Source-linked AI summary
Explaining the Unique Nature of Individual Gait Patterns with Deep Learning
Fabian Horst, Sebastian Lapuschkin, Wojciech Samek, Klaus-Robert Müller, Wolfgang I. Schöllhorn
TL;DR
Machine-learning models can classify gait patterns but often lack transparent decision processes, limiting their interpretability in medical applications. This paper applies DNNs and LRP to individual gait data, finding distinctive, interpretable gait signatures and proposing a framework for understanding nonlinear models in gait analysis.
Problem
Machine-learning gait analyses offer predictive power but often act as opaque black boxes, leaving their decisions difficult to understand in clinical diagnosis.
Method
The study uses DNNs to classify individuals from time-continuous joint angles and ground reaction forces, then applies LRP to attribute predictions to input variables.
Results
Gait patterns contain individual-specific characteristics, and LRP identifies the variables and gait-cycle time windows contributing to each individual’s classification.
Takeaways & Limitations
LRP provides a framework for interpreting nonlinear gait-classification models and may help individualize gait analyses, diagnoses, and interventions.
Abstract
from arXiv · showhide
Machine learning (ML) techniques such as (deep) artificial neural networks (DNN) are solving very successfully a plethora of tasks and provide new predictive models for complex physical, chemical, biological and social systems. However, in most cases this comes with the disadvantage of acting as a black box, rarely providing information about what made them arrive at a particular prediction. This black box aspect of ML techniques can be problematic especially in medical diagnoses, so far hampering a clinical acceptance. The present paper studies the uniqueness of individual gait patterns in clinical biomechanics using DNNs. By attributing portions of the model predictions back to the input variables (ground reaction forces and full-body joint angles), the Layer-Wise Relevance Propagation (LRP) technique reliably demonstrates which variables at what time windows of the gait cycle are most relevant for the characterisation of gait patterns from a certain individual. By measuring the time-resolved contribution of each input variable to the prediction of ML techniques such as DNNs, our method describes the first general framework that enables to understand and interpret non-linear ML methods in (biomechanical) gait analysis and thereby supplies a powerful tool for analysis, diagnosis and treatment of human gait.
Introduction
Biomechanical gait analysis commonly relies on isolated, time-discrete variables, while machine-learning models can use richer time-continuous patterns but are difficult to interpret. This study evaluates DNN-based gait classification and LRP explanations to characterize individual gait signatures.
- Introduction: Conventional gait analyses often examine variables such as gait velocity, step length, or range of motion in isolation.
- Introduction: Machine-learning methods analyze time-continuous gait patterns but commonly operate as black boxes that obscure their decision processes.
- Introduction: Using DNNs, the study tests whether gait patterns are unique enough to predict individual identities.
- Introduction: LRP is used to verify whether nonlinear models learn comprehensible gait characteristics and to identify relevant variables across gait-cycle time windows.
- Introduction: The study investigates whether state-of-the-art machine-learning methods can classify and interpret individual gait patterns.
- Introduction: The broader aim is to support future machine-learning applications for automatic classification of neurological disorders and pathological gait conditions in clinical diagnosis.
Results
Across 57 healthy subjects, gait patterns were classified accurately from time-normalised joint angles and ground reaction forces, while LRP identified the gait features and time windows driving each prediction. Model comparisons showed differences in robustness and relevance reliability despite similarly high accuracy.
- 57 healthy subjects were classified from full-body joint angles and ground reaction forces using deep artificial neural networks.The inputs were time-normalised and concatenated gait vectors, with LRP used to decompose model predictions.
- 100% true-positive rate for LBJAX and 95.23% for GRF were achieved for subject 6 with CNN-A.
- Interpreting and Understanding Model Predictions using Layer-Wise Relevance Propagation: LRP revealed subject-specific gait signatures by locating relevant variables and gait-cycle windows, including ankle, knee, hip, and vertical-force features for subject 6.For subject 6, relevance was highest around terminal stance for vertical ground reaction forces and during initial contact for selected joint angles.
- Above 95.4% accuracy was achieved for ground reaction forces, while full-body and lower-body joint angles reached above 99.9%.Linear SVM showed the highest accuracy across kinetic and kinematic variables, while several neural architectures produced similarly high performance.
- Interpreting and Understanding Model Predictions using Layer-Wise Relevance Propagation: Different subjects were identified through different gait characteristics, magnitudes, or shapes, with relevance peaks occurring at distinct gait-cycle phases.Examples included medial/lateral force near 10% stride for subject 21, vertical force near 90% for subject 28, and swing-phase characteristics for subject 37.
- Interpreting and Understanding Model Predictions using Layer-Wise Relevance Propagation: Most models used complex combinations of gait characteristics, with comparable relevance patterns across architectures and between body sides.The models generally treated most inputs as irrelevant, while increasing model complexity was associated with greater relevance reliability and robustness; Linear (SGD) showed the lowest reliability and noise robustness.
Discussion
The findings support individual gait signatures and show that LRP can identify the variables and time windows underlying DNN-based subject classification. This interpretability is relevant to more individualized gait analysis, diagnosis, and intervention.
- LRP verified that individual gait patterns have distinctive kinematic and kinetic characteristics that can be differentiated across subjects.The analysis used time-continuous ground reaction forces and full-body joint angles.
- LRP identifies which input variables and gait-cycle time windows are most characteristic of a given individual.Relevance values can be determined from single trials to support individualized analyses, diagnoses, and interventions.
- Individual identification generally depends on shapes and interactions among multiple gait variables across the same or different gait-cycle time windows.This contrasts with conventional approaches centered on single pre-selected variables.
- Input relevance values were similar across model architectures in most cases, while LRP also enabled comparison of each model’s classification strategy.The shared relevance patterns indicate that different models often selected features characteristic of the individual subject.
- Relevance decompositions were consistent across test trials and cross-validation splits, with greater reliability for deep convolutional architectures.These findings support the suitability of LRP for explaining machine-learning predictions in clinical biomechanical gait analysis.
- Interpreting machine-learning predictions is presented as essential for overcoming the transparency barrier to clinical gait-analysis applications.The paper links individual gait characteristics with the potential to individualize clinical analyses, diagnoses, and interventions.
Methods
The study analyzed gait from 57 healthy subjects using time-continuous kinetic and kinematic measurements, preprocessing each variable course before classification. It compared linear, fully connected, convolutional, and support-vector models, with LRP decomposing model predictions into input relevance.
- Data collection: 57 healthy subjects provided gait trials with full-body joint angles and ground reaction forces measured during barefoot walking.Each subject completed 20 trials in a single assessment session on a 10 m path.
- Data processing: Each stride was normalized to 101 time points, filtered, z-transformed, and scaled from -1 to 1 before neural-network classification.Scaling was used to prevent numerical difficulties and unequal variable contributions caused by different numeric ranges.
- Classification: Subject identification was evaluated as a multi-class classification problem using time-continuous kinetic and kinematic data from 57 subjects.Prediction accuracies were reported using ten-fold cross-validation with separate training, validation, and test partitions.
- Models: The comparison included linear classifiers, fully connected neural networks, convolutional neural networks, and a linear-kernel support vector machine.Models were compared on prediction accuracy, robustness to test-data noise, and decomposition of input relevance values.
- Layer-Wise Relevance Propagation: LRP distributes a neuron's relevance to its inputs according to their forward-pass contributions and pools incoming relevance at each input neuron.For a single linear layer, the resulting explanation reduces to Ri = xiwi.
Supplementary Results
The supplementary tables report subject-classification rates, perturbation robustness, and coefficient-of-variation results for neural-network models. They provide mean and standard-deviation summaries across the evaluated models and noise conditions.
- Classification rates: Supplementary Table ST1 reports subject-classification rates for artificial-neural-network models with n=57 as mean and standard deviation in percent.The table concerns classification rates across the evaluated models.
- Noise robustness: Supplementary Table ST2 reports area over perturbation curve values for ground-reaction-force classification across multiple noise perturbation runs.Gaussian, salt, pepper, and shot noise conditions are included, with AOPC values reported as mean and standard deviation.
- Variability: Supplementary Table ST3 reports coefficients of variation for subject-classification models as mean and standard deviation.The table summarizes variation in classification results across the evaluated models.
Neural Network Architectures
The study evaluated linear, multilayer-perceptron, and convolutional architectures across feature sets with different channel counts and time dimensions. CNN variants differed in filter sizes, strides, depth, and temporal scope, while dense models used flattened inputs.
- Input representations: The evaluated feature sets use dimensionality D = C × T, where C is the number of channels and T is the number of time points.Architectures were adapted because feature sets contain varying numbers of data channels.
- Linear and dense models: Linear and multilayer-perceptron models receive flattened channel-by-time arrays, with hidden-layer width H and output-label count L defining their structure.The linear baseline has a Dense (D,L) layer, while MLP variants use two or three dense layers with ReLU hidden activations and SoftMax output.
- CNN design groups: CNN-A reads all channels at once with a square temporal scope, whereas CNN-A3 also reads all channels at once but limits its temporal filter scope.CNN-A3 uses two convolutional layers, while the supplied architecture tables specify the corresponding layer configurations for each feature set.
- Feature-set-specific architectures: Supplementary architecture tables specify CNN configurations for ground-reaction forces, full-body joint angles, sagittal-plane joint angles, and lower-body joint angles.The listed groups include CNN-A, CNN-A3, CNN-A6, CNN-C3, CNN-C3-3, and CNN-C6 variants.
Supplementary Subject Specific Analyses
Subject-specific relevance maps identify gait characteristics and gait-cycle windows that contribute to individual classification. Across examples, the highlighted signals differ by subject and include distinct ground-reaction-force features.
- Subject 50: For subject 50, vertical ground-reaction force during mid-stance of right and left foot contact has the highest relevance.The highlighted window corresponds to the phase when one leg is located directly over the foot and the other leg is in the supplied example.
- Subject 54: For subject 54, vertical ground-reaction force during terminal stance of both the right and left leg contributes to identification.The example also includes lower-body sagittal-plane joint-angle relevance maps.
- Subject 56: For subject 56, vertical ground-reaction force during initial and terminal stance of both legs is identified as unique.The example likewise presents lower-body sagittal-plane joint-angle relevance maps.