Abstracts

Radiographic Assessment of Hip Disease in Children with Cerebral Palsy: Development of a Core Measurement Set and Analysis of an Artificial Intelligence System

by Prince Joseph

Cerebral palsy is the most common physical disability during childhood. Cerebral palsy related hip disease is caused by an imbalance of muscle forces, resulting in progressive migration of the hip to complete dislocation. This can decrease function and quality of life. The prevention of hip dislocation is possible if detected early. Therefore, surveillance programmes have been set up to monitor children with cerebral palsy enabling clinicians to intervene early and improve outcomes. Currently, hip disease is assessed by analysing pelvic radiographs with various geometric measurements. This time-consuming task is undertaken frequently when monitoring a child with cerebral palsy. This thesis aimed to identify the key radiographic parameters used by clinicians (the core measurement set), and then build an artificial intelligence system to automate the calculation of this core measurement set. A systematic review was conducted identifying a comprehensive list of previously reported measurements from studies measuring radiographic outcomes in cerebral palsy children with hip pathologies. Fifteen measurements were identified from the systematic review, of which Reimers’ migration percentage was the most commonly reported. These measurements were used to perform a two-round Delphi study among orthopaedic surgeons and physiotherapists. Participants rated the importance of each measurement using a nine-point Likert scale (‘not important’ to critically important’). After the two rounds of the Delphi process, Reimers’ migration percentage was included in the core measurement set. Following the final consensus meeting, the femoral head-shaft angle was also included. The anteroposterior pelvic radiographs of 1650 children were then used to build an artificial intelligence system integrating the core measurement set, in collaboration with engineers from the University of Manchester. The newly developed artificial intelligence system was assessed by comparing its ability to calculate measurements and outline the pelvis and femur on a radiograph. The reliability of the dataset used to train the model was also analysed. The proposed artificial intelligence model achieved a ‘good to excellent’ inter-observer reliability across 450 radiographs when comparing its ability to calculate Reimers’ migration percentage to five clinicians. Its ability to outline the pelvis and proximal femur was ‘adequate’ with the better performance observed in the pelvis than the femur. The reliability of the training dataset used to teach the artificial intelligence model was ‘good’ to ‘very good’. Artificial intelligence systems are feasible solutions to optimise the efficiency of hip radiograph analysis in cerebral palsy. Studies are warranted to include the core measurement set as a minimum when reporting on hip disease in cerebral palsy. Future research should investigate the feasibility of implementing a risk score to predict the likelihood of hip displacement.