Attia, Zachi I.

Background: Identification of systolic heart failure among patients presenting to the emergency department (ED) with acute dyspnea is challenging. The reasons for dyspnea are often multifactorial. A focused physical evaluation and diagnostic testing can lack sensitivity and specificity. The objective of this study was to assess the accuracy of an artificial intelligence-enabled ECG to identify patients presenting with dyspnea who have left ventricular systolic dysfunction (LVSD). Methods: We retrospectively applied a validated artificial intelligence-enabled ECG algorithm for the identification of LVSD (defined as LV ejection fraction ≤35%) to a cohort of patients aged ≥18 years who were evaluated in the ED at a Mayo Clinic site with dyspnea. Patients were included if they had at least one standard 12-lead ECG acquired on the date of the ED visit and an echocardiogram performed within 30 days of presentation. Patients with prior LVSD were excluded. We assessed the model performance using area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity. Results: A total of 1606 patients were included. Median time from ECG to echocardiogram was 1 day (Q1: 1, Q3: 2). The artificial intelligence-enabled ECG algorithm identified LVSD with an area under the receiver operating characteristic curve of 0.89 (95% CI, 0.86–0.91) and accuracy of 85.9%. Sensitivity, specificity, negative predictive value, and positive predictive value were 74%, 87%, 97%, and 40%, respectively. To identify an ejection fraction <50%, the area under the receiver operating characteristic curve, accuracy, sensitivity, and specificity were 0.85 (95% CI, 0.83–0.88), 86%, 63%, and 91%, respectively. NT-proBNP (N-terminal pro-B-type natriuretic peptide) alone at a cutoff of >800 identified LVSD with an area under the receiver operating characteristic curve of 0.80 (95% CI, 0.76–0.84). Conclusions: The ECG is an inexpensive, ubiquitous, painless test which can be quickly obtained in the ED. It effectively identifies LVSD in selected patients presenting to the ED with dyspnea when analyzed with artificial intelligence and outperforms NT-proBNP. Graphic Abstract: A graphic abstract is available for this article.

Artificial intelligence (AI) and machine learning (ML) in medicine are currently areas of intense exploration, showing potential to automate human tasks and even perform tasks beyond human capabilities. Literacy and understanding of AI/ML methods are becoming increasingly important to researchers and clinicians. The first objective of this review is to provide the novice reader with literacy of AI/ML methods and provide a foundation for how one might conduct an ML study. We provide a technical overview of some of the most commonly used terms, techniques, and challenges in AI/ML studies, with reference to recent studies in cardiac electrophysiology to illustrate key points. The second objective of this review is to use examples from recent literature to discuss how AI and ML are changing clinical practice and research in cardiac electrophysiology, with emphasis on disease detection and diagnosis, prediction of patient outcomes, and novel characterization of disease. The final objective is to highlight important considerations and challenges for appropriate validation, adoption, and deployment of AI technologies into clinical practice.

e17535 Background: Androgen Deprivation Therapy (ADT) is commonly used to treat prostate cancer (PC), but has been associated with cardiac morbidity and mortality. The exact mechanisms of this association are unclear. We sought to use an artificial intelligence (AI) enabled algorithm to identify ECG changes in PC patients who received ADT compared to PC patients who did not receive ADT. Methods: From 1,000,000 ECGs performed on 210,414 patients between 1993 and 2017 at our institution, a convolutional neural network was developed to detect predictive signatures for cardiac pathologies. During this process, the ability to predict “estimated sex” of the patient was developed, with output values ranging from 0 (female) to 1 (male). We applied this algorithm to 8,619 ECGs performed on 1,057 men age 75 or younger treated with radiation for high-risk or recurrent PC at our institution, and compared estimated sex after receiving ADT (n = 1,065) to ADT-naive ECGs (n = 7,554). We correlated ECG-identified estimated sex with serum testosterone levels using Spearman rank correlation. Results: Patients who had received ADT had a mean (SD) estimated sex value of 0.81 (0.26) compared to 0.92 (0.17) for those who did not (p < 0.001). Difference between estimated sex in post-ADT ECGs and ADT-naive ECGs remained significant across age groups (Table). Decreased serum total testosterone correlated with decreased estimated sex values in men receiving ADT (R = .57, p < 0.001). Conclusions: ADT for prostate cancer is associated with changes in AI-identified ECG parameters, including lower estimated male sex after receiving ADT. Lower ECG male sex estimate was associated with decreased serum testosterone. In this study, we provide preliminary proof of concept for a potential non-invasive means of monitoring treatment effect and physiologic change using ECGs. [Table: see text]

Background: Deep learning algorithms derived in homogeneous populations may be poorly generalizable and have the potential to reflect, perpetuate, and even exacerbate racial/ethnic disparities in health and health care. In this study, we aimed to (1) assess whether the performance of a deep learning algorithm designed to detect low left ventricular ejection fraction using the 12-lead ECG varies by race/ethnicity and to (2) determine whether its performance is determined by the derivation population or by racial variation in the ECG. Methods: We performed a retrospective cohort analysis that included 97 829 patients with paired ECGs and echocardiograms. We tested the model performance by race/ethnicity for convolutional neural network designed to identify patients with a left ventricular ejection fraction ≤35% from the 12-lead ECG. Results: The convolutional neural network that was previously derived in a homogeneous population (derivation cohort, n=44 959; 96.2% non-Hispanic white) demonstrated consistent performance to detect low left ventricular ejection fraction across a range of racial/ethnic subgroups in a separate testing cohort (n=52 870): non-Hispanic white (n=44 524; area under the curve [AUC], 0.931), Asian (n=557; AUC, 0.961), black/African American (n=651; AUC, 0.937), Hispanic/Latino (n=331; AUC, 0.937), and American Indian/Native Alaskan (n=223; AUC, 0.938). In secondary analyses, a separate neural network was able to discern racial subgroup category (black/African American [AUC, 0.84], and white, non-Hispanic [AUC, 0.76] in a 5-class classifier), and a network trained only in non-Hispanic whites from the original derivation cohort performed similarly well across a range of racial/ethnic subgroups in the testing cohort with an AUC of at least 0.930 in all racial/ethnic subgroups. Conclusions: Our study demonstrates that while ECG characteristics vary by race, this did not impact the ability of a convolutional neural network to predict low left ventricular ejection fraction from the ECG. We recommend reporting of performance among diverse ethnic, racial, age, and sex groups for all new artificial intelligence tools to ensure responsible use of artificial intelligence in medicine.

Background: Sex and age have long been known to affect the ECG. Several biologic variables and anatomic factors may contribute to sex and age-related differences on the ECG. We hypothesized that a convolutional neural network (CNN) could be trained through a process called deep learning to predict a person’s age and self-reported sex using only 12-lead ECG signals. We further hypothesized that discrepancies between CNN-predicted age and chronological age may serve as a physiological measure of health. Methods: We trained CNNs using 10-second samples of 12-lead ECG signals from 499 727 patients to predict sex and age. The networks were tested on a separate cohort of 275 056 patients. Subsequently, 100 randomly selected patients with multiple ECGs over the course of decades were identified to assess within-individual accuracy of CNN age estimation. Results: Of 275 056 patients tested, 52% were males and mean age was 58.6±16.2 years. For sex classification, the model obtained 90.4% classification accuracy with an area under the curve of 0.97 in the independent test data. Age was estimated as a continuous variable with an average error of 6.9±5.6 years (R-squared =0.7). Among 100 patients with multiple ECGs over the course of at least 2 decades of life, most patients (51%) had an average error between real age and CNN-predicted age of <7 years. Major factors seen among patients with a CNN-predicted age that exceeded chronologic age by >7 years included: low ejection fraction, hypertension, and coronary disease ( P <0.01). In the 27% of patients where correlation was >0.8 between CNN-predicted and chronologic age, no incident events occurred over follow-up (33±12 years). Conclusions: Applying artificial intelligence to the ECG allows prediction of patient sex and estimation of age. The ability of an artificial intelligence algorithm to determine physiological age, with further validation, may serve as a measure of overall health.

AbstractIntroductionEmerging medical technology has allowed for monitoring of heart rhythm abnormalities using smartphone compatible devices. The safety and utility of such devices have not been established in patients with cardiac implantable electronic devices (CIEDs). We sought to assess the safety and compatibility of the Food and Drug Administration‐approved AliveCor Kardia device in patients with CIEDs.Methods and ResultsWe prospectively recruited patients with CIED for a Kardia recording during their routine device interrogation. A recording was obtained in paced and nonpaced states. Adverse clinical events were noted at the time of recording. Electrograms (EGMs) from the cardiac device were obtained at the time of recording to assess for any electromagnetic interference (EMI) introduced by Kardia. Recordings were analyzed for quality and given a score of 3 (interpretable rhythm, no noise), 2 (interpretable rhythm, significant noise) or 1 (uninterpretable). A total of 251 patients were recruited (59% with a pacemaker and 41% with ICD). There were no adverse clinical events noted at the time of recording and no changes to CIED settings. Review of all EGMs revealed no EMI introduced by Kardia. Recordings were correctly interpreted in 90% of paced recordings (183 had a score of 3, 43 of 2, and 25 of 1) and 94.7% of nonpaced recordings (147 of 3, 15 of 2, and 9 of 1).ConclusionThe AliveCor Kardia device has an excellent safety profile when used in conjunction with most CIEDs. The quality of recordings was preserved in this population. The device, therefore, can be considered for heart rhythm monitoring in patients with CIEDs.

AbstractObjectivesWe sought to validate a deep learning algorithm designed to predict an ejection fraction (EF) less than or equal to 35% based on the 12‐lead electrocardiogram (ECG) in a large prospective cohort.BackgroundPatients undergoing routine ECG may have undetected left ventricular (LV) dysfunction that warrants further echocardiographic assessment. However, identification of these patients can be challenging.MethodsWe applied the algorithm to all ECGs interpreted by the Mayo Clinic ECG laboratory in September 2018. The performance of the algorithm was tested among patients with recent echocardiographic assessments of LV function. We also applied the algorithm in patients with no recent echocardiographic assessments of LV function to determine the rate of new “positive screens.”ResultsAmong 16 056 adult patients who underwent routine ECG, 8600 (age 67.1 ± 15.2 years, 45.6% male), had a transthoracic echocardiogram (TTE) and 3874 patients had a TTE and ECG less than 1 month apart. Among these patients, the algorithm was able to detect an EF less than or equal to 35% with 86.8% specificity, 82.5% sensitivity, and 86.5% accuracy, (area under the curve, 0.918). Among 474 “false‐positives screens,” 189 (39.8%) had an EF of 36% to 50%. Among patients with no prior TTE, the algorithm identified 3.5% of the patients with suspected EF less than or equal to 35%. Exploratory analysis suggests false positives could be reduced by assessing NT‐pro‐BNP after the initial “positive screen.”ConclusionsA deep learning algorithm detected depressed LV function with good accuracy in routine practice. Further studies are needed to validate the algorithm in patients with no prior echocardiogram and to assess the impact on echocardiography utilization, cost, and clinical outcomes.

PURPOSE Background parenchymal uptake (BPU), which describes the level of radiotracer uptake in normal fibroglandular tissue on molecular breast imaging (MBI), has been identified as a breast cancer risk factor. Our objective was to develop and validate a deep learning model using image convolution to automatically categorize BPU on MBI. METHODS MBI examinations obtained for clinical and research purposes from 2004 to 2015 were reviewed to classify the BPU pattern using a standardized five-category scale. Two expert radiologists provided interpretations that were used as the reference standard for modeling. The modeling consisted of training and validating a convolutional neural network to predict BPU. Model performance was summarized in data reserved to test the performance of the algorithm at the per-image and per-breast levels. RESULTS Training was performed on 24,639 images from 3,133 unique patients. The model performance on the withheld testing data (6,172 images; 786 patients) was evaluated. Using direct matching on the predicted classification resulted in an accuracy of 69.4% (95% CI, 67.4% to 71.3%), and if prediction within one category was considered, accuracy increased to 96.0% (95% CI, 95.2% to 96.7%). When considering the breast-level prediction of BPU, the accuracy remained strong, with 70.3% (95% CI, 68.0% to 72.6%) and 96.2% (95% CI, 95.3% to 97.2%) for the direct match and allowance for one category, respectively. CONCLUSION BPU provided a robust target for training a convolutional neural network. A validated computer algorithm will allow for objective, reproducible encoding of BPU to foster its integration into risk-stratification algorithms.