Automating General Movements Assessment with quantitative deep learning to facilitate early screening of cerebral palsy

Cerebral palsy (CP), the most prevalent childhood motor disability, affects movement, balance, and posture. Early diagnosis is critical for optimal intervention. The Prechtl General Movements Assessment (GMA) is a valuable tool for assessing infant nervous system integrity and predicting CP risk by evaluating the quality of general movements (GMs), particularly fidgety movements (FMs) between 9-20 weeks corrected age. Absence of FMs is a strong predictor of CP. However, GMA requires highly trained professionals, limiting its accessibility and hindering early screening efforts in many regions. This necessitates the development of automated methods for assessing GMs, particularly focusing on the FMs stage. While wearable sensors and motion capture have been explored, these methods can be intrusive and operationally demanding. Video-based approaches offer a non-intrusive alternative. Deep learning, particularly with pose estimation tools, provides a robust solution to automatically extract movement features from videos, overcoming challenges like background interference. Existing automated GMA methods often use qualitative approaches, lacking the objective measurements and numerical data provided by quantitative tools. This research aims to address these limitations by developing a deep learning model that automates GMA at the FMs stage, providing both qualitative and quantitative assessments.

Numerous studies highlight the importance of early CP detection and the value of GMA in this process. Research has explored using wearable sensors and motion capture for automated GM assessment, but these methods face challenges in terms of intrusiveness and operational complexity. Video-based approaches, leveraging techniques like background subtraction and optical flow, offer a less intrusive alternative. Deep learning advancements, particularly in pose estimation, have greatly improved the accuracy of automatically extracting movement features from video data. However, existing automated GMA methods largely focus on qualitative assessments, providing only a final classification without objective quantitative measurements. While some studies have attempted to quantify GMA, their performance hasn't consistently matched qualitative methods, and their emphasis on specific body parts often contradicts the holistic gestalt perception inherent in GMA. The need for more interpretable and accurate automated methods remains a significant challenge in the field. This study builds upon previous work by incorporating both qualitative and quantitative approaches within a deep learning framework, addressing the shortcomings of previous methods.

This study employed a three-cohort design. Cohort 1 (n=906) and Cohort 2 (n=221) were used for internal cross-validation and external validation, respectively. Cohort 3 (n=243) served as a pre-training dataset. The infants' corrected ages ranged from 9 to 20 weeks. Infants with abnormal FMs were excluded due to their rarity and limited predictive power. Infants with continuous and intermittent FMs were categorized as the normal group, while those with sporadic and absent FMs were categorized as the risk group. Videos were recorded according to GMA standards. A 3D pose estimation method (VideoPose3D, fine-tuned on a subset of Cohort 3) was used to predict the coordinates of key joints from the videos. A distance representation approach captured overall motion patterns. A deep learning-based motor assessment model (MAM) was developed using a multi-instance learning (MIL) framework with a Transformer architecture. MAM comprises three branches: a FMs reference branch (Ref Branch), a main branch (Main Branch), and an information branch (Info Branch). The Ref Branch used short FMs and non-FMs clips for pre-training, employing Triplet loss to distinguish between the two. The Main Branch processed entire videos, dividing them into instances for MIL, using an attention-based fusion mechanism to integrate instance predictions. The Info Branch integrated infants' basic characteristics (sex, gestational age, birth weight, assessment age). A Closeness loss function was implemented to ensure the Main Branch's predictions aligned with the Ref Branch. A quantitative GMA method was developed using the predicted frequency of FMs clips in each video. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, PPV, and NPV. The Shapley Additive exPlanations (SHAP) method was used for interpretability analysis. The diagnostic accuracy of GMA beginners with and without MAM assistance was compared.

MAM demonstrated exceptional performance in predicting GMs at the FMs stage. In the external validation dataset, MAM achieved an AUC of 0.967, accuracy of 0.934, sensitivity of 0.925, specificity of 0.936, PPV of 0.802, and NPV of 0.978. Compared to other state-of-the-art methods (EML, STAM, WO-GMA), MAM significantly outperformed them across all metrics. The ablation study showed that while the Info Branch marginally improved performance, the difference was not statistically significant. SHAP analysis indicated that video features were the primary drivers of predictions. MAM accurately identified FMs in videos, showing substantial agreement with expert assessments (median kappa value of 0.601–0.620). The quantitative GMA method based on FMs frequency achieved an AUC of 0.956. Assisting GMA beginners with MAM significantly improved their diagnostic accuracy (average increase of 11.0%). Analysis showed that the distance matrices input construction method and 3D pose estimation outperformed other approaches.

This study successfully developed and validated MAM, a deep learning model for automating GMA at the FMs stage. MAM significantly outperforms existing methods in both accuracy and interpretability. The high AUC values demonstrate the model's strong ability to discriminate between normal and risk groups. The incorporation of a quantitative approach using FMs frequency provides objective, easily interpretable results, improving the diagnostic capabilities of less experienced clinicians. MAM's ability to pinpoint FMs within videos enhances its transparency and reliability. The study's findings demonstrate the potential of MAM to facilitate early and more widespread screening for CP, ultimately improving outcomes for infants at risk. The model's non-intrusive nature (using only video data) makes it a particularly attractive and accessible tool.

This research presents a novel deep learning model (MAM) for automating the Prechtl General Movements Assessment, significantly improving the efficiency and accuracy of early cerebral palsy screening. MAM's superior performance, combined with its quantitative and interpretable features, offers a promising tool for broader clinical implementation. Future research could focus on expanding the dataset, exploring different deep learning architectures, and integrating MAM into existing clinical workflows.

The study's dataset, while substantial, is limited to a specific geographical region and may not fully generalize to diverse populations. The model's performance relies heavily on the quality of video recordings. Further research is needed to evaluate MAM's robustness across various recording conditions and demographic groups. While the model shows high accuracy, it is important to incorporate it into clinical practice carefully and not replace the expert judgment of trained professionals completely.