Blood Pressure Control and Maintenance in a Prospective Cohort of Younger Veterans: Roles of Sex, Race, Ethnicity, and Social Determinants of Health.

Background: Proactive blood pressure (BP) management is particularly beneficial for younger Veterans, who have a greater prevalence and earlier onset of cardiovascular disease than non-Veterans. It is unknown what proportion of younger Veterans achieve and maintain BP control after hypertension onset and if BP control differs by demographics and social deprivation. Methods: Electronic health records were merged from Veterans who enrolled in VA care 10/1/2001-9/30/2017 and met criteria for hypertension - first diagnosis or antihypertensive fill. BP control (140/90 mmHg), was estimated 1, 2, and 5 years post-hypertension documentation, and characterized by sex, race, and ethnicity. Adjusted logistic regressions assessed likelihood of BP control by these demographics and with the Social Deprivation Index (SDI). Results: Overall, 17% patients met criteria for hypertension (n=198,367; 11% of women, median age 41). One year later, 59% of men and 65% of women achieved BP control. After adjustment, women had a 72% greater odds of BP control than men, with minimal change over 5 years. Black adults had a 22% lower odds of BP control than White adults. SDI did not significantly change these results. Conclusions: In the largest study of hypertension in younger Veterans, 41% of men and 35% of women did not have BP control after 1 year, and BP control was consistently better for women through 5 years. Thus, the first year of hypertension management portends future, long-term BP control. As social deprivation did not affect BP control, the VA system may protect against disadvantages observed in the general U.S. population.

A roadmap to artificial intelligence (AI): Methods for designing and building AI ready data to promote fairness.

OBJECTIVES: We evaluated methods for preparing electronic health record data to reduce bias before applying artificial intelligence (AI). METHODS: We created methods for transforming raw data into a data framework for applying machine learning and natural language processing techniques for predicting falls and fractures. Strategies such as inclusion and reporting for multiple races, mixed data sources such as outpatient, inpatient, structured codes, and unstructured notes, and addressing missingness were applied to raw data to promote a reduction in bias. The raw data was carefully curated using validated definitions to create data variables such as age, race, gender, and healthcare utilization. For the formation of these variables, clinical, statistical, and data expertise were used. The research team included a variety of experts with diverse professional and demographic backgrounds to include diverse perspectives. RESULTS: For the prediction of falls, information extracted from radiology reports was converted to a matrix for applying machine learning. The processing of the data resulted in an input of 5,377,673 reports to the machine learning algorithm, out of which 45,304 were flagged as positive and 5,332,369 as negative for falls. Processed data resulted in lower missingness and a better representation of race and diagnosis codes. For fractures, specialized algorithms extracted snippets of text around keywork "femoral" from dual x-ray absorptiometry (DXA) scans to identify femoral neck T-scores that are important for predicting fracture risk. The natural language processing algorithms yielded 98% accuracy and 2% error rate The methods to prepare data for input to artificial intelligence processes are reproducible and can be applied to other studies. CONCLUSION: The life cycle of data from raw to analytic form includes data governance, cleaning, management, and analysis. When applying artificial intelligence methods, input data must be prepared optimally to reduce algorithmic bias, as biased output is harmful. Building AI-ready data frameworks that improve efficiency can contribute to transparency and reproducibility. The roadmap for the application of AI involves applying specialized techniques to input data, some of which are suggested here. This study highlights data curation aspects to be considered when preparing data for the application of artificial intelligence to reduce bias.

A Deep Learning Approach for Automated Extraction of Functional Status and New York Heart Association Class for Heart Failure Patients During Clinical Encounters.

Introduction: Serial functional status assessments are critical to heart failure (HF) management but are often described narratively in documentation, limiting their use in quality improvement or patient selection for clinical trials. We developed and validated a deep learning-based natural language processing (NLP) strategy to extract functional status assessments from unstructured clinical notes. Methods: We identified 26,577 HF patients across outpatient services at Yale New Haven Hospital (YNHH), Greenwich Hospital (GH), and Northeast Medical Group (NMG) (mean age 76.1 years; 52.0% women). We used expert annotated notes from YNHH for model development/internal testing and from GH and NMG for external validation. The primary outcomes were NLP models to detect (a) explicit New York Heart Association (NYHA) classification, (b) HF symptoms during activity or rest, and (c) functional status assessment frequency. Results: Among 3,000 expert-annotated notes, 13.6% mentioned NYHA class, and 26.5% described HF symptoms. The model to detect NYHA classes achieved a class-weighted AUROC of 0.99 (95% CI: 0.98-1.00) at YNHH, 0.98 (0.96-1.00) at NMG, and 0.98 (0.92-1.00) at GH. The activity-related HF symptom model achieved an AUROC of 0.94 (0.89-0.98) at YNHH, 0.94 (0.91-0.97) at NMG, and 0.95 (0.92-0.99) at GH. Deploying the NYHA model among 166,655 unannotated notes from YNHH identified 21,528 (12.9%) with NYHA mentions and 17,642 encounters (10.5%) classifiable into functional status groups based on activity-related symptoms. Conclusions: We developed and validated an NLP approach to extract NYHA classification and activity-related HF symptoms from clinical notes, enhancing the ability to track optimal care and identify trial-eligible patients.

Development and Validation of Case-Finding Algorithms to Identify Pancreatic Cancer in the Veterans Health Administration.

BACKGROUND: Survival in pancreatic ductal adenocarcinoma (PDAC) remains poor due to late diagnosis. Electronic Health Records (EHRs) can be used to study this rare disease, but validated algorithms to identify PDAC in the United States EHRs do not currently exist. AIMS: To develop and validate an algorithm using Veterans Health Administration (VHA) EHR data for the identification of patients with PDAC. METHODS: We developed two algorithms to identify patients with PDAC in the VHA from 2002 to 2023. The algorithms required diagnosis of exocrine pancreatic cancer in either ≥ 1 or ≥ 2 of the following domains: (i) the VA national cancer registry, (ii) an inpatient encounter, or (iii) an outpatient encounter in an oncology setting. Among individuals identified with ≥ 1 of the above criteria, a random sample of 100 were reviewed by three gastroenterologists to adjudicate PDAC status. We also adjudicated fifty patients not qualifying for either algorithm. These patients died as inpatients and had alkaline phosphatase values within the interquartile range of patients who met ≥ 2 of the above criteria for PDAC. These expert adjudications allowed us to calculate the positive and negative predictive value of the algorithms. RESULTS: Of 10.8 million individuals, 25,533 met ≥ 1 criteria (PPV 83.0%, kappa statistic 0.93) and 13,693 individuals met ≥ 2 criteria (PPV 95.2%, kappa statistic 1.00). The NPV for PDAC was 100%. CONCLUSIONS: An algorithm incorporating readily available EHR data elements to identify patients with PDAC achieved excellent PPV and NPV. This algorithm is likely to enable future epidemiologic studies of PDAC.

Identifying incarceration status in the electronic health record using large language models in emergency department settings.

Background: Incarceration is a significant social determinant of health, contributing to high morbidity, mortality, and racialized health inequities. However, incarceration status is largely invisible to health services research due to inadequate clinical electronic health record (EHR) capture. This study aims to develop, train, and validate natural language processing (NLP) techniques to more effectively identify incarceration status in the EHR. Methods: The study population consisted of adult patients (≥ 18 y.o.) who presented to the emergency department between June 2013 and August 2021. The EHR database was filtered for notes for specific incarceration-related terms, and then a random selection of 1,000 notes was annotated for incarceration and further stratified into specific statuses of prior history, recent, and current incarceration. For NLP model development, 80% of the notes were used to train the Longformer-based and RoBERTa algorithms. The remaining 20% of the notes underwent analysis with GPT-4. Results: There were 849 unique patients across 989 visits in the 1000 annotated notes. Manual annotation revealed that 559 of 1000 notes (55.9%) contained evidence of incarceration history. ICD-10 code (sensitivity: 4.8%, specificity: 99.1%, F1-score: 0.09) demonstrated inferior performance to RoBERTa NLP (sensitivity: 78.6%, specificity: 73.3%, F1-score: 0.79), Longformer NLP (sensitivity: 94.6%, specificity: 87.5%, F1-score: 0.93), and GPT-4 (sensitivity: 100%, specificity: 61.1%, F1-score: 0.86). Conclusions: Our advanced NLP models demonstrate a high degree of accuracy in identifying incarceration status from clinical notes. Further research is needed to explore their scaled implementation in population health initiatives and assess their potential to mitigate health disparities through tailored system interventions.

Creation and Validation of an Automated Registry for Outpatient Parenteral Antibiotics.

Existing outpatient parenteral antibiotic therapy (OPAT) registries are resource intensive, and OPAT programs struggle to produce objective data to show the value of their work. We describe the building and validation of an automated OPAT registry within our electronic medical record and provide objective data on the value of the program. Variables and outcomes include age, sex, race, ethnicity, primary insurance payor, antibiotic names, infection syndromes treated, discharge disposition, 30-day all-cause readmission and death rates, complications, and an estimate of the hospital days saved. Records for 146 OPAT episodes were reviewed manually to validate the registry. Data were displayed in a dashboard within the electronic medical record. Over the 4-year time frame, our registry collected 3956 unique patients who completed 4710 episodes (approximately 1200 episodes per year). A total of 400 complications during OPAT were identified. All variables had an accuracy of >90% on validation. The OPAT program resulted in a reduction in hospital length of stay by 88 820 days, or roughly 22 000 days per year. We intend our registry to serve as a blueprint for similar OPAT programs with limited administrative resources. Wider application of our system would allow for easier aggregation and comparisons of OPAT practice and address the lack interinstitutional standardization of OPAT data and outcomes.

Opioid prescription and risk of atrial fibrillation in younger veterans.

BACKGROUND: Opioids may play a part in the development of atrial fibrillation (AF). Understanding the relationship between opioid exposure and AF can help providers better assess the risk and benefits of prescribing opioids. OBJECTIVE: To assess the incidence of AF as a function of prescribed opioids and opioid type. DESIGN: We performed unadjusted and adjusted time-updated Cox regressions to assess the association between opioid exposure and incident AF. PARTICIPANTS: The national study sample was comprised of Veterans enrolled in the Veterans Health Administration (VHA) who served in support of post-9/11 operations. MAIN MEASURES: The main predictor of interest was prescription opioid exposure, which was treated as a time-dependent variable. The first was any opioid exposure (yes/no). Secondary was opioid type. The outcome, incident AF, was identified through ICD-9-CM diagnostic codes at any primary care visit after the baseline period. KEY RESULTS: A total of 609,763 veterans (mean age: 34 years and 13.24% female) were included in our study. Median follow-up time was 4.8 years. Within this cohort, 124,395 veterans (20.40%) were prescribed an opioid. A total of 1,455 Veterans (0.24%) were diagnosed with AF. In adjusted time-updated Cox regressions, the risk of incident AF was higher in the veterans prescribed opioids (hazard ratio [HR]: 1.47; 95% confidence interval [CI]: 1.38-1.57). In adjusted time-updated Cox regressions, both immunomodulating and nonimmunomodulating opioid type was associated with increased risk of incident AF (HR: 1.40; 95% CI: 1.25-1.57 and HR: 1.49; 95% CI: 1.39-1.60), compared to no opioid use, respectively. CONCLUSIONS: Our findings suggest opioid prescription may be a modifiable risk factor for the development of AF.

Relationship Between Pain and LGBT Status Among Veterans in Care in a Retrospective Cross-Sectional Cohort.

Background: Pain assessment is performed in many healthcare systems, such as the Veterans Health Administration, but prior studies have not assessed whether pain screening varies in sexual and gender minority populations that include individuals who identify as lesbian, gay, bisexual, and/or transgender (LGBT). Objective: The purpose of this study was to evaluate pain screening and reported pain of LGBT Veterans compared to non-LGBT Veterans. Methods: Using a retrospective cross-sectional cohort, data from the Corporate Data Warehouse, a national repository with clinical/administrative data, were analyzed. Veterans were classified as LGBT using natural language processing. We used a robust Poisson model to examine the association between LGBT status and binary outcomes of pain screening, any pain, and persistent pain within one year of entry in the cohort. All models were adjusted for demographics, mental health, substance use, musculoskeletal disorder(s), and number of clinic visits. Results: There were 1,149,486 Veterans (218,154 (19%) classified as LGBT) in our study. Among LGBT Veterans, 94% were screened for pain compared to 89% among those not classified as LGBT (non-LGBT) Veterans. In adjusted models, LGBT Veterans' probability of being screened for pain compared to non-LGBT Veterans was 2.5% higher (95% CI 2.3%, 2.6%); risk of any pain was 2.1% lower (95% CI 1.6%, 2.6%); and there was no significant difference between LGBT and non-LGBT Veterans in persistent pain (RR = 1.00, 95% CI (0.99, 1.01), p = 0.88). Conclusions: In a nationwide sample, LGBT Veterans were more likely to be screened for pain but had lower self-reported pain scores, though adjusted differences were small. It was notable that transgender and Black Veterans reported the greatest pain. Reasons for these findings require further investigation.

Dementia risk analysis using temporal event modeling on a large real-world dataset.

The objective of the study is to identify healthcare events leading to a diagnosis of dementia from a large real-world dataset. This study uses a data-driven approach to identify temporally ordered pairs and trajectories of healthcare codes in the electronic health record (EHR). This allows for discovery of novel temporal risk factors leading to an outcome of interest that may otherwise be unobvious. We identified several known (Down syndrome RR = 116.1, thiamine deficiency RR = 76.1, and Parkinson's disease RR = 41.1) and unknown (Brief psychotic disorder RR = 68.6, Toxic effect of metals RR = 40.4, and Schizoaffective disorders RR = 40.0) factors for a specific dementia diagnosis. The associations with the greatest risk for any dementia diagnosis were found to be primarily related to mental health (Brief psychotic disorder RR = 266.5, Dissociative and conversion disorders RR = 169.8), or neurologic conditions or procedures (Dystonia RR = 121.9, Lumbar Puncture RR = 119.0). Trajectory and clustering analysis identified factors related to cerebrovascular disorders, as well as diagnoses which increase the risk of toxic imbalances. The results of this study have the ability to provide valuable insights into potential patient progression towards dementia and improve recognition of patients at risk for developing dementia.

StrokeClassifier: Ischemic Stroke Etiology Classification by Ensemble Consensus Modeling Using Electronic Health Records.

Determining the etiology of an acute ischemic stroke (AIS) is fundamental to secondary stroke prevention efforts but can be diagnostically challenging. We trained and validated an automated classification machine intelligence tool, StrokeClassifier, using electronic health record (EHR) text data from 2,039 non-cryptogenic AIS patients at 2 academic hospitals to predict the 4-level outcome of stroke etiology determined by agreement of at least 2 board-certified vascular neurologists' review of the stroke hospitalization EHR. StrokeClassifier is an ensemble consensus meta-model of 9 machine learning classifiers applied to features extracted from discharge summary texts by natural language processing. StrokeClassifier was externally validated in 406 discharge summaries from the MIMIC-III dataset reviewed by a vascular neurologist to ascertain stroke etiology. Compared with stroke etiologies adjudicated by vascular neurologists, StrokeClassifier achieved the mean cross-validated accuracy of 0.74 (±0.01) and weighted F1 of 0.74 (±0.01). In the MIMIC-III cohort, the accuracy and weighted F1 of StrokeClassifier were 0.70 and 0.71, respectively. SHapley Additive exPlanation analysis elucidated that the top 5 features contributing to stroke etiology prediction were atrial fibrillation, age, middle cerebral artery occlusion, internal carotid artery occlusion, and frontal stroke location. We then designed a certainty heuristic to deem a StrokeClassifier diagnosis as confidently non-cryptogenic by the degree of consensus among the 9 classifiers, and applied it to 788 cryptogenic patients. This reduced the percentage of the cryptogenic strokes from 25.2% to 7.2% of all ischemic strokes. StrokeClassifier is a validated artificial intelligence tool that rivals the performance of vascular neurologists in classifying ischemic stroke etiology for individual patients. With further training, StrokeClassifier may have downstream applications including its use as a clinical decision support system.

Insomnia and Early Incident Atrial Fibrillation: A 16-Year Cohort Study of Younger Men and Women Veterans.

Background There is growing consideration of sleep disturbances and disorders in early cardiovascular risk, including atrial fibrillation (AF). Obstructive sleep apnea confers risk for AF but is highly comorbid with insomnia, another common sleep disorder. We sought to first determine the association of insomnia and early incident AF risk, and second, to determine if AF onset is earlier among those with insomnia. Methods and Results This retrospective analysis used electronic health records from a cohort study of US veterans who were discharged from military service since October 1, 2001 (ie, post-9/11) and received Veterans Health Administration care, 2001 to 2017. Time-varying, multivariate Cox proportional hazard models were used to examine the independent contribution of insomnia diagnosis to AF incidence while serially adjusting for demographics, lifestyle factors, clinical comorbidities including obstructive sleep apnea and psychiatric disorders, and health care utilization. Overall, 1 063 723 post-9/11 veterans (Mean age=28.2 years, 14% women) were followed for 10 years on average. There were 4168 cases of AF (0.42/1000 person-years). Insomnia was associated with a 32% greater adjusted risk of AF (95% CI, 1.21-1.43), and veterans with insomnia showed AF onset up to 2 years earlier. Insomnia-AF associations were similar after accounting for health care utilization (adjusted hazard ratio [aHR], 1.27 [95% CI, 1.17-1.39]), excluding veterans with obstructive sleep apnea (aHR, 1.38 [95% CI, 1.24-1.53]), and among those with a sleep study (aHR, 1.26 [95% CI, 1.07-1.50]). Conclusions In younger adults, insomnia was independently associated with incident AF. Additional studies should determine if this association differs by sex and if behavioral or pharmacological treatment for insomnia attenuates AF risk.

Screening for Intimate Partner Violence Experience and Use in the Veterans Health Administration.

Importance: The practice of screening women for intimate partner violence (IPV) in health care settings has been a critical part of responding to this major public health problem. Yet, IPV prevention would be enhanced with detection efforts that extend beyond screening for IPV experiences to identifying those who use violence in relationships as well. Objective: To determine rates of IPV experiences and use (ie, among perpetrators of IPV) and factors associated with disclosures among adult patients seeking mental health services at the Veterans Health Administration. Design, Setting, and Participants: This cross-sectional study used electronic medical record data drawn from a quality improvement initiative at 5 Veterans Health Administration medical centers conducted between November 2021 and February 2022 to examine IPV disclosures following concurrent screening for IPV experience and use. Participants included patients engaged in mental health services. Data were analyzed in April and May 2023. Exposure: Mental health clinicians were trained to screen for IPV experience and use concurrently and instructed to screen all patients encountered through routine mental health care visits during a 3-month period. Main Outcomes and Measures: Outcomes of interest were past-year prevalence of IPV use and experience, sociodemographic characteristics, and clinical diagnoses among screened patients. Results: A total of 200 patients were offered IPV screening. Of 155 participants (mean [SD] age, 52.45 [15.65] years; 124 [80.0%] men) with completed screenings, 74 (47.7%) denied past-year IPV experience and use, 76 (49.0%) endorsed past-year IPV experience, and 72 (46.4%) endorsed past-year IPV use, including 67 participants (43.2%) who reported IPV experience and use concurrently; only 9 participants (5.8%) endorsed unidirectional IPV experiences and 5 participants (3.2%) endorsed unidirectional IPV use. Patients who reported past-year IPV experience and use were younger than those who denied IPV (experience: mean difference, -7.34 [95% CI, 2.51-12.17] years; use: mean difference, -7.20 [95% CI, 2.40-12.00] years). Patients with a posttraumatic stress disorder diagnosis were more likely to report IPV use (43 patients [59.7%]) than those without a posttraumatic stress disorder diagnosis (29 patients [40.3%]; odds ratio, 2.14; [95% CI, 1.12-4.06]). No other demographic characteristics or clinical diagnoses were associated with IPV use or experience. Conclusions and Relevance: In this cross-sectional study of IPV rates and associated factors, screening for IPV found high rates of both IPV experience and use among patients receiving mental health care. These findings highlight the benefit of screening for IPV experience and use concurrently across gender and age. Additionally, the associations found between PTSD and IPV use underscore the importance of strengthening and developing additional targeted treatment for IPV.

Implementer report: ICD-10 code F44.5 review for functional seizure disorder.

OBJECTIVE: The study aimed to measure the validity of International Classification of Diseases, 10th Edition (ICD-10) code F44.5 for functional seizure disorder (FSD) in the Veterans Affairs Connecticut Healthcare System electronic health record (VA EHR). METHODS: The study used an informatics search tool, a natural language processing algorithm and a chart review to validate FSD coding. RESULTS: The positive predictive value (PPV) for code F44.5 was calculated to be 44%. DISCUSSION: ICD-10 introduced a specific code for FSD to improve coding validity. However, results revealed a meager (44%) PPV for code F44.5. Evaluation of the low diagnostic precision of FSD identified inconsistencies in the ICD-10 and VA EHR systems. CONCLUSION: Information system improvements may increase the precision of diagnostic coding by clinicians. Specifically, the EHR problem list should include commonly used diagnostic codes and an appropriately curated ICD-10 term list for 'seizure disorder,' and a single ICD code for FSD should be classified under neurology and psychiatry.

Identifying suicide documentation in clinical notes through zero-shot learning.

Background and Aims: In deep learning, a major difficulty in identifying suicidality and its risk factors in clinical notes is the lack of training samples given the small number of true positive instances among the number of patients screened. This paper describes a novel methodology that identifies suicidality in clinical notes by addressing this data sparsity issue through zero-shot learning. Our general aim was to develop a tool that leveraged zero-shot learning to effectively identify suicidality documentation in all types of clinical notes. Methods: US Veterans Affairs clinical notes served as data. The training data set label was determined using diagnostic codes of suicide attempt and self-harm. We used a base string associated with the target label of suicidality to provide auxiliary information by narrowing the positive training cases to those containing the base string. We trained a deep neural network by mapping the training documents' contents to a semantic space. For comparison, we trained another deep neural network using the identical training data set labels, and bag-of-words features. Results: The zero-shot learning model outperformed the baseline model in terms of area under the curve, sensitivity, specificity, and positive predictive value at multiple probability thresholds. In applying a 0.90 probability threshold, the methodology identified notes documenting suicidality but not associated with a relevant ICD-10-CM code, with 94% accuracy. Conclusion: This method can effectively identify suicidality without manual annotation.

Intimate partner violence among lesbian, gay, and bisexual veterans.

The present study describes intimate partner violence (IPV) perpetration and victimization alongside theoretically associated variables in a sample of lesbian, gay, and bisexual veterans. We conducted bivariate analyses (chi-square tests and independent t test) to examine whether the frequencies of IPV perpetration and victimization varied by demographic characteristics, military sexual trauma, alcohol use, and mental health symptoms. Out of the 69 lesbian, gay, and bisexual (LGB) veterans who answered the questions on IPV, 16 (23.2%) reported some form of IPV victimization in the past year, and 38 (55.1%) reported past-year perpetration. Among the 43 veterans who reported psychological IPV, roughly half (48.9%) reported bidirectional psychological IPV, 39.5% reported perpetration only, and 11.6% reported victimization only. LGB veterans who reported bidirectional psychological IPV in their relationships were younger and reported greater symptoms of posttraumatic stress disorder symptoms and depression. The results presented here call for universal screening of IPV perpetration and victimization to both accurately assess and ultimately intervene among all veterans. Inclusive interventions are needed for all genders and sexual orientations, specifically interventions that do not adhere to gendered assumptions of perpetrators and victims. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

A Roadmap to Artificial Intelligence (AI): Methods for Designing and Building AI ready Data for Women's Health Studies.

Objectives: Evaluating methods for building data frameworks for application of AI in large scale datasets for women's health studies. Methods: We created methods for transforming raw data to a data framework for applying machine learning (ML) and natural language processing (NLP) techniques for predicting falls and fractures. Results: Prediction of falls was higher in women compared to men. Information extracted from radiology reports was converted to a matrix for applying machine learning. For fractures, by applying specialized algorithms, we extracted snippets from dual x-ray absorptiometry (DXA) scans for meaningful terms usable for predicting fracture risk. Discussion: Life cycle of data from raw to analytic form includes data governance, cleaning, management, and analysis. For applying AI, data must be prepared optimally to reduce algorithmic bias. Conclusion: Algorithmic bias is harmful for research using AI methods. Building AI ready data frameworks that improve efficiency can be especially valuable for women's health.

Detection of Left Ventricular Systolic Dysfunction From Electrocardiographic Images.

BACKGROUND: Left ventricular (LV) systolic dysfunction is associated with a >8-fold increased risk of heart failure and a 2-fold risk of premature death. The use of ECG signals in screening for LV systolic dysfunction is limited by their availability to clinicians. We developed a novel deep learning-based approach that can use ECG images for the screening of LV systolic dysfunction. METHODS: Using 12-lead ECGs plotted in multiple different formats, and corresponding echocardiographic data recorded within 15 days from the Yale New Haven Hospital between 2015 and 2021, we developed a convolutional neural network algorithm to detect an LV ejection fraction <40%. The model was validated within clinical settings at Yale New Haven Hospital and externally on ECG images from Cedars Sinai Medical Center in Los Angeles, CA; Lake Regional Hospital in Osage Beach, MO; Memorial Hermann Southeast Hospital in Houston, TX; and Methodist Cardiology Clinic of San Antonio, TX. In addition, it was validated in the prospective Brazilian Longitudinal Study of Adult Health. Gradient-weighted class activation mapping was used to localize class-discriminating signals on ECG images. RESULTS: Overall, 385 601 ECGs with paired echocardiograms were used for model development. The model demonstrated high discrimination across various ECG image formats and calibrations in internal validation (area under receiving operation characteristics [AUROCs], 0.91; area under precision-recall curve [AUPRC], 0.55); and external sets of ECG images from Cedars Sinai (AUROC, 0.90 and AUPRC, 0.53), outpatient Yale New Haven Hospital clinics (AUROC, 0.94 and AUPRC, 0.77), Lake Regional Hospital (AUROC, 0.90 and AUPRC, 0.88), Memorial Hermann Southeast Hospital (AUROC, 0.91 and AUPRC 0.88), Methodist Cardiology Clinic (AUROC, 0.90 and AUPRC, 0.74), and Brazilian Longitudinal Study of Adult Health cohort (AUROC, 0.95 and AUPRC, 0.45). An ECG suggestive of LV systolic dysfunction portended >27-fold higher odds of LV systolic dysfunction on transthoracic echocardiogram (odds ratio, 27.5 [95% CI, 22.3-33.9] in the held-out set). Class-discriminative patterns localized to the anterior and anteroseptal leads (V2 and V3), corresponding to the left ventricle regardless of the ECG layout. A positive ECG screen in individuals with an LV ejection fraction ≥40% at the time of initial assessment was associated with a 3.9-fold increased risk of developing incident LV systolic dysfunction in the future (hazard ratio, 3.9 [95% CI, 3.3-4.7]; median follow-up, 3.2 years). CONCLUSIONS: We developed and externally validated a deep learning model that identifies LV systolic dysfunction from ECG images. This approach represents an automated and accessible screening strategy for LV systolic dysfunction, particularly in low-resource settings.