Benchmarking implicit neural representations of pediatric gait waveforms against conventional biomechanical descriptors

Machine-learning representations of gait waveforms are increasingly used in movement analysis, but their added value over conventional biomechanical descriptors remains uncertain, particularly in pediatric cohorts. We evaluated whether implicit neural representations (INRs) provide useful encodings of pediatric gait waveforms for reconstruction and developmental modeling. The full cohort comprised 78 healthy children and adolescents, of whom 73 had waveform files conforming to the predefined data structure required for the repeated cross-validation analyses. Selected waveforms were modeled using several INR architectures, and reconstruction quality was assessed using R 2 , RMSE, and MAE. For the INR feature-ablation analyses, one architecture was assigned to each waveform class based exclusively on reconstruction performance, independently of the age targets. Side-specific, bilateral-asymmetry, signal-specific, and descriptor-family feature sets were evaluated using ten repeats of five-fold cross-validation. A separate matched benchmark in the same 73 participants compared compact Fourier-MLP-derived features with conventional biomechanical descriptors and a fold-wise raw-waveform PCA baseline. In the INR ablation analyses, combining side-specific and bilateral-asymmetry features achieved R 2 = 0.700 ± 0.017 for chronological-age regression, and Energy/AUC descriptors were the strongest individual feature family ( R 2 = 0.696 ± 0.031 ). For exploratory three-class age-group classification (4–7, 8–12, and 13–18 years), the combined INR representation reached a balanced accuracy of 0.749 ± 0.034. In the matched representation benchmark, configurations containing conventional biomechanical descriptors performed best. The biomechanics-only baseline achieved regression R 2 = 0.837 ± 0.070 and balanced accuracy = 0.798 ± 0.109 , whereas biomechanics+PCA achieved the highest mean regression performance ( R 2 = 0.842 ± 0.066 ); this difference was not statistically significant. Adding INR-derived features did not improve regression performance and significantly reduced the evaluated classification metrics. INRs should therefore be considered complementary continuous waveform representations rather than replacements for established biomechanical descriptors.