| dc.description.abstract | Objective: The recent development of motion tracking systems offers new possibilities to automate and quantify the assessment of general movements exhibited by infants, which could contribute to early detection and treatment of cerebral palsy (CP) in high-risk infants. The primary aim of this study was to determine how accurately multiple quantified Motor Optimality Score (MOS) items are differentiating between CP and non-CP outcomes. The secondary aim of this study was to determine how the quantified MOS-items correspond with what is scored as “observed” or “not observed” by the clinician or General Movement Assessment (GMA) expert.
Participants & method: video recordings of 557 infants (CP = 84, non-CP = 473) who were enrolled in previous studies and were all at high risk for developing a perinatal brain injury were included. Using MATLAB (version R2021b), a skeleton representation of the infants was constructed, and several variables and features were calculated. A total of nine MOS-items were selected for quantification based on several exclusion criteria. A grid search was performed for each variable in order to find the threshold with the best discriminative performance, based on: 1) the Area Under the Curve (AUC) of the Receiver Operating Characteristic (ROC) curve, 2) specificity, 3) sensitivity, and 4) balanced accuracy was used. The performance of each quantified MOS-item was assessed for all features individually, for a clustering within the items, and for a clustering between the items. The AUC of the ROC curve as well as the distribution differences of the features between CP and non-CP were analyzed to assess the performance of the model. Lastly, the prevalence of each item was determined, split for the CP and non-CP group.
Results: AUC for the within-item feature clusters ranged from 0.5188 (95% CI = 0.3811 - 0.6419) to 0.7623 (95% CI = 0.6557 - 0.8620) and the between-item clustering of all features across the items resulted in an AUC of 0.6207 (95% CI = 0.4918 - 0.7453). Feature importance across the different clusters revealed a high relative weight of body symmetry total time (17.5%), body symmetry frequency (13.4%), and legs lift total time (12.6%). When separated for the different features, the head-centered item had the highest relative weight (50.2%, 60.6%, 62.8% and 54% for respectively the features total time, average time, percentage, and frequency). Deviation in prevalence compared to previous observational research mainly concerned items including joint-angle calculations or movements along the line of gravity, suggesting insufficient quantification of these items.
Conclusion: This study found that individual features and between-item clustering of features had limited to no predictive capability. On the other hand, a within-item clustering of features showed that seven out of nine quantified MOS-items were performing better than a random classifier for the prediction of CP. This suggests that the developed quantification method based on within-item clustering, is a feasible method to differentiate between CP and non-CP outcomes in high-risk infants. | |
