A Multi-Modal Dataset for Ground Reaction Force Estimation Using Consumer Wearable Sensors

This paper introduces a fully open, multi-modal dataset comprising 492 validated trials of time-aligned Apple Watch IMU and force plate data from ten healthy adults performing five activities, designed to support reproducible research and benchmarking for estimating vertical ground reaction forces using consumer-grade wearables.

The Gist

Imagine you want to know how hard your feet are hitting the ground when you walk, jog, or jump. In a high-tech lab, scientists use giant, expensive "force plates" (like super-sensitive bathroom scales built into the floor) to measure this. But you can't carry a giant lab floor in your pocket.

This paper introduces a new digital toolkit that helps scientists figure out how hard your feet are hitting the ground using only the Apple Watch on your wrist or waist.

Here is the breakdown of their work, explained simply:

🧱 The Big Idea: The "Smart Watch vs. The Gold Standard"

Think of the Force Plate as the "Gold Standard" referee. It's the only one that knows the exact truth about how hard you hit the ground. However, it's stuck in a lab and costs a fortune.

The Apple Watch is like a "rookie scout." It's cheap, everyone has one, and it can sense movement, but it doesn't "see" the ground directly.

The Goal: The researchers wanted to teach the "rookie scout" (the Apple Watch) to guess the referee's score (the ground force) so accurately that we could use it anywhere, anytime, without needing the lab.

🏃‍♂️ The Experiment: A Day in the Life of 10 Athletes

To train this AI, they gathered 10 healthy adults (ages 26 to 41) and put them through a "movement obstacle course."

• The Gear: Each person wore two Apple Watches: one on their wrist (like a normal watch) and one on their waist (tucked into a belt near their belly button).
• The Lab: They walked, jogged, and ran across a special floor tile (the Force Plate).
• The Drops: They also did "heel drops" (jumping off their toes) and "step drops" (stepping off a small box) to simulate the impact of landing a jump.

The Result: They recorded 492 different attempts (trials). For every single step, they captured:

1. What the Force Plate felt (The Truth).
2. What the Waist Watch felt (The Core).
3. What the Wrist Watch felt (The Limb).

🧩 The Puzzle: Connecting the Dots

Imagine you are trying to match three different video recordings of the same event, but they started at slightly different times.

• The Force Plate started recording the second the foot touched.
• The Apple Watches were started by hand, so they might be a tiny fraction of a second off.

The researchers built a digital puzzle solver (an algorithm) that looks at the "shape" of the movement. It finds the exact moment the foot hits the ground in the Force Plate data and slides the Watch data until the "peaks" and "valleys" line up perfectly. It's like aligning two songs so the drums hit at the exact same time.

📊 What Did They Find?

1. The Waist is a Better Messenger: The watch on the waist (close to the body's center) was much better at predicting the ground force than the watch on the wrist. This makes sense; your wrist swings around a lot, while your waist moves more like your legs do.
2. The Wrist Still Has Potential: Even though the wrist is far away, it can still give a decent estimate, especially for walking and running. This is huge news because most people wear watches on their wrists, not their waists!
3. It's Reliable: When the same person did the same jump twice, the watches gave very similar results. The data is consistent.

🛠️ Why This Matters (The "So What?")

Before this, if you wanted to study how hard people hit the ground, you had to drag them into a lab with a giant machine.

Now, thanks to this open-source dataset (a giant library of data anyone can download), developers can:

• Build apps that tell runners if they are landing too hard (which causes injuries).
• Create rehab tools for elderly people to monitor their balance at home.
• Train AI models to understand human movement using just a smartwatch.

⚠️ The Fine Print (Limitations)

• Small Class Size: They only had 10 people. It's like testing a new car recipe with only 10 chefs. It's a great start, but we need more chefs to be sure it works for everyone.
• Lab vs. Real World: They did this on a smooth, perfect lab floor. Real life has mud, gravel, and uneven sidewalks. The AI might get confused outside the lab.
• One Brand: They only used Apple Watches. We don't know yet if a Samsung or Garmin watch would work the same way.

🎓 The Takeaway

This paper is like handing the world a recipe book and a set of ingredients. They didn't just say, "You can do this!" They actually did the hard work, collected the ingredients (the data), wrote down the steps (the code), and said, "Here, you can use this to build your own health apps."

It bridges the gap between expensive, stationary science and the wearable tech we carry in our pockets every day.

Technical Summary

1. Problem Statement

Ground Reaction Forces (GRFs) are critical for analyzing human locomotion, assessing musculoskeletal loading, and developing rehabilitation and sports science applications. While laboratory force plates are the "gold standard" for GRF measurement, their high cost, stationary nature, and size limit their use to controlled environments. Wearable inertial sensors offer a solution for everyday monitoring, but existing public datasets often lack:

• Consumer-grade sensors: Most datasets use research-grade IMUs placed on proximal body parts (shank, thigh, pelvis).
• Wrist placement: Wrist-worn devices (like smartwatches) are ubiquitous in consumer health but underrepresented in biomechanical datasets for estimating lower-limb forces.
• Time-aligned multi-modal data: There is a scarcity of datasets that simultaneously record consumer wearables and force plate ground truth with precise temporal alignment across multiple body locations (wrist and waist).

This paper addresses these gaps by introducing a fully open, multi-modal dataset designed to benchmark machine learning models for estimating vertical Ground Reaction Forces (vGRF) using consumer Apple Watch sensors.

2. Methodology

Participants and Protocol

• Subjects: 10 healthy adults (6 male, 4 female; aged 26–41).
• Activities: Five distinct activities were performed: Walking, Jogging, Running, Heel Drops, and Step Drops.
• Sensors:
  • Wearables: Two Apple Watch devices (Series 5 or later) per participant. One on the left wrist (standard position) and one on the waist (secured via elastic belt below the navel, near the center of mass).
  • Ground Truth: An AMTI OR6-7 force plate (1000 Hz sampling) flush-mounted to the floor.
• Data Collection: Participants performed overground trials on a 10m walkway. Trials were repeated 5–12 times per activity. A "heel-tap" marker was occasionally used at the start of trials to aid temporal alignment.

Data Acquisition & Processing

• Sampling Rates: Apple Watch IMU data (accelerometer and gyroscope) sampled at ~100 Hz; Force plate data at 1000 Hz.
• Signal Processing:
  • Filtering: Force plate data was low-pass filtered (4th-order Butterworth, 20 Hz cutoff); IMU data was filtered (2nd-order Butterworth, 10 Hz cutoff). Raw unfiltered data is also provided for impact analysis.
  • Feature Extraction: Acceleration magnitude ($a_{mag}$) was computed to reduce orientation dependence.
• Temporal Alignment: A hybrid alignment strategy was employed to synchronize the asynchronous sensors:
  1. Event-based: Detection of a "heel-tap" impulse if present.
  2. Stance-feature alignment: Using cross-correlation of the gradient of vertical force ($Force\_Z$) and acceleration magnitude during the stance phase.
  3. Validation: Monte Carlo sensitivity analysis confirmed that single-sample timing perturbations ($\pm10$ ms) had minimal impact on correlation metrics.

Quality Control

• Trial Matching: A "trial manifest" links files across modalities using unique UUIDs.
• Validation Framework: A three-phase cross-sensor plausibility check was applied:
  1. Waist $\to$ Force Plate
  2. Wrist $\to$ Waist
  3. Wrist $\to$ Force Plate
• Repeatability: Intraclass Correlation Coefficients (ICC) for peak vGRF were calculated to ensure consistency.

3. Key Contributions

1. First Consumer Wrist-Waist Dataset: The first publicly available dataset pairing consumer-grade Apple Watch sensors (at both wrist and waist) with laboratory force plate ground truth for vGRF estimation.
2. Comprehensive Multi-Modal Structure:
   • 492 validated trials across 10 participants and 5 activities.
   • 395 triad-complete trials (80.3%) containing synchronized data from wrist IMU, waist IMU, and force plate.
   • Includes raw and processed time series, metadata, quality-control flags, and machine-readable data dictionaries.
3. Rigorous Technical Validation:
   • Repeatability: Excellent repeatability for peak vGRF (ICC range: 0.871–0.990).
   • Alignment Robustness: Demonstrated that correlation-based metrics are robust to $\pm10$ ms timing jitter (mean absolute change in correlation $\approx 0.01$).
   • Saturation Analysis: Systematic detection of accelerometer clipping (saturation) in high-impact tasks, providing flags for users to handle data limitations.
4. Open Science & Reproducibility:
   • Data released under CC BY 4.0 on Zenodo.
   • Full analysis scripts (Python 3.9+) for alignment, sensitivity analysis, and figure generation are available on GitHub.
   • Includes specific recommendations for machine learning evaluation (e.g., Leave-One-Participant-Out cross-validation) to prevent data leakage.

4. Results

• Data Availability:
  • Wrist data available for 93.7% of trials.
  • Waist data available for 91.7% of trials.
  • Force plate data available for 91.3% of trials.
• Cross-Sensor Correlations (Triad-Complete Subset, n=395):
  • Waist $\to$ Force Plate: Mean Pearson $r = 0.550$.
  • Wrist $\to$ Force Plate: Mean Pearson $r = 0.486$.
  • Wrist $\to$ Waist: Mean Pearson $r = 0.486$.
  • Note: Locomotor tasks (walking/running) showed stronger correlations than impact tasks (drops), reflecting the complexity of distal-to-proximal force transmission.
• Force Magnitudes:
  • Walking: 1.1–1.2 $\times$ body weight (BW).
  • Running: 2.0–2.5 $\times$ BW.
  • Drop tasks: 2.5–4.0 $\times$ BW.
• Sensor Performance:
  • Accelerometer saturation was detected in 3.5% of wrist trials and 14.6% of waist trials, primarily during step drops and running.
  • Timing gate speeds confirmed consistent self-selected pacing (Walking: 1.32 m/s; Jogging: 2.47 m/s; Running: 3.71 m/s).

5. Significance

This dataset fills a critical void in wearable biomechanics research by enabling the development and benchmarking of consumer-grade vGRF estimation models. Its significance lies in:

• Democratizing Biomechanics: It allows researchers to train models using data from devices that are widely available to the general public, rather than expensive research labs.
• Sensor Placement Insights: By including both wrist and waist data, it facilitates the systematic investigation of how sensor location affects force estimation accuracy, supporting the feasibility of "distal-only" (wrist-only) inference.
• Benchmarking Standard: It provides a standardized, high-quality reference for comparing machine learning algorithms (e.g., deep learning vs. traditional regression) for GRF estimation.
• Methodological Rigor: The inclusion of detailed alignment logs, saturation flags, and sensitivity analyses sets a new standard for transparency in wearable sensor datasets, guiding future researchers on how to handle timing offsets and signal artifacts.

In summary, this resource supports reproducible research in digital health, sports science, and rehabilitation, specifically targeting the transition from laboratory-based force measurement to real-world wearable monitoring.