Source-linked AI summary
Sensor-based Gait Parameter Extraction with Deep Convolutional Neural Networks
Julius Hannink, Thomas Kautz, Cristian F. Pasluosta, Karl-Günter Gaßmann, Jochen Klucken, Bjoern M. Eskofier
TL;DR
Existing double-integration approaches for extracting gait parameters can be limited by assumptions that are easily violated in impaired gait. This paper uses deep convolutional networks to translate wearable-sensor data into eight integration-free spatio-temporal stride characteristics, with the ensemble approach outperforming combined modelling and supporting mobile assessment in clinically critical cases.
Problem
Double-integration approaches for sensor-based gait analysis rely on assumptions that can be violated in everyday clinical practice and impaired gait, limiting applicability.
Method
Deep convolutional neural networks learn mappings from wearable sensor data to eight reference-annotated spatio-temporal stride characteristics, comparing combined and per-parameter ensemble models.
Results
The ensemble approach estimates stride length, stride width, and foot angle with mean accuracy and precision of −0.15±6.09 cm, −0.09±4.22 cm, and 0.13 ± 3.78° respectively, while heel and toe contact-time estimation constitutes state-of-the-art.
Takeaways & Limitations
The proposed method is presented as a suitable substitute for assumption-governed double-integration techniques and could enable mobile gait analysis for severe gait disturbances such as spasticity.
Takeaways & Limitations
Performance depends on the knowledge base because the learned sensor-to-parameter mapping implicitly depends on its collected samples.
Abstract
from arXiv · showhide
Measurement of stride-related, biomechanical parameters is the common rationale for objective gait impairment scoring. State-of-the-art double integration approaches to extract these parameters from inertial sensor data are, however, limited in their clinical applicability due to the underlying assumptions. To overcome this, we present a method to translate the abstract information provided by wearable sensors to context-related expert features based on deep convolutional neural networks. Regarding mobile gait analysis, this enables integration-free and data-driven extraction of a set of 8 spatio-temporal stride parameters. To this end, two modelling approaches are compared: A combined network estimating all parameters of interest and an ensemble approach that spawns less complex networks for each parameter individually. The ensemble approach is outperforming the combined modelling in the current application. On a clinically relevant and publicly available benchmark dataset, we estimate stride length, width and medio-lateral change in foot angle up to ${-0.15\pm6.09}$ cm, ${-0.09\pm4.22}$ cm and ${0.13 \pm 3.78^\circ}$ respectively. Stride, swing and stance time as well as heel and toe contact times are estimated up to ${\pm 0.07}$, ${\pm0.05}$, ${\pm 0.07}$, ${\pm0.07}$ and ${\pm0.12}$ s respectively. This is comparable to and in parts outperforming or defining state-of-the-art. Our results further indicate that the proposed change in methodology could substitute assumption-driven double-integration methods and enable mobile assessment of spatio-temporal stride parameters in clinically critical situations as e.g. in the case of spastic gait impairments.
I. Introduction
The paper addresses the challenge of translating wearable-sensor measurements into clinically meaningful gait features. It proposes deep convolutional neural networks for integration-free, data-driven extraction of stride characteristics.
- Reduced gait quality can severely reduce mobility and quality of life, motivating early quantification and treatment of gait impairments.
- Mobile sensor-based systems are attractive for unobtrusive gait analysis because laboratory-based pressure mats and motion-capture systems have limited availability.
- Wearable sensors measure abstract signals such as foot accelerations and angular rates, whereas clinicians need context-related variables such as stride length and heel-strike angle.
- Double-integration methods depend on a zero-velocity phase within each stride, an assumption that can be violated in clinical practice.
- The proposed framework trains convolutional neural networks by regressing wearable-sensor data to reference expert features.
- The work compares a combined model estimating all outputs with an ensemble of individual models and validates the method on a clinically relevant public dataset.
A. Data Collection and Setup
The study uses inertial sensors attached below the ankles during geriatric walking assessments, with GAITRite providing reference measurements. The dataset is extended with spatial and temporal stride annotations for model evaluation.
- The Shimmer2R platform recorded three-dimensional acceleration and gyroscope data from sensors attached laterally below each ankle.Data were captured at 102.4 Hz with 12-bit resolution.
- GAITRite pressure-mat measurements supplied simultaneous reference data for validating inertial-sensor estimates.The reference system had a spatial resolution of ±1.27 cm.
- The dataset annotation was extended to include stride width, medio-lateral foot-angle change, and heel and toe contact times alongside stride, stance, and swing time.
- After exclusions for medical reasons, sensor malfunction, and measurement errors, 99 patients remained for training and evaluation.
- Gait disorders or fall proneness affected 54% of the study population, supporting the dataset’s clinical relevance.
B. Preprocessing
Sensor strides are extracted, calibrated, aligned, normalized, and padded before network input. Heel-strike-to-heel-strike segmentation avoids requiring a zero-velocity phase, while heel-strike and toe-off events provide some temporal features directly.
- Preprocessing extracts annotated strides, calibrates readings, aligns foot coordinates, normalizes sensor ranges, and pads each stride to 256 samples.
- Strides are defined from heel strike to heel strike, avoiding the zero-velocity re-initialization assumption required by double-integration methods.
- Heel-strike and toe-off events are detected within the sensor data and used to adjust stride borders.
- Stride time, stance time, and swing time are computed directly from heel-strike and toe-off events, leaving five parameters for neural-network estimation.
C. Network Architectures
The study compares a combined convolutional architecture with an ensemble of parameter-specific networks. The ensemble assigns each output its own feature-extraction path, whereas the combined model shares features until the final output stage.
- Both architectures use convolutional, max-pooling, and densely connected layers as their elementary building blocks.
- Model A estimates all output variables jointly with three convolutional, max-pooling, and three densely connected layers.
- Model B uses a separate smaller network for each output variable, with two convolutional layers, max-pooling, and one densely connected layer.
- Max-pooling downsamples feature maps by taking maxima over temporal windows, increasing robustness of extracted features.
- Densely connected layers compute activations from flattened preceding outputs using weight vectors, biases, matrix multiplication, and ReLU activation.
- Model A learns general features for all outputs, while Model B optimizes an individual feature-extraction path for each parameter.
D. Training
The networks are trained by minimizing prediction error on normalized outputs with mini-batch optimization. Adam, dropout, and fixed training settings are used for optimization and over-fitting control.
- Weights and biases are updated by back-propagation to minimize the error on training data or mini-batches.
- Reference outputs are scaled to [0, 1], and test predictions are later rescaled to their physical dimensions.
- Training error is defined as the sum of individual root-mean-square errors across output variables on each mini-batch.
- Adam optimization trains the networks for 4000 iterations with mini-batches of 100 strides.
- Dropout randomly removes hidden-layer nodes during training to reduce over-fitting, with probabilities specified separately for Models A and B.
E. Evaluation Scheme
Evaluation uses patient-wise 10-fold cross-validation on stride-specific sensor data from 99 patients. Training-error behavior is additionally examined on an exemplary 90/10% patient-wise split.
- The study uses 10-fold cross-validation with patient identifiers determining distinct training and test partitions.
- Stride-specific sensor data from 99 patients are evaluated across folds, with fold estimates pooled for average statistics.
- Training error and performance are also examined over iterations using an exemplary patient-wise 90/10% train/test split.
III. Results
Training error stabilized within 4000 iterations for model A and every submodel in model B. The models showed comparable adaptation to the training data under the selected error function.
- 4000 iterations were sufficient for model A and all model B submodels to reach a stable training-error regime.Training was therefore stopped at 4000 iterations.
- E1...5 ≈ 0.02 for model B corresponded to a summed error of E ≈ 0.1 for model A.The comparison indicates similar adaptation under the selected error function.
- Model A and the model B submodels had comparable training-data adaptation in the exemplary 90/10% split.The error was evaluated over training iterations on the entire training set.
B. Stride Parameter Estimation on Unseen Data
On unseen test data, ensemble model B achieved better precision for stride length and width, while the models performed similarly on foot angle and heel and toe contact times. Detailed per-parameter results and state-of-the-art comparisons are provided for model B.
- Ensemble approach B achieved significantly better precision for stride length and width on unseen test data.Its corresponding mean accuracies also exceeded those of model A.
- Model A and model B performed similarly for foot angle, heel contact time, and toe contact time.The reported similarity applies to the remaining three parameters discussed in the comparison.
- Model B was considered superior in this context, with detailed per-parameter error statistics and state-of-the-art comparisons reported in Fig. 7 and Table III.Table II compares average accuracy and precision, with Levene tests used for precision comparisons at the 0.01 significance level.
IV. Discussion
Model B estimates eight stride characteristics on unseen test data, with reported performance compared against reference and state-of-the-art methods. The framework avoids the zero-velocity assumption, but its translation quality depends on the knowledge base and several extensions remain future work.
- Results: −0.15±6.09 cm, −0.09±4.22 cm, and 0.13±3.78° are reported for stride length, stride width, and foot angle, respectively.These are mean accuracy and precision estimates from the superior ensemble approach B.
- Results: ±0.07, ±0.05, ±0.07, ±0.07, and ±0.12 s are reported for stride, swing, stance, heel-contact, and toe-contact times, respectively.The estimates are obtained on unseen test data and compared with GAITRite and prior approaches.
- Comparison: The validation uses 1,185 individual strides from 99 geriatric patients, and stride-length precision outperforms Rampp et al. by 2.3 cm, or 27%.The comparison is statistically significant according to the discussion.
- Comparison: Stride-width estimation reaches an ICC of 0.95, exceeding the ICC of 0.88 reported for a sensor-based estimation by Rebula et al.The discussion characterizes this as outperforming state-of-the-art double-integration methods for stride width.
- Comparison: The reported heel- and toe-contact estimates constitute the state of the art because prior heel-off and toe-on detection was heuristic and unevaluated for precision.The method also enables detection of these events despite their weak manifestation in sensor signals.
- Limitations and scope: The pipeline uses heel-strike-to-heel-strike segments and is independent of the zero-velocity assumption required by double-integration methods.The authors state that this could support mobile gait analysis in clinically critical cases, while the method remains only as good as its knowledge base.