Concept Bottleneck Model for Adolescent Idiopathic Scoliosis Patient Reported Outcomes Prediction

Micky C Nnamdi; Wenqi Shi; Junior Ben Tamo; Henry J Iwinski; Michael J Wattenbarger; May Dongmei Wang

Abstract

Post-surgical patient-reported outcomes (PROs) serve as a crucial subjective measure of surgical success for adolescent idiopathic scoliosis (AIS) patients. Leveraging pre-operative patient information to predict post-operative PROs is instrumental in improving pediatric patient care and providing invaluable insights for clinical decision-making. Recently, deep learning techniques have demonstrated encouraging results in developing predictive models for clinical decision support. However, the inherent black-box nature makes them non-interactive and challenging to troubleshoot during the training phase. To mitigate this issue, our study introduces an interactive concept bottleneck model to predict subjective rehabilitation outcomes for AIS patients. We assess three learning schemas - independent, sequential, and joint - to first comprehend the concepts, which are a set of post-operative radiographic data available during the training phase. Subsequently, these acquired concepts are employed to predict post-operative patient rehabilitation outcomes across five domains: pain, function, general satisfaction, self-image, and mental health. Our results demonstrated improvement compared to the existing baseline, with the joint learning schema yielding the highest F1 score in the function and pain domains, while sequential learning recorded the highest F1 score in the mental health and self-image domains. This proposed framework harbors the immense potential to aid pre-operative surgical planning and further enhance the transparency of AI models, thereby supporting real-world clinical decision-making.

Keywords: concept bottleneck model; explainable artificial intelligence; adolescent idiopathic scoliosis; pediatric healthcare

Links
[Full text PDF][Bibtex]