Individuals with unilateral transfemoral amputation (uTFA) exhibit asymmetric gait pattern during walking due to partial loss of the lower limb and muscles on the amputated side (Winiarski et al., 2021). This gait asymmetry is related to multiple factors such as walking speed (Nolan et al., 2003), type of prosthetic components (Kaufman et al., 2012, Petersen et al., 2010, Schaarschmidt et al., 2012), level of amputation (Keklicek et al., 2019), muscle strength (Heitzmann et al., 2020, Krajbich et al., 2023, Rutkowska-Kucharska et al., 2018), compensatory patterns (Harandi et al., 2020), amputation surgery (Ranz et al., 2017), and person-dependent gait deviations. Additionally, residual femur length (Bell et al., 2013), cardiorespiratory fitness (Gjovaag et al., 2014), prosthetic socket type (Traballesi et al., 2011), prosthetic alignment (Kobayashi et al., 2013, Zhang et al., 2019), and the individual’s walking habits also affect the spatiotemporal parameters of gait. Classifying gait deviations is essential for identifying compensatory patterns and tailoring rehabilitation programs aimed at improving gait symmetry and function. It also enables objective tracking of patients' progress and aids in optimizing prosthetic interventions. However, due to the multifaceted nature of these factors, it is challenging to holistically classify gait in individuals with uTFA.

Limited studies have been conducted to categorize gait in individuals with uTFA (Ichimura et al., 2022). One study successfully identified three distinct gait clusters among individuals with uTFA using an unsupervised machine learning approach—clustering—with each cluster showing significant differences in cadence and step length across various walking speeds (Ichimura et al., 2022). Utilizing an unsupervised machine learning approach is advantageous because it bypasses the need to consider all factors affecting gait, relying instead on direct data-driven pattern recognition to classify gait patterns. This approach has been successfully applied to identify distinct gait patterns in other patient populations, such as those with stroke or cerebral palsy, providing a means of tracking and evaluation of gait variations over time (Abbasi et al., 2021, Chantraine et al., 2022, Roche et al., 2014). Conventional classifiers that incorporate both kinematic and electromyographic (EMG) data can distinguish gait phases with high accuracy (Mobarak et al., 2024, Tigrini et al., 2024), offering valuable insights for gait phases recognition in smart prostheses. Integrating such classifiers with gait data for individuals with uTFA could further improve locomotion and accelerate the development of intelligent prosthetic technologies. However, studies specifically focused on gait pattern classification in individuals with uTFA remain limited. By applying clustering techniques to gait data from this population, our study aims to objectively classify distinct gait types and ultimately inform personalized interventions that address specific gait deviations and rehabilitation needs.

However, using only a clustering algorithm has its limitation, as it does not reveal which specific gait features contribute most to gait classification. Typically, researchers select key features for model training based on their expertise in gait analysis. While expert input is valuable, relying solely on subjective judgement can reduce objectivity and reproducibility in feature selection, potentially limiting the generalizability of the findings. The random forest model can be utilized to calculate feature importance, aiding in identification of the most critical characteristics that distinguish between groups (Luo et al., 2020). By training the model on gait data from both able-bodied individuals and those with uTFA, it can assess which features most effectively differentiate these groups. Features that significantly contribute to the model’s classification accuracy are ranked higher, indicating their importance in capturing the unique aspects of each group's gait pattern.

Symmetry parameters are metrics used to quantify the degree of bilateral symmetry, typically calculated from measured bilateral kinematic or kinetic gait parameters (Viteckova et al., 2018). Different formulas can yield various values, generally represented as percentages within a range, such as 0 to 1, where 0 indicates perfect symmetry and 1 indicates complete asymmetry. For individuals with uTFA, bilateral asymmetry is common (Schaarschmidt et al., 2012), and symmetry parameters could serve as valuable indicators for assessing gait quality and monitoring improvement over time.

In this study, the first objective is to assess the importance of the features (include kinematic parameters, spatiotemporal parameters, and symmetry parameters derived from the spatiotemporal parameters) that distinguish the gait of individuals with uTFA from that of non-disabled using a random forest model. These features are then used to cluster individuals with uTFA. The second objective is to identify the gait patterns of different clusters. Our hypothesis is that the clustering algorithm can identify multiple distinct gait patterns with significant differences in spatiotemporal and symmetry parameters.