Artificial intelligence (AI) algorithms for gait analysis rely on properly identified foot strikes for step-based feature calculation. Smartphone signals collected during movement assessments, such as the 6-minute walk test (6MWT), have been used to train AI models for foot strike identification and fall risk classification in able-bodied populations. However, there is limited research in populations with more asymmetrical gait. People with lower limb amputation can have high gait variability, adversely affecting automatic step detection algorithms. Hence, fall risk models for lower limb amputees have relied on manual foot strike labelling to calculate step-based features for model training, which is inefficient and impractical for clinical use. In this thesis, decision tree and long-short term memory (LSTM) models were developed, optimized, and their performance compared for automated foot strike identification in an amputee population. Eighty people with lower limb amputations (27 fallers, 53 non-fallers) completed a 6MWT with a smartphone at the posterior pelvis. Automated and manually labelled foot strikes from the full 6MWT and from the first two minutes of data were used to calculate step-based features. A random forest model was used to classify fall risk. The best foot strike identification model was an LSTM with 100 hidden nodes in the LSTM layer, 50 hidden nodes in the dense layer, and batch size of 64 (99.0% accuracy, 86.4% sensitivity, 99.4% specificity, 82.7% precision). Automated foot strikes from the full 6MWT data correctly classified more fallers (55.6% versus 48.1%), whereas automated foot strikes from 2-minute data classified more non-fallers (90.6% versus 81.1%). Feature calculation using manually labelled foot strikes resulted in the best overall performance (80.0% accuracy, 55.6% sensitivity, 92.5% specificity). This research created a novel method for automated foot strike identification in lower limb amputees that is equivalent to manual labelling and demonstrated that automated foot strikes can be used to calculate step-based features for fall risk classification. Integration of the foot strike identification model into a smartphone application could allow for immediate stride analysis after completing a 6MWT; however, fall risk classification model improvement is recommended to enhance clinical viability.

Accelerometers and gyroscopes are used to detect foot strike (FS), i.e., the moment when the foot first touches the ground. However, it is unclear whether different conditions (footwear hardness or foot strike pattern) influence the accuracy and precision of different FS detection methods when using such micro-electromechanical sensors (MEMS). This study compared the accuracy of four published MEMS-based FS detection methods with each other and the gold standard (force plate) to establish the most accurate method with regard to different foot strike patterns and footwear conditions. Twenty-three recreational runners (12 rearfoot and 11 forefoot strikers) ran on a 15-m indoor track at their individual running speed in three footwear conditions (low to high hardness). MEMS and a force plate were sampled at a rate of 3750 Hz. Individual accuracy and precision of FS detection methods were found which were dependent on running styles and footwear conditions. Most of the methods were characterized by a delay which generally increased from rearfoot to forefoot strike pattern and from high to low midsole hardness. It can be concluded that only one of the four methods can accurately determine FS in a variety of conditions.

Accelerometers and gyroscopes are used to detect foot strike (FS), i.e., the moment when the foot first touches the ground. However, it is unclear whether different conditions (footwear hardness or foot strike pattern) influence the accuracy and precision of different FS detection methods when using such micro-electromechanical sensors (MEMS). This study compared the accuracy of four published MEMS-based FS detection methods with each other and the gold standard (force plate) to establish the most accurate method with regard to different foot strike patterns and footwear conditions. Twenty-three recreational runners (12 rearfoot and 11 forefoot strikers) ran on a 15-m indoor track at their individual running speed in three footwear conditions (low to high hardness). MEMS and a force plate were sampled at a rate of 3750 Hz. Individual accuracy and precision of FS detection methods were found which were dependent on running styles and footwear conditions. Most of the methods were characterized by a delay which generally increased from rearfoot to forefoot strike pattern and from high to low midsole hardness. It can be concluded that only one of the four methods can accurately determine FS in a variety of conditions.