Karanvir Singh
Independent researcher, Quantix Bio, Cambridge, MA, USA
Correspondence: karanvir.singh@quantixmind.com
Manuscript draft, August 2026. Not yet peer reviewed.
Background. Automated classification of heartbeats into normal, supraventricular ectopic (S) and ventricular ectopic (V) classes is among the most published problems in biomedical signal processing, with reported accuracies routinely above 99 percent. Most of those figures come from evaluations that mix beats from the same patients across training and test sets. The inter patient paradigm, which trains and tests on disjoint patients, is the evaluation that predicts performance on a new patient, and it tells a much less flattering story that the field periodically rediscovers. We provide a fully transparent, minimal feature baseline under the strict paradigm, and use it to make the gap and its structure explicit.
Methods. The standard de Chazal patient division of the MIT-BIH Arrhythmia Database was used: 22 records for training (DS1, 50,511 beats) and 22 disjoint records for testing (DS2, 49,229 beats), AAMI class mapping, paced records excluded. Nine features per beat: four RR context features, four morphology features from the bandpassed signal at the annotated fiducial (width, normalized energy, temporal asymmetry, relative amplitude) and the local rate. One versus rest regularized logistic regression with balancing by minority resampling; decision thresholds chosen on DS1 only, by F1 maximization, and frozen before DS2 was touched. All code is pure numpy and regenerates every number with one command.
Results. Ventricular ectopy was detectable across patients: sensitivity 92.4 percent at positive predictivity 56.9 percent (AUC 0.986), with 3,029 of 3,218 test V beats found at a cost of 1,006 false ventricular calls among 44,177 normal beats. Supraventricular ectopy was not: sensitivity 9.2 percent at positive predictivity 22.1 percent, despite a ranking AUC of 0.836, because the score distribution shifts between patient sets and the DS1 chosen threshold lands in the wrong place on DS2 patients. Per record ventricular sensitivity ranged from 0.55 to 1.00 across the ten test records with at least ten V beats.
Conclusions. Under honest evaluation, a nine feature linear model captures most ventricular ectopy, confirming that V class detection is substantially an RR timing and gross morphology problem, while supraventricular detection fails in a specific and instructive way: the class is rankable but not thresholdable across patients, since what is early for one heart is normal timing for another. Inter patient normalization of rhythm context, not classifier capacity, is the binding constraint, which is consistent with three decades of the mixed paradigm hiding exactly this difficulty. We propose this baseline, released in two hundred lines, as the reference point new methods should be required to beat under the same rules.
Keywords. ECG, arrhythmia, beat classification, inter patient evaluation, AAMI, MIT-BIH
Beat classification papers report some of the highest accuracies in medical machine learning, and some of the least transferable. The central methodological fault line was identified by de Chazal and colleagues in 2004 [1]: if beats from the same recording appear in both training and test sets, the classifier learns the patient, not the physiology, and reported performance bears no relation to deployment on a new patient. They proposed a fixed division of the MIT-BIH Arrhythmia Database [2] into disjoint patient sets, DS1 and DS2, and the field's honest results have been reported on it since [1, 3, 4], typically one to two orders of magnitude more error than the mixed paradigm advertises, a discrepancy documented repeatedly [4, 5].
This study contributes a deliberately minimal baseline under the strict paradigm: nine interpretable features, a linear classifier, thresholds fixed on training patients, everything inspectable. The purpose is threefold: to measure how much of the problem the simplest defensible approach solves; to characterize the structure of what remains, particularly for the supraventricular class, whose inter patient difficulty is notorious [1, 3, 4]; and to provide a reproducible floor that more complex methods, including the deep models now dominant in this literature, should be asked to beat under identical rules.
MIT-BIH Arrhythmia Database, the de Chazal division: DS1 (records 101, 106, 108, 109, 112, 114, 115, 116, 118, 119, 122, 124, 201, 203, 205, 207, 208, 209, 215, 220, 223, 230) for training, DS2 (100, 103, 105, 111, 113, 117, 121, 123, 200, 202, 210, 212, 213, 214, 219, 221, 222, 228, 231, 232, 233, 234) for testing; the four paced records are excluded per AAMI convention. Beat types map to AAMI classes: N (normal, bundle branch block, escape), S (atrial and junctional premature), V (ventricular premature and escape); fusion and unknown classes are excluded from scoring, following common practice. This yields 50,511 training and 49,229 test beats (test composition: 44,177 N, 1,834 S, 3,218 V).
At each annotated fiducial, from the first channel after zero phase 5 to 15 Hz bandpassing: preceding and following RR interval, their ratios to the local six beat mean, the local mean itself, QRS width (duration above one quarter of peak amplitude), normalized energy, temporal asymmetry of absolute amplitude about the fiducial, and peak amplitude normalized by the record's median beat amplitude. The features are the classical clinical discriminators: V beats are wide, early, high amplitude and followed by a pause; S beats are narrow and early, discriminable chiefly by timing.
One versus rest L2 regularized logistic regression, minority class balanced by replication. The decision thresholds for calling S or V were selected by F1 maximization on DS1 predictions only and frozen; beats exceeding both thresholds take the higher scoring class. Nothing about DS2 influenced any modelling choice; DS2 was evaluated once.
Table 1. Confusion matrix, DS2 (rows: truth, columns: prediction).
| N | S | V | |
|---|---|---|---|
| N (44,177) | 42,702 | 469 | 1,006 |
| S (1,834) | 419 | 169 | 1,246 |
| V (3,218) | 116 | 127 | 2,975 |
Table 2. Per class performance on DS2, with ranking AUC.
| Class | Sensitivity | Positive predictivity | AUC |
|---|---|---|---|
| V | 92.4% | 56.9% | 0.986 |
| S | 9.2% | 22.1% | 0.836 |
Nine features and a linear boundary recover 92.4 percent of ventricular beats on unseen patients. The cost, 1,006 false V calls among 44,177 normals (2.3 percent), concentrates in two test records with unusual conduction. The AUC of 0.986 indicates the feature set ranks V beats nearly perfectly; the residual error is threshold placement, a per patient scale problem rather than a representational one. Per record V sensitivity exceeded 0.90 in seven of the ten test records with at least ten V beats, with the worst at 0.55.
S class sensitivity was 9.2 percent, and the confusion matrix shows where the beats went: 1,246 of 1,834 S beats were called V, and 419 were called N. Yet the S score's ranking AUC was 0.836, far above chance. The combination is diagnostic: the model ranks S beats correctly within a patient, but the absolute score scale shifts across patients, so any fixed threshold selected on DS1 patients is wrong for DS2 patients. Prematurity is relative to a heart's own rhythm; a 20 percent early beat is conspicuous in a steady heart and invisible in a variable one. Our per record normalization of amplitude addresses the morphology scale but nothing normalizes the rhythm context, and the S class lives entirely in rhythm context.
Under the same division, the original two lead, many feature system of de Chazal reported V sensitivity 77.7 at positive predictivity 81.9 and S sensitivity 75.9 at positive predictivity 38.5 [1]; subsequent inter patient work spans V sensitivity roughly 80 to 95 with widely varying predictivity, and S sensitivity from single digits to roughly 80 with predictivity rarely above 40 [3, 4, 5]. Our baseline sits at the high sensitivity, low predictivity corner for V, and at the low corner for S, as expected for one lead and nine features. The point of the baseline is not superiority; it is that the gap between it and any proposed method is now exactly measurable under fixed rules, with code.
A classifier of this simplicity scores above 99 percent accuracy under the mixed paradigm, as essentially everything does, because the paradigm rewards memorizing each patient's beat morphology. The inter patient numbers reported here are smaller and true. The literature's recurring inflation is not a scandal of any one paper but a structural consequence of benchmark design, and the remedy is procedural: the patient division and the frozen threshold protocol used here cost nothing to adopt and change reported numbers by an order of magnitude.
The most useful finding is negative and precise: for supraventricular detection, ranking survives patient transfer (AUC 0.836) and thresholding does not (sensitivity 9 percent). This localizes the research problem. It is not feature capacity, which more expressive models address, but inter patient calibration of rhythm context: what is needed is either per patient adaptation of the decision threshold from an initial rhythm sample, or features that are self normalizing in distribution across hearts. This diagnosis is consistent with the fact that the published systems that do well on S under the strict paradigm lean on longer range rhythm context and per record normalization [1, 3], and it connects to our companion work on calibration: the S classifier's scores are discriminative but miscalibrated across patients, and the failure is exactly a calibration failure.
One lead, though the second lead is read by our tooling and fusing it is the obvious extension. Features are computed at annotated fiducials, separating classification from detection error; the end to end pipeline with our companion QRS detector is a released configuration. The F and Q classes are excluded from scoring, as in most of this literature; fusion beats are genuinely ambiguous and their exclusion flatters all methods equally. Class balanced resampling is the simplest imbalance treatment; weighted Newton objectives are equivalent here.
Under honest rules, simple timing and width features nearly solve ventricular ectopy detection and leave supraventricular detection unsolved for an identifiable reason. New beat classification methods should be evaluated under the inter patient paradigm with frozen thresholds, and should report what they add over baselines of this kind, which this paper supplies in reusable form.
The MIT-BIH Arrhythmia Database is openly available from PhysioNet. Complete source regenerates all numbers with one command.
The author declares no competing interests. The work received no funding.
Figure 1. One versus rest ROC curves for the V and S classes on DS2 (roc_interpatient.svg).