Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2025 Mar;31(3):925-931.
doi: 10.1038/s41591-025-03516-x. Epub 2025 Feb 10.

Artificial intelligence for direct-to-physician reporting of ambulatory electrocardiography

L S Johnson  1   2 , P Zadrozniak  3 , G Jasina  3 , A Grotek-Cuprjak  3 , J G Andrade  4 , E Svennberg  5   6 , S Z Diederichsen  7 , W F McIntyre  8   9 , S Stavrakis  10 , J Benezet-Mazuecos  11 , P Krisai  12 , Z Iakobishvili  13   14 , A Laish-Farkash  13 , S Bhavnani  15 , E Ljungström  16 , J Bacevicius  17 , N L van Vreeswijk  18 , M Rienstra  18 , R Spittler  19 , J A Marx  19 , A Oraii  8 , A Miracle Blanco  11 , A Lozano  11 , I Mustafina  10   20 , S Zafeiropoulos  21   22 , R Bennett  4 , J Bisson  23 , D Linz  24   25 , Y Kogan  13 , E Glazer  13 , G Marincheva  13 , M Rahkovich  13 , E Shaked  13 , M H Ruwald  26 , K Haugan  27 , J Węcławski  3 , G Radoslovich  28 , S Jamal  28   29 , A Brandes  30   31 , P T Matusik  32   33 , M Manninger  34 , P B Meyre  12 , S Blum  12 , A Persson  35   36 , A Måneheim  35   36 , P Hammarlund  37 , A Fedorowski  35   5   38 , T Wodaje  5   38 , C Lewinter  5   38   39 , V Juknevicius  17 , R Jakaite  17 , C Shen  15 , T Glotzer  28   29 , P Platonov  16   40 , G Engström  35 , A P Benz  8   19 , J S Healey  8   9
Affiliations

Artificial intelligence for direct-to-physician reporting of ambulatory electrocardiography

L S Johnson et al. Nat Med. 2025 Mar.

Abstract

Developments in ambulatory electrocardiogram (ECG) technology have led to vast amounts of ECG data that currently need to be interpreted by human technicians. Here we tested an artificial intelligence (AI) algorithm for direct-to-physician reporting of ambulatory ECGs. Beat-by-beat annotation of 14,606 individual ambulatory ECG recordings (mean duration = 14 ± 10 days) was performed by certified ECG technicians (n = 167) and an ensemble AI model, called DeepRhythmAI. To compare the performance of the AI model and the technicians, a random sample of 5,235 rhythm events identified by the AI model or by technicians, of which 2,236 events were identified as critical arrhythmias, was selected for annotation by one of 17 cardiologist consensus panels. The mean sensitivity of the AI model for the identification of critical arrhythmias was 98.6% (95% confidence interval (CI) = 97.7-99.4), as compared to 80.3% (95% CI = 77.3-83.3%) for the technicians. False-negative findings were observed in 3.2/1,000 patients for the AI model versus 44.3/1,000 patients for the technicians. Accordingly, the relative risk of a missed diagnosis was 14.1 (95% CI = 10.4-19.0) times higher for the technicians. However, a higher false-positive event rate was observed for the AI model (12 (interquartile range (IQR) = 6-74)/1,000 patient days) as compared to the technicians (5 (IQR = 2-153)/1,000 patient days). We conclude that the DeepRhythmAI model has excellent negative predictive value for critical arrhythmias, substantially reducing false-negative findings, but at a modest cost of increased false-positive findings. AI-only analysis to facilitate direct-to-physician reporting could potentially reduce costs and improve access to care and outcomes in patients who need ambulatory ECG monitoring.

PubMed Disclaimer

Conflict of interest statement

Competing interests: L.S.J. receives consulting fees from Medicalgorithmics. P.Z., G.J. and A.G.C. are Medicalgorithmics employees. J.W. is a former Medicalgorithmics employee. J.S.H. has research grants and speaking fees from BMS/Pfizer, Boehringer Ingelheim, Boston Scientific, Novartis, Medtronic and Servier. A.P.B. has received speaker fees from Bristol Myers Squibb and AstraZeneca, and participates in an educational program supported by Boston Scientific (Fellowship Herzrhythmus). E. Svennberg is supported by the Stockholm County Council (clinical researcher appointment), the Swedish Research Council (DNR 2022-01466), the Swedish Heart and Lung Foundation and CIMED and has received institutional remunerations for lectures from Abbott, AstraZeneca, Bristol-Myers Squibb-Pfizer and Johnson & Johnson. P.H. reports lecture fees from Pfizer and Boehringer Ingelheim. M.H.R. reports speaker fees from Cardiofocus and Boston Scientific. M. Rienstra reports consultancy fees from Bayer (OCEANICAF national PI) and InCarda Therapeutics (RESTORE-SR national PI) to the institution. P.K. reports speaker fees from BMS/Pfizer and research grants from the Swiss National Science Foundation, Swiss Heart Foundation, Foundation for Cardiovascular Research Basel and Machaon Foundation. S.Z.D. is a part-time employee of Vital Beats and has received consultancy fees from Cortrium, Acesion Pharma and Bristol-Myers Squibb-Pfizer, lecture fees from Bayer and Bristol-Myers Squibb-Pfizer and travel grants from Abbott and Boston Scientific. M.M. received speaker fees/honoraria from Abbott, Bayer Healthcare, Biosense Webster, Biotronik, Amomed, AOP Orphan, Boston Scientific, Daiichi Sankyo and BMS/Pfizer and research grants from Biosense Webster and Abbott. P.T.M. has received speech honoraria from Boehringer Ingelheim and participated in educational activities, which were supported by cardiovascular implantable electronic device manufacturers. J. Bacevicius holds a patent for the TeltoHeart technology, consults Teltonika Telemedic and has received travel grants from Abbot and Biosense Webster. V.J. has received speaker fees from BMS/Pfizer, Bayer, Boehringer and Servier. A.B. has received research grants from Theravance, the Zealand Region, the Canadian Institutes of Health Research, the European Union Interreg 5A Programme, the Danish Heart Foundation, the Independent Research Fund Denmark and a lecture honorarium from Bristol Myers Squibb outside the submitted work. T.G. serves on an advisory board for Medtronic and Boston Scientific and has received speaking honoraria from Medtronic, Boston Scientific and Abbott.

Figures

Fig. 1
Fig. 1. False-negative critical arrhythmias per 1,000 patients by AI and technician analysis.
Error bars represent 95% CIs derived using bootstrapping. AVB, AV block.
Fig. 2
Fig. 2. True-positive critical arrhythmias per 1,000 patients by AI and technician analysis.
Error bars represent 95% CIs derived using bootstrapping.
Fig. 3
Fig. 3. False-positive critical arrhythmias per 1,000 patients by AI and technician analysis.
Error bars represent 95% CIs derived using bootstrapping.
Fig. 4
Fig. 4. Diagnoses of patients with critical arrhythmias by DeepRhythmAI and ECG technicians.
Sankey diagram showing arrhythmic event durations for critical arrhythmias as detected by each of the two methods. Cardiologist panel annotations are used to classify DeepRhythmAI and ECG technician annotations into TP, FP or FN. For FP and FN detections, we also report whether these were annotated by the cardiologist panels as another critical arrhythmia class or as a noncritical arrhythmia/noise or NSR. TP, true positives; FP, false positives; FN, false negatives; NSR, normal sinus rhythm.
Extended Data Fig. 1
Extended Data Fig. 1. Age and sex distribution.
Age and sex distribution of the patient sample included in the analyses.
Extended Data Fig. 2
Extended Data Fig. 2. False-negative findings in males and females.
Error bars denote 95% confidence intervals and were derived using bootstrapping.
Extended Data Fig. 3
Extended Data Fig. 3. False-negative findings, excluding events reported as other critical arrhythmias.
Error bars denote 95% confidence intervals and were derived using bootstrapping.
Extended Data Fig. 4
Extended Data Fig. 4. False-positive findings of critical arrhythmias.
Error bars denote 95% confidence intervals and were derived using bootstrapping.
Extended Data Fig. 5
Extended Data Fig. 5. Schematic overview of the DeepRhythmAI model.
The raw ECG signal (in timestamp + mV format) is pre-processed and fed to a single CNN classifier model that identifies the QRS complexes and segments of noisy (non-diagnostic) signals in the raw ECG data. The network components downstream to this module are fed both raw signal and the QRS/noise module output. This combined signal and QRS/noise data are processed by ensemble of a total of 7 models with both wide context (HR trend and morphology of beats) and narrow context (signal details). The wide context module is an ensemble of three custom deep neural network models with both CNN and transformer layers. The narrow context module is an ensemble of three transformer models all based on Vision Transformer ideas but with custom adaptations to 1D multichannel ECG signal. The output from these models is then combined with a wide context asystole filter that has the same architecture as wide context models but with hyperparameters tuned for asystole detection. The asystole filter overrides and replaces the other probabilities when asystoles are detected; otherwise, the output probabilities are averaged.
Extended Data Fig. 6
Extended Data Fig. 6. Description of the strip selection process.
VT, ventricular tachycardia; AF, atrial fibrillation; SVT, supraventricular tachycardia; AIVR, accelerated idioventricular rhythm; IVR, idioventricular rhythm; EAR, ectopic atrial rhythm.
See this image and copyright information in PMC

References

    1. Healey, J. S. et al. Apixaban for stroke prevention in subclinical atrial fibrillation. N. Engl. J. Med.390, 107–117 (2024). - PubMed
    1. Svennberg, E. et al. Clinical outcomes in systematic screening for atrial fibrillation (STROKESTOP): a multicentre, parallel group, unmasked, randomised controlled trial. Lancet398, 1498–1506 (2021). - PubMed
    1. Diederichsen, S. Z. et al. Comprehensive evaluation of rhythm monitoring strategies in screening for atrial fibrillation: insights from patients at risk monitored long term with an implantable loop recorder. Circulation141, 1510–1522 (2020). - PubMed
    1. Steinhubl, S. R. et al. Effect of a home-based wearable continuous ECG monitoring patch on detection of undiagnosed atrial fibrillation: the mSToPS randomized clinical trial. JAMA320, 146–155 (2018). - PMC - PubMed
    1. Dziubinski, M. et al. Diagnostic yield is dependent on monitoring duration. Insights from a full-disclosure mobile cardiac telemetry system. Kardiol. Pol.80, 49–55 (2022). - PubMed

LinkOut - more resources