README.md

metadata

license: mit
task_categories:
  - time-series-forecasting
tags:
  - heart-rate
  - running
  - physiological-modeling
  - lstm
  - endomondo
size_categories:
  - 10K<n<100K

Endomondo Heart Rate Prediction Dataset V2

Dataset Summary

This dataset contains 40,186 running workouts from 761 athletes, designed for heart rate prediction from speed and altitude time-series.

Each workout includes:

• Time-series: Heart rate (target), speed, altitude, timestamps
• Metadata: Workout type, duration, user ID
• Statistics: Pre-computed HR/speed metrics for filtering

Dataset Structure

Splits

Split	Workouts	Description
Train	28,130	Training set (70%)
Validation	6,027	Validation set (15%)
Test	6,029	Test set (15%)

Features

Feature	Type	Description
workout_id	int	Unique workout identifier
user_id	int	Anonymous user identifier
workout_type	string	RECOVERY, STEADY, or INTENSIVE
duration_min	float	Workout duration in minutes
data_points	int	Number of timesteps (max 500)
heart_rate	list[float]	Heart rate time-series [BPM]
speed	list[float]	Speed time-series [km/h]
altitude	list[float]	Altitude time-series [meters]
timestamp	list[float]	Unix timestamps [seconds]
hr_mean	float	Average heart rate [BPM]
hr_std	float	HR standard deviation
hr_min	float	Minimum HR [BPM]
hr_max	float	Maximum HR [BPM]
speed_mean	float	Average speed [km/h]
speed_max	float	Maximum speed [km/h]
altitude_gain	float	Cumulative elevation gain [m]
split	string	train / validation / test

Workout Type Distribution

Type	Count	Description
RECOVERY	15,095	Easy runs (low intensity)
STEADY	22,991	Moderate pace runs
INTENSIVE	2,100	High intensity workouts

Data Quality

All workouts have been:

1. Filtered for quality (removed HR anomalies, corrupted data)
2. Smoothed with 7-point moving average (reduces GPS noise)
3. Validated against physiological constraints:
   • HR mean ≥ 120 BPM
   • HR max ≤ 200 BPM
   • HR std ≥ 5 BPM
   • Speed-HR correlation ≥ -0.3

Removed: 6,064 low-quality workouts

Usage Example

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("rricc22/endomondo-hr-prediction-v2")

# Access splits
train_data = dataset['train']
test_data = dataset['test']

# Example workout
workout = train_data[0]
print(f"Workout Type: {workout['workout_type']}")
print(f"Duration: {workout['duration_min']:.1f} min")
print(f"Avg HR: {workout['hr_mean']:.1f} BPM")
print(f"Avg Speed: {workout['speed_mean']:.1f} km/h")

# Access time-series
heart_rate = workout['heart_rate']  # List of HR values
speed = workout['speed']            # List of speed values

Model Performance

This dataset was used to train an LSTM model achieving:

• 7.42 BPM Mean Absolute Error
• 17% improvement over baseline

See the model card: rricc22/heart-rate-prediction-lstm

Try the demo: Heart Rate Predictor

Source

• Original Data: Endomondo dataset
• Processing Pipeline: Quality filtering → Smoothing → Feature engineering
• Version: V2 (January 2026)
• License: MIT

Citation

If you use this dataset, please cite:

@dataset{endomondo_hr_v2,
  title={Endomondo Heart Rate Prediction Dataset V2},
  author={Riccardo},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/rricc22/endomondo-hr-prediction-v2}
}

Related Resources

• 🤗 Model: heart-rate-prediction-lstm
• 🚀 Demo: Interactive Predictor
• 📊 GitHub: SUB3_V2 Repository