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

AI summary

Overview

Research area: Wearable sensing and machine learning for clinical gait biomechanics, specifically estimating ground reaction forces (GRFs) in Parkinson's disease (PD) from inertial measurement units (IMUs).

Technical level: Advanced. The paper assumes familiarity with deep learning architectures (CNN, BiLSTM, Transformer, TCN), biomechanical outcome measures (vertical ground reaction force, body-weight normalization), and statistical testing conventions.

Scope: The study develops and evaluates a CNN-BiLSTM model that estimates bilateral vertical ground reaction forces (vGRFs) from wearable IMUs in Parkinsonian and healthy gait, then systematically searches for the smallest effective IMU set.

Note on completeness: the provided paper content ends mid-sentence in the Discussion section ("In line with the hypothesis, the results demonstr"). Detailed discussion arguments, limitations, and the Conclusion are therefore not available in the supplied text, and are not summarized below.

What This Paper Is About

Measuring ground reaction forces normally requires laboratory force plates or instrumented insoles, which are impractical for everyday or remote monitoring. Parkinsonian gait is highly variable and asymmetric, so it is unclear whether machine learning models built on healthy people transfer to PD. This paper builds a deep learning framework that estimates left and right vertical ground reaction forces from wearable IMU data in people with PD and healthy controls, and then determines the smallest and best-placed set of IMUs needed to do it accurately.

Key Contributions

1. First deep learning framework for bilateral vGRF estimation in Parkinsonian gait. The authors state that to date no study has investigated deep learning approaches for this specific challenge, and that this is the first such framework using an optimized set of wearable IMUs.

2. Systematic sensor-configuration ablation. Rather than testing a single fixed setup, the authors evaluated all 1,023 possible combinations of the 10 retained IMU placements (from 1 to 10 sensors), out of 8,191 possible combinations across all 13 IMUs, to identify optimal and minimal sensor sets.

3. Disease-specific sensor placement findings. The optimal sensor locations and their ranking differ between PD patients and healthy controls, and side-to-side (left/right) differences are more pronounced in PD, motivating PD-specific—rather than generic—sensor placement strategies.

4. Cross-dataset external validation on an independent PD cohort. Using an external public PD dataset with virtual IMU signals derived from motion capture, the framework was tested with transfer learning and domain adaptation, providing evidence of cross-dataset generalizability.

Main Findings

• Strong intra-subject accuracy, lower inter-subject accuracy: With 13 IMUs, intra-subject estimation reached an R² of 0.98 ± 0.08 for healthy controls (HC, MAE 0.91 ± 0.87 %BW) and 0.98 ± 0.06 for PD (MAE 1.14 ± 0.95 %BW), with no statistically significant difference between groups (p > 0.05). Inter-subject performance dropped significantly (p < 0.05) to R² = 0.93 ± 0.25 for HC (MAE 2.71 ± 2.47 %BW) and R² = 0.91 ± 0.10 for PD (MAE 3.99 ± 2.63 %BW), with higher error in the PD group.

• CNN-BiLSTM outperforms baselines: In the inter-subject PD setting with 13 IMUs, CNN-BiLSTM surpassed TCN, MLP, Transformer, CNN-only, and BiLSTM-only models across all three metrics. The Transformer was the most competitive baseline (MAE 3.80 ± 2.41 %BW; R² 0.90 ± 0.11), while the ablated components of the proposed model reached R² = 0.89 ± 0.09 (CNN-only) and 0.88 ± 0.13 (BiLSTM-only). MLP and TCN performed worst.

• External validation held up: On a matched 12-IMU (non-head) comparison, the primary dataset gave MAE 4.08 ± 2.49 %BW and R² = 0.88 ± 0.11, versus the external dataset's MAE 7.45 ± 3.85 %BW and R² = 0.86 ± 0.17. The R² difference was not statistically significant (p = 0.51).

• Sensor placement matters, differently by group: In HC, right and left dorsal foot (RD, LD) gave the top two median accuracies with no significant difference between them. In PD, the five highest-ranked placements (LD, RS, LS, LB, FH) showed no significant differences among themselves. Localized deviations occurred at FH and LB, where PD mean R² was comparable to or slightly above HC.

• Poor-performing single sensors in PD: The xiphoid process (R² = 0.76 ± 0.58), left wrist (R² = 0.43 ± 1.01), and right wrist (R² = 0.78 ± 0.42) fell below R² = 0.80 and were excluded, leaving 10 IMUs for ablation.

• Four IMUs is optimal; two is minimal: The best overall configuration used four IMUs (dorsal foot LD, right ankle RA, right shank RS, forehead FH), with accuracy plateauing beyond that count. A two-IMU setup at FH and LD still enabled robust estimation and is recommended for minimal sensor use. Median R² exceeded 0.9 across all tested configurations. No significant differences were found between adjacent configurations using 2 to 9 IMUs, but accuracy dropped significantly with only one IMU or with ten IMUs (p ≤ 0.05). The largest upward performance shift occurred when going from 1 to 3 IMUs.

• Laterality in the PD cohort: Lower-limb laterality measured with MDS-UPDRS Part III lower extremity items showed left dominant impairment in 43.1% of the cohort versus right dominant impairment in 32.8%, which the authors suggest may partly explain the cohort-level superiority of the left dorsal foot placement.

Methodology in Plain English

The researchers used a large public multi-site gait dataset with 126 participants: 61 with PD (25 females, 36 males, age 76 ± 9 years) and 65 healthy controls (40 females, 25 males, age 76 ± 9 years). PD severity was predominantly mild (modified Hoehn & Yahr 2.15 ± 0.48; MDS-UPDRS Part III scores ranging from 6 to 46, mean 24 ± 10). Participants wore 13 IMUs at locations including the forehead, xiphoid, lower back, both wrists, both thighs, both shanks, both ankles, and the dorsum of both feet, and performed self-paced walking, hurried-paced walking, and the Timed Up and Go test. Vertical ground reaction forces were derived from a pressure walkway and validated against static standing trials, showing a coefficient of variation below 0.03, then normalized to body weight.

Signals were recorded at 100 Hz after low-pass filtering at 184 Hz. Data were split by subject and trial before being cut into 64-time-point windows with 25% overlap, avoiding temporal leakage between training, validation, and test sets. The model input was triaxial acceleration plus triaxial angular velocity per sensor (78 channels for 13 IMUs; 72 for 12), with magnetometer data deliberately excluded due to sensitivity to magnetic disturbance.

The model, a CNN-BiLSTM, has three parts: a convolutional module that compresses IMU signals into lower-dimensional features, a bidirectional LSTM that captures temporal dependencies in both directions, and fully connected layers that output left and right vGRF sequences. It was compared against MLP, TCN, Transformer, CNN-only, and BiLSTM-only baselines. Hyperparameters (for example, first CNN layer output channels 256, second 128, LSTM hidden size 256 with 3 layers, FC sizes 60 and 30, dropout 0.1) were tuned with the Optuna framework using 5-fold cross validation, and training used the AdamW optimizer with learning rate and weight decay of 0.0001, batch size 64, up to 200 epochs, and early stopping.

For the external dataset of 26 PD participants, only motion capture and force plate data were available, so 12 virtual IMUs were simulated using OpenSim 4.5 inverse kinematics and processed to match the primary dataset. Transfer learning initialized the model with weights from the primary dataset, and external adaptation plus evaluation used the 12 shared non-head sensor locations.

Why This Matters

Impact on research: The paper provides the first systematic evaluation of deep learning-based vGRF estimation specifically in Parkinsonian gait, showing that models validated only on healthy participants may not transfer cleanly. It establishes that sensor optimization is disease-specific—optimal placements for PD differ from those for healthy controls—and that a compact two-sensor configuration can remain viable, which previous fixed-configuration studies had not demonstrated for this population.

Real-world applications:

• Accessible clinical gait assessment where force plates are unavailable, using a small number of wearable sensors.
• Remote and home-based monitoring of gait mechanics and fall risk in people with Parkinson's disease.
• Personalized rehabilitation planning informed by longitudinal vGRF variability, which the authors cite as a potential biomarker for disease progression and treatment response.
• Potential extension of the framework to other pathological gait conditions beyond Parkinson's disease.

Industry relevance: The results point toward practical wearable products with fewer sensors, lower cost, less setup burden, and better patient compliance, which is directly relevant to wearable device manufacturers, digital health and remote-monitoring platforms, and clinical trial designers seeking objective, low-burden gait outcomes.

Future Directions

1. Validate with real IMUs rather than virtual signals. The external validation relied on virtual IMU signals simulated from motion capture because wearable recordings were unavailable, and the authors used transfer learning and domain adaptation to bridge this gap. Confirming results with real wearable sensors in independent PD cohorts is the logical next step.

2. Prospectively test the two-IMU (FH and LD) and four-IMU (LD, RA, RS, FH) configurations in deployment settings. Whether these minimal setups retain accuracy during free-living, unconstrained daily activity—rather than the walking tasks studied here—remains untested.

3. Investigate PD-specific laterality and symptom asymmetry in sensor design. The finding that left/right placement differences are more pronounced in PD, alongside the 43.1% left-dominant versus 32.8% right-dominant impairment distribution, raises the question of whether sensor placement should be individualized per patient based on which side is more affected.

4. Apply the framework to other pathological gait conditions. The authors state the approach could support wearable vGRF-based gait analysis for Parkinsonian gait and potentially other pathological conditions, which would require repeating the disease-specific sensor optimization carried out here.

Target Audience

This paper is most useful for biomechanics and rehabilitation engineering researchers, machine learning researchers working on wearable sensing and time-series regression, and clinical researchers studying Parkinson's disease and gait. It also suits wearable device developers and digital health product teams deciding how many and which sensors a practical gait monitoring system needs, and readers with a background in signal processing or deep learning who want a worked example of disease-specific model and sensor validation.

Authors’ abstract

Accurate gait analysis in Parkinson's disease (PD) typically relies on laboratory-based systems to capture biomechanical data, such as ground reaction forces (GRFs). Estimating GRFs using inertial measurement units (IMUs) provides a feasible alternative. However, this approach remains challenging in pathological gait like PD due to its high variability and complexity. Moreover, existing monitoring approaches often require multiple body-mounted sensors, which limit practicality and reduce patient compliance. To date, no study has investigated the application of deep learning approaches to address this challenge. This study proposes, for the first time, a deep learning framework to estimate bilateral vertical GRFs (vGRFs) in PD using an optimized set of wearable IMUs. A hybrid CNN-BiLSTM model was trained separately on data from 61 PD patients and 65 healthy controls (HC) using 13 IMUs. The model achieved high intra-subject accuracy ($R^2$ = 0.98) and strong inter-subject generalization ($R^2$ = 0.93 for HC, $R^2$ = 0.91 for PD). Sensor configuration was found to significantly influence estimation accuracy, with optimal sensor placement varying between PD patients and HC. For PD patients, estimation accuracy dropped markedly when reducing to a single IMU. The optimal configuration for PD used four IMUs. We identified a minimal setup with only two IMUs still enabled robust estimation. This compact setup offers a practical and scalable solution. Overall, the proposed approach supports the development of wearable vGRF-based gait analysis systems for Parkinsonian gait and potentially other pathological conditions, enabling accessible clinical assessments, remote monitoring, and personalized rehabilitation.