This is the authors' abstract. We don't add key points for this paper.
Computational pulse diagnosis in Traditional Chinese Medicine is constrained by small, imbalanced datasets and by evaluation protocols that leak test information into model selection. We report a leakage-controlled evaluation of three-class pulse recognition (Wiry, Slippery, Tight) on 3,699 beats from 179 healthy young volunteers, recorded at one site on one device, labelled by one practitioner, and available only as chart images. Waveforms were recovered by per-column centroid extraction, which a round-trip test showed to be self-consistent under re-rendering (real-chart median NRMSE 0.023); agreement with the device’s sampled signal could not be tested, and per-trace normalization removes pulse amplitude. Models were compared in one harness: 10 × 5 subject-grouped cross-validation, an inner validation split for early stopping, and a family of contrasts fixed before the repeated runs and tested with the Nadeau–Bengio correctedt-test and Holm adjustment. No new model is proposed. Five alternatives to or enhancements of a plain ResNet1D—Global Context attention, dynamic gating, Inception, a Patch Transformer and a Prototypical Network—brought no detectable benefit (plain ResNet1D macro-F1 0.791, subject-clustered 95% CI [0.731, 0.831]). Using the test fold for early stopping inflated every deep model by +0.015 to +0.037 macro-F1. Every deep model in the final ranking exceeded the best classical model (XGBoost on three-Gaussian features, 0.742) in mean macro-F1, but no contrast survived family-wise correction, and the effect of waveform augmentation was conditional on a weighted sampler. An exploratory Grad-CAM analysis over all held-out beats showed overlapping class patterns. Generalization to patients, other devices and practitioners is untested.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Chiu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: