A method of using a trained model

Abstract

A method of using a trained model to detect cardiovascular disease, the model trained using electrocardiogram signals is described. The method comprises receiving a set of electrocardiogram signals. The method further comprises applying the trained model to the received set of electrocardiogram signals. The method further comprises outputting a likelihood that the received set of electrocardiogram signals are indicative of cardiovascular disease.

Description

[0001] A method of using a trained model

[0002] Field

[0003] The present invention relates to methods of using trained models to detect the presence of a disease or to predict a disease occurring.

[0004] Background

[0005] Artificial intelligence-enabled electrocardiography (AI-ECG) can be used to predict risk of future disease and mortality but has not yet been adopted into clinical practice. Existing model predictions lack actionability at an individual patient level, explainability and biological plausibility. We sought to address these limitations of previous AI-ECG approaches by developing the AI-ECG risk estimator (AIRE) platform.

[0006] References

[0007] 1. Sau A, Ibrahim S, Ahmed A, Handa B, Kramer DB, Waks JW, et al. Artificial intelligence-enabled electrocardiogram to distinguish cavotricuspid isthmus dependence from other atrial tachycardia mechanisms. European Heart Journal -Digital Health. 2022:ztac042.

[0008] 2. Sau A, Ibrahim S, Kramer DB, Waks JW, Qureshi N, Koa-Wing M, et al. Artificial intelligence-enabled electrocardiogram to distinguish atrioventricular re-entrant tachycardia from atrioventricular nodal re-entrant tachycardia. Cardiovascular Digital Health Journal. 2023;4(2):60-7.

[0009] 3. Attia ZI, Kapa S, Lopez-Jimenez F, McKie PM, Ladewig DJ, Satam G, et al. Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med. 2019;25(l):70-4.

[0010] 4. Sau A, Ribeiro AH, McGurk KA, Pastika L, Bajaj N, Ardissino M, et al. Neural network-derived electrocardiographic features have prognostic significance and important phenotypic and genotypic associations. medRxiv.

[0011] 2023:2023.06.15.23291428.

[0012] 5. Raghunath S, Ulloa Cerna AE, Jing L, vanMaanen DP, Stough J, Hartzel DN, et al. Prediction of mortality from 12-lead electrocardiogram voltage data using a deep neural network. Nat Med. 2020;26(6):886-91.

[0013] 6. Sun W, Kalmady SV, Sepehrvand N, Salimi A, Nademi Y, Bainey K, et al.

[0014] Towards artificial intelligence-based learning health system for population-level mortality prediction using electrocardiograms, npj Digit Med. 2023;6(l):21.7. Hughes JW, Tooley J, Torres Soto J, Ostropolets A, Poterucha T, Christensen MK, et al. A deep learning-based electrocardiogram risk score for long term cardiovascular death and disease, npj Digit Med. 2023)6(1): 169.

[0015] 8. Cardoso CS, Sabino EC, Oliveira CDL, de Oliveira LC, Ferreira AM, Cunha-Neto E, et al. Longitudinal study of patients with chronic Chagas cardiomyopathy in Brazil (SaMi-Trop project): a cohort profile. BMJ Open. 2016;6(5):e011181.

[0016] 9. Lima EM, Ribeiro AH, Paixao GMM, Ribeiro MH, Pinto-Filho MM, Gomes PR, et al. Deep neural network-estimated electrocardiographic age as a mortality predictor. Nature Communications. 2021;12(l):5117.

[0017] 10. Schmidt MI, Duncan BB, Mill JG, Lotufo PA, Chor D, Barreto SM, et al. Cohort Profile: Longitudinal Study of Adult Health (ELSA-Brasil). Int J Epidemiol.

[0018] 2015;44(l):68-75.

[0019] 11. Sudlow C, Gallacher J, Allen N, Beral V, Burton P, Danesh J, et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 2015;12(3):el001779.

[0020] 12. Ribeiro AH, Ribeiro MH, Paixao GMM, Oliveira DM, Gomes PR, Canazart JA, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat Commun. 2020)11(1): 1760.

[0021] 13. Gensheimer MF, Narasimhan B. A scalable discrete-time survival model for neural networks. Peer], 2019;7:e6257.

[0022] 14. Suchard MA, Schuemie MJ, Krumholz HM, You SC, Chen R, Pratt N, et al. Comprehensive comparative effectiveness and safety of first-line antihypertensive drug classes: a systematic, multinational, large-scale analysis. Lancet.

[0023] 2019 394(10211): 1816-26.

[0024] 15. Stabenau HF, Waks JW. BRAVEHEART: Open-source software for automated electrocardiographic and vectorcardiographic analysis. Computer Methods and Programs in Biomedicine. 2023)242:107798.

[0025] 16. Kubota Y. tf-keras-vis 2020 [Available from: https: / / github.com / keisen / tf-keras-vis.

[0026] 17. Meyer HV, Dawes TJW, Serrani M, Bai W, Tokarczuk P, Cai J, et al. Genetic and functional insights into the fractal structure of the heart. Nature.

[0027] 2020)584(7822): 589-94.

[0028] 18. Jonsson BA, Bjornsdottir G, Thorgeirsson TE, Ellingsen LM, Walters GB, Gudbjartsson DF, et al. Brain age prediction using deep learning uncovers associated sequence variants. Nature Communications. 2019;10(l):5409.

[0029] 19. Agarwal SK, Chambless LE, Ballantyne CM, Astor B, Bertoni AG, Chang PP, et al. Prediction of incident heart failure in general practice: the Atherosclerosis Risk in Communities (ARIC) Study. Circ Heart Fail. 2012)5(4) :422-9.20. van der Harst P, van Setten J, Verweij N, Vogler G, Franke L, Maurano MT, et al. 52 Genetic Loci Influencing Myocardial Mass. J Am Coll Cardiol. 2016)68(13): 1435-48.

[0030] 21. Young WJ, Haessler J, Benjamins JW, Repetto L, Yao J, Isaacs A, et al. Genetic architecture of spatial electrical biomarkers for cardiac arrhythmia and relationship with cardiovascular disease. Nat Commun. 2023;14(l):1411.

[0031] 22. Jespersen T, Grunnet M, Olesen S-P. The KCNQ1 potassium channel: from gene to physiological function. Physiology. 2005;20(6):408-16.

[0032] 23. Verweij N, Benjamins JW, Morley MP, van de Vegte YJ, Teumer A, Trenkwalder T, et al. The Genetic Makeup of the Electrocardiogram. Cell Syst. 2020;ll(3):229-38. e5.

[0033] 24. Zhu X, Zhu L, Wang H, Cooper RS, Chakravarti A. Genome-wide pleiotropy analysis identifies novel blood pressure variants and improves its polygenic risk scores. Genet Epidemiol. 2022)46(2): 105-21.

[0034] 25. Sakaue S, Kanai M, Tanigawa Y, Karjalainen J, Kurki M, Koshiba S, et al. A cross-population atlas of genetic associations for 220 human phenotypes. Nat Genet.

[0035] 2021)53(10): 1415-24.

[0036] 26. Christakoudi S, Evangelou E, Riboli E, Tsilidis KK. GWAS of allometric bodyshape indices in UK Biobank identifies loci suggesting associations with morphogenesis, organogenesis, adrenal cell renewal and cancer. Sci Rep.

[0037] 2021)11(1): 10688.

[0038] 27. Libiseller-Egger J, Phelan JE, Attia ZI, Benavente ED, Campino S, Friedman PA, et al. Deep learning-derived cardiovascular age shares a genetic basis with other cardiac phenotypes. Scientific Reports. 2022;12(l):22625.

[0039] 28. Graff M, Scott RA, Justice AE, Young KL, Feitosa MF, Barata L, et al. Genomewide physical activity interactions in adiposity - A meta-analysis of 200,452 adults. PLoS Genet. 2017;13(4):el006528.

[0040] 29. Altman DG, Bland JM. Diagnostic tests 2: Predictive values. BMJ.

[0041] 1994 309(6947) : 102.

[0042] 30. Khurshid S, Friedman S, Reeder C, Di Achille P, Diamant N, Singh P, et al. ECG-Based Deep Learning and Clinical Risk Factors to Predict Atrial Fibrillation. Circulation.

[0043] 2022)145(2): 122-33.

[0044] 31. Visseren FLJ, Mach F, Smulders YM, Carballo D, Koskinas KC, Back M, et al. 2021 ESC Guidelines on cardiovascular disease prevention in clinical practice. Eur Heart J. 2021;42(34):3227-337.

[0045] 32. Bachtiger P, Kelshiker MA, Petri CF, Gandhi M, Shah M, Kamalati T, et al.

[0046] Survival and health economic outcomes in heart failure diagnosed at hospitaladmission versus community settings: a propensity-matched analysis. BMJ Health Care Inform. 2023;30(l).

[0047] 33. Bachtiger P, Petri CF, Scott FE, Ri Park S, Kelshiker MA, Sahemey HK, et al. Point-of-care screening for heart failure with reduced ejection fraction using artificial intelligence during ECG-enabled stethoscope examination in London, UK: a prospective, observational, multicentre study. The Lancet Digital Health.

[0048] 2022:52589750021002569.

[0049] 34. Istolahti T, Lyytikainen LP, Huhtala H, Nieminen T, Kahbnen M, Lehtimaki T, et al. The prognostic significance of T-wave inversion according to ECG lead group during long-term follow-up in the general population. Ann Noninvasive Electrocardiol.

[0050] 2021;26(l):el2799.

[0051] 35. Schroder LC, Holkeri A, Eranti A, Haukilahti MAE, Kerola T, Kentta TV, et al. Poor R-wave progression as a predictor of sudden cardiac death in the general population and subjects with coronary artery disease. Heart Rhythm. 2022;19(6):952-9.

[0052] 36. Imanishi R, Seto S, Ichimaru S, Nakashima E, Yano K, Akahoshi M. Prognostic Significance of Incident Complete Left Bundle Branch Block Observed Over a 40-Year Period. The American Journal of Cardiology. 2006;98(5):644-8.

[0053] 37. Ghassemi M, Oakden-Rayner L, Beam AL. The false hope of current approaches to explainable artificial intelligence in health care. Lancet Digit Health.

[0054] 2021;3(ll):e745-e50.

[0055] Supplementary References

[0056] 1. Lima EM, Ribeiro AH, Paixao GMM, Ribeiro MH, Pinto-Filho MM, Gomes PR, et al. Deep neural network-estimated electrocardiographic age as a mortality predictor. Nature Communications. 2021;12(l):5117.

[0057] 2. Schmidt MI, Duncan BB, Mill JG, Lotufo PA, Chor D, Barreto SM, et al. Cohort Profile: Longitudinal Study of Adult Health (ELSA-Brasil). Int J Epidemiol.

[0058] 2015;44(l):68-75.

[0059] 3. Cardoso CS, Sabino EC, Oliveira CDL, de Oliveira LC, Ferreira AM, Cunha-Neto E, et al. Longitudinal study of patients with chronic Chagas cardiomyopathy in Brazil (SaMi-Trop project): a cohort profile. BMJ Open. 2016;6(5):e011181.

[0060] 4. Cardoso CS, Ribeiro ALP, Oliveira CDL, Oliveira LC, Ferreira AM, Bierrenbach AL, et al. Beneficial effects of benznidazole in Chagas disease: NIH SaMi-Trop cohort study. PLoS Negl Trap Dis. 2018;12(ll):e0006814.5. Sudlow C, Gallacher J, Allen N, Beral V, Burton P, Danesh J, et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 2015;12(3):el001779.

[0061] 6. Fry A, Littlejohns TJ, Sudlow C, Doherty N, Adamska L, Sprosen T, et al.

[0062] Comparison of Sociodemographic and Health-Related Characteristics of UK Biobank Participants With Those of the General Population. Am J Epidemiol.

[0063] 2017)186(9): 1026-34.

[0064] 7. Bai W, Suzuki H, Huang J, Francis C, Wang S, Tarroni G, et al. A populationbased phenome-wide association study of cardiac and aortic structure and function. Nature Medicine. 2020)26(10): 1654-62.

[0065] 8. Gensheimer MF, Narasimhan B. A scalable discrete-time survival model for neural networks. Peer], 2019;7:e6257.

[0066] 9. Ribeiro AH, Ribeiro MH, Paixao GMM, Oliveira DM, Gomes PR, Canazart JA, et al. Automatic diagnosis of the 12-lead ECG using a deep neural network. Nat Commun. 2020)11(1): 1760.

[0067] 10. Chollet F. Keras 2015 [Available from: https: / / keras.io.

[0068] 11. Martin A, Ashish A, Paul B, Eugene B, Zhifeng C, Craig C, et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. 2015.

[0069] 12. van de Leur RR, Bos MN, Taha K, Sammani A, Yeung MW, van Duijvenboden S, et al. Improving explainability of deep neural network-based electrocardiogram interpretation using variational auto-encoders European Heart Journal - Digital Health.

[0070] 2022;3(3):390-404.

[0071] 13. Stensrud MJ, Hernan MA. Why Test for Proportional Hazards? JAMA.

[0072] 2020;323(14): 1401-2.

[0073] 14. Ribeiro AH. Predicting age from the electrocardiogram and its usage as a mortality predictor 2021 [

[0074] 15. Meyer HV, Dawes TJW, Serrani M, Bai W, Tokarczuk P, Cai J, et al. Genetic and functional insights into the fractal structure of the heart. Nature.

[0075] 2020)584(7822): 589-94.

[0076] 16. Yang J, Lee SH, Goddard ME, Visscher PM. GCTA: A Tool for Genome-wide Complex Trait Analysis. The American Journal of Human Genetics. 2011;88(l):76-82.Summary

[0077] According to a first aspect of the invention, there is provided a method of using a trained model to detect cardiovascular disease, the model trained using electrocardiogram signals, the method comprising: receiving a set of electrocardiogram signals, applying the trained model to the received set of electrocardiogram signals, and outputting a likelihood that the received set of electrocardiogram signals are indicative of cardiovascular disease.

[0078] According to a second aspect of the invention, there is provided a method of using a trained model to predict cardiovascular disease occurring, the model trained using electrocardiogram signals, the method comprising: receiving a set of electrocardiogram signals, applying the trained model to the received set of electrocardiogram signals, and outputting a likelihood that the received electrocardiogram signals are indicative of cardiovascular disease occurring.

[0079] The cardiovascular disease of the first or second aspect may be any one or more from: structural cardiovascular disease; hypertension; valvular heart disease; pulmonary hypertension; cardiomyopathy; arrhythmia; and stroke.

[0080] According to a third aspect of the invention, there is provided a method of using a trained model to detect a disease, the model trained using electrocardiogram signals, the method comprising: receiving a set of electrocardiogram signals, applying the trained model to the received set of electrocardiogram signals, and outputting a likelihood that the received set of electrocardiogram signals are indicative of the disease.

[0081] According to a fourth aspect of the invention, there is provided a method of using a trained model to predict a disease occurring, the model trained using electrocardiogram signals, the method comprising: receiving a set of electrocardiogram signals, applying the trained model to the received set of electrocardiogram signals, and outputting a likelihood that the received set of electrocardiogram signals are indicative of the disease occurring.

[0082] That is, the signals are indicative of the disease occurring in the future, for example, for a particular subject.

[0083] The disease may be diabetes, kidney disease, or liver disease.According to a fifth aspect of the invention there is provided a method of training a model to detect disease, the method comprising: receiving a plurality of sets of electrocardiogram signals, at least one set of electrocardiogram signals being from a subject, a time from the at least one set of electrocardiogram signals being recorded, to the onset of a disease from the subject, the death of the subject, or the last time the subject was seen with no disease.

[0084] The disease maybe a cardiovascular disease.

[0085] At least one of the sets of electrocardiogram signals may be labelled with either an existing disease, or labelled with the prediction of an existing disease.

[0086] The subject may be a human.

[0087] The at least some of the plurality of sets of electrocardiogram signals are from the same subject recorded at different times.

[0088] That is, the plurality of sets of electrocardiogram signals comprise longitudinal data.

[0089] The methods may be computer-implemented.

[0090] According to a sixth aspect of the invention, there is provided a computing device comprising: a memory; and at least one processor configured for performing the method of any of the first to fifth aspects.

[0091] According to a seventh aspect of the invention, there is provided a non-transitory computer readable medium, storing one or more programs for execution by one or more processors of a computing device, the one or more programs including instructions for performing the method of any of the first to fifth aspects.Brief description of the drawings

[0092] Certain embodiments of the present invention will now be described, by way of example, with reference to the accompanying drawings, in which:

[0093] Figure 1

[0094] Example subject-specific survival predictions

[0095] AIRE outputs subject-specific survival curves. Two examples are shown for subjects who died during follow up (A) and two for subjects who survived through the follow up period (B). Dashed red lines indicate the date of death and dashed black lines indicate AIRE predicted date of death. (C) Examples of subjects with many ECGs during the study period, each blue dot is a survival prediction from a single ECG. AIRE predicted survival trends down over time and predicted probability of survival is particularly low prior to actual time of death (red dashed line).

[0096] Figure 2

[0097] Mortality prediction performance - BIDMC Test set

[0098] Kaplan-Meier curves of AIRE predicted all-cause mortality by risk quartile in the whole BIDMC test set (A) and a subset of normal ECGs (B). (C) Comparison of AIRE performance across sex and major ethnic groups, AIRE performs well across all demographic groups. Using Cox models, AIRE was compared with existing risk factors and ECG parameters. In all comparisons, AIRE had a higher C-index than all comparators for both all-cause mortality (C), and cardiovascular mortality (D). Including AIRE combined with other parameters provided further improvements. ECG parameters: heart rate, PR interval, QRS duration, QTc interval, CV risk factors: diabetes mellitus, hypertension, smoking history, hyperlipidaemia, ethnicity. ASCVD risk factors: systolic blood pressure, total cholesterol, HDL cholesterol, hypertension, smoking history, diabetes mellitus, ethnicity. 10-year ASCVD risk assessed using the pooled cohort equation.

[0099] Figure 3

[0100] Survival analysis in four diverse, transnational, external validation datasets

[0101] In all cohorts, AIRE successfully identified groups at higher risk of all-cause mortality. (A) Sao Paulo-Minas Gerais Tropical Medicine Research Center (SaMi-TROP) cohort of subjects with Chagas disease, (B) CODE cohort of primary care subjects in Brazil, (C) The Brazilian Longitudinal Study of Adult Health (ELSA-Brasil) volunteer cohort, (D) UK Biobank volunteer cohort.Figure 4

[0102] Mortality prediction in high-risk disease groups and prediction of actionable end-points. Using Cox models, AIRE was compared with existing risk factors, ECG and imaging parameters in subgroups with severe aortic stenosis (A) and primary pulmonary hypertension (B). We also evaluated disease specific models for ASCVD (C), ventricular arrhythmia (D) and heart failure prediction (E). Echocardiographic parameters for severe aortic stenosis: left ventricular ejection fraction (LVEF), aortic valve area, peak gradient and mean gradient. Echocardiographic parameters for primary pulmonary hypertension: LVEF, LV end diastolic diameter, tricuspid regurgitation (TR) pressure gradient, TR severity, right ventricular (RV) function, RV diameter. ECG parameters: heart rate, PR interval, QRS duration, QTc interval. ASCVD risk factors: systolic blood pressure, total cholesterol, HDL cholesterol, hypertension, smoking history, diabetes mellitus, ethnicity.

[0103] 10-year ASCVD risk assessed using the pooled cohort equation. ARIC-HF risk factors: body mass index, systolic blood pressure, prevalent ASCVD, diabetes mellitus, smoking history, previous myocardial infarction, hypertension, ethnicity.

[0104] Figure 5

[0105] AIRE model explainability: (A) A variational auto-encoder was used to identify the most important morphological features in AIRE predicted mortality, three features are shown, identifying the importance of a broad QRS complex in a left bundle morphology as well as biphasic and inverted T waves. (B) Average ± standard deviation ECG waveforms for the 10,000 highest and lowest predicted survival ECGs from the BIDMC test set. This analysis identified poor R wave progression, low QRS amplitude and T wave flattening / inversion as important features in AIRE predicted survival.

[0106] Figure 6

[0107] Exploration of underlying biology through Phenome and Genome-wide association studies.

[0108] (A) Genome-wide association study (GWAS) Manhattan plots of genomic loci associated with predicted survival. Nearest genes to significant single nucleotide polymorphisms are shown. The red line depicts the genome-wide significant threshold (P<5 x 10-8). (B) Phenome-wide association study in the UK Biobank. Cardiac associations include left ventricular ejection fraction (LVEF), atrial and right ventricular phenotypes. Non-cardiac associations included brain phenotypes such as total volume of white matter hyperintensities and pack years of smoking. (C) Association of AIRE predicted survival with echocardiographic parameters in the BIDMC test set. LA: left atrium, LAEF: LAejection fraction, TR: tricuspid regurgitation, MV: mitral valve, LVESD: LV end-systolic diameter, LVEDD: LV end-diastolic diameter, RA: right atrium, AV: aortic valve.

[0109] Figure SI

[0110] Using Cox models, AIRE was compared with existing risk factors and ECG parameters. Sensitivity analysis using one ECG per subject, selected at random

[0111] Figure S2

[0112] Sensitivity analyses using quartile threshold derived in model training cohort validation sets. AIRE model with quartiles from BIDMC (panel A and B). AIRE primary care model with quartiles from CODE (C and D).

[0113] Figure S3

[0114] Using Cox models, AIRE was compared with existing risk factors and ECG parameters. Sensitivity analysis using one ECG per subject, selected at random

[0115] Figure S4

[0116] Mortality prediction performance - BIDMC Test set, Single ECG lead

[0117] Kaplan-Meier curves of AIRE predicted all-cause mortality by risk quartile in the whole BIDMC test set using only ECG lead 1

[0118] Figure S5

[0119] ASCVD prediction sensitivity analyses using blood and BP results within 30 days (A) and 90 days (B). CV death prediction sensitivity analyses using blood test and blood pressure results within 30 days (A) and 90 days (B)

[0120] Figure S6

[0121] Two example Grad-CAM plots depicting most important areas for model predictions in red and least important as blue.

[0122] Detailed description of certain embodiments

[0123] IntroductionThe electrocardiogram (ECG) has been a fundamental tool in clinical medicine for over a century. With the recent advent of artificial intelligence (Al) the potential applications of the ECG have significantly expanded, including both diagnostic and predictive capabilities (1-3). Recent studies have demonstrated the remarkable predictive capabilities of AI-ECG models, not only in terms of predicting mortality, but also cardiac diseases (4-7).

[0124] Existing mortality prediction models are limited by prediction of survival at one, or a small number of set time points and lack information for clinicians on specific actionable pathways. A high-risk prediction is unhelpful to a clinician if there is no accompanying information on how to affect the survival trajectory of their patient. To make AI-ECG predictions more actionable, it is essential to also consider time-to-event predictions and specific predictions for diseases with established preventive and disease modifying treatments.

[0125] Furthermore, the adoption of Al into clinical practice is significantly limited by concerns regarding explainability and biological plausibility. Just as knowledge of the mechanisms of action of drugs are important for physicians to have confidence in their application, biological plausibility of Al predictions ensures their credibility and acceptance.

[0126] To address these limitations of existing risk prediction models, we aimed to develop and perform transnational validation on an AI-ECG risk prediction platform that is not only accurate, but also actionable, explainable, and biologically plausible.

[0127] Methods

[0128] In this study, we first developed the AI-ECG risk estimation (AIRE) model for prediction of all-cause mortality. We subsequently developed seven additional sub models. The eight models together are referred to as the AIRE platform. A model development and validation flow chart is shown in Figure 1.

[0129] Ethical approvals

[0130] This study complies with all relevant ethical regulations, further details are provided in the Supplementary Methods.

[0131] Cohorts

[0132] We studied five, intentionally diverse, cohorts from a wide range of patient groups. While each cohort consists of individuals from a specific subset of the population (primary care,secondary care, cardiomyopathy and volunteers), these cohorts combined can be considered representative of a wide range of patient groups and volunteers. The Beth Israel Deaconess Medical Center (BIDMC) cohort is a secondary care dataset comprised of routinely collected data from, Boston, USA. The Sao Paulo-Minas Gerais Tropical Medicine Research Center (SaMi-Trop) is a cohort of patients with chronic Chagas cardiomyopathy (8). The Clinical Outcomes in Digital Electrocardiography (CODE) cohort is a Brazilian database of ECGs recorded in primary care (9) containing ECGs of both 10s and 7s duration. The subset of this dataset with only 10s ECGs is referred to as CODE-10s. The Longitudinal Study of Adult Health (ELSA-Brasil) cohort consists of Brazilian public servants (10). The UK Biobank (UKB) is longitudinal study of volunteers (11). Further details are provided in the Supplementary Methods.

[0133] AI-ECG risk estimation platform development

[0134] We developed the AIRE platform using the BIDMC cohort as the derivation dataset. As leads III, aVL, aVR, aVF are linear combinations of leads I and II, these leads were not used for model development or evaluation. We confirmed this by training a model with 12 leads, which had no better performance than the 8-lead model (Supplementary Results). For mortality end-points, ECGs without paired life status at 30 days were excluded. The data was split at a ratio of 50 / 10 / 40% for training, validation and internal test, respectively. Data was split by patient ID stratified by presence of ECGs with paired 5-year life status. Therefore, ECGs from the same patient could only be in one of training, validation or internal test sets. ECGs were treated independently and multiple ECGs per patient (where available) were utilised. We used a previously described convolutional neural network architecture based on residual blocks (12) and modified the architecture such that the final layer accommodated a discrete-time survival approach (13). The discrete-time survival approach allows the model to account for both time to outcome (mortality) and censorship (i.e., loss to follow up). Further details including ECG pre-processing Supplementary Methods.

[0135] The single-lead (lead I) model, AIRE-IL, was developed using the same methodology above, using the same BIDMC data split but using just lead I as the model input. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis Checklist for Prediction Model Development and Validation was followed.

[0136] Model fine-tuning for primary care population

[0137] Using the CODE dataset, we finetuned the model to be more representative of a primary care population (AIRE-primary care) with a much lower risk of adverse events. We used75% of the CODE dataset for finetuning, with 5 % as a validation set. The final 20% was used for internal validation of AIRE-primary care. Data was split by patient ID.

[0138] Model fine-tuning for other endpoints

[0139] We also developed five other subsequent models by fine-tuning the AIRE model separately for cardiovascular (CV) death, non-CV death, ventricular arrhythmia (VA), atherosclerotic cardiovascular disease (ASCVD) and heart failure (HF). These models were trained in the BIDMC dataset, and named AIRE-CV death, AIRE-NCV death, AIREVA, AIRE-ASCVD and AIRE-HF, respectively. The same splits were used as for training the original model. Fine-tuning was performed by loading the previous model and training using a low learning rate without freezing any layers. Diverse external validation was performed to explore validity in diverse cohorts and ensure the models were not overfit to the development cohort. Internal validation and external validation datasets are shown in Figure 1. Further details in Supplementary Methods.

[0140] Comparison of AIRE with other models

[0141] We compared AIRE-primary care and AIRE-ASCVD to the recently described SEER (Stanford Estimator of ECG Risk), which had a similar goal of predicting cardiovascular mortality and ASCVD (7). The model code and weights were downloaded, and performance evaluated in the UKB, as this was a dataset external to AIRE-primary care, AIRE-ASCVD and SEER, with cause of death and ASCVD event data available.

[0142] Survival and statistical analyses

[0143] In the test set, we generated predictions for all ECGs, for the primary analyses of allcause mortality. Sensitivity analyses, including a single random ECG per subject were also performed. For analyses requiring a single predictor value, the probability of survival at 5 years was used. Risk quartiles were defined (low, intermediate-low, intermediate-high, high) based on values in the validation set. Given the diverse populations and event rates evaluated, risk quartiles were redefined in each dataset. In each case, where categorical risk levels were required, 5% of the dataset was used to define the quartiles and evaluation was performed in the remaining 95%. Kaplan-Meier curves comparing the risk quartiles were plotted and statistical significance assessed using the log rank test. Sensitivity analyses using the original validation set quartiles were also performed.

[0144] Where binary predictions were required for metrics, Youden's index was used to identify the threshold that maximised the sensitivity and specificity in the validation set.Cox models were fit using the test dataset comparing demographics, clinical variables, imaging parameters and AIRE platform predictions. For the Cox models incorporating AIRE platform predictions, all model outputs (I. e. , predicted probabilities of death at each timepoint) were used as inputs, as well as age, sex, heart rate, PR interval, QRS duration and QTc interval. These models are designated AIRE-Cox for the AIRE model and AIRE-CV Death-Cox, AIRE-ASCVD-Cox, AIRE-VA-Cox, AIRE-HF-Cox for the other models. Comparator Cox model components are stated for each analysis, they included, LV ejection fraction (LVEF), ECG parameters (heart rate, PR interval, QRS duration and QTc interval), CV risk factors: diabetes mellitus, hypertension, smoking history, hyperlipidaemia, ethnicity. ASCVD risk factors: systolic blood pressure, total cholesterol, HDL cholesterol, hypertension, smoking history, diabetes mellitus, ethnicity. 10-year ASCVD risk assessed using the pooled cohort equation. ARIC-HF risk factors: body mass index, systolic blood pressure, prevalent ASCVD, diabetes mellitus, smoking history, previous myocardial infarction, hypertension, ethnicity. For the Cox model comparisons complete case analysis was used. Although the clinical, demographic and imaging data is unlikely to be missing at random, they are likely to be available for the patient groups in whom these predictions are likely to be relevant. Nested Cox models were compared with the Likelihood Ratio test, while non-nested Cox models were compared with the Partial Likelihood Ratio test. Statistical analyses were performed with R 4.2.0 statistical package (R Core Team, Vienna, Austria) or Python (version 3.9). Primary analysis used all ECGs meeting inclusion criteria for each analysis, sensitivity analyses used a single ECG per subject.

[0145] Diagnostic and imaging data

[0146] ICD-9 and ICD-10 codes were used to define presence / absence of disease in the BIDMC and UKB cohorts. Cardiovascular death in the BIDMC cohort was defined as mortality occurring within 30 days of a diagnostic code for acute myocardial infarction, ischaemic stroke, intracranial haemorrhage, sudden cardiac death, or heart failure as previously described (7, 14). In the UKB, cause of death was ascertained based on the ICD10 code stated as the primary cause of death. Diagnostic codes were not available in the SaMi-Trop, CODE and ELSA-Brasil datasets. Echocardiograms within 60 days of an ECG were linked and used for analyses incorporating echocardiographic parameters. Medication usage was not available in the BIDMC cohort, therefore ICD9 and ICD10 codes consistent with a diagnosis of hypertension were used to code for antihypertensive medication use for calculation of the pooled cohort equation. Blood results and blood pressure (BP) readings taken within 180 days of the ECG were averaged. Sensitivity analyses wereperformed using 90 days and 30 days results. Normal ECG definition is described in the Supplementary Methods.

[0147] Explainability

[0148] In order to understand the ECG morphologies associated with predicted survival, we used two approaches. Median beats were extracted using the BRAVEHEART ECG analysis software as previously described (15).

[0149] First, we trained a variational autoencoder (VAE) as using median ECG beats. Further details in Supplementary Methods. In preliminary analyses, models based on only the VAE latent features were found to be inferior to the supervised deep learning approach described above (Supplementary Results), therefore the VAE was used for explainability only, and not used for AIRE model training or any of the prediction models described in this manuscript. VAE latent features were input into a linear regression with predicted survival as the output. The top 3 most important features as assessed by the t-value were visualised by latent traversal.

[0150] Second, using the median beats we calculated the average waveform from the 10,000 ECGs with the lowest and highest AIRE predicted mortality. The mean and standard deviation of these waveforms was then plotted.

[0151] Third, gradient-weighted class activation maps (Grad-CAM) were plotted using tf-keras-vis vO.8.7 (16).

[0152] PheWAS

[0153] To better understand the biology underlying AIRE predictions, we performed phenome-wide association studies (PheWAS). We performed PheWAS analysis in the UKB, which contains data from over 3000 phenotypes derived from patient measurements, surveys, and investigations. Univariate correlation was performed to investigate the association between ECG predicted survival and phenotypes, adjusted for age, sex and age2. We additionally investigated the association of predicted survival with continuous echo traits in the BIDMC dataset. Left ventricular trabeculation was calculated as previously described (17). Deep learning-derived brain age was calculated as previously described (18).

[0154] Further methods including for GWAS are described in the Supplementary Methods. The funder had no role in data collection, analysis, interpretation, writing of the manuscript or the decision to submit.Results

[0155] AIRE accurately predicts mortality across diverse timepoints

[0156] In the BIDMC cohort, 1,163,401 ECGs were available from 189,539 subjects. Mean follow-up period was 5.46±5.81 years on a per ECG basis, 3.41 (4.08) years taking a random ECG per subject. 34,851 (18.4%) subjects died during follow-up (Table SI).

[0157] AIRE produces subject-specific survival curves from only a single ECG and can predict time-to-death (Figure 2A and 2B). Figure 2C demonstrates the evolution of AIRE-predicted survival based on multiple ECGs performed over several years of follow up.

[0158] In the hold out test set, AIRE predicted all-cause mortality with a concordance-index of 0.775 (0.773-0.776). Detailed performance metrics are shown in Table S2-4. Figure 3A shows the marked separation of survival curves of risk quartiles in the test set. Table S5 shows age- and sex-adjusted hazard ratios for high-risk vs low-risk quartiles for all cohorts. When considering only ECGs labelled as normal by cardiologists, there remained a significant difference in mortality between high-risk and low-risk subjects (Figure 3B). Importantly, AIRE had similar performance in both men and women and in major ethnic groups (Figure 4A and Table S6). Further results are reported in the Supplementary Results.

[0159] AIRE is superior to demographic data and traditional risk factors for mortality prediction We compared the ability of AIRE to predict mortality, against using demographic data, risk factors and risk scores in the BIDMC test set (Figure 4B). AIRE-Cox had a significantly higher C-index than all other parameters combined (0.794 (0.792-0.795) vs 0.759 (0.758-0.761), p < 0.0001).

[0160] In the BIDMC test set, AIRE-CV Death predicted CV death with a C-index of 0.832 (0.831-0.834). AIRE-CV Death-Cox had a significantly higher C-index for prediction of CV death than all other parameters combined (0.844 (0.839-0.849) vs 0.795 (0.789-0.801), p < 0.0001) (Figure 4C). Finally, AIRE-Non-CV Death predicted non-CV death with a C-index of 0.749 (0.747-0.751). Figure SI shows sensitivity analyses using a single random ECG per subject.

[0161] AIRE predicts mortality in transnational external datasetsWe tested if AIRE is applicable to a wide range of settings, from volunteers to primary care to patients with cardiomyopathy. First, we evaluated the performance of AIRE in the SaMi-Trop cohort of patients with chronic Chagas cardiomyopathy (8). Dataset demographics for all cohorts are shown in Table SI, results summary in Table S5 and Sil. The C-index was 0.773 (0.733-0.813, Figure 3C, Figure S2).

[0162] The CODE cohort is a Brazilian database of ECGs recorded in primary care (9). In order to investigate external validity of AIRE, we first evaluated the performance of AIRE without any fine-tuning. As AIRE was trained exclusively on 10s ECGs, we evaluated the model on the 10s subset (CODE-lOs), AIRE had a C-index of 0.762 (0.759-0.765) for allcause mortality prediction. AIRE-primary care more accurately predicted mortality with an improved C-index of 0.802 (95% CI 0.799-0.805, Figure 3D). When considering only the ECGs labelled as normal (26766 ECGs from 21897 subjects), there remained a significant difference in mortality between high-risk and low-risk subjects based on model predictions (Table S5).

[0163] Further evaluation of AIRE-primary care was performed in another independent external dataset, ELSA-Brasil (n = 13739) a volunteer cohort of civil servants from Brazil (10)). The C-index was 0.713 (0.691-0.735 Figure 3E). Again, when considering normal ECGs only there was a significant difference in mortality between high and low risk subjects (Table S5).

[0164] Finally, we additionally evaluated the performance of AIRE-primary care in the UK Biobank, a relatively healthy volunteer population (n = 42386) with only 526 (1.2%) deaths during follow-up. The C-index was 0.638 (0.608-0.668 (Figure 3F)) for all-cause mortality. As cause of death was available in the UKB, we additionally examined the ability of AIRE-primary care to predict CV death. C-index for CV death was 0.695 (0.636-0.754)

[0165] Hughes et al recently reported SEER (Stanford Estimator of ECG Risk) with the similar goal of predicting cardiovascular mortality (7). AIRE-primary care was superior to SEER at predicting CV death in the UKB (SEER C-index 0.572 (0.514-0.630) p values for comparison to AIRE-primary care <0.001).

[0166] Actionable predictions: (1) Future ASCVD

[0167] AIRE-ASCVD was able to predict future ASCVD in subjects without known ASCVD (C-index 0.696 (0.694-0.698) n = 227588 ECGs from 56598 subjects). Detailedperformance metrics are shown in Table S7-9. We externally validated these findings in the UKB Biobank (UKB); a healthy, volunteer population. AIRE-ASCVD had reduced performance in predicting ASCVD, C-index 0.643 (0.624-0.662) but was significantly better than the SEER model (7) (SEER C-index 0.547 (0.527-0.567), p <0.0001 for comparison with AIRE-ASCVD).

[0168] We compared AIRE-ASCVD to other risk parameters, including the pooled cohort equation (PCE) and ASCVD risk factors, in a subset of outpatients (4580 ECGs from 2926 subjects) in the BIDMC test set with appropriate available data. AIRE-ASCVD-Cox had a significantly higher C-index than all other factors combined (0.679 (0.651-0.708) vs 0.642 (0.613-0.672), Figure 4D, p < 0.00001). Figure S3 shows sensitivity analyses using a single random ECG per subject for all three actionable prediction endpoints. Numerical results for all Cox models are shown in Table S10.

[0169] Actionable predictions: (2) Future Ventricular arrhythmia

[0170] AIRE-VA was able to accurately predict future VA (C-index 0.760 (0.756-0.763) n = 393203 ECGs from 62443 subjects) in subjects without a previous history of VA. When compared to other conventional risk parameters, including ECG parameters and left ventricular ejection fraction (101935 ECGs from 21093 subjects), AIRE-VA-Cox, had a significantly higher C-index than all other factors combined (0.722 (0.716-0.728) vs 0.699 (0.693-0.704), Figure 4E, p < 0.0001). The Supplementary Results describe performance in subgroups of LVEF <50% and dilated cardiomyopathy.

[0171] In the UKB, (n = 34400 with both ECG and cardiac magnetic resonance imaging (CMR) and without previous VA, 44 events), AIRE-VA had similar performance in predicting first occurrence of VA 0.719 (0.635-0.803) and performed at least equivalent to LVEF from cardiac magnetic resonance imaging (CMR) (C-index 0.595 (0.494-0.697), p for comparison 0.11).

[0172] Actionable predictions: (3) Future heart failure

[0173] AIRE-HF was able to accurately predict future HF in subjects without a previous history of HF (C-index 0.787 (0.785-0.789) n = 310200, from 61747 unique subjects). In a subset of patients with the available data, we compared AIRE-HF-Cox to other conventional risk parameters in Cox models (36486 ECGs from 12288 subjects), including HF risk factors identified in the Atherosclerotic Risk in Communities (ARIC) study (19). AIRE-HF-Cox had a significantly higher C-index than all other factors combined (0.761 (0.755-0.767) vs 0.716 (0.710-0.722), Figure 4F, p < 0.00001). In the external validation cohort, UKB,AIRE-HF-Cox had similar performance at predicting future HF (C-index 0.768 (0.733 -0.802).

[0174] Single-lead ECG model performance

[0175] We trained a single-lead version of AIRE using lead I only (AIRE-IL). The performance of AIRE-IL (C-index 0.751 (0.750-752) Figure S4) was only slightly inferior in discrimination compared to the 8-lead AIRE model.

[0176] Explainable ECG morphologies associate with adverse prognosis

[0177] Using a VAE, we found features of QRS morphology, particularly broader and more left bundle branch block morphologies, inverted and biphasic T waves as well as ST segment changes were identified as the most significant morphological features associated with high predicted mortality (Figure 5A). In a second approach, using median beats the BIDMC test set we found poor precordial R wave progression, low QRS amplitude and T wave flattening / inversion as important features in AIRE-predicted survival (Figure 5B). The third approach, Grad-CAM, confirms the contribution of the QRS complex and T waves to AIRE predictions (Figure S6).

[0178] Biological plausibility: Genetic associations of AIRE-predicted survival

[0179] We performed a genome-wide association study to identify genetic loci associated with high-risk AIRE predictions (Figure 6A, Table S12). We found significant loci adjacent to TBX3, VGLL2, CCDC91 and KCNQ1. The identified genes reinforce our ECG explainability analyses, which includes our VAE analysis. TBX3 has been associated with QRS duration, QRS voltage and QRS-T angle (20, 21), abnormalities of which were shown to be associated with high-risk predictions on our analysis of VAE latent features. KCNQ1 is associated with Long QT Syndrome 1 and QT interval (22). VGLL2 has additionally been associated with ECG morphologies (23). These genes were additionally associated with non-ECG traits. TBX3 has been associated with blood pressure (24), myocardial mass (20) and trabecular development (17). VGLL2 has been associated with blood pressure, atrial fibrillation, BMI and AI-ECG derived delta-age (23-27). KCNQ1 is also associated with metabolic syndrome phenotypes (28). Finally, CCDC91 (Coiled-Coil Domain Containing 91) associates with BMI (28).

[0180] Biological plausibility: Phenotypic associations of AIRE-predicted survival

[0181] We performed a PheWAS in the UK Biobank, Figure 6B. In particular, CMR associations included reduced left ventricular (LV) ejection fraction (EF), more positive, i.e., abnormal, global longitudinal strain, increased LV mass and increased left atrial size,which were correlated with reduced AIRE-predicted survival. We additionally specifically examined the association between predicted survival and left ventricular trabeculation and found a significant negative correlation (Table S13).

[0182] In BIDMC echocardiographic analyses (Figure 6C), we found LVEF was positively correlated with predicted survival while LA volume and surrogate measures of pulmonary pressure (TR velocity) and right ventricular diameter were both negatively correlated with predicted survival.

[0183] Finally, investigation of non-cardiac imaging phenotypes identified associations with multiple multi-model brain imaging phenotypes including the total volume of white matter hyperintensities and deep learning-derived brain age (18) (Table S14).

[0184] Discussion

[0185] We describe, for the first time, an actionable, explainable, and biologically plausible mortality and risk prediction AI-ECG platform of eight AI-ECG models. Importantly, our platform was externally validated across ethnically and demographically diverse transnational cohorts.

[0186] AIRE predicts time-to-mortality in diverse cohorts

[0187] This paper substantially extends the work of others on mortality protection using the ECG. Raghunath et al described the use of deep learning for mortality prediction (5), while Sun et al more recently built upon this work (6). Our study has several significant differences from these previous publications. Firstly, the use of a survival neural network architecture, provides our model with the ability to predict time of death without being constrained to a small number of time points. Additionally, this allows us to use training data from subjects that were censored, for whom the time for death was not known, this allows the use of more real-world data (which is almost always censored) for model training. Furthermore, by comparing our model to existing clinical risk factors and imaging parameters, we have demonstrated the significant additive value of our model beyond traditional approaches. Finally, we performed external validation across diverse populations, demonstrating the wide applicability of our model platform.

[0188] Model positive predictive value (PPV) was reduced in the volunteer cohorts. This is unsurprising, given the very low event rate in these populations and the dependence of PPV on disease prevalence, or in this case incidence (29). Our findings are similar to other AI-ECG studies in this area, which have reported comparable PPVs in specificcohorts (7, 30). We propose the AIRE platform would be best utilised in different ways depending on the clinical population. In low-risk populations, the high negative predictive value would allow confident reassurance of individuals at lowest risk, while in higher risk populations, the high positive predictive value could allow clinicians to identify patients at increased risk of adverse events.

[0189] Event-specific risk prediction

[0190] The AIRE platform can also predict future cardiovascular events such as ASCVD, HF and VA, in addition to predicting mortality. ASCVD prediction is currently used extensively in international guidelines for decision-making around lipid lowering therapies (31). In this study, we have shown that AIRE-ASCVD provides additional information that could improve ASCVD risk prediction and is superior and additive to the existing PCE. AIRE-primary care and AIRE-ASCVD were superior to SEER, a recently described AI-ECG model, for CV death and ASCVD prediction, respectively.

[0191] Similarly, predicting future VA is a particularly important endpoint, as there are clear preventative, and therapeutic options. Current guidelines advocate LVEF as the primary factor in determining eligibility for a primary prevention ICD. In our study, we demonstrated that AIRE-VA is a better predictor of future VT / VF than LVEF and could therefore potentially be incorporated into this decision-making paradigm.

[0192] Lastly, predicting future heart failure is important given the high number of unplanned hospital admissions due to undiagnosed heart failure (32). Through prediction of heart failure with AIRE-HF, early clinical assessment, echocardiography and institution of appropriate therapies may reduce adverse events, particularly in HF with reduced ejection fraction.

[0193] Single-lead ECG applications

[0194] Extending risk prediction to the single-lead ECG is particularly important given the rapidly increasing number of single lead devices, including consumer products (33) . Our study highlights the excellent performance of AIRE at mortality prediction on only a single lead. This could be particularly applicable for inpatient cardiac monitoring, where frequent AIRE predictions may be employed for early detection of risk. Single-lead AIRE models could also be used for remote monitoring in the outpatient setting using wearable devices, for example in patients with chronic diseases such as heart failure, where high-risk predictions could trigger pre-emptive treatments to prevent hospital admissions.

[0195] Explainability and biological plausibilityA significant challenge in deep learning is explainability of the model predictions. Our ECG explainability findings are in line with prior studies that highlight these features as being prognostically importantly (34-36) and are reinforced by our GWAS findings, with consistent associations with QRS duration, voltage, QRS-T angle and QT interval.

[0196] We identified phenotypic associations with cardiac chamber structure and function, including trabeculation and myocardial mass. These plausible biological pathways were reinforced by GWAS associations with TBX3. VGLL2 has been additionally described in relation to AI-ECG derived delta-age, which is a marker of accelerated biological aging (9). We also identified AIRE-predicted survival as inversely correlated with deep learning-derived brain-age. Finally, we identified variants in KCNQ1 and CCDC91 that suggest AIRE may capture metabolic risk as an additional mechanism. These findings suggest AIRE-predicted survival is a biomarker of overall health, including biological age and the presence of clinical and subclinical disease.

[0197] Limitations

[0198] There are limitations to the accuracy and granularity of ICD diagnostic codes that are used in this study to ascertain disease status. In particular, ventricular arrhythmias as reported by ICD codes are not necessarily sustained or haemodynamically significant and therefore patients predicted to have these events would not necessarily benefit from an implantable cardioverter defibrillator. There are drawbacks to each of the explainability methods presented in this work, and it is well recognised that there is no perfect explainability technique for deep learning (37). Although deep learning models are not fully explainable, our aim is that, by providing multiple complementary explainability analyses, the reader can gain some insight into the ECG morphologies associated with higher risk. The ultimate goal of developing AIRE is to use it to guide treatment decisions on patients, by integrating it fully into an electronic health records system or medical device, though the work presented is only one of many steps towards that goal. Other future steps required would include testing it prospectively in clinical studies and obtaining the appropriate regulatory approvals.

[0199] Conclusion

[0200] In conclusion, we describe the AIRE platform, an actionable, explainable and biologically plausible AI-ECG risk estimation platform that has the potential for use worldwide acrossa wide range of clinical contexts, including primary and secondary care, for short- and long-term risk prediction at a population and disease-specific levels.

[0201] ECG datasets

[0202] The datasets used in this study were selected in order to maximise the diversity of populations in which AIRE was tested. As this was a retrospective study, no a priori sample size calculations were performed. Missing data was handled by complete-case analysis. Cohort inclusion required ECG and mortality data, therefore there was no missing data for predictor (AIRE) or outcome (mortality).

[0203] (i) The BIDMC cohort

[0204] The BIDMC cohort is a dataset comprised of routinely collected data from Beth Israel Deaconess Medical Center, Boston, USA. Subjects over 16 years old with a valid ECG performed from 2014 to 2023 were included. Prior ECGs back to 2000 were included for these subjects. Mortality was determined via the Massachusetts Department of Public Health (DPH) and / or review of the BIDMC electronic medical record, while diagnostic International Classification of Diseases (ICD) codes were used to determine disease status. Subjects were censored at time of death or last in-person hospital contact.

[0205] (ii) The CODE Cohort

[0206] The CODE cohort is a database of 2,322,513 ECG records from 1,676,384 different patients of 811 counties in the state of Minas Gerais / Brazil from the Telehealth Network of Minas Gerais (TNMG). The cohort is linked to public mortality databases. Patients over 16 years old with a valid ECG performed from 2010 to 2017 were included. Clinical data, including medical diagnoses, were self-reported. In a 15% stratified sample of the original cohort (CODE-15) an ECG was labelled "normal" according to conventional clinical reporting and based on automated interval measurements (1).

[0207] (iii) ELSA-Brasil Cohort

[0208] ELSA-Brasil is a cohort study of 15,105 Brazilian public servants, aged 35 to 74 at enrolment. All active or retired employees of six participating institutions were eligible for the study. The full inclusion criteria and protocol have been previously described (2). 13,739 subjects had ECG and outcome data available for analysis.

[0209] (iv) The SaMi-Trop CohortThe SaMi-Trop cohort is a prospective cohort of 1,631 patients with chronic Chagas cardiomyopathy and has been previously described in detail (3). Briefly, the inclusion criteria were: self-reported Chagas disease and aged 19 years or more. Digital ECGs were performed in 2011-2012 by TNMG. 83% of this cohort had abnormal ECGs (4).

[0210] (v) The UK Biobank Cohort

[0211] The UK Biobank is longitudinal study of over 500,000 volunteers aged 40-69 at the time of enrolment in 2006-2010 (5). At baseline assessment participants provided information on health and lifestyle via questionnaire, had physical measures taken (including height, weight, and blood pressure) and donated samples of blood urine and saliva. A subgroup of participants were invited back for subsequent visits for additional investigations, including for detailed studies including cardiac magnetic resonance imaging (MRI), brain MRI and digital ECGs. 42,386 subjects with digital ECGs taken at the instance 2 visit were available for analysis. There is evidence of healthy volunteer selection bias (6). Outcomes were linked to cancer and death registry data, hospital admissions and primary care records. Detailed phenotyping using the cardiac MRI data has been previously described (7).

[0212] Supplementary Methods

[0213] ECG pre-processing

[0214] 12-lead ECGs were pre-processed with a bandpass filter 0.5 to 100Hz, a notch filter at 60Hz and re-sampling to 400Hz. Zero padding was added to make the input shape a power of 2. This resulted in 4096 samples for each lead for a 10s recording (4000 samples + 48 zeros at the start and end), that was used as input to the neural network model. As leads III, aVL, aVR, aVF are linear combinations of leads I and II, these leads were not used for model development or evaluation. Therefore, the final input shape of a single ECG was 4096 x 8.

[0215] Model architecture and loss function

[0216] The model was trained using the discrete-time survival approach described in (8).

[0217] Specifically, the patients' follow-up time is divided into a set of fixed intervals, and an estimated conditional hazard probability for each interval is calculated (probability offailure in that interval, given that the individual has survived at least to the beginning of the interval). For each time interval j, the loss function is defined as:

[0218] 2_(i = l)A(d_j)|| ln(h jAi ) + (i-d j + l)A(r_ J)|ln( [1-hl jAi )

[0219] where h_jAi is the hazard probability for individual I during time interval j, r_j is the number of individuals that have not experienced failure or censoring before the beginning of the interval j, and d_j is the number of individuals who have during the interval j. The overall loss function is the sum of the losses for each time interval.

[0220] The output of the neural network is an n-dimensional vector, with each element representing the predicted conditional probability of surviving that time interval ( [1-h _jAi ). An individual's probability of surviving through the end of time interval j is given by:

[0221] S j- n (i 1)Aj (1-h i)

[0222] Model training

[0223] The output of the model is a predicted probability of survival within each discrete timeinterval. The model was trained to account for events occurring 10 years from the time of the ECG. The model was selected based on the lowest validation loss and was evaluated on the unseen test set. Hyperparameter optimisation was performed using the BIDMC validation set. Models were trained for up to 50 epochs and the lowest validation loss of each training run used to evaluate model performance and select optimal hyperparameters. Selecting the model iteration with lowest validation loss helps prevent overfitting in a manner similar to early stopping. Models generally converged in under 20 epochs. The hyperparameters tuned were the learning rate, batch size and discrete-time survival timepoints. Remaining hyperparameters were used as previously described without further tuning (9). Models were trained using a single Nvidia RTX 6000 on Imperial College London's high performance computing cluster. The Keras framework with a TensorFlow backend was used for neural network training and inference (10, 11).

[0224] Model fine-tuning for other endpoints supplement

[0225] For non-mortality endpoints (ASCVD, VA, HF), subjects were coded as having prevalent disease, future disease or neither at the time of the ECG. Although the goal of these models was to predict future events, we hypothesised that information on prevalent disease would be helpful for model training. In order to include prevalent disease in the discrete-time survival model, we encoded prevalent disease at the first timepoint in the discrete-time survival labels. When evaluating model performance, subjects withprevalent disease were excluded from evaluation for that specific model, for example for evaluating the VA model subjects with VA were excluded, but subjects with HF or ASCVD remained.

[0226] Variational auto-encoder training

[0227] We trained a variational autoencoder (VAE) as previously described (12) using median ECG beats. The VAE consists of three components: the encoder, the decoder and the latent space. The encoder and the decoder are made up of one-dimensional convolutional layers with increasing filters and decreasing kernel sizes closer to the latent space. The latent space was restricted to 30 features at a maximum, although typically only a subset of these features was used by the model for the reconstruction. The model was trained to minimize both the median ECG reconstruction loss, defined by a symmetric mean absolute percentage error function, and the Kullback-Leibler divergence (KL loss). This second term is specifically added to the VAE model to ensure that the features generated by the model are generative and disentangled. An additional 0-parameter was included as a weight on the KL-term to optimize the balance between the reconstruction loss and the latent factor interpretability. We tested beta values of 0.1, 0.25, 0.5, 1, 3, 5 and 10 and defined the best model at a 0-parameter of 0.25 based on the Pearson correlation between the median and its reconstruction in the validation dataset, as well as a visual inspection of the latent vector transversals.

[0228] Cox models proportional hazards

[0229] Recent work suggests virtually all real-world clinical datasets will violate the proportional hazards assumptions if sufficiently powered and that statistical tests for the proportional hazards assumption may be unnecessary (13). In line with these recommendations, the proportional hazards assumption was not evaluated and the hazard ratio from our Cox models should be interpreted as a weighted average of the true hazard ratios over the follow-up period.

[0230] Normal ECG definition

[0231] We evaluated the performance of AIRE in clinician reported normal ECGs. In the BIDMC dataset a subset of ECGs had Cardiologist reports. Normal ECGs in BIDMC were determined by searching for 'normal ecg' in the free text reports, a whole word match was required in order to exclude 'abnormal ecg'. ECGs with the phrase 'otherwise' werealso excluded from the normal definition. Additionally we filtered by heart rate (60-100 bpm), PR interval (less than 200ms), QRS duration (less than 120ms) and QTc interval (less than 470ms). Normal ECGs in ELSA-Brasil and CODE were defined as previously described (14).

[0232] GWAS methods

[0233] To identify genetic associations with the ECG mortality predictions, we performed a genome-wide association study (GWAS) in the UKB. As the predicted survival trait was skewed, the data were normalized by rank-based inverse normal transform prior to the analysis. The GWAS analysis was adjusted for the following covariates: age at imaging visit, sex, height, body mass index (BMI), imaging assessment centre and the first 10 genetic principal components.

[0234] Standard quality control was undertaken. Included single nucleotide polymorphisms (SNPs) had a minor allele frequency (MAF) >0.1% and an imputation INFO score of >0.4. Unrelated individuals of genetically-determined European ancestry were included as previously described (15). The GWAS was undertaken using FastGWA mixed linear model association analysis through the Genome-wide Complex Trait Analysis software using a genetic relationship matrix (GRM) to adjust for population structure (16). The Manhattan plots depict the nearest gene. Top SNPs were identified that had a P-value <5x10-8. Associations of the detected SNPs were evaluated using the NHGRI-EBI GWAS Catalog, PhenoScanner, GTEx and the GeneAtlas UK Biobank PheWAS browser.

[0235] Supplementary Results

[0236] 12-lead vs 8-lead performance

[0237] In order to confirm no additional information was provided by the four calculated leads of the ECG (III, aVR, aVL, aVF) we performed a sensitivity analysis, training models using a full 12-lead ECG as input, everything else was kept constant in this experiment.

[0238] Performance was evaluated in the BIDMC validation set. The 8-lead AIRE model had a c-index of 0.771 (0.769- 0.774), while for the 12-lead model this was 0.770 (0.767-0.772). As the 12-lead model was not superior to the 8 lead model, we proceeded with 8-lead models for the analyses.

[0239] Models based on VAE features for mortality predictionWe compared AIRE to models developed using the VAE features. We used Cox proportional hazards models (C-index 0.682), and XGBoost (C-index 0.715) models as comparators. These were inferior to AIRE (C-index 0.775) but had reasonable performance indicating the value of the VAE features for mortality prediction.

[0240] AIRE accurately predicts mortality across diverse timepoints supplementary results

[0241] In sensitivity analysis including only patients resident in Massachusetts, (for whom death status would be more accurate) model performance was unchanged (C-index 0.775 (0.773-0.776).

[0242] Actionable predictions: (2) Ventricular arrhythmia supplementary results

[0243] In a subgroup of patients (27889 ECGs from 4371 subjects) with an LVEF of <50%, AIRE alone had a higher C-index than LVEF (0.635 (0.628 - 0.643) vs 0.622 (0.614 - 0.630)) in predicting first episode of VA. In a subgroup of patients with dilated cardiomyopathy (DCM) (8252 ECGs from 1171 subjects), AIRE alone had a C-index of 0.666 (0.652- 0.681) compared to 0.613 (0.598 - 0.629) for LVEF alone. Incorporating LVEF and AIRE together in a Cox model yielded marginal improvements in C-index of 0.657 (0.649 - 0.665) and 0.673 (0.659 - 0.687) for patients with LVEF <50% or DCM respectively.

[0244] Table SI

[0245] Dataset demographics

[0246] Data at the timepoint of a randomly selected ECG per subject is shown for the BIDMC and CODE datasets.

[0247] Categorical variables n (%), continuous variables mean (SD)

[0248]

[0249]

[0250] Table S2

[0251] Model performance metrics at fixed timepoints using continuous predicted outcomes

[0252]

[0253] Table S3

[0254] Model performance metrics at fixed timepoints using binary predicted outcomes

[0255]

[0256] Table S4

[0257] Model performance - time-dependent area under the receiver operating characteristics curve

[0258]

[0259] Table S5

[0260] Summary of age and sex adjusted hazard ratios (high risk quartile vs low risk)

[0261]

[0262] Table S6

[0263] Summary table of sex and ethnicity specific AIRE performance for mortality prediction in BIDMC Test set

[0264]

[0265]

[0266] le S7

[0267] E disease specific submodel metrics at fixed timepoints using binary predicted outcomes

[0268]

[0269]

[0270] le S8

[0271] E disease specific submodel metrics at fixed timepoints using continuous predicted outcomes

[0272]

[0273] Table S9

[0274] AIRE disease specific submodel performance - time-dependent area under the receiver operating characteristics curve

[0275]

[0276] Table S10 Summary table of cox model results

[0277] C-index (95% CI) is reported

[0278]

[0279] Table Sil

[0280] Mortality prediction results summary table

[0281] Model performance as assessed by C-index (95% CI) is shown

[0282] AIRE: Artificial-intelligence enhanced ECG risk estimator

[0283]

[0284]

[0285] le S12 GWAS Lead variants

[0286] ome build 37

[0287]

[0288] Table S13

[0289] Association of left ventricular trabeculation with predicted survival (n = 33,333)

[0290] <

[0291] <

[0292]

[0293] Table S14

[0294] Association of deep learning-derived brain age with predicted survival (n = 17348)

[0295] <

[0296] < &

[0297]

[0298] Modifications

[0299] It will be appreciated that various modifications may be made to the embodiments hereinbefore described. Such modifications may involve equivalent and other features which are already known in the design and use of methods of using a trained model to detect the presence of a disease or to predict the likelihood of a disease occurring and which may be used instead of or in addition to features already described herein. Features of one embodiment may be replaced or supplemented by features of another embodiment.

[0300] Although claims have been formulated in this application to particular combinations of features, it should be understood that the scope of the disclosure of the present invention also includes any novel features or any novel combination of features disclosed herein either explicitly or implicitly or any generalization thereof, whether or not it relates to the same invention as presently claimed in any claim and whether or not it mitigates any or all of the same technical problems as does the present invention. The applicants hereby give notice that new claims may be formulated to such features and / or combinations of such features during the prosecution of the present application or of any further application derived therefrom.

Claims

1. Claims1. A method of using a trained model to detect cardiovascular disease, the model trained using electrocardiogram signals, the method comprising:receiving a set of electrocardiogram signals;applying the trained model to the received set of electrocardiogram signals; andoutputting a likelihood that the received set of electrocardiogram signals are indicative of cardiovascular disease.

2. A method of using a trained model to predict cardiovascular disease occurring, the model trained using electrocardiogram signals, the method comprising:receiving a set of electrocardiogram signals;applying the trained model to the received set of electrocardiogram signals; andoutputting a likelihood that the received electrocardiogram signals are indicative of cardiovascular disease occurring.

3. The method of claims 1 or 2, wherein the cardiovascular disease is any one or more from:structural cardiovascular disease;hypertension;valvular heart disease;pulmonary hypertension;cardiomyopathy;arrhythmia; andstroke.

4. A method of using a trained model to detect a disease, the model trained using electrocardiogram signals, the method comprising:receiving a set of electrocardiogram signals;applying the trained model to the received set of electrocardiogram signals; andoutputting a likelihood that the received set of electrocardiogram signals are indicative of the disease.

5. A method of using a trained model to predict a disease occurring, the model trained using electrocardiogram signals, the method comprising:receiving a set of electrocardiogram signals;applying the trained model to the received set of electrocardiogram signals; andoutputting a likelihood that the received set of electrocardiogram signals are indicative of the disease occurring.

6. The method of claims 4 or 5, wherein the disease is diabetes, kidney disease, or liver disease.

7. A method of training a model to detect disease, the method comprising:receiving a plurality of sets of electrocardiogram signals, at least one set of electrocardiogram signals being from a subject;a time from the at least one set of electrocardiogram signals being recorded, to the onset of a disease from the subject, the death of the subject, or the last time the subject was seen with no disease.

8. The method of claim 7, wherein the disease maybe a cardiovascular disease.

9. The method of claim 7 or 8, wherein at least one of the sets of electrocardiogram signals is labelled with either an existing disease, or labelled with the prediction of an existing disease.

10. The method of any of claims 7 to 9 wherein the at least some of the plurality of sets of electrocardiogram signals are from the same subject recorded at different times.

11. A computing device comprising:a memory; andat least one processor configured for performing the method of any preceding claim.

12. A non-transitory computer readable medium, storing one or more programs for execution by one or more processors of a computing device, the one or more programs including instructions for performing the method of any of claims 1 to 10.

Images