Predicting gait kinetics using 3-degrees of freedom acceleration data and artificial neural networks.

BACKGROUND: Motion analysis plays an important role in clinical decision-making and biomechanical research. Conventional assessments rely on laboratory-based motion capture and force plates, which are accurate but resource-intensive. Wearable sensors combined with artificial neural networks offer a promising alternative for use outside the laboratory, but they must be validated against established systems. The aim of this study was to predict important kinetic gait parameters, such as ground reaction forces, knee joint moments and knee power, using wearable sensors and neural networks. Such an approach may support gait assessment in orthopaedic conditions and enable more responsive adjustment of exoprostheses. METHODS: Thirty-two healthy adults provided informed consent and were instrumented. Standard kinematic and kinetic data were captured using a conventional motion analysis setup, while linear accelerations were recorded synchronously using wearable sensors. These acceleration signals were used as inputs for two neural network models (long short-term memory and multilayer perceptron) to predict kinetic parameters. FINDINGS: Both models achieved good to very good agreement with the reference system. Pearson correlation coefficients ranged from 0.79 for the knee joint power to 0.99 for anteroposterior ground reaction force. The normalised root mean square errors between 3.1% and 10.3% further demonstrated promising predictive accuracy. INTERPRETATION: The findings indicate that wearable sensors combined with neural networks can estimate clinically relevant kinetic gait parameters with high accuracy outside a laboratory setting. In the future, such predictions may help standardize dynamic adjustments of exoprostheses, support the diagnosis of orthopaedic conditions and evaluate treatment outcomes.