Background Artificial intelligence (AI) diagnostic models are typically developed in hospital-based populations enriched for disease prevalence and severity, where distinctions between health and disease are pronounced. When deployed in community-dwelling populations with lower prevalence and milder phenotypes, these distinctions may become less well-defined, raising questions about transportability and the impact of disease spectrum on model performance. Objectives The purpose of this study was to evaluate how differences in disease spectrum influence the transportability of AI-based diagnostic models into community-dwelling populations, using artificial intelligence electrocardiogram analysis (AI-ECG) for structural heart disease (SHD) as a test case. Methods EchoNext is an AI-ECG model trained in a multicenter, hospital-based cohort to detect SHD. We evaluated its performance in PREVUE-VALVE (Age and Sex-Specific PREValence of AcqUirEd VALVular Heart DiseasE Study), a community-based study of individuals aged 65 to 85 years undergoing in-home ECG and transthoracic echocardiography. Performance was assessed using area under the receiver operating characteristic curve (AUC) and predictive value and compared with derivation and external hospital-based cohorts, using propensity matching to account for differences in disease prevalence and case mix. Subgroup analyses were performed in clinically relevant populations. Results Among 3,000 PREVUE-VALVE participants, 2,402 met criteria for analysis. Compared with hospital-based cohorts, PREVUE-VALVE had lower SHD prevalence (8% vs 43%), less severe disease, and a shift in phenotype, including more moderate tricuspid regurgitation and less systolic heart failure. Consistent with these differences, model discrimination was lower in PREVUE-VALVE than in the hospital-based cohort (AUC: 71% [95% CI: 66%-76%] vs 83% [95% CI: 82%-83%]). Propensity matching attenuated but did not eliminate this difference, while performance was similar across external hospital-based cohorts, supporting disease spectrum and clinical context as key drivers. Performance was modestly better in PREVUE-VALVE subgroups with higher SHD prevalence and greater disease severity, such as individuals with an abnormal ECG (AUC: 79% [95% CI: 75%-83%]) or impaired health status (AUC: 76% [95% CI: 70%-82%]). Conclusions In a community-dwelling population with lower disease prevalence and milder phenotypes, AI-ECG performance was attenuated relative to hospital-based cohorts, driven by differences in disease spectrum. These findings underscore the impact of both disease prevalence and case mix on AI performance and highlight the importance of model evaluation within intended use populations. (Age and Sex-Specific PREValence if AcqUirEd VALVular Heart DiseasE; NCT05357404)

BACKGROUND AND AIMS:Classification and risk stratification in aortic (AR), mitral (MR), and tricuspid regurgitation (TR) remains a significant clinical challenge. This study aimed to develop an artificial intelligence (AI) system to assess valvular regurgitation and stratify MR-progression risk. METHODS:Using transthoracic echocardiograms (TTEs) at two sites (internal development/test, external test), the DELINEATE-Regurgitation system was developed to classify AR, MR, and TR severity using colour Doppler videos. Methods of summating video-level classifications into study-level predictions were tested, comparing single-view with multiview approaches integrating predictions across multiple videos. Model agreement with cardiologists was assessed by weighted kappa. A separate AI system (DELINEATE-MR-Progression) analysing colour Doppler videos was developed to predict which patients with mild, mild-moderate, and moderate MR were most likely to progress to moderate-severe or severe MR with analysis by Kaplan-Meier and Cox proportional hazards models. RESULTS:A total of 71 660 TTEs with 1 203 980 colour Doppler videos were included. The weighted kappa in internal/external test sets for regurgitation classification was 0.81/0.76 for AR, 0.76/0.72 for MR, and 0.73/0.64 for TR using a multiview approach taking all colour Doppler videos in a study, demonstrating substantial agreement with cardiologist interpretation with superiority of multiview over single view approaches. In the progression analysis, the AI score stratified MR-progression risk even when controlled for clinical factors known to be associated with MR progression [internal test set hazard ratio 4.1 (95% confidence interval 2.5-6.6)]. CONCLUSIONS:An AI system can accurately classify AR, MR, and TR and predict MR progression beyond currently known risk factors.

Early detection of structural heart disease is critical to improving outcomes, but widespread screening remains limited by the cost and accessibility of imaging tools such as echocardiography1,2. Recent advances in machine learning applied to heart rhythm recordings have shown promise in identifying disease3,4, although previous work has been limited by development in narrow populations or targeting only select heart conditions5. Here we introduce a deep learning model, EchoNext, trained on more than 1 million heart rhythm and imaging records across a large and diverse health system to detect many forms of structural heart disease. The model demonstrated high diagnostic accuracy in internal and external validation, outperforming cardiologists in a controlled evaluation and showing consistent performance across different care settings and racial and/or ethnic groups. The models were prospectively evaluated in a clinical trial of patients without previous cardiac imaging, successfully identifying previously undiagnosed heart disease. These findings support the potential of artificial intelligence to expand access to heart disease screening at scale. To enable further development and transparency, we have publicly released model weights and a large, annotated dataset linking heart rhythm data to imaging-based diagnoses.

Nearly two hundred common-variant depression risk loci have been identified by genome-wide association studies (GWAS). However, the impact of rare coding variants on depression remains poorly understood. Here, we present whole-exome sequencing analyses of depression with seven different definitions based on survey, questionnaire, and electronic health records in 320,356 UK Biobank participants. We showed that the burden of rare damaging coding variants in loss-of-function intolerant genes is significantly associated with risk of depression with various definitions. We compared the rare and common genetic architecture across depression definitions by genetic correlation and showed different genetic relationships between definitions across common and rare variants. In addition, we demonstrated that the effects of rare damaging coding variant burden and polygenic risk score on depression risk are additive. The gene set burden analyses revealed overlapping rare genetic variant components with developmental disorder, autism, and schizophrenia. Our study provides insights into the contribution of rare coding variants, separately and in conjunction with common variants, on depression with various definitions and their genetic relationships with neurodevelopmental disorders.

Background: Transthyretin cardiac amyloidosis (ATTR-CA) is an underdiagnosed cause of heart failure and early disease detection is essential for improving outcomes. We previously developed a multimodal deep learning model, named ATTRact-Net, leveraging electrocardiogram (ECG) and echocardiogram data that can identify patients with ATTR-CA. Patients at risk for ATTR-CA often have multiple ECGs and echos available for analysis, leading to many different risk predictions by AI models. There is little evidence on how to go from study-level prediction to patient-level decision. We studied varying methods to integrate multiple ECG/echo pair predictions to maximize model performance in the detection of ATTR-CA. We hypothesized that retaining the patient-level mean risk across the prior 2 years would improve model performance compared to using maximum predicted risk. Methods and Results: ATTRact-Net was originally trained on 799 patients with 22,344 ECG/echo pairs completed within 2 years of PYP scan. Previously published performance of the model showed an AUROC of 0.83 in internal testing. For this study, a new test set composed of an outside hospital not used for training, had 422 patients with 12,387 pairs and ATTR-CA prevalence of 23%. In this new hospital’s population, we tested model performance using four patient-level aggregation methods (mean, median, max, and no aggregation) for integrating predictions across multiple ECG/echo pairs. Performance was compared based on the area under the receiver operating characteristic curve (AUROC), Diagnostic Odds Ratio (DOR), and F1 score using the Youden index as the optimal threshold. In the external validation dataset, the model achieved AUROCs of 0.78 (max), 0.82 (mean), 0.82 (median) and 0.82 (none) shown in Figure 1A. Mean and median aggregation outperformed both max and no aggregation by F1-score and DOR (Figure 1B). Conclusion: For a multimodal AI model detecting ATTR-CA, aggregating ECG/echo risk predictions using mean or median from the prior 2 years considerably improved diagnostic yield compared to using patient’s maximum score or no aggregation. A prospective clinical trial is underway using this strategy for early diagnosis of ATTR-CA across our healthcare system.

Bipolar disorder is a leading contributor to disability, premature mortality, and suicide. Early identification of risk for bipolar disorder using generalizable predictive models trained on diverse cohorts around the United States could improve targeted assessment of high risk individuals, reduce misdiagnosis, and improve the allocation of limited mental health resources. This observational case-control study intended to develop and validate generalizable predictive models of bipolar disorder as part of the multisite, multinational PsycheMERGE Consortium across diverse and large biobanks with linked electronic health records (EHRs) from three academic medical centers: in the Northeast (Massachusetts General Brigham), the Mid-Atlantic (Geisinger) and the Mid-South (Vanderbilt University Medical Center). Predictive models were developed and validated with multiple algorithms at each study site: random forests, gradient boosting machines, penalized regression, including stacked ensemble learning algorithms combining them. Predictors were limited to widely available EHR-based features agnostic to a common data model including demographics, diagnostic codes, and medications. The main study outcome was bipolar disorder diagnosis as defined by the International Cohort Collection for Bipolar Disorder, 2015. In total, the study included records for 3,529,569 patients including 12,533 cases (0.3%) of bipolar disorder. After internal and external validation, algorithms demonstrated optimal performance in their respective development sites. The stacked ensemble achieved the best combination of overall discrimination (AUC = 0.82 - 0.87) and calibration performance with positive predictive values above 5% in the highest risk quantiles at all three study sites. In conclusion, generalizable predictive models of risk for bipolar disorder can be feasibly developed across diverse sites to enable precision medicine. Comparison of a range of machine learning methods indicated that an ensemble approach provides the best performance overall but required local retraining. These models will be disseminated via the PsycheMERGE Consortium website.

BACKGROUND:Artificial intelligence, particularly deep learning (DL), has immense potential to improve the interpretation of transthoracic echocardiography (TTE). Mitral regurgitation (MR) is the most common valvular heart disease and presents unique challenges for DL, including the integration of multiple video-level assessments into a final study-level classification. METHODS:A novel DL system was developed to intake complete TTEs, identify color MR Doppler videos, and determine MR severity on a 4-step ordinal scale (none/trace, mild, moderate, and severe) using the reading cardiologist as a reference standard. This DL system was tested in internal and external test sets with performance assessed by agreement with the reading cardiologist, weighted κ, and area under the receiver-operating characteristic curve for binary classification of both moderate or greater and severe MR. In addition to the primary 4-step model, a 6-step MR assessment model was studied with the addition of the intermediate MR classes of mild-moderate and moderate-severe with performance assessed by both exact agreement and ±1 step agreement with the clinical MR interpretation. RESULTS:A total of 61 689 TTEs were split into train (n=43 811), validation (n=8891), and internal test (n=8987) sets with an additional external test set of 8208 TTEs. The model had high performance in MR classification in internal (exact accuracy, 82%; κ=0.84; area under the receiver-operating characteristic curve, 0.98 for moderate or greater MR) and external test sets (exact accuracy, 79%; κ=0.80; area under the receiver-operating characteristic curve, 0.98 for moderate or greater MR). Most (63% internal and 66% external) misclassification disagreements were between none/trace and mild MR. MR classification accuracy was slightly higher using multiple TTE views (accuracy, 82%) than with only apical 4-chamber views (accuracy, 80%). In subset analyses, the model was accurate in the classification of both primary and secondary MR with slightly lower performance in cases of eccentric MR. In the analysis of the 6-step classification system, the exact accuracy was 80% and 76% with a ±1 step agreement of 99% and 98% in the internal and external test set, respectively. CONCLUSIONS:This end-to-end DL system can intake entire echocardiogram studies to accurately classify MR severity and may be useful in helping clinicians refine MR assessments.

Introduction: Recent studies using deep learning techniques have demonstrated promising left ventricular ejection fraction (LVEF) assessment from transthoracic echocardiograms (TTEs). However, most prior studies have focused on videos from a single apical view, a technique known to be subject to limitations given the regionality of LV systolic function. We hypothesized that a deep learning model trained to include echocardiographic video clips from multiple views from a large dataset will improve accuracy in LVEF assessment. Methods: We identified all adult TTEs with a clinically reported LVEF at Columbia University between 2019-2024. A view classification model was trained to identify apical 4 and 2-chamber and parasternal long and short-axis views for LVEF assessment. The internal dataset was split into train, validation and test sets to train spatiotemporal convolutional models for each of the 4 views to assess LVEF for each video clip. The median clip-level LVEF within a study was used to derive a study-level LVEF. The model was evaluated on an internal test set and a large external test set, which included all available adult TTEs from Weill Cornell Medical Center since 2011. As benchmark comparison, the previously published EchoNet-Dynamic model was also evaluated on the external test set. Results: The model was trained and validated on 97,566 internal studies, comprising 1,424,265 videos from 60,741 unique patients. The model achieved state of the art performance on the internal test set (16,396 studies), with mean absolute error (MAE) of 3.4% and root mean squared error (RMSE) of 4.6%. Multi-view results were superior to all single-view models. Model showed robust predictions on external test set (179,298 studies), with MAE of 5.6% and RMSE of 7.1% and outperformed EchoNet-Dynamic (Table). Conclusions: We developed a deep learning model trained on multiple echocardiographic views using the largest dataset to date. Our model achieved state-of-the-art accuracy in assessing LVEF with a level of agreement between the AI and cardiologist LVEF assessments comparable to cardiologist interobserver variability. Further studies are underway to study the implementation of these models within clinical systems.

Objective: Treatment-resistant depression (TRD) occurs in roughly one-third of all individuals with major depressive disorder (MDD). Although research has suggested a significant common variant genetic component of liability to TRD, with heritability estimated at 8% when compared with nontreatment-resistant MDD, no replicated genetic loci have been identified, and the genetic architecture of TRD remains unclear. A key barrier to this work has been the paucity of adequately powered cohorts for investigation, largely because of the challenge in prospectively investigating this phenotype. The objective of this study was to perform a wellpowered genetic study of TRD. Methods: Using receipt of electroconvulsive therapy (ECT) as a surrogate for TRD, the authors applied standard machine learning methods to electronic health record data to derive predicted probabilities of receiving ECT. These probabilities were then applied as a quantitative trait in a genome-wide association study of 154,433 genotyped patients across four large biobanks. Results: Heritability estimates ranged from 2% to 4.2%, and significant genetic overlap was observed with cognition, attention deficit hyperactivity disorder, schizophrenia, alcohol and smoking traits, and body mass index. Two genome-wide significant loci were identified, both previously implicated in metabolic traits, suggesting shared biology and potential pharmacological implications. Conclusions: This work provides support for the utility of estimation of disease probability for genomic investigation and provides insights into the genetic architecture and biology of TRD.

Abstract Background Valvular heart diseases, including aortic regurgitation (AR), mitral regurgitation (MR), and tricuspid regurgitation (TR) are significant contributors to cardiovascular morbidity and mortality worldwide. Artificial intelligence (AI) may improve echocardiographic assessment, but no models to date have successfully evaluated valvular regurgitation. Purpose The goal of this study was to develop an AI system to improve the accuracy of echocardiographic AR, MR, and TR interpretation. Methods Complete transthoracic echocardiogram (TTE) studies performed at a single center were split into train (n=65,301), validation (14,018), and test sets (13,746). The Deep Learning for Echo Analysis, Tracking, and Evaluation of Valvular Regurgitation (DELINEATE-Regurgitation) system intakes entire TTE studies, identifies color Doppler clips showing each valve, and yields a study level classification of AR, MR, and TR on a 6-grade scale (none/trace, mild, mild-moderate, moderate, moderate-severe, severe) using the cardiologist interpretation as a reference standard. This system used a hybrid neural network, leveraging the spatiotemporal feature learning capabilities of convolutional neural networks with the sequential processing strengths of transformer networks. The end-to-end hybrid approach allows the model to learn nuanced representations of the data including features within and across color Doppler videos. Model accuracy was assessed using quadratically weighted Cohen’s kappa (k), area under the receiver operator characteristic curve (AUROC) for the detection of both "moderate or greater" and "severe" regurgitation, and AI-cardiologist agreement the 6-grade scale. Additionally, an analysis was conducted using a 4-grade regurgitation scale (none/trace, mild, moderate, and severe) by rolling up intermediate classes (mild-moderate and moderate-severe) into the next higher class for both cardiologist determinations and model predictions. Results The DELINEATE-Regurgitation system demonstrated a high accuracy in the classification of AR (k=0.857), MR (k=0.867), and TR (k=0.847) (Figure 1). DELINEATE predicted the same severity as cardiologists (AR 91.8%, MR 79.9%, TR 77.7%) or within ±1 grade accuracy (AR 99.5%, MR 98.6%, TR 98.6%). The models had excellent binary discrimination in the detection of moderate or greater (AR AUROC 99.4, MR 98.6, TR 98.1) and severe (AR 99.7, MR 99.3, TR 99.6) regurgitation (Figure 2). In the 4-grade scale, the model had exact agreement with cardiologists in 92.3% of AR, 82.4% of MR, and 80.6% of TR cases. Panel adjudication of AI-cardiologist disagreement and correlation with quantitative metrics at higher levels of regurgitation will be presented. Conclusions The DELINEATE-Regurgitation AI system demonstrates a high level of accuracy in the assessment of AR, MR, and TR using color Doppler TTE videos. Future work will focus on automating quantification and developing strategies for optimal clinical deployment.Figure 1 - Confusion MatricesFigure 2 - ROC Curves

Purpose To estimate the association of psychiatric polygenic scores with healthcare utilization and comorbidity burden. Methods Observational cohort study (N = 118,882) of adolescent and adult biobank participants with linked electronic health records (EHRs) from three diverse study sites; (Massachusetts General Brigham, Vanderbilt University Medical Center, Geisinger). Polygenic scores (PGS) were derived from the largest available GWAS of major depressive depression, bipolar disorder, and schizophrenia at the time of analysis. Negative binomial regression models were used to estimate the association between each psychiatric PGS and healthcare utilization and comorbidity burden. Healthcare utilization was measured as frequency of emergency department (ED), inpatient (IP), and outpatient (OP) visits. Comorbidity burden was defined by the Elixhauser Comorbidity Index and the Charlson Comorbidity Index. Results Participants had a median follow-up duration of 12 years in the EHR. Individuals in the top decile of polygenic score for major depressive disorder had significantly more ED visits (RR=1.22, 95% CI; 1.17, 1.29) compared to those the lowest decile. Increases were also observed with IP and comorbidity burden. Among those diagnosed with depression and in the highest decile of the PGS, there was an increase in all utilization types (ED: RR=1.56, 95% CI 1.41, 1.72; OP: RR=1.16, 95% CI 1.08, 1.24; IP: RR=1.23, 95% CI 1.12, 1.36) post-diagnosis. No clinically significant results were observed with bipolar and schizophrenia polygenic scores. Conclusions Polygenic score for depression is modestly associated with increased healthcare resource utilization and comorbidity burden, in the absence of diagnosis. Following a diagnosis of depression, the PGS was associated with further increases in healthcare utilization. These findings suggest that depression genetic risk is associated with utilization and burden of chronic disease in real-world settings.

EHR-linked biobanks provide longitudinal data on diverse clinical populations that can fuel clinical, genetic, and epidemiologic research. Assigning case status is an important step in conducting psychiatric research using EHR, however it is often defined using phenotypic algorithms which have been developed from majority white individuals. Bias in assigning diagnosis codes across groups could lead to poorer performance and exacerbate disparities.

Introduction: Early detection of structural heart diseases (SHD) can help improve patient outcomes. Electrocardiogram (ECG) based deep learning (DL) models have the potential to identify patients (pts) with undiagnosed disease in need of further testing such as echocardiography. Most models have focused on detecting a single abnormality. A few studies have attempted detection on a comprehensive aggregate set of outcomes, but with limited scope and external validation. Hypothesis: An ECG DL model can identify pts at risk for SHDs, with stable external validation performance regardless of hospital type (academic/community), clinical setting (inpatient/outpatient/emergency department), or demographics (race/ethnicity). Methods: We identified 426,659 ECG-echo study pairs since 2008 from 5 NewYork-Presbyterian-affiliated (NYP) hospitals, including Columbia, to train the EchoNext DL model. ECGs were associated with SHD if acquired within one year prior to an echo reporting left ventricular (LV) ejection fraction ≤45%, LV wall thickness ≥1.3cm, moderate or severe valvular disease, moderate or large pericardial effusion, moderate or severe right ventricular dysfunction, or pulmonary artery systolic pressure ≥45 mmHg. EchoNext was tested on unseen data from four other NYP hospitals, including Cornell. Results: In the external dataset (n=21,421 pts), 33% were identified with SHD. EchoNext achieved an area under the receiver operating characteristic curve (AUROC) of 0.85, with robust results with respect to individual hospital (AUROC range: 0.82-0.87), clinical setting (0.78-0.84), and patient race/ethnicity (0.84-0.86; Figure). Conclusion: EchoNext demonstrated excellent performance in detecting SHD in the largest multi-institutional, multi-racial external validation of such a model to date. This model may allow AI-enabled reflex referral to echocardiography from ECG. Prospective trials are ongoing to validate.

Introduction: Early detection of structural heart diseases (SHD) can help improve patient outcomes. Electrocardiogram (ECG) based deep learning (DL) models have the potential to identify patients (pts) with undiagnosed disease in need of further testing such as echocardiography. Most models have focused on detecting a single abnormality. A few studies have attempted detection on a comprehensive aggregate set of outcomes, but with limited scope and external validation. Hypothesis: An ECG DL model can identify pts at risk for SHDs, with stable external validation performance regardless of hospital type (academic/community), clinical setting (inpatient/outpatient/emergency department), or demographics (race/ethnicity). Methods: We identified 426,659 ECG-echo study pairs since 2008 from 5 NewYork-Presbyterian-affiliated (NYP) hospitals, including Columbia, to train the EchoNext DL model. ECGs were associated with SHD if acquired within one year prior to an echo reporting left ventricular (LV) ejection fraction ≤45%, LV wall thickness ≥1.3cm, moderate or severe valvular disease, moderate or large pericardial effusion, moderate or severe right ventricular dysfunction, or pulmonary artery systolic pressure ≥45 mmHg. EchoNext was tested on unseen data from four other NYP hospitals, including Cornell. Results: In the external dataset (n=21,421 pts), 33% were identified with SHD. EchoNext achieved an area under the receiver operating characteristic curve (AUROC) of 0.85, with robust results with respect to individual hospital (AUROC range: 0.82-0.87), clinical setting (0.78-0.84), and patient race/ethnicity (0.84-0.86; Figure). Conclusion: EchoNext demonstrated excellent performance in detecting SHD in the largest multi-institutional, multi-racial external validation of such a model to date. This model may allow AI-enabled reflex referral to echocardiography from ECG. Prospective trials are ongoing to validate.