⚡ electrical engineering

Deep Learning-Based Estimation of Ground Reaction Forces in Parkinsonian Gait Using an Optimized Set of IMU Data

This study introduces a novel deep learning framework using a hybrid CNN-BiLSTM model and an optimized four-sensor IMU configuration to accurately estimate bilateral vertical ground reaction forces in Parkinson's disease patients, offering a practical and scalable solution for wearable gait analysis and remote monitoring.

Original authors: Run Lin, Yingtian Tang, Jiawen Xu, Dongfei Huo, Lefan Wang, Helen Dawes, Dominic J. Farris, Dong Wang, Xijin Hua

Published 2026-08-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Run Lin, Yingtian Tang, Jiawen Xu, Dongfei Huo, Lefan Wang, Helen Dawes, Dominic J. Farris, Dong Wang, Xijin Hua

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to understand how a car engine works just by listening to the sound it makes while driving down a bumpy road. That's essentially what scientists do when they study how people walk. For decades, the "gold standard" for measuring the forces our feet hit the ground with—called Ground Reaction Forces (GRFs)—has been a giant, expensive pressure mat found only in high-tech laboratories. It's like having a super-accurate speed trap, but you can only use it in one specific spot. This makes it impossible to see how someone walks in their own home, at the park, or while rushing to catch a bus.

Enter the humble Inertial Measurement Unit (IMU). Think of these as the tiny, super-smart accelerometers inside your smartphone or smartwatch. They can measure movement and tilt anywhere, anytime. The big question scientists have been asking is: Can we use these little sensors, combined with a powerful type of computer brain called "Deep Learning," to guess exactly how hard a person's feet are hitting the ground, even when that person is walking in a strange or unsteady way? This is especially tricky for people with Parkinson's disease, whose walking patterns can be as unpredictable as a jazz solo, full of sudden stops, shuffles, and wobbles. If we can crack this code, we could give doctors a way to monitor patients' health from their living rooms, spotting problems before they lead to dangerous falls.

The Paper's Big Idea: Teaching Computers to "Feel" the Walk

This study is like a detective story where the investigators are trying to figure out the best way to use a set of tiny motion sensors to guess the invisible forces of a Parkinsonian walk. The researchers built a special computer model—a hybrid brain made of two parts working together. One part, the CNN, is like a pattern-spotter that looks at the raw data from the sensors to find local clues. The other part, the BiLSTM, is like a time-traveling storyteller that remembers the sequence of movements, looking at the walk both forward and backward to understand the full story. They trained this model on data from 61 people with Parkinson's disease and 65 healthy older adults, using a full set of 13 sensors strapped all over their bodies (from their foreheads down to their feet).

What They Found: The "Sweet Spot" of Sensors

The results were surprisingly good. When the computer model was tested on people it had already "met" (the same people it was trained on), it guessed the ground forces with incredible accuracy, getting it right 98% of the time. But the real test was seeing if it could guess the walk of a new person it had never seen before. Even then, it performed strongly, getting it right about 91% of the time for Parkinson's patients and 93% for healthy people.

However, the most exciting discovery wasn't just that the model worked, but how it worked best. The researchers realized that you don't need to strap sensors to every single body part to get a great answer. In fact, they found that the "best" places to put the sensors were different for people with Parkinson's compared to healthy people. For the healthy group, the sensors on the feet worked best. But for the Parkinson's group, the sensors on the feet were great, but adding a sensor on the forehead and the lower back made a huge difference.

The Magic Number: Two vs. Four

Here is the playful part: The researchers played a game of "how few sensors can we use?" They tested every possible combination of sensors, from just one up to ten.

The One-Sensor Trap: They found that using just one sensor was a bad idea for Parkinson's patients. The accuracy dropped significantly, like trying to guess a whole movie plot from a single frame.
The Four-Sensor Champion: The absolute best setup for the highest accuracy used four sensors: one on the left foot, one on the right shin, one on the right thigh, and one on the forehead. This setup gave the model a "score" of 0.93 (where 1.0 is perfect).
The Two-Sensor Hero: But here is the cool twist. They found that a setup with just two sensors—one on the left foot and one on the forehead—was almost as good as the four-sensor setup. It still got an accuracy score of 0.91.

What This Means

The paper suggests that we don't need a full-body suit of sensors to monitor Parkinson's gait effectively. A simple, lightweight setup with just two sensors could be enough to give doctors a reliable picture of how a patient is walking. This is a big deal because it means the technology could be much more comfortable and practical for patients to wear every day. The study also explicitly argues against the idea that the same sensor setup works for everyone; what works for a healthy person might not work for someone with Parkinson's, and the computer model needs to be tuned specifically for the condition.

While the study used a mix of real sensor data and some computer-generated "virtual" sensors for extra testing, the results strongly suggest that this deep learning approach is a robust way to turn simple movement data into a powerful tool for understanding and helping people with Parkinson's disease. It's a step toward making high-tech medical monitoring as easy and unobtrusive as wearing a pair of smart shoes.

Technical Summary: Deep Learning-Based Estimation of Ground Reaction Forces in Parkinsonian Gait

Problem Statement
Accurate gait analysis in Parkinson's disease (PD) is critical for diagnosis, treatment planning, and monitoring, yet it traditionally relies on laboratory-based force plates which are impractical for continuous, real-world use. While inertial measurement units (IMUs) offer a portable alternative, estimating Ground Reaction Forces (GRFs) in pathological gait remains challenging due to the high variability, asymmetry, and complexity of PD movements. Existing approaches often rely on physics-based models (e.g., inverse dynamics, musculoskeletal modeling) that struggle with the non-periodic and asymmetric nature of PD gait, or data-driven methods that have been predominantly validated only on healthy populations. Furthermore, the optimal number and placement of IMUs for PD patients remain unclear, with many current setups requiring multiple sensors that reduce patient compliance.

Methodology
The study proposes a deep learning framework utilizing a hybrid Convolutional Neural Network–Bidirectional Long Short-Term Memory (CNN-BiLSTM) architecture to estimate bilateral vertical GRFs (vGRFs) from wearable IMU data.

Datasets: The primary dataset comprised 126 participants (61 PD patients, 65 Healthy Controls) wearing 13 IMUs across various walking tasks (self-paced, hurried-paced, and Timed Up and Go). An independent external dataset of 26 PD patients was used for cross-dataset validation, where virtual IMU signals were generated from motion capture data.
Model Architecture: The proposed CNN-BiLSTM model processes triaxial acceleration and angular velocity signals. The CNN module extracts spatial features from the IMU data, while the BiLSTM module captures temporal dependencies bidirectionally. The output is a sequence of estimated left and right vGRFs.
Experimental Design:
- Evaluation Protocols: Performance was assessed under intra-subject (subject-specific training/testing) and inter-subject (training on some subjects, testing on unseen subjects) settings.
- Baseline Comparisons: The proposed model was compared against Multi-Layer Perceptron (MLP), Temporal Convolutional Network (TCN), Transformer, and ablated variants (CNN-only, BiLSTM-only).
- Ablation and Configuration Study: To determine optimal sensor setups, the authors systematically evaluated 1,023 unique combinations of 1 to 10 IMUs (selected from 13 initial positions after filtering low-performing single sensors). This included analyzing the impact of sensor count and specific anatomical placement on estimation accuracy.

Key Contributions

First Deep Learning Application for PD vGRF: This study is the first to investigate and apply a deep learning framework specifically for estimating bilateral vGRFs in Parkinsonian gait, addressing the gap in existing literature which focuses largely on healthy populations or physics-based models.
Optimized Sensor Configuration: The research identifies that optimal sensor placement differs significantly between PD patients and healthy controls. It provides a systematic evaluation of sensor counts, demonstrating that while four IMUs yield the highest accuracy, a minimal two-IMU setup offers a practical trade-off.
Robust Inter-Subject Generalization: The framework demonstrates strong generalization capabilities across unseen subjects and external datasets, validating its potential for real-world deployment without requiring extensive subject-specific calibration.

Results

Accuracy: The CNN-BiLSTM model achieved high intra-subject accuracy ( $R^2 = 0.98$ ) and strong inter-subject generalization ( $R^2 = 0.93$ for Healthy Controls, $R^2 = 0.91$ for PD patients).
Model Comparison: The hybrid CNN-BiLSTM architecture outperformed all baseline models (MLP, TCN, Transformer) and ablated variants, confirming the necessity of combining spatial feature extraction with bidirectional temporal modeling for this task.
Sensor Placement: Single-IMU analysis revealed that the Left Dorsum of the foot (LD) was the most effective single sensor for PD patients, whereas Healthy Controls performed best with sensors on the foot dorsum (both sides). Notably, PD patients showed greater sensitivity to sensor placement and side-specific asymmetry compared to controls.
Sensor Count: Estimation accuracy improved as the number of IMUs increased from 1 to 4, after which gains plateaued.
- Optimal Configuration: Four IMUs placed at the Left Dorsum (LD), Right Ankle (RA), Right Shank (RS), and Forehead (FH) achieved the highest performance.
- Minimal Configuration: A two-IMU setup (LD and FH) achieved comparable accuracy ( $R^2 = 0.91 \pm 0.09$ ) to the four-sensor setup, offering a streamlined solution.
External Validation: Cross-dataset validation on the external PD dataset (using 12 virtual IMUs) yielded results consistent with the primary dataset ( $R^2 = 0.86$ ), supporting the model's generalizability.

Significance and Claims
The paper claims that the proposed approach provides a practical and scalable solution for wearable vGRF-based gait analysis in Parkinson's disease. By identifying a minimal two-sensor configuration that maintains robust estimation accuracy, the study addresses the critical balance between measurement precision and patient compliance. The authors assert that this framework supports the development of accessible clinical assessments, remote monitoring, and personalized rehabilitation systems for neurological populations. The study emphasizes that disease-specific validation is crucial, as sensor strategies effective for healthy individuals may not translate directly to pathological gait. Ultimately, the work lays the groundwork for intelligent, wearable gait analysis systems tailored to the unique biomechanical characteristics of Parkinsonian movement.

Technical Summary: Deep Learning-Based Estimation of Ground Reaction Forces in Parkinsonian Gait

More like this