Cardiology · Personal ECG rhythm analysis (deep-neural-network determinations)
KardiaAI
AliveCor, Inc.
The algorithm suite behind AliveCor's Kardia personal ECG has the rare thing in this catalogue: randomised evidence. In REHEARSE-AF, older patients who self-recorded twice a week with an AliveCor device were diagnosed with atrial fibrillation almost four times as often as controls. The honest caveat sits beside it — that trial validated a screening pathway with clinician overread, and when investigators tested the algorithm alone on cardiology inpatients, its accuracy fell well short of the humans reading the same traces.
Performance Metrics
Clinical Evidence
The evidence base is unusually deep for a consumer-facing algorithm, and unusually instructive about the difference between a screening pathway and a standalone classifier. The anchor is REHEARSE-AF (Halcox and colleagues, Circulation 2017), a randomised controlled trial in 1,001 ambulatory patients aged 65 and over with a CHADS-VASc score of 2 or higher. Participants randomised to twice-weekly 30-second recordings on an AliveCor device — with algorithmic AF detection plus physiologist and cardiologist overread — were diagnosed with atrial fibrillation 3.9 times as often as routine care over 12 months (19 versus 5 diagnoses; hazard ratio 3.9, 95 percent confidence interval 1.4 to 10.4). NICE's 2022 assessment of the same device family counted six comparative studies, including three randomised trials, showing improved AF detection. The 2023 extended follow-up of the REHEARSE-AF cohort (Ellins and colleagues, European Heart Journal Open) is the necessary counterweight: over a median 4.2 years, AF diagnoses were 43 in the former screening arm versus 31 in controls (hazard ratio 1.37, 95 percent confidence interval 0.86 to 2.19, not significant), with no difference in stroke or systemic embolism or in death. The detection advantage seen during the trial was not maintained once screening ceased. Standalone algorithm accuracy is more mixed, and the two key studies pull in different directions. In the iREAD study (William and colleagues, Heart Rhythm 2018), 225 near-simultaneous Kardia and 12-lead ECG pairs from 52 patients gave the automated algorithm 96.6 percent sensitivity and 94.1 percent specificity for AF — but only after excluding the 62 recordings (27.6 percent of 225) the algorithm declined to classify. In PEAK-AF (Wegner and colleagues, University Hospital Münster, European Journal of Internal Medicine 2020), blinded electrophysiologists reading Kardia traces from 99 cardiology inpatients achieved 100 percent sensitivity and 94 percent specificity, while the automated algorithm on the same recordings managed only 55 to 70 percent sensitivity and 60 to 69 percent specificity, with a negative predictive value of 100 percent. The consistent reading: the recordings are diagnostic-quality, the pathway works, and the algorithm is a conservative first filter whose unclassified and negative calls are more dependable than its positive ones. Both standalone studies predate the 2020 V2 algorithms, whose cleared update was aimed precisely at reducing unclassified results; comparable independent accuracy data for V2 were not located.
| Study | Design | n | Sensitivity | Specificity | AUC | Published |
|---|---|---|---|---|---|---|
| Halcox JPJ, Wareham K, Cardew A, et al. (REHEARSE-AF; Swansea University) | RCTRCT | 1,001 | — | — | — | Circulation 2017;136(19):1784–1794 (PMID 28851729). 1,001 patients ≥65 with CHADS-VASc ≥2 randomised to twice-weekly AliveCor iECG (with algorithmic detection plus physiologist/cardiologist overread) or routine care; AF diagnoses 19 vs 5 at 12 months, HR 3.9 (95% CI 1.4–10.4), P=0.007; cost per additional diagnosis $10,780. |
| William AD, Kanbour M, Callahan T, et al. (iREAD; Cleveland Clinic) | ProspectiveProspective | 52 | 96.6% (AF, after excluding unclassified recordings) | 94.1% (κ = 0.89 vs physician-read 12-lead ECG) | — | Heart Rhythm 2018;15(10):1561–1565 (PMID 30143448). 225 near-simultaneous Kardia and 12-lead ECG pairs from 52 AF patients, adjudicator-blinded; headline accuracy applies after exclusion of the 62 recordings (27.6% of 225) the algorithm reported as unclassified. |
| Wegner FK, Kochhäuser S, Ellermann C, et al., Dechering DG (PEAK-AF; University Hospital Münster; blinded inpatient evaluation) | ProspectiveProspective | 99 | 55–70% (algorithm alone; 100% for blinded electrophysiologists on the same traces) | 60–69% (algorithm alone; NPV 100%) | — | Eur J Intern Med 2020;73:72–75 (PMID 31806411). 296 recordings from 99 consecutive cardiology inpatients; the human overread of Kardia lead-I traces reached 100% sensitivity / 94% specificity (κ = 0.90) while the automated algorithm was markedly less accurate (κ = 0.4–0.53) — the standalone-vs-pathway gap in one cohort. Predates the 2020 V2 algorithms. |
Inside the algorithm
Editorial featureHow KardiaAI reaches a determination.
Five stages — from raw input to verdict — drawn from manufacturer documentation and the public regulatory record.
- INGEST
- NORMALISE
- DETECT
- LOCALISE
- VERDICT
Stage 01 · INGEST
Thirty seconds, one or six leads, anywhere.
KardiaAI works on short ECGs recorded outside the clinic. A user holds a KardiaMobile pad or the six-lead 6L device, and 30 seconds of trace stream to the Kardia app. The recording context is the platform's defining constraint: acquisition is unsupervised, electrode contact is variable, and the signal is a fraction of what a 12-lead machine sees — the algorithms are built for exactly that input.
Input
30s personal ECG recording
Inference
Per recording, near-instant in-app
Inside the Auris+ Listing
Five more sections complete this device’s Auris+ Listing.
Decision Ledger
ProPro unlocks a private, cross-vendor log of every case you read on this device — what the AI called, what you concluded, and a one-line reason if you overrode it. Ready for the EU AI Act's deployer logging obligations when they land in 2028.
Clinical Evidence Deep Dive
ProPro unlocks the structured clinical-evidence summary — study count, target patient population, and a tabular accuracy-metrics view drawn from peer-reviewed sources.
Peer-Reviewed Publications
ProPro unlocks the curated peer-reviewed publication list with PubMed cross-links — the citation backbone of every editorial verdict.
Post-Market & Regulatory Conditions
ProPro unlocks the post-market surveillance summary, recall record, and the conditions of approval that bound real-world use.
AI Algorithm Version History
ProPro unlocks the chronological record of algorithm version changes — what changed when, drawn from manufacturer changelogs and regulatory filings.
Regulatory Approvals
Safety Record
MAUDE reports describe KardiaMobile determinations of possible atrial fibrillation during episodes later identified as ventricular tachycardia — in one report delaying emergency care by a reported 30–40 minutes — and heart-rate miscounting from T-wave over-sensing. Individual reports, not an audited rate; no recall or FDA safety communication for the device family was located in indexed public sources as of August 2026.
No recall or FDA safety communication for KardiaAI or the Kardia device family was located in indexed public sources as of August 2026. The MAUDE adverse-event database carries reports that mark the algorithm's risk boundary: in separate reports, users experiencing ventricular tachycardia received a "possible atrial fibrillation" determination, in one case delaying emergency care by a reported 30 to 40 minutes, and another report describes heart-rate miscounting from T-wave over-sensing. These are individual reports, not an audited rate, and rhythm determinations outside the cleared set are by design reported as unclassified rather than named. MAUDE could not be queried directly from this environment, so the absence of further reports is "none found in public reporting" rather than an exhaustive audit.
Intended Use & Indications
KardiaAI is the interpretive software library behind AliveCor's Kardia products and services. A user records a 30-second ECG on a KardiaMobile (single-lead) or KardiaMobile 6L (six-lead) device paired with the Kardia smartphone app; the algorithm suite analyses the trace and returns a rhythm determination within the app. The cleared indication is narrower than that consumer context suggests: the 2019 clearance describes a software analysis library intended to assess ambulatory ECG rhythms from adult subjects, when prescribed or used under the care of a physician, and states that it is not intended for OTC-use-only devices. The November 2020 second-generation clearance (Kardia AI V2) expanded the determinations a user can receive to include sinus rhythm with premature ventricular contractions, sinus rhythm with supraventricular ectopy, and sinus rhythm with wide QRS, alongside atrial fibrillation, bradycardia and tachycardia. Recordings that fall outside the algorithms' operating parameters are reported as unclassified rather than forced into a category. The determinations are screening aids for detection: they do not constitute a diagnosis, and AliveCor's cleared labelling routes positive and unclear results to clinician review.