A single-token class-discriminant codebook improved event classification AUC by 0.056 (95% CI +0.051 to +0.060) over an unsupervised codebook, though its anomaly monitor did not beat a fixed rule.
In on-device ECG analysis, a learned single-token codebook improves event classification but does not provide a superior anomaly trigger compared to hand-tuned rules.
Mean Difference: 0.056 (95% CI 0.051–0.06)
Version 2 (2026-07-03): Reproducibility archive completed- added the analysis code, result files, and frozen pre-registration scopes for the four monitor-versus-rule follow-up axes (detection latency, unseen-anomaly generalization, noise robustness, and cross-dataset transfer) that back the five-axis comparison in Section 3. The manuscript, figures, and all reported results are unchanged. Description This record accompanies a manuscript that scopes, under strict pre-registration, where the *learning* in a single-token class-discriminant codebook actually helps. The encoder reduces each window of a sensor signal to one quantization index computed on the device, and that same index exposes two read-outs for free: a decision ("what is this window?") and an anomaly monitor (the distance from the token's centroid). A natural, commercially attractive claim is that the learned monitor is a smarter trigger than a hand-tuned rule. On real public electrocardiography (MIT-BIH Arrhythmia), real noise (MIT-BIH Noise Stress Test), and a second institution's database (St Petersburg INCART), with patient-disjoint splits, five seeds, paired-bootstrap confidence intervals, and outcome bands frozen before any data were examined, that claim fails: the learned monitor does not beat a fairly-tuned fixed rule on any of five axes — discrimination, detection latency, unseen-anomaly generalization, noise robustness, or cross-dataset transfer — and is worse on noise and timing. These negatives are reported verbatim, and the "smarter trigger" claim is dropped. In the same experimental frame, however, the learning is decisively valuable in the decision read-out: a discriminant codebook classifies a held-out event from a single token at AUC +0.056 above an otherwise-identical unsupervised codebook (95% CI +0.051 to +0.060; every seed; three decision heads). That dividend is shown genuine (a shuffled-label control collapses it), generalizing (it grows to +0.102 under transfer to a second institution), robust across an eight-fold range of codebook sizes, dominant over three unsupervised baselines, calibrated on par, and personalizable to the individual (+0.090), in a 6.2 KB (eight-bit) on-device footprint at roughly 14,600 multiply-accumulates per window. The conclusion: the value of a learned single-token codebook is the integrated decision substrate — decision, monitoring, extreme compression, and privacy from one token — not a categorically superior trigger; the anomaly monitor is a comparable-to-a-rule by-product, and the learning that matters lives in the decision. Method companion: Paper 19 (doi:10.5281/zenodo.20788187). This is an application and validation of previously-filed and previously-published methods and introduces no new subject matter. Keywords: class-discriminant codebook; vector quantization; anomaly detection; rule-based detection; pre-registration; on-device machine learning; learned versus hand-crafted features; electrocardiography; arrhythmia; personalization; cross-dataset generalization; honest negatives; edge AI References 1. N. Tishby, F. C. Pereira, and W. Bialek, "The information bottleneck method," in Proc. 37th Allerton Conf. Communication, Control, and Computing, 1999, pp. 368-377. 2. R. M. Gray, "Vector quantization," IEEE ASSP Magazine, vol. 1, no. 2, pp. 4-29, 1984. 3. R. A. Fisher, "The use of multiple measurements in taxonomic problems," Annals of Eugenics, vol. 7, no. 2, pp. 179-188, 1936. 4. G. B. Moody and R. G. Mark, "The impact of the MIT-BIH Arrhythmia Database," IEEE Engineering in Medicine and Biology Magazine, vol. 20, no. 3, pp. 45-50, 2001. 5. A. L. Goldberger et al., "PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals," Circulation, vol. 101, no. 23, pp. e215-e220, 2000. 6. G. B. Moody, W. K. Muldrow, and R. G. Mark, "A noise stress test for arrhythmia detectors," Computers in Cardiology, vol. 11, pp. 381-384, 1984. 7. Association for the Advancement of Medical Instrumentation, "Testing and reporting performance results of cardiac rhythm and ST-segment measurement algorithms," ANSI/AAMI EC57, 2012. 8. V. Chandola, A. Banerjee, and V. Kumar, "Anomaly detection: A survey," ACM Computing Surveys, vol. 41, no. 3, pp. 1-58, 2009. 9. B. Schölkopf, J. C. Platt, J. Shawe-Taylor, A. J. Smola, and R. C. Williamson, "Estimating the support of a high-dimensional distribution," Neural Computation, vol. 13, no. 7, pp. 1443-1471, 2001. 10. C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger, "On calibration of modern neural networks," in Proc. 34th Int. Conf. Machine Learning (ICML), 2017, pp. 1321-1330. 11. B. A. Nosek, C. R. Ebersole, A. C. DeHaven, and D. T. Mellor, "The preregistration revolution," Proc. National Academy of Sciences, vol. 115, no. 11, pp. 2600-2606, 2018. 12. B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap. New York: Chapman patent rights are separate from the copyright license. This manuscript reports pre-registered experiments, with negatives reported verbatim, and introduces no new subject matter. Public datasets (PhysioNet MIT-BIH Arrhythmia, MIT-BIH Noise Stress Test, and St Petersburg INCART) retain their own licenses and are not redistributed in the archive. Licensing inquiries: randolphf@fieldstoneanalyticsllc.com. Randolph James Ferlic, M.D. · Kimberly Kate Ferlic · Fieldstone Analytics, LLC · DOI 10.5281/zenodo.21148683
Ferlic et al. (Fri,) conducted a other in Arrhythmia. Single-token class-discriminant codebook vs. Unsupervised codebook and hand-tuned fixed rule was evaluated on Event classification (AUC) (AUC +0.056, 95% CI +0.051 to +0.060). A single-token class-discriminant codebook improved event classification AUC by 0.056 (95% CI +0.051 to +0.060) over an unsupervised codebook, though its anomaly monitor did not beat a fixed rule.