An internal methodological evaluation of a multimodal VIDEO-IMU metric-learning framework for fidgety movement classification in preterm neonates: a controlled sample-level study

Abstract Background Preterm neonates have a substantially higher risk of developing cerebral palsy than term neonates, and early identification of abnormal fidgety movement (FM) is clinically important for early risk screening. Methods This pilot study investigated whether a combined metric-learning loss can improve FM/non-FM (nFM) representation learning in a neonatal multimodal dataset. Data were obtained from 11 preterm neonates at 3–5 months of corrected age and segmented into 1957 multimodal samples consisting of motion history image (MHI) features and inertial measurement unit (IMU) signals, with FM/nFM labels and segment metadata organized in a unified manifest. The proposed ImageBind-based framework was trained using MHI and IMU inputs, with FM/nFM labels used for supervised loss optimization, and final test-stage inference was likewise performed using only MHI and IMU data. To address the complementary limitations of triplet loss and tuned constellation loss, we designed a combined loss that jointly enforces local distance constraints and global embedding structure. Results Within this predefined internal sample-level evaluation setting, the combined-loss model outperformed single-loss baselines, achieving an accuracy of 0.838, F1-score of 0.828, AUROC of 0.909, and AUPRC of 0.894 on the internal test partition. These findings suggest that the combined loss mitigated the failure modes of single-loss training and improved class-wise balance under the present pilot methodological setting. Conclusion Overall, this study provides foundational evidence that multimodal metric learning can support FM/nFM representation learning in preterm neonates and motivates further validation in larger subject-held-out cohorts.