred/)
4. Initialize dataset:

   from pyhealth.datasets import COVIDREDDataset
   dataset = COVIDREDDataset(root="/data/covidred/", split="train", task="prediction")
   

Usage Example

from pyhealth.datasets import COVIDREDDataset
from pyhealth.tasks import covidred_prediction_fn
from torch.utils.data import DataLoader

# Load dataset for early COVID-19 prediction
dataset = COVIDREDDataset(
    root="/path/to/covidred",
    split="train",
    window_days=7,
    task="prediction"
)

# Apply task function
samples = [covidred_prediction_fn(dataset[i]) for i in range(len(dataset))]

# Create dataloader
dataloader = DataLoader(samples, batch_size=32, shuffle=True)

# Train your model
for batch in dataloader:
    signals = batch["signal"]  # Shape: (batch_size, n_features, window_days)
    labels = batch["label"]    # Shape: (batch_size,)
    # ... training code

Files Changed

This PR adds three new files to PyHealth:

1. pyhealth/datasets/covidred.py - Dataset loader class
2. pyhealth/tasks/covidred.py - Task functions for COVID-19 detection/prediction
3. examples/covidred_example.py - Complete usage example with LSTM classifier

Citation

If you use this dataset implementation, please cite the original COVID-RED study:

@data{FW9PO7_2021,
  author = {Olthof, A.W. and Schut, A. and van Beijnum, B.F. and others},
  publisher = {DataverseNL},
  title = {{Remote Early Detection of SARS-CoV-2 infections (COVID-RED)}},
  year = {2021},
  version = {V1},
  doi = {10.34894/FW9PO7},
  url = {https://doi.org/10.34894/FW9PO7}
}