Can we use lower extremity joint moments predicted by the artificial intelligence model during walking in patients with cerebral palsy in the clinical gait analysis?

Several studies have highlighted the advantages of employing artificial intelligence (AI) models in gait analysis. However, the credibility and practicality of integrating these models into clinical gait routines remain uncertain. This study critically evaluates an AI model's ability to predict lower extremity joint moments during gait in patients with cerebral palsy (CP). We employed a three-step approach to assess the feasibility of a previously developed AI model that predicted joint moments during walking for 622 patients with CP, using joint kinematics as input. First, we established clinically relevant thresholds for lower extremity joint moments, categorizing into three labels: acceptable (Green), acceptable with caution (Yellow), and unacceptable (Red). This categorization was based on the normalized root mean square error (nRMSE) between lab-measured and predicted joint moments. We explored the relationship between gait kinematics and joint moments by correlating the kinematic inputs with their respective output labels. Finally, we developed a linear discrimination analysis (LDA) model to predict labels for newly predicted joint. Assessing the validity of thresholds, an ANOVA one-way analysis and Bonferroni post-hoc statistical tests were performed to find significant differences between the nRMSE values for each label. The hip joint exhibited the largest population of Green labels (84%), while the ankle joint had the smallest (50%). Regressive differences in joint kinematics and gait profile scores were observed across all labels. The LDA model achieved an accuracy of 85.2% and an F-score of 92% for predicting Green label in hip joint moment. Additionally, more severe patient conditions were associated with an increase in Red-labeled predictions. Our findings highlight significant differences in nRMSE among labels, demonstrating the effectiveness of the proposed thresholds for labeling joint moments. Overall, the AI model's performance was rated as moderate, and the three-step approach proved valuable for assessing the feasibility of AI models in clinical settings.