{"paper_id":"43dba3e8-afb9-4aa2-bf6c-2cf3764f1bb6","body_text":"An Artificial Intelligence Model to Detect Abnormal Ejection Fraction from Non-Contrast Chest Computed Tomography: The CT-LVEF study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article An Artificial Intelligence Model to Detect Abnormal Ejection Fraction from Non-Contrast Chest Computed Tomography: The CT-LVEF study Jayant Raikhelkar, Zilong Bai, Ashley Beecy, Fengbei Liu, Nusrat Nizam, and 16 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5677688/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Heart failure (HF), a major global health challenge, affects millions worldwide and poses substantial healthcare and economic burdens. The left ventricular ejection fraction (LVEF) is a critical dynamic parameter used to characterize HF and guide treatment. In this study, we developed and validated an artificial intelligence (AI) model capable of predicting abnormal LVEF directly from static, non-gated, non-contrast chest computed tomography (CT) scans, a novel application for an imaging modality typically used for unrelated indications. Using a multi-institutional dataset of 34,058 paired CT and echocardiogram studies from two academic centers, we trained our model on over 25,000 studies and validated it on 8,110 studies from a separate institution. Remarkably, our model demonstrated robust performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.786 on the hold-out test set and 0.755 on external validation. Our findings are particularly promising given the widespread availability of CT scans—over 80 million performed annually in the U.S.—making this opportunistic screening approach highly practical. Beyond strong predictive performance, the AI model outperformed expert radiologists in both accuracy and efficiency and provided interpretable visualizations highlighting imaging features linked to reduced LVEF. By enabling the identification of HF from routine chest CTs performed for other indications, this technology holds significant promise for early detection, reducing the diagnostic gap, and improving outcomes in asymptomatic HF. Health sciences/Diseases/Cardiovascular diseases/Heart failure Health sciences/Health care/Medical imaging Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Heart Failure (HF) has been declared a global pandemic with over 64 million individuals afflicted worldwide. 1 About 6.7 million of these individuals are in the USA and this number is expected to rise to 8.5 million by 2030. 1 , 2 Despite improvements in medical and surgical treatment, HF remains a leading cause of hospitalizations and mortality. 3 In addition, the worldwide economic burden of HF is estimated to be $ 346 billion US dollars annually. 4 The left ventricular ejection fraction (EF) is a critical parameter which is used to phenotype heart failure into several types, namely heart failure with reduced EF (HFrEF), Heart Failure with mid-range EF( HFmrEF) and heart failure with preserved EF (HFpEF). 5 There is evidence for early initiation of lifestyle modifications and medications that constitute guideline directed medical therapy (GDMT) for both HfrEF and HFmrHF to decrease both HF hospitalizations and mortality. 6 – 8 It is estimated that a significant number of patients remain undiagnosed with HfmrEF and HfrEF and thus are unable to be started on appropriate HF therapies. 2 There is an unmet need for earlier detection of these patients with the goal of improving outcomes by means of initiation of lifestyle modification and GDMT at an earlier stage of disease. Computed tomography (CT) scans are a common medical imaging modality, with over 80 million studies being performed yearly in the USA 9 . Chest CT scans are frequently performed in both inpatient and outpatient settings for a wide range of indications, such as screening for lung cancer, acute pulmonary embolism detection, evaluation for infectious diseases like pneumonia and trauma. These studies contain a wealth of biometric information, and several studies have recently investigated the use of artificial intelligence (AI) based automated algorithms applied to CT scans for identifying patients at higher risk for adverse events, including cardiovascular adverse events 10 , 11 . In this study, we developed and validated a novel AI model capable of predicting abnormal left-ventricle ejection fraction (LVEF), a dynamic physiological parameter, directly from non-gated, non-contrast chest CT scans. This represents a surprising and transformative application, as CT scans are typically used for static imaging of non-cardiac structures. Using a multi-institutional dataset of over 34,000 paired CT and echocardiogram studies, our model demonstrated robust performance, achieving AUROCs of 0.786 on the primary test set and 0.755 in external validation. Beyond strong predictive performance, the AI model outperformed radiologists in both accuracy and efficiency and provided interpretable visualizations highlighting imaging biomarkers linked to reduced LVEF. By unlocking new diagnostic possibilities from one of the most commonly performed imaging modalities our approach reveals the potential of opportunistic screening for heart failure. Our findings underscore how this technology could bridge the diagnostic gap in asymptomatic HF, enabling earlier detection, personalized interventions, and improved patient outcomes. Results Dataset curation and study design We curated a multimodal dataset of paired chest CT imaging data and echocardiogram (echo) study reports from Columbia University Irving Medical Center (CU) and Weill Cornell Medical Center (WCM). We focused on non-contrast, non-gated chest CT volumetric data in this study. The CT images and left ventricle ejection fraction (LVEF) values derived from echo reports were paired by patient using our data linkage protocol (e.g., closest in study times within 6 months, Supplementary Fig. 1). The dataset included 25,948 CT-echo pairs from the Columbia University (CU) cohort and 8,110 from the Weill Cornell Medicine (WCM) cohort, with CT images serving as inputs to the AI model and LVEF derived from echo reports providing the target variable. CT-echo pairs were matched based on proximity in study times (within six months). Using an LVEF threshold of 50%, we categorized patients into normal (LVEF ≥ 50%) and abnormal (LVEF < 50%) groups, with 15.94% of pairs classified as abnormal in the CU cohort and 17.44% in the WCM cohort. This curated dataset covered diverse demographics, with a balanced gender distribution of male (n = 14,452, 55.69% of CU, n = 4,364, 53.81% of WCM) and female (n = 11,995, 44.31% of CU, n = 4,115, 46.19% of WCM) patients. We divided the study cohort into three age groups: 18–40, 40–65, and 65–120 years. We also investigated three racial groups: Asian/Native Hawaiian/Other Pacific Islands (ANO), White, and Black or African American (BAA). Table 1 presents the detailed summary statistics for study pairs, stratified by the institution of CT data collection and patient demographics, and compared normal vs abnormal EF status. Table 1 Population Description. The datasets analyzed are compiled from Columbia University Irving Medical Center (CU) and Weill Cornell Medicine (WCM). Each instance in the study consists of an input-label pair, where the input is preprocessed non-contrast chest CT volumetric data, and the label reflects normal or abnormal left ventricle ejection fraction (LVEF) status derived from the associated echocardiography report. The AI model is trained on 80% of the instances with CT images from the Columbia dataset. Subpopulation statistics are calculated by stratifying the data according to the sites where the CT studies were collected and the demographics of the patients. The numbers indicate the pairs with CTs from different sites. One patient may be associated with different LVEF statuses by different CTs. * Percentages for Patient Sex or Race subgroups do not sum to 100% due to additional categories with insufficient sample sizes. These categories, labeled as “Others” or “Not provided,” were omitted from the population statistics for clarity. Chi-square test of independence was conducted to compare the distributional differences between CU and WCM in terms of patient sex, age groups, and race. The analyses were performed on all patients as well as the normal and abnormal subgroups for each variable. Columbia University Irving Medical Center Total: 25,948 chest CTs Weill Cornell Medical College Total: 8,110 chest CTs Between-site Chi-square test of independence P-value Normal (LVEF ≥ 50) - 21,787 (83.96%) Abnormal (LVEF < 50) − 4,161 (16.04%) Normal (LVEF ≥ 50) − 6695 (82.70%) Abnormal (LVEF < 50) − 1415 (17.29%) Overall Normal Abnormal Patient Sex* Male 10351 (47.51%) 2616 (62.87%) 3207 (47.90%) 942 (66.57%) < 0.0001 < 0.0001 0.5784 Female 9799 (44.98%) 1278 (30.71%) 3486 (52.07%) 472 (33.36%) Patient Age(years) 18–39 2011 (9.23%) 218 (5.23%) 351 (5.24%) 56 (3.96%) < 0.0001 < 0.0001 0.0102 40–64 8276 (37.99%) 1249 (30.02%) 2204 (32.92%) 387 (27.35%) 65 and above 11500 (52.78%) 2694 (64.74%) 4140 (61.84%) 972 (68.69%) Patient Race* Asian/Native Hawaiian/Other Pacific Islander 784 (3.60%) 90 (2.16%) 403 (6.02%) 90 (6.36%) < 0.0001 < 0.0001 < 0.0001 White 10761 (49.39%) 1754 (42.15%) 3596 (53.71%) 699 (42.33%) Black or African American 3140 (14.41%) 654 (15.72%) 897 (13.40%) 212 (14.98%) For CT data acquisition, non-contrast non-gated CT imaging data from the two institutions were collected using equipment from different vendors. In the CU cohort, CT studies were distributed among four vendors as follows: TOSHIBA (4,142), SIEMENS (1,487), Canon Medical Systems (351), and GE Medical Systems (19,968). In the WCM cohort, CT studies were distributed between two vendors as: SIEMENS (105) and GE Medical Systems (8,005). Volumetric CT scans, performed along the axial orientation and focused on the chest region, encompassing the entire cardiac structure and including marginal portions of the lower neck and upper abdomen, were selected. We used CT volumetric data collected through a smooth kernel (See Supplementary Table 1 for thresholds of smooth vs sharp kernels). AI Model Training and Evaluation We used the study pairs from the CU cohort as the primary dataset to develop our AI-based models, and those from the WCM cohort for external validation. These models were designed to predict the binarized LVEF status of a patient using their CT volume data. We performed a 70%-10%-20% train-validation-test split, divided by patients, within the primary dataset from the CU cohort. This split was used to train an ensemble of vision-transformer (ViT) neural networks 12 (AI model), initialized with pretrained weights, on the training set (see Methods for details). We then evaluated the model's performance on the test subset using accuracy and F1 score. This involved using a total of 20,774 study pairs from the primary dataset for the AI model training and 5,174 study pairs as the hold-out test set for performance evaluation. We randomly selected 100 study pairs from the CU hold-out test set and 100 study pairs from the WCM external validation set for radiologist review to compare with our AI-based model prediction results (see Methods). On the CU hold-out test set, the AI model achieved an AUROC of 0.786 (95% confidence interval (CI): 0.772–0.801) and F1 score of 0.817 (95% CI: 0.807–0.827) for predicting LVEF-derived status labels (n = 5,174, 20%) (Fig. 2 ). We investigated the sensitivity of our model given different levels of specificity (Supplementary Table 2). Performance metrics were further analyzed across demographic subgroups. (Table c in Fig. 2 ). The number of instances in each demographic subgroup is summarized in Table 1 . For each demographic attribute, our model performed best within the male subgroup with an AUROC of 0.790 (95% CI: 0.777–0.802), within the senior (65 ~ 120 years) age subgroup with an AUROC of 0.791 (95% CI: 0.778–0.804), and within the ANO racial subgroup with an AUROC of 0.792 (95% CI: 0.742–0.842)(Table b in Fig. 2 ). For each demographic subgroup, we further investigated the performance of our AI model with respect to LVEF measurements from different time intervals. In particular, we studied the CT-to-echo study time interval being 1 day, 1 week and 1 month. In each demographic subgroup, our model performed the best for 1-day interval targets, compared to the 1-week and 1-month scenarios, as well as the entire demographic subgroup. This pattern was consistent across all the demographic subgroups (Supplementary Table 3). External validation To evaluate the generalizability of our AI model across different institutions with potentially varied CT data collection protocols, we validated using external data from WCM. This dataset consisted of n = 6,695 normal study pairs with echo-derived LVEF ≥ 50% and n = 1,415 abnormal study pairs with echo-derived LVEF < 50%. Our AI model achieved an AUROC of 0.755 (95% CI: 0.741–0.769) and an F1 score of 0.816 (95% CI: 0.806–0.825). For a range of sensitivity and specificity scores on the full ROC curve, see Supplementary Table 2. In the demographic-stratified analysis, our AI model exhibited a similar pattern to that observed on the CU hold-out test set. Specifically, our model performed best within the male subgroup with an AUROC of 0.748 (95% CI: 0.73–0.767), within the senior (65 and above) age subgroup with an AUROC of 0.752 (95% CI: 0.733–0.771), and within the ANO racial subgroup with an AUROC of 0.768 (95% CI: 0.721–0.814) (Table c in Fig. 2 ). These findings highlight the model's ability to maintain strong predictive performance across diverse demographic subgroups. When analyzing performance based on the time interval between CT and echo studies, the model consistently performed best for pairs with a 1-day interval, compared to those with 1-week or 1-month intervals, as well as the broader subgroup without time-interval-specific selection. This pattern was observed across all demographic subgroups on the WCM validation set (Supplementary Table 3). These results emphasize the importance of temporal proximity between imaging and echocardiography in achieving optimal prediction accuracy. Model interpretation Model Interpretability for Imaging Biomarker Discovery We employed guided gradient-weighted class activation mapping (Grad-CAM) 13 , a visualization technique that highlights the most influential regions in the input data. This method generates heatmaps, providing interpretable insights into our AI model’s decision-making. We applied Grad-CAM to the CT volume data from both the CU hold-out test set and the WCM external validation set. Figure 4 displays representative heatmaps overlaid on axial CT slices, illustrating the regions of greatest model activation. For detailed analysis, we randomly selected 10 normal subjects (LVEF ≥ 50% in their echo report) and 9 abnormal subjects (LVEF < 50% in their echo report) from the hold-out CU test set, and conducted rigorous manual examinations with a heart failure cardiologist and a thoracic radiologist. For the abnormal EF patients, the Grad-CAM heatmaps generally highlighted the lower neck, right or left lung, and chest wall (Fig. 4 ). This pattern was consistently observed across the subjects with abnormal EF in the CU test set (Supplementary Data 1). The Grad-CAM heatmaps of normal cases are included in Supplementary Data 2 and did not reveal an anatomically meaningful pattern. These observations highlight the potential of Grad-CAM not only to enhance AI model interpretability but also to uncover previously unrecognized imaging biomarkers that could provide novel insights into heart failure pathophysiology. Comparison with radiologist manual analysis and inference We compared LVEF status predictions made by our AI model against those made by dedicated thoracic radiologists using CT imaging data to benchmark accuracy, ensure clinical relevance, and identify strengths and weaknesses. Specifically, we randomly selected 90 non-contrast CT scans from the CU hold-out test set and 100 from the WCM external validation set, with similar rates of abnormal EF status as their respective full test sets. LVEF status information was withheld before sending the patient lists to the radiologists to ensure unbiased manual interpretation. The manual prediction process was done by two highly experienced thoracic radiologists, one (RAD1) from CU and the other (RAD2) from WCM. Both radiologists independently reviewed the CU 90 CT scans and made predictions regarding LVEF status. RAD2 also reviewed the 100 WCM scans. Using the LVEF status derived from echo reports as the ground-truth, our AI model outperformed the radiologists at both sites. On the CU scans, the AI model achieved an F1 score of 0.83, compared to 0.645 for RAD1 and 0.8 for RAD2. On the WCM scans, the AI model’s F1 score was 0.833, surpassing RAD2’s score of 0.622 (Table 2 ). Table 2 Performance comparison between our AI model and radiologists. Performance evaluation using F1 score for RAD1, RAD2 and our AI model on randomly selected CT scans from the CU test set and the WCM external validation set. CU N = 90 sampled CT scans WCM N = 100 sampled CT scans RAD1 0.645 - RAD2 0.8 0.622 AI Model 0.83 0.833 This is achieved by our AI model at a faster speed of analysis and inference: our AI model completed LVEF status prediction tasks within approximately 1 minute on the 90 CU CT volumetric data and the 100 WCM samples, respectively. In contrast, it took RAD1 ~ 2.2m per CT scan to interpret 90 CT volumetric data sampled from the CU hold-out test set, and ~ 2m per CT scan by RAD2 on the 100 WCM scans sampled from the external validation set. Detailed error analysis revealed complementary strengths between the AI model and radiologists. Among the 90 CU scans, the numbers of incorrect predictions were as follows: 13 by the AI model, 37 by RAD1, and 18 by RAD2. Notably, 10 errors overlapped between the AI model and one radiologist, while only 6 errors were shared by the AI model and both radiologists. On the WCM scans, the AI model made 15 incorrect predictions, compared to 43 by RAD2. Errors shared by the AI model and RAD2 totaled 7. These patterns suggest that the AI model and radiologists leverage different aspects of the imaging data, potentially enabling complementary interpretations in clinical practice. We summarized the false-positive and false-negative predictions of our AI model and the two radiologists as Venn diagrams in Fig. 4 , offering further insights into the nature of the errors. Validation on independent consecutive test sets To further evaluate the performance of our developed AI model in a real-world clinical setting, we constructed two independent testing sets using the CT studies and Echo reports collected from an overall 2,228 subjects admitted to CU (n = 1,411 subjects) and WCM (n = 817 subjects) of the NYP healthcare systems. The CU data spanned August to December 2023, while the WCM data covered August 2023 to July 2024. These independent testing sets included subjects that were not used for developing our AI model. Their CT and echo data were collected after the primary (retrospective) dataset, ensuring temporal separation from the training data. Evaluation of AI model on CU independent consecutive test set. From the 1,895 subjects with 2,417 CT studies and 4,448 Echo reports collected from CU, we curated a CU independent consecutive test set for the LVEF status prediction model comprising 1,251 subjects with 1,411 pairs of CT-echo studies. Filtering and pairing protocols excluded 484 CT studies and 1,006 echo reports to ensure data quality and consistency (see Methods). Notably, the LVEF status distributions stratified by the patient demographics in the CU prospective test set (Supplementary Table 4) were highly consistent with the ones on the primary dataset for our AI model development. The CT series were distributed among 3 vendors as follows: GE MEDICAL SYSTEMS (1,141), Canon Medical Systems (196), SIEMENS (74). Our AI model trained on the CU primary dataset demonstrated robust performance on the CU independent consecutive test set, achieving an AUROC of 0.784 (95% CI: 0.762–0.805), an F1 score of 0.854 (95% CI: 0.846–0.862) and a balanced accuracy of 0.594 (95% CI: 0.577–0.611) for LVEF status prediction. Performance metrics stratified by age, gender, and race are presented in Supplementary Fig. 1c. Evaluation of AI model on WCM independent consecutive test set. Using the 1,201 subjects with 1,366 WCM CT studies and their associated 2,377 Echo study reports, we curated a WCM prospective test set for the LVEF status prediction model comprising 755 subjects with 817 pairs of CT-echo studies. Based on our filtering and pairing protocols (see Methods), 549 CT studies and 1,560 Echo study reports were excluded. Notably, the LVEF status distributions stratified by the patient demographics in the WCM independent consecutive test set (Supplementary Table 4) are highly consistent with the ones on the WCM external validation dataset (Supplementary Table 4). The CT series were distributed among two vendors as follows: GE Medical Systems (737) and SIEMENS (80). Our AI model trained on the CU primary dataset demonstrated robust performance on the WCM prospective test set, achieving an AUROC of 0.775 (95% CI: 0.756–0.794), an F1 score of 0.855 (95% CI: 0.849–0.861) and a balanced accuracy of 0.625 (95% CI: 0.606–0.644) for LVEF status prediction. The model performance is detailed in the Supplementary Fig. 1c stratified by age, gender and race. Across both independent consecutive test sets, the AI model demonstrated strong and consistent performance, closely aligning with results from retrospective datasets. Importantly, these findings were consistent across demographic subgroups, imaging sites, and CT vendors, further supporting the robustness and generalizability of the model for LVEF status prediction in diverse real-world settings. Opportunistic Screening We continued to pull 712 chest CT studies of 702 patients, without requiring paired echocardiograms, to assess our model across a broader patient population undergoing routine CT scans. After filtering and preprocessing based on the DICOM metadata and imaging files of the CT data (see Methods), we obtain 633 studies of 627 patients. Our model identified 23 patients for further evaluation. Discussion In this study, we demonstrate that an AI model can accurately detect abnormal LVEF using non-contrast, non-gated chest CTs ordered for standard clinical indications from two separate sites. The model achieved AUROCs of 0.786 at CU and 0.755 at WCM, on retrospective test data. In our independent consecutive test set study with data collected after August 2023, our AI model demonstrated similar performance, achieving AUROC 0.784 of CU, 0.775 of WCM. To our knowledge this is the first reported study to identify abnormal LVEF, which is traditionally detected from targeted cardiac testing such as echocardiography, multi-gated acquisition scan (MUGA) or cardiac MRI, from standard non gated non contrast chest CTs ordered for other medical indications. The lifetime risk of developing HF is now estimated to be 24% and the prevalence 1.9–2.6% 2 . Unfortunately, initial gains in heart failure management in the US have reversed with mortality rates due to HF higher than those in 1999, thought to be driven by increasing incidence of obesity, diabetes mellitus and hypertension. 14 . About 24–34% of patients are estimated to have Stage B HF (asymptomatic LV dysfunction). 2 Lifestyle modifications and GDMT have been shown to improve quality of life as well as mortality in this cohort of patients. 15 . There is an unmet need to detect these patients earlier to initiate lifestyle modifications and GDMT. There are currently no clinical screening guidelines for HF recommended in the general population, likely due to the clinical cost of screening using targeted cardiac testing. Opportunistic screening has been described as the practice of systematically leveraging imaging data that are incidental to the clinical indication of study 16 . There have been several studies that have attempted to identify abnormal LV structure or function using opportunistic screening. Several studies have used ECG based AI algorithms to detect LV dysfunction and risk of HF. 17 – 21 . Bhave et al used a deep learning model to identify patients with severe left ventricular hypertrophy or dilated left ventricle, harbingers of HF, from chest X rays with a composite AUC of 0.79. 22 CT provides detailed cross sectional information and a wealth of biometric data compared to chest X-rays. In a study by Miller et al a combined AI model was able to quantify coronary calcium, left atrial volume and left ventricular mass index to predict cardiac death or MI with an AUC of 0.792. 10 To our knowledge our study is the first to utilize AI to determine abnormal LVEF. LVEF has been traditionally determined by cardiac specific testing in which cardiac dimensions are able to be calculated in both systole and diastole such as in echocardiography or cardiac MRI to calculate the percentage of blood ejected with each beat. Calculation of EF using non- gated, non-contrast chest CT presents a novel application of non-gated chest CT that has not been explored previously, and in a previous era potentially not thought to be possible. It is estimated that over 15 million chest CTs are performed in the US yearly. 23 Opportunistic screening for abnormal EF may present a unique opportunity in the future for early identification of Stage B heart failure with a goal to initiate lifestyle measures and initiation of GDMT to prevent progression to Stage C and Stage D heart failure, heart failure admissions and mortality. Deploying this AI-based opportunistic screening tool to identify reduced LVEF introduces important considerations for healthcare delivery. The model, trained on retrospective data, was evaluated across various operating points (Supplementary Table 2). If we optimize for specificity, at the CU site, it achieved a specificity of 80% and a sensitivity of 67%, and at the WCM site, it reached a specificity of 77% and a sensitivity of 67%. Using the same threshold, applied across a broader patient population undergoing routine CT scans, without requiring paired echocardiograms, the model identified 23 out of 627 patients for further evaluation. These operating characteristics applied to the prevalence in the test population, equate to a number needed to evaluate (NNE) of 3. Integrating model outputs into standard CT reporting workflows can streamline adoption, particularly among radiologists and clinicians, although optimal implementation may necessitate new operational pathways, including centralized processes to ensure prompt referrals for confirmatory echocardiograms. Notably, 3.63% of patients flagged by the model received CT scans for indications that may not justify insurance coverage for follow-up echocardiography, potentially leading to significant out-of-pocket costs. This financial burden is concerning, given that only 37% of U.S. adults can cover an unexpected $ 400 expense without borrowing, making it vital to consider financial protections for lower-income patients. 24 , 25 Finally, training protocols should incorporate these healthcare delivery considerations alongside model information to ensure equitable and effective deployment across diverse populations. Our AI model outperformed 2 board certified thoracic radiologists in predicting LVEF status from CT imaging, achieving higher F1 scores on both the CU and WCM test sets Notably, error analysis revealed that the AI model and radiologists’ mistakes were largely non-overlapping, suggesting that their predictions captured different aspects of the imaging data. Averaging prediction scores from the AI model and radiologists yielded only modest improvements in AUROC (0.74 on CU, 0.737 on WCM) compared to radiologists but fell short compared to pure AI predictions (0.757 on CU, 0.763 on WCM), indicating limited synergy from a simple score combination. In this study we utilized Grad-CAM mapping to evaluate potential imaging biomarkers detected by the AI model to identify abnormal EF. The mapping of the heart itself appears intuitive as overt structural changes in cardiac structure are associated with LV dysfunction. However, in addition to the heart, biomarkers noted included the lower neck region, lower lung regions and chest wall. It is unclear why these regions predict abnormal EF at this stage and an intriguing question which merits further study. Possible explanations to consider could be early signs of vascular congestion (great vessels of the neck and lungs) and structural adaptations of the thoracic cavity in the setting of HF (chest wall). Our study has limitations, for which we outline potential areas for future work. Firstly, the applicability of our AI model is limited due to its retrospective nature. We plan to conduct a prospective study (in contrast to the independent consecutive test set study we conducted) with follow-up sessions to monitor EF status over time. In addition EF calculation by 2D echocardiography has been reported to show intraobserver and interobserver variations of 8–21% and 6–13% respectively 26 . Although we attempted to pair the echo report closely associated with chest CTs in time, variation in EF overtime is also well described. We plan to extend our study by considering different training data set building strategies when data collected from different real-life scenarios are available (e.g., sufficiently large cohort of patients with same day CT session followed by echo session). Our study only considered non-contrast CT scans to include a broader range of patients and increase dataset size. However, given the distinct features provided by contrast-enhanced CT, such as clearer visualization of blood vessels, we intend to expand our research by training the model on contrast-enhanced CT data. Additionally, we used Grad-CAM 13 , a saliency map approach, as the method for visualizing imaging biomarkers, applied to the individual scans. This offers qualitative interpretation in case studies but was limited in exploring population-level insights. We attribute this limitation to the lack of registration between different CT scans for comparing between different CT scans at the voxel-level, and plan to address it by integrating advanced CT registration approaches 27 in our future research. In conclusion, we developed and validated an AI model that can successfully predict abnormal LVEF status, directly from non-gated non-contrast CT volumetric imaging data. Our study curated a large multi-modality dataset of paired CT-echo studies based on non-contrast CT scans and echo study reports, using data collected from two independent academic medical centers. As the first study of such cross-modality prediction of cardiac disease indicators using chest CT, our model demonstrated promising performance in both the internal validation and the external validation. Our AI model also outperformed experienced radiologists as human interpreters on selected subsets from both cohorts. We also identified possible imaging biomarkers through study of selected representative CT scans overlaid by saliency maps. This initiative provides empirical and numerical evidence supporting the use of AI-based opportunistic screening in real-world clinical settings to facilitate earlier diagnosis and possible intervention for patients with HfmEF and HFrEF. By analyzing chest CT imaging data, this approach may not only optimize diagnostic accuracy for structural cardiac diseases but also lead to the discovery of novel biomarkers in structural features, which could significantly improve patient care and deepen our understanding of cardiac pathophysiology. Methods Ethics approval This study was approved by the institutional review boards (IRB) at Weill Cornell Medicine and Columbia University Irving Medical Center. A waiver for informed consent was obtained. Data Curation and Preprocessing All patients aged 18 years or older who had at least one echocardiogram (echo) within 6 months before or after a CT session at either Columbia University Irving Medical Center (CU) or Weill Cornell Medical Center (WCM) between July 2005 and August 2023 were identified. This initial data collection process identified 31,576 patients. The echo study reports and CT imaging data were then filtered, linked, and preprocessed according to the protocols detailed in the following sections. Eventually, we identified a cohort of 19,410 patients with a total of 34,058 qualified echo-CT data pairs. From this cohort, two study groups were formed based on the source of the CT data: CU (14,083 unique patients with 25,948 echo-CT pairs) and WCM (5,327 unique patients with 8,110 echo-CT pairs). The CU group served as the primary dataset for AI model development, while the WCM group was reserved for external validation. The CU dataset was further randomly split at the patient level into training, validation, and test subsets (i.e., each patient was assigned to only 1 of these 3 subsets). For our AI models, we created CT-echo study pairs using a CT volume image as the input and the LVEF value derived from an echo report as the target label for each pair. These pairs are formed and curated through data linkage , data filtering (see Supplementary Materials) , and the 3D CT volume preprocessing . The data curation and preprocessing flowchart is in Fig. 5 . 3D non-contrast chest CT volume preprocessing DICOM images of CT scans were transformed into Hounsfield Units (HU) to provide a consistent and standard measurement for data interpretation and analysis. The 2D DICOM images of the same CT series were merged into a 3D Numpy array to represent this 3D CT volume.. A window of [-1,000, 1,000] was applied to the intensity values to exclude air (<-1,000 HU) and bones (> 1,000 HU). The interpolation function from the Python library Scipy ndimage 28 was used to rescale the 3D voxel spacing of all the 3D CT volumes into a standardized resolution of \\(\\:2\\times\\:2\\times\\:2\\:m{m}^{3}\\) ,where the spline interpolation was applied. Finally, 3D cropping and zero padding were applied where applicable, to resize the 3D volume into a volume of size 164 x 164 x 164. AI model: Classifier based on a Pretrained CT-ViT Encoder We used a vision transformer architecture 12 : the encoder of the CT-ViT framework from GenerateCT 29 , as our backbone model for feature learning from the 3D CT images. We utilized the pre-trained weights of this encoder 29 and further trained it on our training cohort where both CT and echo studies were from CU. This encoder takes a preprocessed 3D CT volume image of size 164 x 164 x 164 and randomly cropped into \\(\\:164\\times\\:144\\times\\:144\\) (dimensions are depth, height and width: \\(\\:Z\\times\\:H\\times\\:W\\) ) as input and output a \\(\\:512\\) -dimensional feature vector. The CT-ViT encoder consists of three modules: the patch embedding layer, the spatial transformer module and the causal transformer module. The patch embedding layer first extracts non-overlapping patches of \\(\\:2\\times\\:16\\times\\:16\\) from the 3D CT volume input. Each patch is then transformed into a \\(\\:512\\) -dimensional feature vector (dimension in tensor: \\(\\:D\\) ) with a fully connected layer. This transformation yields a \\(\\:\\left(\\frac{164}{2}\\right)\\:\\times\\:\\left(\\frac{144}{16}\\right)\\times\\:\\left(\\frac{144}{16}\\right)\\:\\times\\:512\\) feature tensor for each input 3D image. This tensor is then fed into the spatial and temporal transformer. Subsequently in the spatial transformer module, multiple transformer layers were applied with self-attention along the spatial dimensions (i.e., W and H) of the reshaped tensor of \\(\\:\\left(\\frac{164}{2}\\right)\\:\\times\\:(\\frac{144}{16}\\times\\:\\frac{144}{16})\\:\\times\\:512\\) (dimensions: \\(\\:Z\\times\\:(H\\times\\:W)\\times\\:D\\) ). This is followed by the causal transformer module where multiple transformer layers were applied over the temporal (i.e., Z) dimension of the reshaped tensor of \\(\\:(\\frac{144}{16}\\times\\:\\frac{144}{16})\\:\\times\\:\\left(\\frac{164}{2}\\right)\\:\\times\\:512\\) (dimensions: \\(\\:(H\\times\\:W)\\times\\:Z\\times\\:D\\) ) with causal self-attention such that each spatial token only observes spatial tokens from previous slices in an auto-regressive manner. The dimensionality of output is retained after each spatial and causal transformer layer, ensuring that the volumetric information is preserved throughout the model fine-tuning. Finally, an average pooling over spatial (W and H) and temporal (Z) dimension of output tensor was applied to obtain the output \\(\\:512\\) -dimensional feature vector for each input. A fully connected layer is used as a classification head and outputs a univariate prediction as a probability for the input 3D image to be a positive sample. For predicting binarized LVEF, the AI model were trained to minimize the Binary Cross Entropy loss between the prediction and the binarized LVEF derived from echo report in Eq. 1 , where the i-th input preprocessed 3D CT volume is denoted \\(\\:{x}_{i}\\) , its binarized LVEF derived from echo report \\(\\:{y}_{i}\\) , and the prediction made by our framework based on CT-ViT encoder \\(\\:{\\widehat{y}}_{i}\\) , the total number of training instances \\(\\:N\\) . $$\\:{L}_{BCE}({x}_{i},{y}_{i},\\:{\\widehat{y}}_{i})\\:=\\:-{\\sum\\:}_{i=1}^{N}{y}_{i}log{(\\widehat{y}}_{i})\\:+\\:(1\\:-\\:{y}_{i}\\left)log\\right(1-{\\widehat{y}}_{i})$$ 1 We used the AdamW optimizer with an initial learning rate of \\(\\:1{0}^{-5}\\) , weight decay of \\(\\:1{0}^{-4}\\) and batch size of 8 for 20 epochs. The learning rate decayed to half at epoch 15. During training, random cropping from \\(\\:164\\times\\:164\\times\\:164\\:\\) to \\(\\:164\\times\\:144\\times\\:144\\) and random horizontal flipping were used for data augmentation. During testing, center cropping from \\(\\:164\\times\\:164\\times\\:164\\:\\) to \\(\\:164\\times\\:144\\times\\:144\\) were used. To assess the model's behavior and reliability, we trained it using the CU training and validation sets five times using different random seeds, resulting in five distinct sets of model weights. The AI model was trained and evaluated with a single NVIDIA A100 GPU. Evaluation metrics LVEF-based classification metric LVEF values were collected from the echo report to form labels for both training and testing. They were binarized according to a clinician-determined threshold (50%), where higher than or equal to this threshold is considered normal and lower than this threshold is considered to be indication of reduced ejection fraction (HFmrEF and/or HFrEF). A CT-echo pair where the echo-derived LVEF \\(\\:<50\\%\\) was defined as the “positive” class for calculation of ROC, AUROC, sensitivity, specificity, accuracy and balanced accuracy. In addition, we evaluated the LVEF-based classification performance stratified by sites and demographics on the mixed-site test cohort. Particularly, site-wise we investigated CU-CT and WCM-echo, WCM-CT and CU-echo, as well as WCM-CT and WCM-echo. Each sub-cohort is further stratified by demographics, such as patient age, gender and race. In each stratified test subset, we investigated the LVEF-classification performance regarding different ordering of CT and echo studies in each pair - i.e., CT followed by echo or echo followed by CT, and different time gaps between the two studies - i.e., same day, one week, 1 month. Refer to the Supplementary Table S2 for the detailed breakdown of the number of study pairs in each site-specific test sub cohort. Interpretation of the Model Visualization of contributive regions of AI model To understand the predictions made by the deep learning model's predictions, we used Gradient-weighted Class Activation Mapping (Grad-CAM) 13 , a popular technique for visualizing the regions within an input image that contributes most significantly to the model's output prediction. Grad-CAM computes the gradient of the target class with respect to feature maps and generates a heatmap overlay that visually indicates which parts of the image are most influential in the model's decision-making process. Radiologist manual prediction of EF status from CT scans In predicting (EF) status from CT imaging data, the radiologists reported to focus on LV chamber dilatation and various ancillary findings. Vue PACS, a software product by Philips, was used for viewing and interpreting CT scans. While Vue PACS allows for detailed imaging review and manual measurements, it does not integrate AI-driven tools for evaluating chamber dilatation or advanced cardiac metrics, such as left atrial or ventricular size. The radiologists primarily assessed whether the left-sided heart chambers were significantly enlarged. They also looked for signs of calcified apical aneurysms or fatty infiltration in the apex, which suggest past myocardial infarction and scarring, further indicating compromised EF, or sternal wires indicating prior sternotomy. Additionally, they examined features such as pleural and pericardial effusions, interlobular septal thickening (a sign of interstitial edema), and lung parenchymal edema. These findings, along with cardiac dilatation, suggested heart failure and support the prediction of an abnormal EF. Although Vue PACS facilitates these evaluations, the radiologists emphasized the subjective nature of the predictions, as the software does not provide automated tools for assessing chamber size or making advanced cardiac measurements. Many patients with altered EF may not show overt signs on non-contrast CT scans. While heavy coronary artery calcification might suggest myocardial damage, it does not always correlate with reduced EF. The radiologists relied on their expertise in combining manual observations within Vue PACS but acknowledged that echocardiography or cardiac MRI provide more precise assessments of EF and are the standard of care. Statistical analysis The performance of the AI model, with the output being probability scores, were evaluated using the Receiver Operating Characteristic (ROC) curve, and measured with the Area Under ROC (AUROC). Sensitivity and specificity pairs tables were evaluated with either one of the two configured with different thresholds and computed the other using the corresponding point on the ROC curve. The binary classification results, both from the AI models and from the radiologist manual reading results, were evaluated using the F1-score (i.e., harmonic mean of Precision and Recall) and balanced accuracy (i.e., the average of sensitivity and specificity) metrics. All metrics were computed through an open-source python library Scikit-learn (sklearn) 30 . The mean results, standard deviations, and confidence intervals were calculated using model weights trained with five different random seeds on the CU training set. The computations used Python's native statistics library, specifically the mean and stdev methods. Declarations Author Contributions Initiation and design of the study : Jayant Raikhelkar, Zilong Bai, Ashley N. Beecy, Deborah Estrin, Mert Sabuncu, Nir Uriel. Data collection : Chris Kelsey, David vanMaanan, Jeffrey Ruhl. Deep learning and statistical analysis : Zilong Bai, Ashley N. Beecy, Fengbei Liu, Nusrat Binta Nizam, Varsha Kishore. Radiologist review : Jay Leb, Alan Legasto. Supervision of research : Ashley N. Beecy, Deborah Estrin, Mert Sabuncu, Nir Uriel. Writing the first draft of the manuscript : Jayant Raikhelkar, Zilong Bai, Mert Sabuncu. All authors contributed to the writing and editing of the manuscript and approved the manuscript. References Bozkurt B et al (2021) Universal Definition and Classification of Heart Failure: A Report of the Heart Failure Society of America, Heart Failure Association of the European Society of Cardiology, Japanese Heart Failure Society and Writing Committee of the Universal Definition of Heart Failure. J Card Fail. 10.1016/j.cardfail.2021.01.022 Bozkurt B et al (2024) HF STATS. : Heart Failure Epidemiology and Outcomes Statistics An Updated 2024 Report from the Heart Failure Society of America. Journal of cardiac failure (2024) 10.1016/j.cardfail.2024.07.001 Bozkurt B et al (2023) Heart Failure Epidemiology and Outcomes Statistics: A Report of the Heart Failure Society of America. J Card Fail 29 Lippi G, Sanchis-Gomar F (2020) Global epidemiology and future trends of heart failure. AME Med J 5 Authors/Task Force Members: et al. (2021) ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: Developed by the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC). With the special contribution of the Heart Failure Association (HFA) of the ESC. Eur J Heart Fail 24, 4–131 (2022) Heidenreich PA et al (2022) 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol 79 Savarese G, Stolfo D, Sinagra G, Lund LH (2022) Heart failure with mid-range or mildly reduced ejection fraction. Nat Rev Cardiol 19 Mc Causland FR et al (2024) Finerenone and Kidney Outcomes in Patients with Heart Failure: The FINEARTS-HF Trial. J Am Coll Cardiol. 10.1016/j.jacc.2024.10.091 Website https://www.consumerreports.org/cro/magazine/2015/01/the-surprising-dangers-of-ct-sans-and-x-rays/index.htm Miller RJH et al (2024) Predicting mortality from AI cardiac volumes mass and coronary calcium on chest computed tomography. Nat Commun 15 Sandhu AT et al (2023) Incidental Coronary Artery Calcium: Opportunistic Screening of Previous Nongated Chest Computed Tomography Scans to Improve Statin Rates (NOTIFY-1 Project). Circulation 147:703–714 Dosovitskiy A et al (2020) An Image is Worth 16x16 Words. Transformers for Image Recognition at Scale Selvaraju RR et al Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. https://ieeexplore.ieee.org/document/8237336 Sayed A et al (2024) Reversals in the Decline of Heart Failure Mortality in the US, 1999 to 2021. JAMA Cardiol 9:585–589 Maddox TM et al (2024) 2024 ACC Expert Consensus Decision Pathway for Treatment of Heart Failure With Reduced Ejection Fraction: A Report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 83 Pickhardt PJ et al (2023) Opportunistic Screening: Radiology Scientific Expert Panel. Radiology 307 Attia ZI et al (2019) Prospective validation of a deep learning electrocardiogram algorithm for the detection of left ventricular systolic dysfunction. J Cardiovasc Electrophys 30 Kashou AH et al (2021) Artificial Intelligence-Augmented Electrocardiogram Detection of Left Ventricular Systolic Dysfunction in the General Population. Mayo Clinic proceedings 96 Attia ZI et al (2019) Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med 25 Kwon JM et al (2019) Development and Validation of Deep-Learning Algorithm for Electrocardiography-Based Heart Failure Identification. Korean circulation J 49 Yao X et al (2021) Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat Med 27 Bhave S et al (2024) Deep learning to detect left ventricular structural abnormalities in chest X-rays. Eur Heart J 45 de González A (2009) B. Projected cancer risks from computed tomographic scans performed in the United States in 2007. Arch Intern Med 169 Report on the Economic Well-Being of U.S. Households in 2023 - (2024) May Board of Governors of the Federal Reserve System https://www.federalreserve.gov/publications/2024-economic-well-being-of-us-households-in-2023-executive-summary.htm Jain SS, Mello MM, Shah NH (2024) Avoiding Financial Toxicity for Patients from Clinicians’ Use of AI. N Engl J Med. 10.1056/NEJMp2406135 Kim Y-J, Kim RJ (2011) The role of cardiac MR in new-onset heart failure. Curr Cardiol Rep 13:185–193 Dida H, Charif F, Benchabane A (2022) Registration of computed tomography images of a lung infected with COVID-19 based in the new meta-heuristic algorithm HPSGWO. Multimedia Tools Appl 81:18955–18976 scipy.ndimage.interpolation .zoom — SciPy v0.14.0 Reference Guide. https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.ndimage.interpolation.zoom.html Hamamci IE et al (2023) GenerateCT: Text-Conditional Generation of 3D Chest CT Volumes scikit-learn https://scikit-learn.org/stable/ Additional Declarations There is NO Competing Interest. Supplementary Files SupplementaryTable1.docx SupplementaryTable2.docx SupplementaryTable3.docx SupplementaryTable4.docx SupplementaryFigure1.png SupplementaryData1.zip SupplementaryData2.zip Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-5677688\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Article\",\"associatedPublications\":[],\"authors\":[{\"id\":398520643,\"identity\":\"52adf9be-4acf-47f3-a86b-c50470049f9f\",\"order_by\":0,\"name\":\"Jayant Raikhelkar\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCUlEQVRIie2QsUoDQRCG5ziIzaDtBS7JK2y4dvFZZhE2hVlrixRXmWZj2vMtIgHrPRdSJa0IsbjKSuHsIoi6ASMnuLG12K+anX8/mBmAQOAfEmmIKkLe3TVSAASzTXK/ErMqldmugX8qLm+1K34r8qYC+5RYr6qEhmYwLU5FDSOO7G5Vmg3wzsx4BpuMGaPlgyru1TyBhUS2PqNSg8y8yrQFJPSjmiXqxg1jnTJkBsGKfYoR73bAmkr5Bh9+ZXIR5YSWfijWHc2v6EUMhLJfLJ/nCbld2m4Xm7KT7Mqj9LU8eN0g7x2N1XVdj3j3cK3sy9P5cefSp+TNF31X7PfvW3r+KBAIBAJffAJ7m2Qn595U+gAAAABJRU5ErkJggg==\",\"orcid\":\"\",\"institution\":\"Columbia University Irving Medical Center/New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Jayant\",\"middleName\":\"\",\"lastName\":\"Raikhelkar\",\"suffix\":\"\"},{\"id\":398520644,\"identity\":\"28ec9686-4129-4b61-a5bc-a228d843c41a\",\"order_by\":1,\"name\":\"Zilong Bai\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0002-3891-8015\",\"institution\":\"Weill Cornell Medicine\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Zilong\",\"middleName\":\"\",\"lastName\":\"Bai\",\"suffix\":\"\"},{\"id\":398520645,\"identity\":\"df411e5d-6327-403c-9c91-f2c5ca6f77dc\",\"order_by\":2,\"name\":\"Ashley Beecy\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Division of Cardiology, Department of Medicine, Weill Cornell Medicine, New York, NY Information Technology Data Science, New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Ashley\",\"middleName\":\"\",\"lastName\":\"Beecy\",\"suffix\":\"\"},{\"id\":398520646,\"identity\":\"3d575633-79ac-4f4a-9c3c-dc6279eb34df\",\"order_by\":3,\"name\":\"Fengbei Liu\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Cornell Tech\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Fengbei\",\"middleName\":\"\",\"lastName\":\"Liu\",\"suffix\":\"\"},{\"id\":398520647,\"identity\":\"7da0da11-be1a-4b20-a355-968e7cf7e42a\",\"order_by\":4,\"name\":\"Nusrat Nizam\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Cornell Tech\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Nusrat\",\"middleName\":\"\",\"lastName\":\"Nizam\",\"suffix\":\"\"},{\"id\":398520648,\"identity\":\"40e77f73-a0a3-45fa-bafd-37e07e21d7ae\",\"order_by\":5,\"name\":\"Varsha Kishore\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Department of Computer Science, Cornell University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Varsha\",\"middleName\":\"\",\"lastName\":\"Kishore\",\"suffix\":\"\"},{\"id\":398520649,\"identity\":\"71f84e33-b318-442b-b8c0-89d796f5f052\",\"order_by\":6,\"name\":\"Chris Kelsey\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Information Technology Data Science, New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Chris\",\"middleName\":\"\",\"lastName\":\"Kelsey\",\"suffix\":\"\"},{\"id\":398520650,\"identity\":\"bf92ce04-ce14-4408-b6ff-4e6ccab0cc43\",\"order_by\":7,\"name\":\"David vanMaanan\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Information Technology Data Science, New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"David\",\"middleName\":\"\",\"lastName\":\"vanMaanan\",\"suffix\":\"\"},{\"id\":398520651,\"identity\":\"20ed636d-c540-45cd-b26b-338d79bfbf65\",\"order_by\":8,\"name\":\"Jeffrey Ruhl\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Information Technology Data Science, New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jeffrey\",\"middleName\":\"\",\"lastName\":\"Ruhl\",\"suffix\":\"\"},{\"id\":398520652,\"identity\":\"ee73ce39-df1f-4386-9899-2885300619ba\",\"order_by\":9,\"name\":\"Naomi Tesfuzigta\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Cornell Tech\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Naomi\",\"middleName\":\"\",\"lastName\":\"Tesfuzigta\",\"suffix\":\"\"},{\"id\":398520653,\"identity\":\"05466852-cd9c-40c6-9861-1d8d2f3b4bc7\",\"order_by\":10,\"name\":\"Erica Lancet\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"New York-Presbyterian Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Erica\",\"middleName\":\"\",\"lastName\":\"Lancet\",\"suffix\":\"\"},{\"id\":398520654,\"identity\":\"f939e15d-d5ab-4a26-ae5f-b0488d533e46\",\"order_by\":11,\"name\":\"Jay Leb\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Department of Radiology, Columbia University Irving Medical Center\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Jay\",\"middleName\":\"\",\"lastName\":\"Leb\",\"suffix\":\"\"},{\"id\":398520655,\"identity\":\"f452fe84-8ee1-406c-ad33-c94ffd5ece6e\",\"order_by\":12,\"name\":\"Alan Legasto\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Department of Radiology, Weill Cornell Medicine\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Alan\",\"middleName\":\"\",\"lastName\":\"Legasto\",\"suffix\":\"\"},{\"id\":398520656,\"identity\":\"c9c47ae5-6855-40a0-a440-67ba1ce917a2\",\"order_by\":13,\"name\":\"Pierre Elias\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0002-9643-3024\",\"institution\":\"Columbia University Irving Medical Center\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Pierre\",\"middleName\":\"\",\"lastName\":\"Elias\",\"suffix\":\"\"},{\"id\":398520657,\"identity\":\"e789bbd5-7f86-4672-b151-77d44fd53793\",\"order_by\":14,\"name\":\"Timothy Poterucha\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0001-7284-3937\",\"institution\":\"Columbia University Medical Center\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Timothy\",\"middleName\":\"\",\"lastName\":\"Poterucha\",\"suffix\":\"\"},{\"id\":398520658,\"identity\":\"e180bcc7-d392-4c20-aa02-29344c03b247\",\"order_by\":15,\"name\":\"Deepa Kumaraiah\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Seymour, Paul, and Gloria Milstein Division of Cardiology, Department of Medicine, Columbia University Irving Medical Center/New York–Presbyterian Hospital, NY\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Deepa\",\"middleName\":\"\",\"lastName\":\"Kumaraiah\",\"suffix\":\"\"},{\"id\":398520659,\"identity\":\"aa8b79ab-ce3d-4d27-8b0d-de8aa2e3ac47\",\"order_by\":16,\"name\":\"Fei Wang\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0001-9459-9461\",\"institution\":\"Weill Cornell Medicine\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Fei\",\"middleName\":\"\",\"lastName\":\"Wang\",\"suffix\":\"\"},{\"id\":398520660,\"identity\":\"1881764e-78e4-4188-8b13-4e26b54bb822\",\"order_by\":17,\"name\":\"Gabriel Sayer\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"NewYork–Presbyterian Hospital\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Gabriel\",\"middleName\":\"\",\"lastName\":\"Sayer\",\"suffix\":\"\"},{\"id\":398520661,\"identity\":\"8707123f-37ff-478d-bae4-42120dead969\",\"order_by\":18,\"name\":\"Deborah Estrin\",\"email\":\"\",\"orcid\":\"https://orcid.org/0000-0001-6477-0096\",\"institution\":\"Cornell Tech, Cornell University, WCM Population Health\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Deborah\",\"middleName\":\"\",\"lastName\":\"Estrin\",\"suffix\":\"\"},{\"id\":398520662,\"identity\":\"c0bb2a25-6bac-4cb7-aee1-8a86bde400b5\",\"order_by\":19,\"name\":\"Mert Sabuncu\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Cornell\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mert\",\"middleName\":\"\",\"lastName\":\"Sabuncu\",\"suffix\":\"\"},{\"id\":398520663,\"identity\":\"4c11a4cc-637a-4edf-a841-905ea8e70d26\",\"order_by\":20,\"name\":\"Nir Uriel\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Columbia University\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Nir\",\"middleName\":\"\",\"lastName\":\"Uriel\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2024-12-19 14:40:38\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-5677688/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-5677688/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":75520086,\"identity\":\"25c73ab3-9287-4588-9941-79e9614ffb10\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:36\",\"extension\":\"png\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":439318,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eStudy Design.\\u003c/strong\\u003e We collected CT imaging data and echo reports from two independent sites, Columbia University Irving Medical Center (CU) and Weill Cornell Medicine (WCM). The CT and echo studies were matched per patient based on specific inclusion and exclusion criteria to create input-label pairs for developing our AI model. Preprocessed CT scans served as model inputs, while LVEF values from echo studies provided binary labels for training and ground-truth performance evaluation. We conducted sub-cohort evaluations by stratifying the patient cohort based on demographic attributes and varying time intervals between CT and echo studies to explore different application scenarios. Additionally, a radiologist reviewed randomly selected CT scans without being informed of the model's prediction results.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure1StudyDesign.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/6b922f58cd2e9be6191b1b49.png\"},{\"id\":75518560,\"identity\":\"084a2c26-5ace-4505-8558-68d9e0db43ef\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:37\",\"extension\":\"png\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":658287,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003ePerformance Evaluation: Detection of abnormal EF using non-gated non-contrast Chest CT. a \\u003c/strong\\u003eand\\u003cstrong\\u003e b.\\u003c/strong\\u003e Receiver operating characteristic (ROC) curves and area under the ROC (AUROC) scores for site-specific test set performance: for CU (a) and WCM (b) respectively. \\u003cstrong\\u003ec. \\u003c/strong\\u003eAverage\\u003cstrong\\u003e \\u003c/strong\\u003eperformance metrics (in AUROC, F1-score and Balanced Accuracy) for subpopulations stratified from the test set according to site and demographic attributes: gender, age and race. See \\u003cstrong\\u003eTable 1\\u003c/strong\\u003efor corresponding population statistics.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure2PerformanceEvaluation.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/cb4dcad908454148ee5f7dee.png\"},{\"id\":75518564,\"identity\":\"f8b4bdcb-fecc-49c9-b99c-13739525eb63\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:38\",\"extension\":\"png\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":2588122,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eModel Interpretability.\\u003c/strong\\u003e Sample visualizations of Grad-CAM saliency maps overlaid on axial-oriented CT scans for four patients with abnormal LVEF. These maps highlight regions within the CT scans that contributed most significantly to the AI model’s predictions, providing insights into the features used for classification.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure3ModelInterpretability.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/0b9ca1f199dcd72d9d0e119d.png\"},{\"id\":75518550,\"identity\":\"f7955127-f622-49a0-918a-0857f3382501\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:36\",\"extension\":\"png\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":126941,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eVenn Diagrams of\\u003c/strong\\u003e \\u003cstrong\\u003eError Analysis. \\u003c/strong\\u003eFalse positive and false negative results detected by the AI model and the two board certified thoracic radiologists. \\u003cstrong\\u003ea \\u003c/strong\\u003eand\\u003cstrong\\u003e b\\u003c/strong\\u003e:\\u003cstrong\\u003e \\u003c/strong\\u003eFalse positive results and false negative results\\u003cstrong\\u003e \\u003c/strong\\u003eamong the\\u003cstrong\\u003e \\u003c/strong\\u003e\\u0026nbsp;90 CU scans. \\u003cstrong\\u003ec \\u003c/strong\\u003eand\\u003cstrong\\u003e d\\u003c/strong\\u003e: False positive results and false negative results among the 100 WCM scans.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure4ErrorAnalysis.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/5763d770450aab10f28a75ed.png\"},{\"id\":75520092,\"identity\":\"4d937a8a-9177-4f0e-9bb7-4587e689c1c5\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:38\",\"extension\":\"png\",\"order_by\":5,\"title\":\"Figure 5\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":444519,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003e\\u003cstrong\\u003eData curation and preprocessing flowchart.\\u003c/strong\\u003e\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"Figure5Datacurationandpreprocessingflowchart.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/7de8ca306f3892ebc7fd234e.png\"},{\"id\":75520424,\"identity\":\"ff2995ed-72f4-44c3-a406-e91ed8e5da68\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 12:06:52\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":5294433,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/485106a6-0781-402d-b589-9bd1380c2b26.pdf\"},{\"id\":75520091,\"identity\":\"0891ea6a-f9b5-4b26-90c0-6fcc9b09a378\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:38\",\"extension\":\"docx\",\"order_by\":1,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":36540,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryTable1.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/1cd8f5abac17080617b74b93.docx\"},{\"id\":75520089,\"identity\":\"4222ecc1-8b62-462c-98cc-5894e99d533f\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:38\",\"extension\":\"docx\",\"order_by\":2,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":17932,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryTable2.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/4df63e41020626151e5d7293.docx\"},{\"id\":75518563,\"identity\":\"e6e8573f-0653-427e-ac49-ca5384578387\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:37\",\"extension\":\"docx\",\"order_by\":3,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":16477,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryTable3.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/1ff228aacc08b1d7bca2bf68.docx\"},{\"id\":75518567,\"identity\":\"66d9977e-25e0-4cad-b5d2-fc272239638b\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:38\",\"extension\":\"docx\",\"order_by\":4,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":15094,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryTable4.docx\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/59db9e8765f1246c9dcfb26c.docx\"},{\"id\":75518595,\"identity\":\"51b4b120-261f-424c-be29-5b37f941fcb1\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:50:39\",\"extension\":\"png\",\"order_by\":5,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":858005,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryFigure1.png\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/77ff731cad7cf3732c2f05ce.png\"},{\"id\":75520096,\"identity\":\"e46a0969-1514-4871-aaab-3bce54a08776\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:39\",\"extension\":\"zip\",\"order_by\":7,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":8183394,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryData1.zip\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/8d7349af4b810e1bc85ef52a.zip\"},{\"id\":75520094,\"identity\":\"69f65bae-ba48-4190-a33e-f420f8f64954\",\"added_by\":\"auto\",\"created_at\":\"2025-02-05 11:58:38\",\"extension\":\"zip\",\"order_by\":8,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"supplement\",\"size\":10049120,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"SupplementaryData2.zip\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-5677688/v1/b6ef3015a6bdbd1aa88dea87.zip\"}],\"financialInterests\":\"There is \\u003cb\\u003eNO\\u003c/b\\u003e Competing Interest.\",\"formattedTitle\":\"An Artificial Intelligence Model to Detect Abnormal Ejection Fraction from Non-Contrast Chest Computed Tomography: The CT-LVEF study\",\"fulltext\":[{\"header\":\"Introduction\",\"content\":\"\\u003cp\\u003eHeart Failure (HF) has been declared a global pandemic with over 64\\u0026nbsp;million individuals afflicted worldwide.\\u003csup\\u003e \\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e \\u003c/sup\\u003e About 6.7\\u0026nbsp;million of these individuals are in the USA and this number is expected to rise to 8.5\\u0026nbsp;million by 2030.\\u003csup\\u003e \\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e \\u003c/sup\\u003e Despite improvements in medical and surgical treatment, HF remains a leading cause of hospitalizations and mortality.\\u003csup\\u003e \\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e \\u003c/sup\\u003e In addition, the worldwide economic burden of HF is estimated to be \\u003cspan\\u003e$\\u003c/span\\u003e346\\u0026nbsp;billion US dollars annually.\\u003csup\\u003e \\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e \\u003c/sup\\u003e \\u003c/p\\u003e \\u003cp\\u003eThe left ventricular ejection fraction (EF) is a critical parameter which is used to phenotype heart failure into several types, namely heart failure with reduced EF (HFrEF), Heart Failure with mid-range EF( HFmrEF) and heart failure with preserved EF (HFpEF).\\u003csup\\u003e\\u003cspan citationid=\\\"CR5\\\" class=\\\"CitationRef\\\"\\u003e5\\u003c/span\\u003e\\u003c/sup\\u003e There is evidence for early initiation of lifestyle modifications and medications that constitute guideline directed medical therapy (GDMT) for both HfrEF and HFmrHF to decrease both HF hospitalizations and mortality.\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR7\\\" citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e\\u003c/sup\\u003e It is estimated that a significant number of patients remain undiagnosed with HfmrEF and HfrEF and thus are unable to be started on appropriate HF therapies.\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e There is an unmet need for earlier detection of these patients with the goal of improving outcomes by means of initiation of lifestyle modification and GDMT at an earlier stage of disease.\\u003c/p\\u003e \\u003cp\\u003eComputed tomography (CT) scans are a common medical imaging modality, with over 80\\u0026nbsp;million studies being performed yearly in the USA\\u003csup\\u003e\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e\\u003c/sup\\u003e. Chest CT scans are frequently performed in both inpatient and outpatient settings for a wide range of indications, such as screening for lung cancer, acute pulmonary embolism detection, evaluation for infectious diseases like pneumonia and trauma. These studies contain a wealth of biometric information, and several studies have recently investigated the use of artificial intelligence (AI) based automated algorithms applied to CT scans for identifying patients at higher risk for adverse events, including cardiovascular adverse events\\u003csup\\u003e\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e\\u003c/sup\\u003e.\\u003c/p\\u003e \\u003cp\\u003eIn this study, we developed and validated a novel AI model capable of predicting abnormal left-ventricle ejection fraction (LVEF), a dynamic physiological parameter, directly from non-gated, non-contrast chest CT scans. This represents a surprising and transformative application, as CT scans are typically used for static imaging of non-cardiac structures. Using a multi-institutional dataset of over 34,000 paired CT and echocardiogram studies, our model demonstrated robust performance, achieving AUROCs of 0.786 on the primary test set and 0.755 in external validation. Beyond strong predictive performance, the AI model outperformed radiologists in both accuracy and efficiency and provided interpretable visualizations highlighting imaging biomarkers linked to reduced LVEF.\\u003c/p\\u003e \\u003cp\\u003eBy unlocking new diagnostic possibilities from one of the most commonly performed imaging modalities our approach reveals the potential of opportunistic screening for heart failure. Our findings underscore how this technology could bridge the diagnostic gap in asymptomatic HF, enabling earlier detection, personalized interventions, and improved patient outcomes.\\u003c/p\\u003e\"},{\"header\":\"Results\",\"content\":\"\\u003cp\\u003eDataset curation and study design\\u003c/p\\u003e \\u003cp\\u003eWe curated a multimodal dataset of paired chest CT imaging data and echocardiogram (echo) study reports from Columbia University Irving Medical Center (CU) and Weill Cornell Medical Center (WCM). We focused on non-contrast, non-gated chest CT volumetric data in this study. The CT images and left ventricle ejection fraction (LVEF) values derived from echo reports were paired by patient using our data linkage protocol (e.g., closest in study times within 6 months, Supplementary Fig.\\u0026nbsp;1). The dataset included 25,948 CT-echo pairs from the Columbia University (CU) cohort and 8,110 from the Weill Cornell Medicine (WCM) cohort, with CT images serving as inputs to the AI model and LVEF derived from echo reports providing the target variable. CT-echo pairs were matched based on proximity in study times (within six months). Using an LVEF threshold of 50%, we categorized patients into normal (LVEF\\u0026thinsp;\\u0026ge;\\u0026thinsp;50%) and abnormal (LVEF\\u0026thinsp;\\u0026lt;\\u0026thinsp;50%) groups, with 15.94% of pairs classified as abnormal in the CU cohort and 17.44% in the WCM cohort.\\u003c/p\\u003e \\u003cp\\u003eThis curated dataset covered diverse demographics, with a balanced gender distribution of male (n\\u0026thinsp;=\\u0026thinsp;14,452, 55.69% of CU, n\\u0026thinsp;=\\u0026thinsp;4,364, 53.81% of WCM) and female (n\\u0026thinsp;=\\u0026thinsp;11,995, 44.31% of CU, n\\u0026thinsp;=\\u0026thinsp;4,115, 46.19% of WCM) patients. We divided the study cohort into three age groups: 18\\u0026ndash;40, 40\\u0026ndash;65, and 65\\u0026ndash;120 years. We also investigated three racial groups: Asian/Native Hawaiian/Other Pacific Islands (ANO), White, and Black or African American (BAA). Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e presents the detailed summary statistics for study pairs, stratified by the institution of CT data collection and patient demographics, and compared normal vs abnormal EF status.\\u003c/p\\u003e \\u003cp\\u003e \\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab1\\\" border=\\\"1\\\"\\u003e \\u003ccaption language=\\\"En\\\"\\u003e \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 1\\u003c/div\\u003e \\u003cdiv class=\\\"CaptionContent\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003ePopulation Description.\\u003c/b\\u003e The datasets analyzed are compiled from Columbia University Irving Medical Center (CU) and Weill Cornell Medicine (WCM). Each instance in the study consists of an input-label pair, where the input is preprocessed non-contrast chest CT volumetric data, and the label reflects normal or abnormal left ventricle ejection fraction (LVEF) status derived from the associated echocardiography report. The AI model is trained on 80% of the instances with CT images from the Columbia dataset. Subpopulation statistics are calculated by stratifying the data according to the sites where the CT studies were collected and the demographics of the patients. The numbers indicate the pairs with CTs from different sites. One patient may be associated with different LVEF statuses by different CTs. * Percentages for Patient Sex or Race subgroups do not sum to 100% due to additional categories with insufficient sample sizes. These categories, labeled as \\u0026ldquo;Others\\u0026rdquo; or \\u0026ldquo;Not provided,\\u0026rdquo; were omitted from the population statistics for clarity. Chi-square test of independence was conducted to compare the distributional differences between CU and WCM in terms of patient sex, age groups, and race. The analyses were performed on all patients as well as the normal and abnormal subgroups for each variable.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"8\\\"\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c7\\\" colnum=\\\"7\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c8\\\" colnum=\\\"8\\\"\\u003e\\u003c/div\\u003e \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e\\u0026nbsp;\\u003c/th\\u003e \\u003cth align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c3\\\" namest=\\\"c2\\\"\\u003e \\u003cp\\u003eColumbia University Irving Medical Center\\u003c/p\\u003e \\u003cp\\u003eTotal: 25,948 chest CTs\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colspan=\\\"2\\\" nameend=\\\"c5\\\" namest=\\\"c4\\\"\\u003e \\u003cp\\u003eWeill Cornell Medical College\\u003c/p\\u003e \\u003cp\\u003eTotal: 8,110 chest CTs\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colspan=\\\"3\\\" nameend=\\\"c8\\\" namest=\\\"c6\\\"\\u003e \\u003cp\\u003eBetween-site\\u003c/p\\u003e \\u003cp\\u003eChi-square test of independence\\u003c/p\\u003e \\u003cp\\u003eP-value\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eNormal (LVEF\\u0026thinsp;\\u0026ge;\\u0026thinsp;50) -\\u003c/p\\u003e \\u003cp\\u003e21,787 (83.96%)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003eAbnormal (LVEF\\u0026thinsp;\\u0026lt;\\u0026thinsp;50) \\u0026minus;\\u0026thinsp;4,161 (16.04%)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003eNormal (LVEF\\u0026thinsp;\\u0026ge;\\u0026thinsp;50) \\u0026minus;\\u0026thinsp;6695 (82.70%)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003eAbnormal (LVEF\\u0026thinsp;\\u0026lt;\\u0026thinsp;50) \\u0026minus;\\u0026thinsp;1415 (17.29%)\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003eOverall\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c7\\\"\\u003e \\u003cp\\u003eNormal\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c8\\\"\\u003e \\u003cp\\u003eAbnormal\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colspan=\\\"8\\\" nameend=\\\"c8\\\" namest=\\\"c1\\\"\\u003e \\u003cp\\u003ePatient Sex*\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eMale\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e10351 (47.51%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e2616 (62.87%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e3207 (47.90%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e942 (66.57%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c7\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c8\\\" morerows=\\\"1\\\" rowspan=\\\"2\\\"\\u003e \\u003cp\\u003e0.5784\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eFemale\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e9799 (44.98%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e1278 (30.71%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e3486 (52.07%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e472 (33.36%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colspan=\\\"8\\\" nameend=\\\"c8\\\" namest=\\\"c1\\\"\\u003e \\u003cp\\u003ePatient Age(years)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e18\\u0026ndash;39\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e2011 (9.23%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e218 (5.23%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e351 (5.24%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e56 (3.96%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c7\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c8\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e0.0102\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e40\\u0026ndash;64\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e8276 (37.99%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e1249 (30.02%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2204 (32.92%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e387 (27.35%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e65 and above\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e11500 (52.78%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e2694 (64.74%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e4140 (61.84%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e972 (68.69%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colspan=\\\"8\\\" nameend=\\\"c8\\\" namest=\\\"c1\\\"\\u003e \\u003cp\\u003ePatient Race*\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eAsian/Native Hawaiian/Other Pacific Islander\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e784 (3.60%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e90 (2.16%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e403 (6.02%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e90 (6.36%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c6\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c7\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c8\\\" morerows=\\\"2\\\" rowspan=\\\"3\\\"\\u003e \\u003cp\\u003e\\u0026lt;\\u0026thinsp;0.0001\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eWhite\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e10761 (49.39%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e1754 (42.15%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e3596 (53.71%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e699 (42.33%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eBlack or African American\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e3140 (14.41%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e654 (15.72%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e897 (13.40%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e212 (14.98%)\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003c/tbody\\u003e \\u003c/colgroup\\u003e \\u003c/table\\u003e\\u003c/div\\u003e \\u003c/p\\u003e \\u003cp\\u003eFor CT data acquisition, non-contrast non-gated CT imaging data from the two institutions were collected using equipment from different vendors. In the CU cohort, CT studies were distributed among four vendors as follows: TOSHIBA (4,142), SIEMENS (1,487), Canon Medical Systems (351), and GE Medical Systems (19,968). In the WCM cohort, CT studies were distributed between two vendors as: SIEMENS (105) and GE Medical Systems (8,005). Volumetric CT scans, performed along the axial orientation and focused on the chest region, encompassing the entire cardiac structure and including marginal portions of the lower neck and upper abdomen, were selected. We used CT volumetric data collected through a smooth kernel (See Supplementary Table\\u0026nbsp;1 for thresholds of smooth vs sharp kernels).\\u003c/p\\u003e \\u003cp\\u003eAI Model Training and Evaluation\\u003c/p\\u003e \\u003cp\\u003eWe used the study pairs from the CU cohort as the primary dataset to develop our AI-based models, and those from the WCM cohort for external validation. These models were designed to predict the binarized LVEF status of a patient using their CT volume data. We performed a 70%-10%-20% train-validation-test split, divided by patients, within the primary dataset from the CU cohort. This split was used to train an ensemble of vision-transformer (ViT) neural networks\\u003csup\\u003e\\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e\\u003c/sup\\u003e (AI model), initialized with pretrained weights, on the training set (see Methods for details). We then evaluated the model's performance on the test subset using accuracy and F1 score. This involved using a total of 20,774 study pairs from the primary dataset for the AI model training and 5,174 study pairs as the hold-out test set for performance evaluation. We randomly selected 100 study pairs from the CU hold-out test set and 100 study pairs from the WCM external validation set for radiologist review to compare with our AI-based model prediction results (see Methods).\\u003c/p\\u003e \\u003cp\\u003eOn the CU hold-out test set, the AI model achieved an AUROC of 0.786 (95% confidence interval (CI): 0.772\\u0026ndash;0.801) and F1 score of 0.817 (95% CI: 0.807\\u0026ndash;0.827) for predicting LVEF-derived status labels (n\\u0026thinsp;=\\u0026thinsp;5,174, 20%) (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). We investigated the sensitivity of our model given different levels of specificity (Supplementary Table\\u0026nbsp;2). Performance metrics were further analyzed across demographic subgroups. (Table c in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). The number of instances in each demographic subgroup is summarized in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e. For each demographic attribute, our model performed best within the male subgroup with an AUROC of 0.790 (95% CI: 0.777\\u0026ndash;0.802), within the senior (65\\u0026thinsp;~\\u0026thinsp;120 years) age subgroup with an AUROC of 0.791 (95% CI: 0.778\\u0026ndash;0.804), and within the ANO racial subgroup with an AUROC of 0.792 (95% CI: 0.742\\u0026ndash;0.842)(Table b in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). For each demographic subgroup, we further investigated the performance of our AI model with respect to LVEF measurements from different time intervals. In particular, we studied the CT-to-echo study time interval being 1 day, 1 week and 1 month. In each demographic subgroup, our model performed the best for 1-day interval targets, compared to the 1-week and 1-month scenarios, as well as the entire demographic subgroup. This pattern was consistent across all the demographic subgroups (Supplementary Table\\u0026nbsp;3).\\u003c/p\\u003e \\u003cp\\u003eExternal validation\\u003c/p\\u003e \\u003cp\\u003eTo evaluate the generalizability of our AI model across different institutions with potentially varied CT data collection protocols, we validated using external data from WCM. This dataset consisted of n\\u0026thinsp;=\\u0026thinsp;6,695 normal study pairs with echo-derived LVEF\\u0026thinsp;\\u0026ge;\\u0026thinsp;50% and n\\u0026thinsp;=\\u0026thinsp;1,415 abnormal study pairs with echo-derived LVEF\\u0026thinsp;\\u0026lt;\\u0026thinsp;50%. Our AI model achieved an AUROC of 0.755 (95% CI: 0.741\\u0026ndash;0.769) and an F1 score of 0.816 (95% CI: 0.806\\u0026ndash;0.825). For a range of sensitivity and specificity scores on the full ROC curve, see Supplementary Table\\u0026nbsp;2.\\u003c/p\\u003e \\u003cp\\u003eIn the demographic-stratified analysis, our AI model exhibited a similar pattern to that observed on the CU hold-out test set. Specifically, our model performed best within the male subgroup with an AUROC of 0.748 (95% CI: 0.73\\u0026ndash;0.767), within the senior (65 and above) age subgroup with an AUROC of 0.752 (95% CI: 0.733\\u0026ndash;0.771), and within the ANO racial subgroup with an AUROC of 0.768 (95% CI: 0.721\\u0026ndash;0.814) (Table c in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). These findings highlight the model's ability to maintain strong predictive performance across diverse demographic subgroups.\\u003c/p\\u003e \\u003cp\\u003eWhen analyzing performance based on the time interval between CT and echo studies, the model consistently performed best for pairs with a 1-day interval, compared to those with 1-week or 1-month intervals, as well as the broader subgroup without time-interval-specific selection. This pattern was observed across all demographic subgroups on the WCM validation set (Supplementary Table\\u0026nbsp;3). These results emphasize the importance of temporal proximity between imaging and echocardiography in achieving optimal prediction accuracy.\\u003c/p\\u003e \\u003cp\\u003eModel interpretation\\u003c/p\\u003e \\u003cp\\u003eModel Interpretability for Imaging Biomarker Discovery\\u003c/p\\u003e \\u003cp\\u003eWe employed guided gradient-weighted class activation mapping (Grad-CAM)\\u003csup\\u003e\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e\\u003c/sup\\u003e, a visualization technique that highlights the most influential regions in the input data. This method generates heatmaps, providing interpretable insights into our AI model\\u0026rsquo;s decision-making. We applied Grad-CAM to the CT volume data from both the CU hold-out test set and the WCM external validation set. Figure\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e displays representative heatmaps overlaid on axial CT slices, illustrating the regions of greatest model activation.\\u003c/p\\u003e \\u003cp\\u003eFor detailed analysis, we randomly selected 10 normal subjects (LVEF\\u0026thinsp;\\u0026ge;\\u0026thinsp;50% in their echo report) and 9 abnormal subjects (LVEF\\u0026thinsp;\\u0026lt;\\u0026thinsp;50% in their echo report) from the hold-out CU test set, and conducted rigorous manual examinations with a heart failure cardiologist and a thoracic radiologist. For the abnormal EF patients, the Grad-CAM heatmaps generally highlighted the lower neck, right or left lung, and chest wall (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e). This pattern was consistently observed across the subjects with abnormal EF in the CU test set (Supplementary Data 1). The Grad-CAM heatmaps of normal cases are included in Supplementary Data 2 and did not reveal an anatomically meaningful pattern.\\u003c/p\\u003e \\u003cp\\u003eThese observations highlight the potential of Grad-CAM not only to enhance AI model interpretability but also to uncover previously unrecognized imaging biomarkers that could provide novel insights into heart failure pathophysiology.\\u003c/p\\u003e \\u003cp\\u003eComparison with radiologist manual analysis and inference\\u003c/p\\u003e \\u003cp\\u003eWe compared LVEF status predictions made by our AI model against those made by dedicated thoracic radiologists using CT imaging data to benchmark accuracy, ensure clinical relevance, and identify strengths and weaknesses. Specifically, we randomly selected 90 non-contrast CT scans from the CU hold-out test set and 100 from the WCM external validation set, with similar rates of abnormal EF status as their respective full test sets. LVEF status information was withheld before sending the patient lists to the radiologists to ensure unbiased manual interpretation.\\u003c/p\\u003e \\u003cp\\u003eThe manual prediction process was done by two highly experienced thoracic radiologists, one (RAD1) from CU and the other (RAD2) from WCM. Both radiologists independently reviewed the CU 90 CT scans and made predictions regarding LVEF status. RAD2 also reviewed the 100 WCM scans. Using the LVEF status derived from echo reports as the ground-truth, our AI model outperformed the radiologists at both sites.\\u003c/p\\u003e \\u003cp\\u003eOn the CU scans, the AI model achieved an F1 score of 0.83, compared to 0.645 for RAD1 and 0.8 for RAD2. On the WCM scans, the AI model\\u0026rsquo;s F1 score was 0.833, surpassing RAD2\\u0026rsquo;s score of 0.622 (Table\\u0026nbsp;\\u003cspan refid=\\\"Tab2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e).\\u003c/p\\u003e \\u003cp\\u003e \\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab2\\\" border=\\\"1\\\"\\u003e \\u003ccaption language=\\\"En\\\"\\u003e \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 2\\u003c/div\\u003e \\u003cdiv class=\\\"CaptionContent\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003ePerformance comparison between our AI model and radiologists.\\u003c/b\\u003e Performance evaluation using F1 score for RAD1, RAD2 and our AI model on randomly selected CT scans from the CU test set and the WCM external validation set.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"3\\\"\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e\\u0026nbsp;\\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eCU N\\u0026thinsp;=\\u0026thinsp;90 sampled CT scans\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003eWCM N\\u0026thinsp;=\\u0026thinsp;100 sampled CT scans\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eRAD1\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e0.645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e-\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eRAD2\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e0.8\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.622\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eAI Model\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e0.83\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.833\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003c/tbody\\u003e \\u003c/colgroup\\u003e \\u003c/table\\u003e\\u003c/div\\u003e \\u003c/p\\u003e \\u003cp\\u003eThis is achieved by our AI model at a faster speed of analysis and inference: our AI model completed LVEF status prediction tasks within approximately 1 minute on the 90 CU CT volumetric data and the 100 WCM samples, respectively. In contrast, it took RAD1\\u0026thinsp;~\\u0026thinsp;2.2m per CT scan to interpret 90 CT volumetric data sampled from the CU hold-out test set, and ~\\u0026thinsp;2m per CT scan by RAD2 on the 100 WCM scans sampled from the external validation set.\\u003c/p\\u003e \\u003cp\\u003eDetailed error analysis revealed complementary strengths between the AI model and radiologists. Among the 90 CU scans, the numbers of incorrect predictions were as follows: 13 by the AI model, 37 by RAD1, and 18 by RAD2. Notably, 10 errors overlapped between the AI model and one radiologist, while only 6 errors were shared by the AI model and both radiologists. On the WCM scans, the AI model made 15 incorrect predictions, compared to 43 by RAD2. Errors shared by the AI model and RAD2 totaled 7. These patterns suggest that the AI model and radiologists leverage different aspects of the imaging data, potentially enabling complementary interpretations in clinical practice. We summarized the false-positive and false-negative predictions of our AI model and the two radiologists as Venn diagrams in Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e, offering further insights into the nature of the errors.\\u003c/p\\u003e \\u003cp\\u003eValidation on independent consecutive test sets\\u003c/p\\u003e \\u003cp\\u003eTo further evaluate the performance of our developed AI model in a real-world clinical setting, we constructed two independent testing sets using the CT studies and Echo reports collected from an overall 2,228 subjects admitted to CU (n\\u0026thinsp;=\\u0026thinsp;1,411 subjects) and WCM (n\\u0026thinsp;=\\u0026thinsp;817 subjects) of the NYP healthcare systems. The CU data spanned August to December 2023, while the WCM data covered August 2023 to July 2024. These independent testing sets included subjects that were not used for developing our AI model. Their CT and echo data were collected after the primary (retrospective) dataset, ensuring temporal separation from the training data.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eEvaluation of AI model on CU independent consecutive test set.\\u003c/b\\u003e \\u003c/p\\u003e \\u003cp\\u003eFrom the 1,895 subjects with 2,417 CT studies and 4,448 Echo reports collected from CU, we curated a CU independent consecutive test set for the LVEF status prediction model comprising 1,251 subjects with 1,411 pairs of CT-echo studies. Filtering and pairing protocols excluded 484 CT studies and 1,006 echo reports to ensure data quality and consistency (see Methods). Notably, the LVEF status distributions stratified by the patient demographics in the CU prospective test set (Supplementary Table\\u0026nbsp;4) were highly consistent with the ones on the primary dataset for our AI model development. The CT series were distributed among 3 vendors as follows: GE MEDICAL SYSTEMS (1,141), Canon Medical Systems (196), SIEMENS (74). Our AI model trained on the CU primary dataset demonstrated robust performance on the CU independent consecutive test set, achieving an AUROC of 0.784 (95% CI: 0.762\\u0026ndash;0.805), an F1 score of 0.854 (95% CI: 0.846\\u0026ndash;0.862) and a balanced accuracy of 0.594 (95% CI: 0.577\\u0026ndash;0.611) for LVEF status prediction. Performance metrics stratified by age, gender, and race are presented in Supplementary Fig.\\u0026nbsp;1c.\\u003c/p\\u003e \\u003cp\\u003e \\u003cb\\u003eEvaluation of AI model on WCM independent consecutive test set.\\u003c/b\\u003e Using the 1,201 subjects with 1,366 WCM CT studies and their associated 2,377 Echo study reports, we curated a WCM prospective test set for the LVEF status prediction model comprising 755 subjects with 817 pairs of CT-echo studies. Based on our filtering and pairing protocols (see Methods), 549 CT studies and 1,560 Echo study reports were excluded. Notably, the LVEF status distributions stratified by the patient demographics in the WCM independent consecutive test set (Supplementary Table\\u0026nbsp;4) are highly consistent with the ones on the WCM external validation dataset (Supplementary Table\\u0026nbsp;4). The CT series were distributed among two vendors as follows: GE Medical Systems (737) and SIEMENS (80). Our AI model trained on the CU primary dataset demonstrated robust performance on the WCM prospective test set, achieving an AUROC of 0.775 (95% CI: 0.756\\u0026ndash;0.794), an F1 score of 0.855 (95% CI: 0.849\\u0026ndash;0.861) and a balanced accuracy of 0.625 (95% CI: 0.606\\u0026ndash;0.644) for LVEF status prediction. The model performance is detailed in the Supplementary Fig.\\u0026nbsp;1c stratified by age, gender and race.\\u003c/p\\u003e \\u003cp\\u003eAcross both independent consecutive test sets, the AI model demonstrated strong and consistent performance, closely aligning with results from retrospective datasets. Importantly, these findings were consistent across demographic subgroups, imaging sites, and CT vendors, further supporting the robustness and generalizability of the model for LVEF status prediction in diverse real-world settings.\\u003c/p\\u003e \\u003cp\\u003eOpportunistic Screening\\u003c/p\\u003e \\u003cp\\u003eWe continued to pull 712 chest CT studies of 702 patients, without requiring paired echocardiograms, to assess our model across a broader patient population undergoing routine CT scans. After filtering and preprocessing based on the DICOM metadata and imaging files of the CT data (see Methods), we obtain 633 studies of 627 patients. Our model identified 23 patients for further evaluation.\\u003c/p\\u003e\"},{\"header\":\"Discussion\",\"content\":\"\\u003cp\\u003eIn this study, we demonstrate that an AI model can accurately detect abnormal LVEF using non-contrast, non-gated chest CTs ordered for standard clinical indications from two separate sites. The model achieved AUROCs of 0.786 at CU and 0.755 at WCM, on retrospective test data. In our independent consecutive test set study with data collected after August 2023, our AI model demonstrated similar performance, achieving AUROC 0.784 of CU, 0.775 of WCM. To our knowledge this is the first reported study to identify abnormal LVEF, which is traditionally detected from targeted cardiac testing such as echocardiography, multi-gated acquisition scan (MUGA) or cardiac MRI, from standard non gated non contrast chest CTs ordered for other medical indications.\\u003c/p\\u003e \\u003cp\\u003eThe lifetime risk of developing HF is now estimated to be 24% and the prevalence 1.9\\u0026ndash;2.6%\\u003csup\\u003e2\\u003c/sup\\u003e. Unfortunately, initial gains in heart failure management in the US have reversed with mortality rates due to HF higher than those in 1999, thought to be driven by increasing incidence of obesity, diabetes mellitus and hypertension.\\u003csup\\u003e\\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e\\u003c/sup\\u003e. About 24\\u0026ndash;34% of patients are estimated to have Stage B HF (asymptomatic LV dysfunction).\\u003csup\\u003e\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e\\u003c/sup\\u003e Lifestyle modifications and GDMT have been shown to improve quality of life as well as mortality in this cohort of patients.\\u003csup\\u003e\\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e\\u003c/sup\\u003e. There is an unmet need to detect these patients earlier to initiate lifestyle modifications and GDMT. There are currently no clinical screening guidelines for HF recommended in the general population, likely due to the clinical cost of screening using targeted cardiac testing.\\u003c/p\\u003e \\u003cp\\u003eOpportunistic screening has been described as the practice of systematically leveraging imaging data that are incidental to the clinical indication of study\\u003csup\\u003e\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e\\u003c/sup\\u003e. There have been several studies that have attempted to identify abnormal LV structure or function using opportunistic screening. Several studies have used ECG based AI algorithms to detect LV dysfunction and risk of HF.\\u003csup\\u003e\\u003cspan additionalcitationids=\\\"CR18 CR19 CR20\\\" citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e\\u003c/sup\\u003e. Bhave et al used a deep learning model to identify patients with severe left ventricular hypertrophy or dilated left ventricle, harbingers of HF, from chest X rays with a composite AUC of 0.79.\\u003csup\\u003e\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e\\u003c/sup\\u003eCT provides detailed cross sectional information and a wealth of biometric data compared to chest X-rays. In a study by Miller et al a combined AI model was able to quantify coronary calcium, left atrial volume and left ventricular mass index to predict cardiac death or MI with an AUC of 0.792.\\u003csup\\u003e\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u003c/sup\\u003e To our knowledge our study is the first to utilize AI to determine abnormal LVEF. LVEF has been traditionally determined by cardiac specific testing in which cardiac dimensions are able to be calculated in both systole and diastole such as in echocardiography or cardiac MRI to calculate the percentage of blood ejected with each beat. Calculation of EF using non- gated, non-contrast chest CT presents a novel application of non-gated chest CT that has not been explored previously, and in a previous era potentially not thought to be possible. It is estimated that over 15\\u0026nbsp;million chest CTs are performed in the US yearly.\\u003csup\\u003e\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e\\u003c/sup\\u003e Opportunistic screening for abnormal EF may present a unique opportunity in the future for early identification of Stage B heart failure with a goal to initiate lifestyle measures and initiation of GDMT to prevent progression to Stage C and Stage D heart failure, heart failure admissions and mortality.\\u003c/p\\u003e \\u003cp\\u003eDeploying this AI-based opportunistic screening tool to identify reduced LVEF introduces important considerations for healthcare delivery. The model, trained on retrospective data, was evaluated across various operating points (Supplementary Table\\u0026nbsp;2). If we optimize for specificity, at the CU site, it achieved a specificity of 80% and a sensitivity of 67%, and at the WCM site, it reached a specificity of 77% and a sensitivity of 67%. Using the same threshold, applied across a broader patient population undergoing routine CT scans, without requiring paired echocardiograms, the model identified 23 out of 627 patients for further evaluation. These operating characteristics applied to the prevalence in the test population, equate to a number needed to evaluate (NNE) of 3. Integrating model outputs into standard CT reporting workflows can streamline adoption, particularly among radiologists and clinicians, although optimal implementation may necessitate new operational pathways, including centralized processes to ensure prompt referrals for confirmatory echocardiograms. Notably, 3.63% of patients flagged by the model received CT scans for indications that may not justify insurance coverage for follow-up echocardiography, potentially leading to significant out-of-pocket costs. This financial burden is concerning, given that only 37% of U.S. adults can cover an unexpected \\u003cspan\\u003e$\\u003c/span\\u003e400 expense without borrowing, making it vital to consider financial protections for lower-income patients.\\u003csup\\u003e \\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e,\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e \\u003c/sup\\u003e Finally, training protocols should incorporate these healthcare delivery considerations alongside model information to ensure equitable and effective deployment across diverse populations.\\u003c/p\\u003e \\u003cp\\u003e Our AI model outperformed 2 board certified thoracic radiologists in predicting LVEF status from CT imaging, achieving higher F1 scores on both the CU and WCM test sets Notably, error analysis revealed that the AI model and radiologists\\u0026rsquo; mistakes were largely non-overlapping, suggesting that their predictions captured different aspects of the imaging data. Averaging prediction scores from the AI model and radiologists yielded only modest improvements in AUROC (0.74 on CU, 0.737 on WCM) compared to radiologists but fell short compared to pure AI predictions (0.757 on CU, 0.763 on WCM), indicating limited synergy from a simple score combination.\\u003c/p\\u003e \\u003cp\\u003eIn this study we utilized Grad-CAM mapping to evaluate potential imaging biomarkers detected by the AI model to identify abnormal EF. The mapping of the heart itself appears intuitive as overt structural changes in cardiac structure are associated with LV dysfunction. However, in addition to the heart, biomarkers noted included the lower neck region, lower lung regions and chest wall. It is unclear why these regions predict abnormal EF at this stage and an intriguing question which merits further study. Possible explanations to consider could be early signs of vascular congestion (great vessels of the neck and lungs) and structural adaptations of the thoracic cavity in the setting of HF (chest wall).\\u003c/p\\u003e \\u003cp\\u003eOur study has limitations, for which we outline potential areas for future work. Firstly, the applicability of our AI model is limited due to its retrospective nature. We plan to conduct a prospective study (in contrast to the independent consecutive test set study we conducted) with follow-up sessions to monitor EF status over time. In addition EF calculation by 2D echocardiography has been reported to show intraobserver and interobserver variations of 8\\u0026ndash;21% and 6\\u0026ndash;13% respectively\\u003csup\\u003e26\\u003c/sup\\u003e. Although we attempted to pair the echo report closely associated with chest CTs in time, variation in EF overtime is also well described. We plan to extend our study by considering different training data set building strategies when data collected from different real-life scenarios are available (e.g., sufficiently large cohort of patients with same day CT session followed by echo session). Our study only considered non-contrast CT scans to include a broader range of patients and increase dataset size. However, given the distinct features provided by contrast-enhanced CT, such as clearer visualization of blood vessels, we intend to expand our research by training the model on contrast-enhanced CT data. Additionally, we used Grad-CAM\\u003csup\\u003e\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e\\u003c/sup\\u003e, a saliency map approach, as the method for visualizing imaging biomarkers, applied to the individual scans. This offers qualitative interpretation in case studies but was limited in exploring population-level insights. We attribute this limitation to the lack of registration between different CT scans for comparing between different CT scans at the voxel-level, and plan to address it by integrating advanced CT registration approaches \\u003csup\\u003e\\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e\\u003c/sup\\u003e in our future research.\\u003c/p\\u003e \\u003cp\\u003eIn conclusion, we developed and validated an AI model that can successfully predict abnormal LVEF status, directly from non-gated non-contrast CT volumetric imaging data. Our study curated a large multi-modality dataset of paired CT-echo studies based on non-contrast CT scans and echo study reports, using data collected from two independent academic medical centers. As the first study of such cross-modality prediction of cardiac disease indicators using chest CT, our model demonstrated promising performance in both the internal validation and the external validation. Our AI model also outperformed experienced radiologists as human interpreters on selected subsets from both cohorts. We also identified possible imaging biomarkers through study of selected representative CT scans overlaid by saliency maps. This initiative provides empirical and numerical evidence supporting the use of AI-based opportunistic screening in real-world clinical settings to facilitate earlier diagnosis and possible intervention for patients with HfmEF and HFrEF. By analyzing chest CT imaging data, this approach may not only optimize diagnostic accuracy for structural cardiac diseases but also lead to the discovery of novel biomarkers in structural features, which could significantly improve patient care and deepen our understanding of cardiac pathophysiology.\\u003c/p\\u003e \"},{\"header\":\"Methods\",\"content\":\"\\u003cp\\u003e\\u003cstrong\\u003eEthics approval\\u003c/strong\\u003e\\u003c/p\\u003e\\n\\u003cp\\u003eThis study was approved by the institutional review boards (IRB) at Weill Cornell Medicine and Columbia University Irving Medical Center. A waiver for informed consent was obtained.\\u003c/p\\u003e\\n\\u003cp\\u003eData Curation and Preprocessing\\u003c/p\\u003e\\n\\u003cp\\u003eAll patients aged 18 years or older who had at least one echocardiogram (echo) within 6 months before or after a CT session at either Columbia University Irving Medical Center (CU) or Weill Cornell Medical Center (WCM) between July 2005 and August 2023 were identified. This initial data collection process identified 31,576 patients. The echo study reports and CT imaging data were then filtered, linked, and preprocessed according to the protocols detailed in the following sections. Eventually, we identified a cohort of 19,410 patients with a total of 34,058 qualified echo-CT data pairs. From this cohort, two study groups were formed based on the source of the CT data: CU (14,083 unique patients with 25,948 echo-CT pairs) and WCM (5,327 unique patients with 8,110 echo-CT pairs). The CU group served as the primary dataset for AI model development, while the WCM group was reserved for external validation. The CU dataset was further randomly split at the patient level into training, validation, and test subsets (i.e., each patient was assigned to only 1 of these 3 subsets). For our AI models, we created CT-echo study pairs using a CT volume image as the input and the LVEF value derived from an echo report as the target label for each pair. These pairs are formed and curated through \\u003cem\\u003edata linkage\\u003c/em\\u003e, \\u003cem\\u003edata filtering (see Supplementary Materials)\\u003c/em\\u003e, and the \\u003cem\\u003e3D CT volume preprocessing\\u003c/em\\u003e. The data curation and preprocessing flowchart is in Fig. \\u003cspan class=\\\"InternalRef\\\"\\u003e5\\u003c/span\\u003e.\\u003c/p\\u003e\\n\\u003ch3\\u003e3D non-contrast chest CT volume preprocessing\\u003c/h3\\u003e\\n\\u003cp\\u003eDICOM images of CT scans were transformed into Hounsfield Units (HU) to provide a consistent and standard measurement for data interpretation and analysis. The 2D DICOM images of the same CT series were merged into a 3D Numpy array to represent this 3D CT volume.. A window of [-1,000, 1,000] was applied to the intensity values to exclude air (\\u0026lt;-1,000 HU) and bones (\\u0026gt;\\u0026thinsp;1,000 HU). The interpolation function from the Python library Scipy ndimage \\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e\\u003c/sup\\u003e was used to rescale the 3D voxel spacing of all the 3D CT volumes into a standardized resolution of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:2\\\\times\\\\:2\\\\times\\\\:2\\\\:m{m}^{3}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e,where the spline interpolation was applied. Finally, 3D cropping and zero padding were applied where applicable, to resize the 3D volume into a volume of size 164 x 164 x 164.\\u003c/p\\u003e\\n\\u003cp\\u003eAI model: Classifier based on a Pretrained CT-ViT Encoder\\u003c/p\\u003e\\n\\u003cp\\u003eWe used a vision transformer architecture\\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e\\u003c/sup\\u003e: the encoder of the CT-ViT framework from GenerateCT\\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e\\u003c/sup\\u003e, as our backbone model for feature learning from the 3D CT images. We utilized the pre-trained weights of this encoder\\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e\\u003c/sup\\u003e and further trained it on our training cohort where both CT and echo studies were from CU. This encoder takes a preprocessed 3D CT volume image of size 164 x 164 x 164 and randomly cropped into \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:164\\\\times\\\\:144\\\\times\\\\:144\\\\)\\u003c/span\\u003e\\u003c/span\\u003e (dimensions are depth, height and width: \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:Z\\\\times\\\\:H\\\\times\\\\:W\\\\)\\u003c/span\\u003e\\u003c/span\\u003e) as input and output a \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e-dimensional feature vector. The CT-ViT encoder consists of three modules: the patch embedding layer, the spatial transformer module and the causal transformer module. The patch embedding layer first extracts non-overlapping patches of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:2\\\\times\\\\:16\\\\times\\\\:16\\\\)\\u003c/span\\u003e\\u003c/span\\u003e from the 3D CT volume input. Each patch is then transformed into a \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e-dimensional feature vector (dimension in tensor: \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:D\\\\)\\u003c/span\\u003e\\u003c/span\\u003e) with a fully connected layer. This transformation yields a \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:\\\\left(\\\\frac{164}{2}\\\\right)\\\\:\\\\times\\\\:\\\\left(\\\\frac{144}{16}\\\\right)\\\\times\\\\:\\\\left(\\\\frac{144}{16}\\\\right)\\\\:\\\\times\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e feature tensor for each input 3D image. This tensor is then fed into the spatial and temporal transformer. Subsequently in the spatial transformer module, multiple transformer layers were applied with self-attention along the spatial dimensions (i.e., W and H) of the reshaped tensor of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:\\\\left(\\\\frac{164}{2}\\\\right)\\\\:\\\\times\\\\:(\\\\frac{144}{16}\\\\times\\\\:\\\\frac{144}{16})\\\\:\\\\times\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e (dimensions: \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:Z\\\\times\\\\:(H\\\\times\\\\:W)\\\\times\\\\:D\\\\)\\u003c/span\\u003e\\u003c/span\\u003e). This is followed by the causal transformer module where multiple transformer layers were applied over the temporal (i.e., Z) dimension of the reshaped tensor of\\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:(\\\\frac{144}{16}\\\\times\\\\:\\\\frac{144}{16})\\\\:\\\\times\\\\:\\\\left(\\\\frac{164}{2}\\\\right)\\\\:\\\\times\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e (dimensions: \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:(H\\\\times\\\\:W)\\\\times\\\\:Z\\\\times\\\\:D\\\\)\\u003c/span\\u003e\\u003c/span\\u003e) with causal self-attention such that each spatial token only observes spatial tokens from previous slices in an auto-regressive manner. The dimensionality of output is retained after each spatial and causal transformer layer, ensuring that the volumetric information is preserved throughout the model fine-tuning. Finally, an average pooling over spatial (W and H) and temporal (Z) dimension of output tensor was applied to obtain the output \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:512\\\\)\\u003c/span\\u003e\\u003c/span\\u003e-dimensional feature vector for each input. A fully connected layer is used as a classification head and outputs a univariate prediction as a probability for the input 3D image to be a positive sample.\\u003c/p\\u003e\\n\\u003cp\\u003eFor predicting binarized LVEF, the AI model were trained to minimize the Binary Cross Entropy loss between the prediction and the binarized LVEF derived from echo report in Eq. \\u003cspan class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e, where the i-th input preprocessed 3D CT volume is denoted \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{x}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, its binarized LVEF derived from echo report \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{y}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, and the prediction made by our framework based on CT-ViT encoder \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:{\\\\widehat{y}}_{i}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, the total number of training instances \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:N\\\\)\\u003c/span\\u003e\\u003c/span\\u003e.\\u003c/p\\u003e\\n\\u003cdiv id=\\\"Equ1\\\" class=\\\"Equation\\\"\\u003e\\n \\u003cdiv class=\\\"mathdisplay\\\" id=\\\"FileID_Equ1\\\" name=\\\"EquationSource\\\"\\u003e$$\\\\:{L}_{BCE}({x}_{i},{y}_{i},\\\\:{\\\\widehat{y}}_{i})\\\\:=\\\\:-{\\\\sum\\\\:}_{i=1}^{N}{y}_{i}log{(\\\\widehat{y}}_{i})\\\\:+\\\\:(1\\\\:-\\\\:{y}_{i}\\\\left)log\\\\right(1-{\\\\widehat{y}}_{i})$$\\u003c/div\\u003e\\n \\u003cdiv class=\\\"EquationNumber\\\"\\u003e1\\u003c/div\\u003e\\n\\u003c/div\\u003e\\n\\u003cp\\u003eWe used the AdamW optimizer with an initial learning rate of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:1{0}^{-5}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e, weight decay of \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:1{0}^{-4}\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and batch size of 8 for 20 epochs. The learning rate decayed to half at epoch 15. During training, random cropping from \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:164\\\\times\\\\:164\\\\times\\\\:164\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e to \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:164\\\\times\\\\:144\\\\times\\\\:144\\\\)\\u003c/span\\u003e\\u003c/span\\u003e and random horizontal flipping were used for data augmentation. During testing, center cropping from \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:164\\\\times\\\\:164\\\\times\\\\:164\\\\:\\\\)\\u003c/span\\u003e\\u003c/span\\u003e to \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:164\\\\times\\\\:144\\\\times\\\\:144\\\\)\\u003c/span\\u003e\\u003c/span\\u003e were used. To assess the model\\u0026apos;s behavior and reliability, we trained it using the CU training and validation sets five times using different random seeds, resulting in five distinct sets of model weights. The AI model was trained and evaluated with a single NVIDIA A100 GPU.\\u003c/p\\u003e\\n\\u003cp\\u003eEvaluation metrics\\u003c/p\\u003e\\n\\u003ch3\\u003eLVEF-based classification metric\\u003c/h3\\u003e\\n\\u003cp\\u003eLVEF values were collected from the echo report to form labels for both training and testing. They were binarized according to a clinician-determined threshold (50%), where higher than or equal to this threshold is considered normal and lower than this threshold is considered to be indication of reduced ejection fraction (HFmrEF and/or HFrEF). A CT-echo pair where the echo-derived LVEF \\u003cspan class=\\\"InlineEquation\\\"\\u003e\\u003cspan class=\\\"mathinline\\\"\\u003e\\\\(\\\\:\\u0026lt;50\\\\%\\\\)\\u003c/span\\u003e\\u003c/span\\u003e was defined as the \\u0026ldquo;positive\\u0026rdquo; class for calculation of ROC, AUROC, sensitivity, specificity, accuracy and balanced accuracy. In addition, we evaluated the LVEF-based classification performance stratified by sites and demographics on the mixed-site test cohort. Particularly, site-wise we investigated CU-CT and WCM-echo, WCM-CT and CU-echo, as well as WCM-CT and WCM-echo. Each sub-cohort is further stratified by demographics, such as patient age, gender and race. In each stratified test subset, we investigated the LVEF-classification performance regarding different ordering of CT and echo studies in each pair - i.e., CT followed by echo or echo followed by CT, and different time gaps between the two studies - i.e., same day, one week, 1 month. Refer to the Supplementary Table \\u003cspan class=\\\"InternalRef\\\"\\u003eS2\\u003c/span\\u003e for the detailed breakdown of the number of study pairs in each site-specific test sub cohort.\\u003c/p\\u003e\\n\\u003cp\\u003eInterpretation of the Model\\u003c/p\\u003e\\n\\u003ch3\\u003eVisualization of contributive regions of AI model\\u003c/h3\\u003e\\n\\u003cp\\u003eTo understand the predictions made by the deep learning model\\u0026apos;s predictions, we used Gradient-weighted Class Activation Mapping (Grad-CAM)\\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e\\u003c/sup\\u003e, a popular technique for visualizing the regions within an input image that contributes most significantly to the model\\u0026apos;s output prediction. Grad-CAM computes the gradient of the target class with respect to feature maps and generates a heatmap overlay that visually indicates which parts of the image are most influential in the model\\u0026apos;s decision-making process.\\u003c/p\\u003e\\n\\u003ch3\\u003eRadiologist manual prediction of EF status from CT scans\\u003c/h3\\u003e\\n\\u003cp\\u003eIn predicting (EF) status from CT imaging data, the radiologists reported to focus on LV chamber dilatation and various ancillary findings. Vue PACS, a software product by Philips, was used for viewing and interpreting CT scans. While Vue PACS allows for detailed imaging review and manual measurements, it does not integrate AI-driven tools for evaluating chamber dilatation or advanced cardiac metrics, such as left atrial or ventricular size.\\u003c/p\\u003e\\n\\u003cp\\u003eThe radiologists primarily assessed whether the left-sided heart chambers were significantly enlarged. They also looked for signs of calcified apical aneurysms or fatty infiltration in the apex, which suggest past myocardial infarction and scarring, further indicating compromised EF, or sternal wires indicating prior sternotomy. Additionally, they examined features such as pleural and pericardial effusions, interlobular septal thickening (a sign of interstitial edema), and lung parenchymal edema. These findings, along with cardiac dilatation, suggested heart failure and support the prediction of an abnormal EF.\\u003c/p\\u003e\\n\\u003cp\\u003eAlthough Vue PACS facilitates these evaluations, the radiologists emphasized the subjective nature of the predictions, as the software does not provide automated tools for assessing chamber size or making advanced cardiac measurements. Many patients with altered EF may not show overt signs on non-contrast CT scans. While heavy coronary artery calcification might suggest myocardial damage, it does not always correlate with reduced EF. The radiologists relied on their expertise in combining manual observations within Vue PACS but acknowledged that echocardiography or cardiac MRI provide more precise assessments of EF and are the standard of care.\\u003c/p\\u003e\\n\\u003cdiv id=\\\"Sec8\\\" class=\\\"Section2\\\"\\u003e\\n \\u003ch2\\u003eStatistical analysis\\u003c/h2\\u003e\\n \\u003cp\\u003eThe performance of the AI model, with the output being probability scores, were evaluated using the Receiver Operating Characteristic (ROC) curve, and measured with the Area Under ROC (AUROC). Sensitivity and specificity pairs tables were evaluated with either one of the two configured with different thresholds and computed the other using the corresponding point on the ROC curve. The binary classification results, both from the AI models and from the radiologist manual reading results, were evaluated using the F1-score (i.e., harmonic mean of Precision and Recall) and balanced accuracy (i.e., the average of sensitivity and specificity) metrics. All metrics were computed through an open-source python library Scikit-learn (sklearn)\\u003csup\\u003e\\u003cspan class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e\\u003c/sup\\u003e. The mean results, standard deviations, and confidence intervals were calculated using model weights trained with five different random seeds on the CU training set. The computations used Python\\u0026apos;s native \\u003cem\\u003estatistics\\u003c/em\\u003e library, specifically the \\u003cem\\u003emean\\u003c/em\\u003e and \\u003cem\\u003estdev\\u003c/em\\u003e methods.\\u003c/p\\u003e\\n\\u003c/div\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003ch2\\u003eAuthor Contributions\\u003c/h2\\u003e \\u003cp\\u003e \\u003cb\\u003eInitiation and design of the study\\u003c/b\\u003e: Jayant Raikhelkar, Zilong Bai, Ashley N. Beecy, Deborah Estrin, Mert Sabuncu, Nir Uriel. \\u003cb\\u003eData collection\\u003c/b\\u003e: Chris Kelsey, David vanMaanan, Jeffrey Ruhl. \\u003cb\\u003eDeep learning and statistical analysis\\u003c/b\\u003e: Zilong Bai, Ashley N. Beecy, Fengbei Liu, Nusrat Binta Nizam, Varsha Kishore. \\u003cb\\u003eRadiologist review\\u003c/b\\u003e: Jay Leb, Alan Legasto. \\u003cb\\u003eSupervision of research\\u003c/b\\u003e: Ashley N. Beecy, Deborah Estrin, Mert Sabuncu, Nir Uriel. \\u003cb\\u003eWriting the first draft of the manuscript\\u003c/b\\u003e: Jayant Raikhelkar, Zilong Bai, Mert Sabuncu. \\u003cb\\u003eAll authors contributed to the writing and editing of the manuscript and approved the manuscript.\\u003c/b\\u003e\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\u003cli\\u003e\\u003cspan\\u003eBozkurt B et al (2021) Universal Definition and Classification of Heart Failure: A Report of the Heart Failure Society of America, Heart Failure Association of the European Society of Cardiology, Japanese Heart Failure Society and Writing Committee of the Universal Definition of Heart Failure. J Card Fail. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.cardfail.2021.01.022\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.cardfail.2021.01.022\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBozkurt B et al (2024) HF STATS. : Heart Failure Epidemiology and Outcomes Statistics An Updated 2024 Report from the Heart Failure Society of America. \\u003cem\\u003eJournal of cardiac failure\\u003c/em\\u003e (2024) \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.cardfail.2024.07.001\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.cardfail.2024.07.001\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBozkurt B et al (2023) Heart Failure Epidemiology and Outcomes Statistics: A Report of the Heart Failure Society of America. J Card Fail 29\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eLippi G, Sanchis-Gomar F (2020) Global epidemiology and future trends of heart failure. AME Med J 5\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAuthors/Task Force Members: et al. (2021) ESC Guidelines for the diagnosis and treatment of acute and chronic heart failure: Developed by the Task Force for the diagnosis and treatment of acute and chronic heart failure of the European Society of Cardiology (ESC). With the special contribution of the Heart Failure Association (HFA) of the ESC. \\u003cem\\u003eEur J Heart Fail\\u003c/em\\u003e 24, 4\\u0026ndash;131 (2022)\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHeidenreich PA et al (2022) 2022 AHA/ACC/HFSA Guideline for the Management of Heart Failure: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. J Am Coll Cardiol 79\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSavarese G, Stolfo D, Sinagra G, Lund LH (2022) Heart failure with mid-range or mildly reduced ejection fraction. Nat Rev Cardiol 19\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMc Causland FR et al (2024) Finerenone and Kidney Outcomes in Patients with Heart Failure: The FINEARTS-HF Trial. J Am Coll Cardiol. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1016/j.jacc.2024.10.091\\u003c/span\\u003e\\u003cspan address=\\\"10.1016/j.jacc.2024.10.091\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eWebsite \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.consumerreports.org/cro/magazine/2015/01/the-surprising-dangers-of-ct-sans-and-x-rays/index.htm\\u003c/span\\u003e\\u003cspan address=\\\"https://www.consumerreports.org/cro/magazine/2015/01/the-surprising-dangers-of-ct-sans-and-x-rays/index.htm\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMiller RJH et al (2024) Predicting mortality from AI cardiac volumes mass and coronary calcium on chest computed tomography. Nat Commun 15\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSandhu AT et al (2023) Incidental Coronary Artery Calcium: Opportunistic Screening of Previous Nongated Chest Computed Tomography Scans to Improve Statin Rates (NOTIFY-1 Project). Circulation 147:703\\u0026ndash;714\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDosovitskiy A et al (2020) An Image is Worth 16x16 Words. Transformers for Image Recognition at Scale\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSelvaraju RR et al Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://ieeexplore.ieee.org/document/8237336\\u003c/span\\u003e\\u003cspan address=\\\"https://ieeexplore.ieee.org/document/8237336\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eSayed A et al (2024) Reversals in the Decline of Heart Failure Mortality in the US, 1999 to 2021. JAMA Cardiol 9:585\\u0026ndash;589\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eMaddox TM et al (2024) 2024 ACC Expert Consensus Decision Pathway for Treatment of Heart Failure With Reduced Ejection Fraction: A Report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 83\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003ePickhardt PJ et al (2023) Opportunistic Screening: Radiology Scientific Expert Panel. \\u003cem\\u003eRadiology\\u003c/em\\u003e 307\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAttia ZI et al (2019) Prospective validation of a deep learning electrocardiogram algorithm for the detection of left ventricular systolic dysfunction. J Cardiovasc Electrophys 30\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKashou AH et al (2021) Artificial Intelligence-Augmented Electrocardiogram Detection of Left Ventricular Systolic Dysfunction in the General Population. \\u003cem\\u003eMayo Clinic proceedings\\u003c/em\\u003e 96\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eAttia ZI et al (2019) Screening for cardiac contractile dysfunction using an artificial intelligence-enabled electrocardiogram. Nat Med 25\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKwon JM et al (2019) Development and Validation of Deep-Learning Algorithm for Electrocardiography-Based Heart Failure Identification. Korean circulation J 49\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eYao X et al (2021) Artificial intelligence-enabled electrocardiograms for identification of patients with low ejection fraction: a pragmatic, randomized clinical trial. Nat Med 27\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eBhave S et al (2024) Deep learning to detect left ventricular structural abnormalities in chest X-rays. Eur Heart J 45\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003ede Gonz\\u0026aacute;lez A (2009) B. Projected cancer risks from computed tomographic scans performed in the United States in 2007. Arch Intern Med 169\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eReport on the Economic Well-Being of U.S. Households in 2023 - (2024) May \\u003cem\\u003eBoard of Governors of the Federal Reserve System\\u003c/em\\u003e \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://www.federalreserve.gov/publications/2024-economic-well-being-of-us-households-in-2023-executive-summary.htm\\u003c/span\\u003e\\u003cspan address=\\\"https://www.federalreserve.gov/publications/2024-economic-well-being-of-us-households-in-2023-executive-summary.htm\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eJain SS, Mello MM, Shah NH (2024) Avoiding Financial Toxicity for Patients from Clinicians\\u0026rsquo; Use of AI. N Engl J Med. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003e10.1056/NEJMp2406135\\u003c/span\\u003e\\u003cspan address=\\\"10.1056/NEJMp2406135\\\" targettype=\\\"DOI\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eKim Y-J, Kim RJ (2011) The role of cardiac MR in new-onset heart failure. Curr Cardiol Rep 13:185\\u0026ndash;193\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eDida H, Charif F, Benchabane A (2022) Registration of computed tomography images of a lung infected with COVID-19 based in the new meta-heuristic algorithm HPSGWO. Multimedia Tools Appl 81:18955\\u0026ndash;18976\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003escipy.ndimage.interpolation .zoom \\u0026mdash; SciPy v0.14.0 Reference Guide. \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.ndimage.interpolation.zoom.html\\u003c/span\\u003e\\u003cspan address=\\\"https://docs.scipy.org/doc/scipy-0.14.0/reference/generated/scipy.ndimage.interpolation.zoom.html\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003eHamamci IE et al (2023) GenerateCT: Text-Conditional Generation of 3D Chest CT Volumes\\u003c/span\\u003e\\u003c/li\\u003e \\u003cli\\u003e\\u003cspan\\u003escikit-learn \\u003cspan class=\\\"ExternalRef\\\"\\u003e\\u003cspan class=\\\"RefSource\\\"\\u003ehttps://scikit-learn.org/stable/\\u003c/span\\u003e\\u003cspan address=\\\"https://scikit-learn.org/stable/\\\" targettype=\\\"URL\\\" class=\\\"RefTarget\\\"\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/span\\u003e\\u003c/li\\u003e\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":true,\"hideJournal\":true,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true},\"keywords\":\"\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-5677688/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-5677688/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003cp\\u003eHeart failure (HF), a major global health challenge, affects millions worldwide and poses substantial healthcare and economic burdens. The left ventricular ejection fraction (LVEF) is a critical dynamic parameter used to characterize HF and guide treatment. In this study, we developed and validated an artificial intelligence (AI) model capable of predicting abnormal LVEF directly from static, non-gated, non-contrast chest computed tomography (CT) scans, a novel application for an imaging modality typically used for unrelated indications. Using a multi-institutional dataset of 34,058 paired CT and echocardiogram studies from two academic centers, we trained our model on over 25,000 studies and validated it on 8,110 studies from a separate institution. Remarkably, our model demonstrated robust performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.786 on the hold-out test set and 0.755 on external validation. Our findings are particularly promising given the widespread availability of CT scans\\u0026mdash;over 80\\u0026nbsp;million performed annually in the U.S.\\u0026mdash;making this opportunistic screening approach highly practical. Beyond strong predictive performance, the AI model outperformed expert radiologists in both accuracy and efficiency and provided interpretable visualizations highlighting imaging features linked to reduced LVEF. By enabling the identification of HF from routine chest CTs performed for other indications, this technology holds significant promise for early detection, reducing the diagnostic gap, and improving outcomes in asymptomatic HF.\\u003c/p\\u003e\",\"manuscriptTitle\":\"An Artificial Intelligence Model to Detect Abnormal Ejection Fraction from Non-Contrast Chest Computed Tomography: The CT-LVEF study\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2025-02-05 11:50:30\",\"doi\":\"10.21203/rs.3.rs-5677688/v1\",\"editorialEvents\":[],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"5475298c-d599-4657-b8e4-18a8bc462920\",\"owner\":[],\"postedDate\":\"February 5th, 2025\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[{\"id\":42456075,\"name\":\"Health sciences/Diseases/Cardiovascular diseases/Heart failure\"},{\"id\":42456076,\"name\":\"Health sciences/Health care/Medical imaging\"}],\"tags\":[],\"updatedAt\":\"2025-08-08T18:50:25+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2025-02-05 11:50:30\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-5677688\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-5677688\",\"identity\":\"rs-5677688\",\"version\":[\"v1\"]},\"buildId\":\"8U1c8b4HqxoKbykW_rLl7\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}