Towards Data-Driven Gait Analysis with a Special Focus on Individuals with Multiple Sclerosis

The healthcare sector is witnessing a paradigm shift, driven by the transformative potential of digital technologies. This digital revolution fundamentally redefines data collection, interpretation, and utilization, impacting both clinical practice and research endeavours. Within this evolving landscape, wearable sensors have emerged as a promising tool for unobtrusive monitoring of physiological kinematic data. This holds particular promise for the understanding of the progression of neurological conditions, where mobility assessments play a significant role. One such neurological condition where data-driven sensory assessments can be particularly impactful is multiple sclerosis (MS). MS is a slowly progressive heterogeneous neurodegenerative disease which primarily manifests through reduced mobility. Considering the complexity of the disease, traditional clinical evaluations, relying on subjective observation and intermittent testing, may overlook subtle gait changes over time. This highlights the need for more objective and sensitive assessment methods. Recent advancements in machine learning technology—which have been specifically developed to extract meaningful information from high-dimensional data and model complex non-linear biomechanical signals—appear well-suited to augment clinical assessment with data-driven insights. As such, this thesis proposes a data-driven assessment pipeline, which includes detection of gait impairment, severity assessment, as well as methodologies for longitudinal monitoring.

Within the context of gait impairment detection, the first contribution of this thesis consists in the proposal of a novel gait anomaly detection technique, using the Mahalanobis squared distance, along with the minimum covariance determinant. This approach offers robust estimates of the true healthy participant condition and improves sensitivity of gait impairment detection for the MS population.

The second contribution of this thesis is the proposal of a novel neural network-based framework for disease severity assessment using a single wearable sensor worn on the lower back. The thesis introduces contrastive learning approaches, which aim to effectively cluster individuals with similar gait patterns, improving model generalization. Additionally, the thesis employs layer-wise relevance propagation to identify key gait features associated with severity assessment, aiding interpretability and building trust into the model predictions.

The third contribution addresses a critical gap in the current practice: the lack of reliable methods for monitoring disease progression over time. Therefore, the thesis introduces a novel technique for assessing the consistency of movement patterns between clinical visits. Using residual patterns as a sensitive feature, computed as the difference between the predictions of autoregressive with eXogenous inputs models and the true measured data, together with kernel two-sample hypothesis testing using the maximum mean discrepancy, this thesis provides some fundamental considerations and identifies the challenges for longitudinal data analysis, highlighting the need for more generalizable models.

The final developments in this thesis approach gait analysis from a different perspective by introducing a novel Bayesian framework for probabilistically modelling the shank angular velocity as a proxy for lower limb distal motion. Given the heterogeneous MS gait pattern, a probabilistic approach can offer valuable insights in the challenging problem of assessing and quantifying the degree of gait impairment and its changes over time, especially in the context of neurological disorders, such as MS, which is marked by intrinsically unpredictable disease progression. As such, the last contribution presented here is the extension of hierarchical Gaussian process models to effectively handle heteroscedasticity and facilitate scalability to large datasets through sparse inference. By acknowledging the hierarchical nature of wearable sensor data—collected from contralateral limbs, individuals, and groups of individuals comprising a population- this modelling approach allows a granular analysis of the gait patterns. The idea is to make a departure from understanding gait with respect to a set of summary features. Instead, the shank angular velocity is modelled functionally, across the entire gait cycle, with automatic uncertainty estimation.

The methodological development of the algorithms presented in this thesis leaves the user with a toolbox of methods which can facilitate not only a better understanding of the gait patterns exhibited by people with MS, but can be also extrapolated to other pathological conditions affecting gait.