Comparison of raw-signal, feature-based and hybrid representations for machine-learning classification of retinal...

Pattern electroretinography (PERG) is a non-invasive test of macular and retinal ganglion cell function, but its low amplitude and inter-subject variability limit its diagnostic use. This study compares three representations of the transient PERG waveform for the detection of six retinal and optic nerve conditions - Birdshot chorioretinopathy, cone-rod dystrophy, hereditary optic atrophy, macular dystrophy, retinitis pigmentosa and Stargardt disease - against healthy controls, using the publicly available PERG-IOBA dataset. A total of 672 recordings from 336 examinations of 304 participants were preprocessed in six stages, and three representations were evaluated under one identical protocol: the raw 255-point amplitude-preserving waveform; 60 engineered clinical, spectral, wavelet and complexity descriptors; and a hybrid representation of 343 variables combining both, with multi-wavelet features from db4, sym2 and coif3. Performance was estimated by 10 × 5-fold participant-grouped cross-validation, in which both eyes and all repeat visits of an individual were kept in the same partition, with nested hyperparameter selection and all data-derived preprocessing fitted inside each training fold. Under the pre-specified primary classifier, the raw waveform gave the highest pooled out-of-fold AUC in four of the six classes (0.906 in cone-rod dystrophy, 0.842 in retinitis pigmentosa, 0.755 in macular dystrophy, 0.673 in hereditary optic atrophy) and the engineered descriptors in the remaining two (0.803 in Stargardt disease, 0.657 in Birdshot chorioretinopathy). The hybrid representation was highest in none, falling below the raw waveform in five of six classes by 0.002 to 0.018, and no non-linear classifier showed a hybrid advantage above 0.006 in mean AUC. A sensitivity analysis identified what does matter: removing the absolute amplitude scale by per-signal z-scoring cost the raw representation 0.055 AUC on average but the engineered descriptors only 0.023, and reversed the ordering of the three representations. Engineered descriptors did not improve on the waveform, but they made the problem tractable for linear models, which gained 0.128 in mean AUC from the reduction in dimensionality. These internally validated findings indicate that absolute PERG amplitude should be preserved rather than normalised away; participant-independent, multi-centre validation is required before clinical application.