A Pre-Registered, Open Pipeline for Early Cerebral Palsy Risk Assessment from Infant Videos
Melanie Segado1,2,3, Laura A Prosser4,3, Andrea F Duncan4,5
1Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, United States.
Insights
This study developed an automated pipeline for predicting Cerebral Palsy (CP) risk using General Movements Assessment (GMA) from infant videos. The approach enables broader data sharing and model training for more accessible CP screening tools.
Area of Science:
- Neurology
- Developmental Pediatrics
- Biomedical Engineering
Background:
- Cerebral Palsy (CP) affects 1 in 500 children, impacting motor control due to abnormal brain development.
- Early detection via General Movements Assessment (GMA) at 3-4 months is crucial for CP risk assessment.
- Current machine learning models for GMA prediction lack generalizability and require dataset-specific tuning, hindering clinical application and data sharing.
Purpose of the Study:
- To develop a generalizable, end-to-end machine learning pipeline for predicting GMA scores from infant videos.
- To overcome limitations of existing models, such as dataset-specific tuning and privacy constraints, enabling multi-site data aggregation.
- To establish groundwork for robust, globally accessible CP screening tools, particularly for low-resource settings.
Main Methods:
- An end-to-end pipeline was created using off-the-shelf pose estimation and general-purpose feature extraction.
- Automated machine learning (AutoML) was employed, with no dataset-specific tuning.
- The approach was applied to a new dataset of 1063 infants from a high-risk cohort, with evaluation on a locked validation set.
Main Results:
- The model achieved moderate predictive accuracy for clinician-assessed GMA scores.
- Key performance metrics included an Area Under the Receiver Operating Characteristic Curve (ROC-AUC) of 0.79 and an Area Under the Precision-Recall Curve (PR-AUC) of 0.34.
- The performance is noteworthy given the 12% positive class prevalence in the dataset.
Conclusions:
- The developed pipeline offers a generalizable approach for GMA-based CP risk prediction.
- Open-sourcing code and de-identified data facilitates further research and development of CP screening tools.
- This work supports the creation of robust, accessible CP screening solutions for diverse clinical settings.
Abstract:
Cerebral Palsy (CP), affecting approximately 1 in 500 children due to abnormal brain development, impacts movement control. Early risk assessment via the General Movements Assessment (GMA) at 3-4 months is highly predictive for CP but relies on trained clinicians. Machine-learning-based approaches for predicting GMA score from video have shown considerable promise, but are not openly available and rely on fine-tuned pre-processing steps, hand-crafted feature sets, and experimenter-driven hyperparameter selection. This, combined with strict privacy constraints on sharing data, limits the extent to which models can be trained and tested across datasets, thus reducing clinical impact. There is therefore a need to develop approaches that will work across different datasets to enable multi-site dataset aggregation and model training. To address this gap, we developed an end-to-end pipeline that uses off-the-shelf pose estimation, general-purpose feature extraction, and automated machine learning-none of which are tuned to a specific dataset. We applied this approach to a newly generated large dataset of 1063 infants (with approximately 12% positive class for adverse GMA outcome, drawn from a high-risk clinical cohort) within a preregistered study design. Model performance was evaluated on a strict "lock-box" validation set, which remained untouched during any phase of model development or pre-processing optimization. The developed model achieved moderate predictive accuracy for clinician-assessed GMA scores (Area Under the Receiver Operating Characteristic Curve, ROC-AUC = 0.79; Area Under the Precision-Recall Curve, PR-AUC = 0.34). The moderate accuracy is noteworthy given the 12% positive class prevalence. By releasing de-identified feature data and open-source code, and simplifying the training pipeline using automated machine learning, our work establishes essential groundwork for future robust, globally relevant CP screening tools suitable for low-resource settings.
