Karanvir Singh
Independent researcher, Quantix Bio, Cambridge, MA, USA
Correspondence: karanvir.singh@quantixmind.com
Manuscript draft, August 2026. Not yet peer reviewed.
Background. Consumer wearables screen for atrial fibrillation (AF) from pulse interval irregularity, and published interval only detectors report excellent accuracy on databases in which AF alternates with clean sinus rhythm. Real ambulatory populations carry frequent ventricular and supraventricular ectopy, which also produces irregular interbeat intervals. We quantified, at the level of individual beats and under strict patient level separation, how much of interval based AF detection survives contact with an ectopy rich population.
Methods. All 48 records of the MIT-BIH Arrhythmia Database were used, with every beat labelled AF or non AF from the cardiologists' rhythm annotations, parsed from the aux fields of the MIT annotation format (108,534 scorable beats, 11,386 in AF across eight records). Eight features were computed on a causal window of twenty RR intervals: five irregularity measures (normalized RMSSD, the fraction of successive differences over 50 ms, Shannon entropy of the RR histogram, coefficient of variation, normalized median absolute successive difference) and three rhythm structure measures designed to separate patterned from patternless irregularity (lag one and lag two autocorrelation of successive differences, and sign alternation rate, which is near one in bigeminy and near half in AF). A logistic model was evaluated leave one record out, so the tested patient never contributes training beats, and performance was stratified by whether the trailing window contained any non sinus conducted beat.
Results. Pooled leave one record out AUC was 0.885, and at 95 percent specificity only 20 percent of AF beats were detected. Stratification located the failure precisely: on windows free of ectopy, AUC was 0.935 with 59 percent sensitivity at 95 percent specificity, while on the 50 percent of windows containing ectopy AUC fell to 0.826. Per record results were extreme: two AF records with controlled, comparatively regular ventricular response ranked below chance (within record AUC 0.27 and 0.39), while AF records with typical irregularity were well detected (AUC 0.86 to 0.89). The structure features reduced false positives on bigeminy but did not repair the pooled result.
Conclusions. Interval irregularity is a genuinely strong AF signal in ectopy free rhythm and a misleading one otherwise, and half the windows in this standard ambulatory population contained ectopy. Reported accuracies of interval only AF screening transfer to clean cohorts, not to the Holter population where screening decisions are made. Waveform information, or at minimum explicit ectopy handling, is not an optional refinement for pulse based AF screening; it is load bearing.
Keywords. atrial fibrillation, RR intervals, wearable screening, ectopy, MIT-BIH, generalization
Atrial fibrillation screening has moved to the wrist: photoplethysmography based devices flag AF from irregular pulse intervals, and the interval only approach is attractive because pulse timing is cheap while waveform morphology is fragile in ambulatory optical signals [1, 2]. The algorithmic literature supports it with strong numbers, typically obtained on the MIT-BIH Atrial Fibrillation Database, whose long records alternate between AF and largely clean sinus rhythm [3, 4].
The clinical concern with this evidence base is a population mismatch. The people screened for AF, older, with structural heart disease, are exactly the people with frequent premature beats, and a premature beat plus its compensatory pause is, to an interval sequence, an irregularity. False AF alerts driven by ectopy are a documented nuisance of consumer devices [5]. What has been less clearly quantified is the size of the effect on a database where ectopy is abundant and every beat carries both a rhythm label and a beat type label, permitting exact attribution.
The MIT-BIH Arrhythmia Database is that database: not an AF benchmark, but an ectopy rich ambulatory population in which eight records contain annotated AF episodes alongside every variety of premature beat [6]. This study evaluates interval only AF detection there, at beat level, under leave one record out separation, and uses the beat type annotations to decompose exactly where the approach holds and where it fails.
All 48 half hour records. Rhythm annotations were parsed from the aux strings of the MIT annotation format (a reader for which was written for this study and is released); each beat was labelled AF if the prevailing rhythm string at its time was "(AFIB". Eight records contain AF (201, 202, 203, 210, 217, 219, 221, 222), contributing 11,386 AF beats among 108,534 scorable beats. Atrial flutter was labelled non AF, following the annotation distinction.
For each beat, on the trailing window of twenty RR intervals derived from the reference beat annotations: normalized RMSSD; fraction of successive differences exceeding 50 ms; Shannon entropy of the eight bin RR histogram; coefficient of variation; normalized median absolute successive difference; and three structure features motivated by the clinical description of AF as irregularly irregular: lag one and lag two autocorrelation of the successive difference sequence, and the sign alternation rate of successive differences. Bigeminy, the classic confounder, produces strongly negative lag one autocorrelation and alternation near one; AF produces neither. Reference annotations rather than detected beats were used deliberately, to isolate rhythm classification from detection error; the released code can substitute the detector of our companion study.
Logistic regression, leave one record out: the model scoring a record's beats is trained on the other 47 records only. Pooled AUC, sensitivity at pooled 90 and 95 percent specificity, per record analysis, and stratification of the pooled result by whether the trailing 20 beat window contained any beat whose type is not sinus conducted (normal, bundle branch block, or escape).
Pooled AUC was 0.885; at 95 percent specificity, sensitivity was 20 percent, and at 90 percent specificity, 52 percent. For a screening context where specificity below the high nineties generates unmanageable false positive volume, these are not deployable numbers.
Table 1. Performance stratified by ectopy in the feature window.
| Stratum | Share of windows | AUC | Sensitivity at 95% specificity |
|---|---|---|---|
| No ectopic beat in window | 50.4% | 0.935 | 59% |
| At least one ectopic beat | 49.6% | 0.826 | collapsed |
On ectopy free windows the interval signal performs close to what the AF screening literature reports on clean databases. With ectopy in the window, discrimination drops eleven points, and because half of all windows in this population contain ectopy, the pooled result inherits the failure.
Table 2. Within record results for the eight AF records.
| Record | AF beats | Within record AUC | Character |
|---|---|---|---|
| 201 | 891 | 0.39 | AF with ectopy rich sinus segments |
| 202 | 940 | 0.87 | typical paroxysmal AF |
| 203 | 2,338 | 0.53 | AF plus abundant ventricular ectopy |
| 210 | 2,572 | 0.27 | AF with controlled ventricular response |
| 217 | 338 | 0.89 | AF against paced background |
| 219 | 1,782 | 0.86 | typical AF episodes |
| 221 | 2,337 | 0.48 | AF with ventricular ectopy |
| 222 | 188 | 0.66 | AF among atrial arrhythmias |
| overall | 11,386 | 0.885 pooled |
The two below chance records are the two failure modes in pure form. Record 210 is AF with a comparatively regular ventricular response: the disease is present, the interval signature is not. Records 201, 203 and 221 have the converse problem in their non AF segments: sinus rhythm salted with ectopy that out irregularizes the AF elsewhere in the database.
The lag autocorrelation and alternation features behaved as designed on bigeminy, suppressing the patterned false positives, but did not materially change pooled performance (AUC 0.885 versus 0.891 for irregularity features alone), because much real ectopy is not patterned: isolated premature beats at irregular times are, to any interval statistic, small doses of AF.
The interval only premise of consumer AF screening was directly measurable on this database, and it splits cleanly: excellent on the rhythm contexts the benchmark databases contain, broken on the contexts an ambulatory screening population actually presents. The decomposition matters more than the pooled number: the method does not degrade uniformly, it fails specifically where ectopy lives, and succeeds specifically where it does not.
For device evaluation: AF detector validation should report performance stratified by ectopy burden, and validation databases should be chosen or augmented to contain realistic ectopy; a detector validated only on AF versus clean sinus has been validated for a population that does not present for screening. For device design: either waveform evidence (P wave absence, fibrillatory activity) or an explicit ectopy recognition stage is required equipment, consistent with the design of the better performing published algorithms [4, 7]. For clinical interpretation: an interval flagged AF episode in a patient with known ectopy carries much less evidence than the same flag in a patient without.
Beat level evaluation is stricter than episode level; requiring sustained detection over minutes would improve all numbers while blurring the attribution this study is for. The database's AF sample is eight patients; the per record heterogeneity we report is exactly why interval methods should not be summarized by pooled numbers, ours included. Features were computed from reference beat positions, isolating rhythm classification from QRS detection error; end to end evaluation with our companion detector is the released code's default alternative. The MIT-BIH Atrial Fibrillation Database, where interval methods excel, was not rerun here; our claim is not that published results are wrong on their own benchmark, but that the benchmark is unrepresentative of the screening population.
Interval irregularity is a real AF signal with a specific, prevalent, and measurable failure mode. On the standard ectopy rich ambulatory database, that failure mode affects half of all analysis windows and removes the method's screening utility. The result argues for stratified validation as a reporting standard in pulse based AF detection.
The MIT-BIH Arrhythmia Database is openly available from PhysioNet. Complete source, including the rhythm annotation parser, regenerates every number with one command.
The author declares no competing interests. The work received no funding.
Figure 1. Leave one record out ROC, pooled (roc.svg).