📄 bioengineering

Edge-First Ground Reaction Force Estimation with Consumer Smartwatches

This paper presents an edge-first wearable system using consumer smartwatches and a compact temporal convolutional network to estimate vertical ground reaction forces locally, achieving strong correlation with laboratory measurements, particularly for cyclic locomotion.

Original authors: Ghaffarzadeh, P., Chakraborty, D., Aslansefat, K., Dostan, A., Papadopoulos, Y.

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Ghaffarzadeh, P., Chakraborty, D., Aslansefat, K., Dostan, A., Papadopoulos, Y.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to understand how hard your feet hit the ground every time you take a step. Scientists call this the "Ground Reaction Force" (GRF). It's like a secret handshake between your body and the Earth, telling us exactly how much weight you're putting down and how your muscles are working to keep you upright. Usually, to measure this, you have to visit a high-tech lab and walk over a giant, expensive pressure-sensitive floor tile. It's accurate, but it's also stuck in one place, like a museum exhibit you can't take home.

Now, imagine if your smartwatch could do the same job. That's the big question this paper tackles. The researchers are asking: Can we turn the tiny sensors inside a regular consumer watch into a portable force-measuring machine? They aren't trying to replace the lab entirely; instead, they want to see if we can get a "good enough" reading right from your wrist or waist, without needing to send your data to the cloud or wait for a supercomputer. It's about taking a complex physics problem and shrinking it down to fit in your pocket, all while keeping your private movement data safe on your own phone.

The "Edge-First" Detective Team

The authors of this paper built a system they call "Edge-First," which is a fancy way of saying they made the smartwatch and the phone do all the heavy lifting themselves, without asking the internet for help. Think of it like a detective squad where the two Apple Watches are the spies gathering clues, and the iPhone is the detective's office where all the solving happens.

Here's how the team works: You wear two Apple Watch Series 6 devices—one on your wrist and one on your waist. As you walk, jog, or run, these watches act like high-speed cameras, snapping 100 pictures of your movement every single second. They don't just take photos; they record how you're shaking and spinning in three dimensions. These watches then whisper this data to your iPhone via a short-range radio link. The iPhone acts as the brain, organizing the data, cleaning up the static, and running a special AI model called GRFNet-MultiScale to guess what the ground force looks like. The best part? The whole process happens right there on your phone. No Wi-Fi, no cloud servers, just you and your device.

The Magic Model: A Multi-Scale Time Machine

The secret sauce is the AI model, GRFNet-MultiScale. Imagine trying to understand a song. You need to hear the quick, sharp drum hits (the moment your heel strikes the ground) and the long, flowing melody of the rest of the step. A simple model might only hear the drum or only the melody, but this model is like a super-listener that hears both at once.

The researchers designed this model to look at the data in different "time scales." It has special layers that zoom in on tiny, split-second impacts and other layers that zoom out to see the bigger picture of your entire step. By combining these views, the model can reconstruct the force curve—the story of how hard you push off the ground—just from the shaking of your arm and waist.

What They Found: The Wrist is a Surprising Hero

When the team tested this system on 10 healthy people, the results were quite promising, though with some clear limits.

The Dual-Sensor Power: When they used both the wrist and waist watches together, the system was pretty good at guessing the force. It matched the real lab measurements with a correlation of 0.798 (where 1.0 is a perfect match) and had an error margin of 257 N (Newtons, the unit of force).
The Wrist-Only Surprise: Here's the fun twist. Even if you only wear the watch on your wrist and forget the waist one, the system still works! It retained 82.5% of the accuracy of the two-watch setup. This suggests that your arm swing carries a lot of the "footprint" of your walking, even though your foot is far away.
The Activity Gap: The system is a champion at rhythmic activities like walking, jogging, and running. It got the best scores there. However, it struggled with sudden, jerky movements. If you did a "heel drop" (jumping off your toes and landing hard), the system got confused. It's like the model is great at understanding a steady drumbeat but gets lost when the music suddenly stops and starts with a crash.

Why It Matters: Trusting the Machine

One of the coolest things the researchers did was ask the AI, "How do you know this?" Usually, AI is a "black box"—it gives an answer, but you don't know why. The authors used a technique called "Temporal Attribution" to peek inside the black box. They found that the model was looking at the wrist acceleration during the early part of the step (when your heel hits the ground) to make its guess.

Even better, this wasn't a fluke. No matter which person they tested or how they split the data, the AI kept pointing to the same spot: the wrist moving when the foot hits. This consistency makes the system feel more trustworthy, like a detective who always finds the same clue, rather than guessing randomly.

The Bottom Line

This paper doesn't claim that your Apple Watch can perfectly replace a million-dollar lab machine. It explicitly says that for sudden, explosive jumps, the current setup isn't ready yet. But for the everyday stuff—walking, jogging, and running—it shows that we can get a solid, privacy-safe estimate of how our bodies are loading up, right from our wrists.

The study proves that the "edge" (your phone and watch) is powerful enough to handle complex biomechanics without needing the cloud. It's a step toward a future where we can track our physical health trends over months and years, not just in a single lab visit, using devices we already wear. The door is open, but the researchers are careful to say: "We can walk through it for steady steps, but we need to learn how to run before we try to jump."

Technical Summary: Edge-First Ground Reaction Force Estimation with Consumer Smartwatches

Problem Statement
Ground reaction force (GRF) is a critical signal for movement analysis, gait assessment, and rehabilitation monitoring. However, accurate GRF measurement is currently confined to instrumented laboratory force plates, which are fixed, expensive, and unsuitable for longitudinal monitoring in daily life. While prior research has demonstrated that inertial measurement unit (IMU) data can estimate GRF, significant gaps remain in translating these proofs-of-concept into deployable wearable systems. Specifically, existing literature often lacks rigorous evaluation of subject-independent generalization, detailed edge-system architectures (communication, buffering, local inference), and reproducible time-resolved explainability.

Methodology
The authors propose an end-to-end, edge-first wearable system using consumer-grade hardware and local processing:

Hardware Architecture: The system utilizes two Apple Watch Series 6 devices (one at the dominant wrist, one at the waist) as sensor nodes and an iPhone as the edge node. The watches stream 12-channel inertial data (3-axis acceleration and 3-axis angular velocity per device) at 100 Hz.
Data Pipeline: Data is transmitted via WatchConnectivity in 5-second chunks (500 samples) to the iPhone. The iPhone handles local buffering, packet reordering, and preprocessing (alignment, low-pass filtering at 10 Hz for IMU and 20 Hz for force plate, z-score normalization, and segmentation into 198-sample fixed windows). Crucially, the entire inference path executes locally on the phone without cloud dependence, preserving privacy and reducing latency sensitivity.
Model Architecture (GRFNet-MultiScale): The core predictive model is a compact 1D temporal convolutional network designed for edge deployment. It features:
- Four dilated residual blocks with dilation rates $d \in \{1, 2, 4, 8\}$ and kernel size 3, creating an effective receptive field of 127 samples (1.27 s).
- A global context branch using adaptive average pooling to capture window-level structure.
- A lightweight decoder for per-timestep vertical GRF (vGRF) prediction.
- The dual-sensor model contains ~1.2M parameters; a wrist-only variant contains ~605k parameters.
Explainability: To ensure trustworthiness, the study employs two complementary temporal explainability methods:
- Temporal SMILE: A masking-based method that perturbs short contiguous windows to measure output changes.
- TimeSHAP-style Analysis: A segment-level attribution method based on Shapley values.
- Both methods are evaluated for stability across validation folds.

Key Contributions

Edge-First Pipeline: An end-to-end system demonstrating local preprocessing, storage, and inference on consumer devices, eliminating the need for cloud connectivity during operation.
Compact Multi-Scale Model: Introduction of GRFNet-MultiScale, which balances predictive accuracy with the computational constraints of mobile edge devices.
Strict Generalization Evaluation: Comprehensive evaluation under a 10-fold Leave-One-Subject-Out (LOSO) protocol, testing generalization to unseen users rather than subject-specific tuning.
Reproducible Temporal Attribution: Demonstration that temporal attribution patterns (identifying when and where signals matter) remain stable across different validation splits, moving beyond single-split visualizations.

Results

Predictive Performance: Under LOSO validation on 539 stance windows from 10 healthy participants, the dual-sensor system achieved a mean Pearson correlation ( $r$ ) of 0.798 with a Root Mean Square Error (RMSE) of 257 N.
Sensor Configuration: A wrist-only configuration retained 82.5% of the dual-sensor correlation ( $r=0.658$ ), validating the utility of standard smartwatch-only setups.
Activity Dependence: Performance varied significantly by activity. The system performed best on cyclic locomotion: jogging ( $r=0.867$ ) and running ( $r=0.829$ ), followed by walking. Performance was moderate for step-downs and low for heel drops ( $r=0.301$ ), attributed to the difficulty of reconstructing brief, impulsive impacts from low-pass filtered 100 Hz signals.
Derived Metrics: The system showed high accuracy for integrated metrics, with impulse ( $r=0.983$ ) and effective contact duration ( $r=0.982$ ) correlating strongly with ground truth. Peak vGRF ( $r=0.741$ ) and loading rate ( $r=0.588$ ) were less accurate.
Explainability: Both Temporal SMILE and TimeSHAP consistently identified early-stance wrist acceleration as the dominant reproducible signal. This attribution pattern remained stable across all 45 pairwise LOSO fold comparisons (median Spearman $\rho=0.755$ ).

Significance and Claims
The paper claims that consumer wearables can support meaningful, edge-first estimation of vertical GRF under controlled sensing conditions, provided that claims are aligned with measured capabilities. The system is positioned not as a replacement for laboratory force plates in all settings, but as a tool for repeated longitudinal load monitoring during cyclic locomotion (walking, jogging, running).

The authors emphasize that the contribution lies in the full-stack integration of sensing, communication, edge execution, and reproducible explainability. They explicitly define the current boundaries: the system is strongest for cyclic tasks, does not yet support impulsive heel-drop events, relies on force-plate timing for evaluation (requiring future IMU-based stance detection for full deployment), and has not yet been characterized for on-device energy consumption or battery life. The work establishes a design pattern for wearable AI that prioritizes privacy, local inference, and mechanistic interpretability.

More like this