A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography

A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (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];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography D M Anisuzzaman , View ORCID Profile Jeffrey G. Malins , John I. Jackson , Eunjung Lee , Jwan A. Naser , Behrouz Rostami , Jared G. Bird , Dan Spiegelstein , Talia Amar , Christie C. Ngo , Jae K. Oh , Patricia A. Pellikka , Jeremy J. Thaden , Francisco Lopez-Jimenez , Timothy J. Poterucha , Paul A. Friedman , Sorin V. Pislaru , Garvan C. Kane , Zachi I. Attia doi: https://doi.org/10.1101/2025.07.10.25331304 D M Anisuzzaman 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jeffrey G. Malins 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Jeffrey G. Malins John I. Jackson 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Eunjung Lee 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jwan A. Naser 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MBBS Find this author on Google Scholar Find this author on PubMed Search for this author on this site Behrouz Rostami 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jared G. Bird 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Dan Spiegelstein 2 UltraSight Ltd. , Rehovot, Israel MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Talia Amar 2 UltraSight Ltd. , Rehovot, Israel MSc Find this author on Google Scholar Find this author on PubMed Search for this author on this site Christie C. Ngo 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN BSc Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jae K. Oh 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Patricia A. Pellikka 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Jeremy J. Thaden 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Francisco Lopez-Jimenez 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD, MSc, MBA Find this author on Google Scholar Find this author on PubMed Search for this author on this site Timothy J. Poterucha 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Paul A. Friedman 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Sorin V. Pislaru 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD, PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Garvan C. Kane 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN MD, PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site Zachi I. Attia 1 Department of Cardiovascular Medicine, Mayo Clinic , Rochester, MN PhD Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: attia.itzhak{at}mayo.edu Abstract Full Text Info/History Metrics Supplementary material Data/Code Preview PDF Abstract Echocardiography traditionally requires experienced operators to select and interpret clips from specific viewing angles. Clinical decision-making is therefore limited for handheld cardiac ultrasound (HCU), which is often collected by novice users. In this study, we developed a view-agnostic deep learning framework to estimate left ventricular ejection fraction (LVEF), patient age, and patient sex from any of several views containing the left ventricle. Model performance was: (1) consistently strong across retrospective transthoracic echocardiography (TTE) datasets; (2) comparable between prospective HCU versus TTE (625 patients; LVEF r 2 0.80 vs. 0.86, LVEF [> or ≤40%] AUC 0.981 vs. 0.993, age r 2 0.85 vs. 0.87, sex classification AUC 0.985 vs. 0.996); (3) comparable between prospective HCU data collected by experts versus novice users (100 patients; LVEF r 2 0.78 vs. 0.66, LVEF AUC 0.982 vs. 0.966). This approach may broaden the clinical utility of echocardiography by lessening the need for user expertise in image acquisition. Introduction Echocardiography stands as a fundamental diagnostic tool in cardiology, traditionally reliant on the meticulous selection of specific viewing angles and the interpretative expertise of human operators. This conventional approach, while effective, harbours inherent limitations due to its dependency on the skill and experience of the echocardiographer. For example, certain views are particularly challenging for inexperienced operators to acquire and select 1 . As a result, the conventional approach potentially leads to variability in assessments and constraints in the scope of analysis 2 . Moreover, due to a shortage of trained sonographers globally 3 , significant delays in obtaining a transthoracic echocardiogram (TTE) exist at many medical centres. In response to these challenges, this study presents a view-agnostic deep learning framework for echocardiographic analysis that reduces reliance on specific views. A view-agnostic learning framework may be especially useful for handheld cardiac ultrasound (HCU) images. HCU, which when deployed in point-of-care settings such as patient bedside is termed point-of-care ultrasound or POCUS, typically focuses on clinically relevant questions regarding the presence or absence of specific cardiovascular abnormalities and is often acquired by individuals with modest training. This variability in operator experience can translate to an even greater possibility for variability in image quality and study interpretation compared to a standard TTE, and consequently a greater need for a robust framework that is less susceptible to this potential variability. The objective of this study is to introduce and validate an imaging approach that eliminates the traditional requirements of view-specific echocardiography, enabling acquisition of data by users with minimal training. We hypothesized that by leveraging the capabilities of advanced convolutional neural networks trained using multiple TTE views, we could analyse echocardiographic data (either TTE or HCU) from any 2D view designated as valid (that is, containing the left ventricle). To test this hypothesis, we employed a range of retrospective and prospective echocardiographic datasets to develop, train, and validate a set of view-agnostic models. The first model demonstrates the power of this approach in estimating left ventricular ejection fraction (LVEF) – a basic measurement in echocardiography – whereas the second and third models respectively estimate a patient’s age and sex from echocardiographic images, thereby further demonstrating the power of this approach in extracting physiologic information. Methods Data acquisition and selection All study procedures were approved by the Mayo Clinic Institutional Review Board, and we have complied with all relevant ethical regulations. Patient cohorts only included individuals who had provided prior authorization for the use of their data in research. The model training cohort consisted entirely of TTE patients from Mayo Clinic Rochester and surrounding Mayo Clinic Health System sites in Minnesota and Wisconsin. In this dataset, there was one exam per patient from 19,627 patients total, with all echocardiogram study dates between January 1, 2007 and September 30, 2022. Each patient was randomly assigned to either the training, validation, or testing datasets to avoid cross-contamination of the datasets. The validation dataset was used to make decisions regarding model architecture as well as hyperparameter tuning. Clinically indicated echocardiography was recorded in clips of three cardiac cycles and performed using one of the following machines: GE Healthcare E95 or S70, Philips Epiq, or Siemens Acuson Sequoia. The reference measured LVEF was based on the clinical measure using the following hierarchy 4 : 3D volumetric measure, followed by a 2D biplane measurement (apical views), 2D modified Quinones, M-mode modified Quinones, and lastly, a 2D visual estimate by a level III trained expert echocardiographer. In addition to the randomly selected set of patients from Minnesota and Wisconsin that we used for internal testing, we also evaluated models using several additional testing datasets, as described below. More details regarding inclusion and exclusion criteria can be found in the Supplementary Material. Randomly selected patients with TTE exams collected at the Mayo Clinic site in Scottsdale, Arizona (1,695 patients) between January 1, 2022 and February 29, 2024 Randomly selected patients with TTE exams collected at the Mayo Clinic site in Jacksonville, Florida (1,862 patients) between January 1, 2022 and February 29, 2024 The EchoNet Dynamic dataset, which is a publicly available dataset of TTE clips from 10,030 patients collected at Stanford University Hospital (data from 10,015 patients used for analysis; one A4C clip per study). Use of this dataset complies with the Stanford University School of Medicine Research Use Agreement and Terms of Use. A prospective cohort of 625 patients who had TTE and HCU data collected during the same session. These patients visited Mayo Clinic Rochester for a clinically indicated TTE exam between November 1, 2022 and September 30, 2023. Immediately following their TTE exam (i.e., within the same session), if patients provided verbal consent for additional clips to be collected for research purposes, the following five HCU 2D video clips were collected using a Philips Lumify device: PLAX, parasternal short axis, A4C, A3C, A2C. An HCU dataset of 100 patients collected across three external sites (35 patients from Aurora St. Luke’s Medical Center in Milwaukee, WI; 32 patients from the University of Chicago Medical Center; 33 patients from Sheba Medical Center in Israel). The study protocol was approved by each respective institution’s IRB committee, and all patients were consented according to the study protocol. From each patient, two operators each collected up to 10 views of the heart using a Philips Lumify device. One of the operators was an experienced sonographer, whereas the other operator was one of 9 novice users who used real-time guidance AI software (UltraSight Ltd.) to optimize the acquired views 1 . From the full exam, if more than one of the A2C and A4C views were obtained, the highest quality clip was selected for each view, as determined by the same real-time guidance AI software used for acquisition (UltraSight Ltd.). Data preprocessing DICOM (Digital Imaging and Communications in Medicine) data were preprocessed using a framework previously described 5 , which included steps such as identification of the echocardiographic imaging sector, removal of portions of the electrocardiogram trace overlapping this sector (using opencv version 4.5.5 in Python 6 ), and view classification of video clips. Clips were included for analysis only if they were B-mode, had at least 48 frames, and belonged to one of four ‘valid’ view categories containing the left ventricle: parasternal long axis (PLAX), apical 2-chamber (A2C), apical 3-chamber (A3C), and apical 4-chamber (A4C). Figure 1 provides an overview of the data processing workflow. More details regarding the preprocessing workflow can be found in the Supplementary Material. Download figure Open in new tab Figure 1: An overview of the view-agnostic deep learning framework employed in the current study. Note that the LVEF model estimates are hypothetical and do not constitute actual data points from the study. The icon used to denote the deep learning model was taken from BioRender.com. Model training We used the S3D model architecture developed by Xie et al. 7 , but adapted it to take in single channel (grayscale images) input. We first trained a model for LVEF estimation from scratch (i.e., without using the pre-trained weights for the S3D model) and saved the best model based on validation loss. This saved model’s weights were then used for transfer learning to train age estimation and sex classification models. We selected 24 frames to constitute a single input video by sampling every other frame from a fixed-length segment. To increase model generalizability, we applied five augmentations to the training dataset: random rotation between-10 and +10 degrees, Gaussian blurring, central and random cropping, and horizontal flipping 8 . With temporal sliding windows and augmentations, the models were trained on 473,803 clips from 19,627 patients. More details regarding model training and selection can be found in the Supplementary Material. Evaluating model performance To evaluate performance of the LVEF estimation model, in addition to assessing continuous output from the regression model, we also performed a binary classification between significantly reduced LVEF (clinically-calculated LVEF ≤ 40%) versus normal or mildly reduced LVEF (clinically-calculated LVEF > 40%). To do this, we first obtained the threshold LVEF value that balanced sensitivity and specificity for the validation dataset. We then applied this threshold (45.72%) to the testing datasets. For age, we performed a continuous regression, whereas for sex, we performed a binary classification between “male” and “female”. For the retrospective TTE datasets, we averaged model estimates across all valid clips containing the left ventricle. For the prospective dataset of simultaneously collected TTE and HCU clips, we instead selected one clip per view for the TTE dataset (i.e., the clip with the highest view classifier inference score for each view 9 ) to enable a direct comparison with the HCU dataset, for which only clip per view was collected. Details regarding the model evaluation metrics that we used can be found in the Supplementary Material. Results Sample characteristics Patient demographics and clinical characteristics are presented in Table 1 . As shown in Table 1 , (1) mean LVEF was lower in the model development datasets compared to the model evaluation datasets because the model development datasets were specifically enriched for the middle-to-lower end of the LVEF distribution, and (2) demographic characteristics and comorbidities differed according to geographic location, and were also influenced by patient selection criteria for the various cohorts. View this table: View inline View popup Table 1. Demographic and clinical characteristics of the patient cohorts for model development and model evaluation. LVEF estimation for TTE We evaluated the LVEF regression model on TTE datasets from Minnesota-Wisconsin, Arizona, and Florida, as well as the publicly available EchoNet dataset ( Figure 2 ). For Minnesota-Wisconsin, Arizona, and Florida we used all available valid-class clips, and so the result for each patient represents the average of the model outputs for each clip. In contrast, the EchoNet dataset contains only apical 4-chamber (A4C) clips, so there was only one single estimate from one clip per patient. Performance was strong in all cases, which was further confirmed by the Bland-Altman plots shown in Figure 3 . As shown in Figure 3 , for all four cohorts, the difference between model-estimated and clinically-calculated LVEF was less than 10% in at least 90% of cases. Download figure Open in new tab Figure 2: LVEF estimation and classification for the (A) Minnesota-Wisconsin, (B) Arizona, (C) Florida, and (D) EchoNet datasets. In each subplot, the left plot shows the correlation for LVEF estimation (i.e., comparison of model-estimated LVEF to clinically-calculated LVEF), whereas the right plot shows the ROC curve and confusion matrix for reduced versus normal LVEF classification. Download figure Open in new tab Figure 3: Bland-Altman plots showing the discrepancy between clinically-calculated and model-estimated LVEF for the (A) Minnesota-Wisconsin, (B) Arizona, (C) Florida, and (D) EchoNet datasets. The text in blue shows the number of cases with errors within certain bands (e.g., less than 5%, between 5% and 10%, and so on). The decision to use all available valid-class clips was based on a preliminary analysis evaluating the LVEF regression model using different numbers of randomly selected clips and all available valid-class clips for the Minnesota-Wisconsin dataset ( Figure 4 ). Experimental setups consisted of pulling between one and five random clips from all available valid-class videos for each patient and then averaging estimates across the clips that were pulled. This procedure of pulling random clips and averaging estimates was repeated 15 times, giving rise to the distributions shown in the leftmost five boxplots in each subplot. As shown in Figure 4 , as the number of valid-class clips increased, the median RMSE and MAE decreased, whereas the median r 2 and AUC increased. This analysis revealed that the LVEF model performs well even with a single valid-class clip, but performs even better as the number of clips increases. Download figure Open in new tab Figure 4: The effect of the number of clips on LVEF estimation performance for the Minnesota-Wisconsin testing dataset. The leftmost five boxplots within each panel are for one through five random clips selected from the full set of valid-class clips for a patient’s exam, whereas “All valid clips” represents all valid-class clips from a patient’s exam. In each boxplot, the center line corresponds to the median, the box limits correspond to the upper and lower quartiles, the whiskers correspond to 1.5 times the interquartile range, and the points correspond to outliers. In addition to evaluating performance based on the number of clips (regardless of which view was selected), we also assessed model performance based on selecting specific views or combinations of views ( Figure 5 ). Similar to the finding reported in Figure 4 , this analysis revealed that model performance was generally superior when more clips were selected, but additionally revealed that when the number of views was held constant, the specific views that were taken did not have a systematic impact on performance. This finding suggests the model is indeed view-agnostic. Download figure Open in new tab Figure 5: LVEF model performance for the four ‘valid’ views and combinations of these views. For combinations of views, model performance was evaluated using the average model estimate per patient across views, with only one clip taken per view (i.e., the clip with the highest score from the view classifier for that view). The legend on the left was plotted using version 0.9.0 of the upsetplot package 38 in Python version 3.9.20. LVEF estimation for HCU We evaluated the LVEF regression model on HCU datasets from Mayo-Rochester and an external dataset collected by UltraSight Ltd. ( Figure 6 ). For the Mayo-Rochester dataset, we observed very comparable results between TTE and HCU for both LVEF estimation and classification of reduced versus normal LVEF. A limitation of this experiment was that both TTE and HCU clips were collected by experienced sonographers. This limitation is addressed by the strong correlation that was observed between LVEF estimates for clips collected from expert sonographers versus novice users in the Ultrasight Ltd. dataset. In only 6% of cases, the difference in model-estimated LVEF between novice versus expert-acquired clips was greater than 10%. Download figure Open in new tab Figure 6: LVEF estimation and classification performance for (A) the simultaneously collected TTE-HCU dataset, and (B) the HCU dataset collected by both experts and novice users using UltraSight real-time AI guidance technology. The green lines between each set of points in the leftmost plots (i) correspond to the difference between estimates. The red points in the middle plots (ii) show the outliers for which the difference between estimates is greater than or equal to 10. For the outliers, in the legend, the true LVEF value is shown with “T” and the difference is shown with “D”. The rightmost plots (iii) show the ROC curves and metrics for reduced versus normal LVEF classification. Age estimation and sex classification We observed a very strong correlation between chronological age and model-estimated age for all three of the Minnesota-Wisconsin, Arizona, and Florida cohorts, as well as a tight overlap with small differences (short green lines between each estimate pair) between AI-age estimates for HCU and TTE ( Figure 7 ). Download figure Open in new tab Figure 7: Model performance for the view-agnostic age estimation and sex classification models. The left three plots show age estimation (A) and (B) sex classification results for the TTE cohorts for both the internal (i) and external ((ii) and (iii)) testing datasets. The rightmost plot (iv) shows the overlapping correlation for HCU and TTE data collected simultaneously from the same set of patients at Mayo-Rochester, with green line between each corresponding set of estimates showing the difference between HCU versus TTE model-estimated age. The 95% confidence interval for each corresponding metric is shown in square brackets. Sex classification model performance (Figure 8) was excellent, with an AUC greater than 0.981 for all three of the Minnesota-Wisconsin, Arizona, and Florida cohorts. Nearly overlapping ROC curves were also observed for HCU and TTE exams. Discussion In this study, we created an AI framework using convolutional neural networks that are view agnostic, meaning that they do not depend on specific views to generate clinically useful results. We demonstrate that this framework can automatically derive a clinically important parameter – left ventricular ejection fraction – reliably and with a level of accuracy similar to that of human cardiologists 10 , 11 . In addition, we provide evidence that this framework can also extract physiologic information beyond that which expert human readers can report (age and sex), showing the feasibility of this framework to potentially extend the capabilities of cardiac ultrasound. As we observed, this framework is not only applicable for images collected using comprehensive TTE, but also for images acquired utilizing HCU by both experienced sonographers and novice users. Consequently, the view-agnostic approach may lessen the need for user expertise in image acquisition and stands to make echocardiography a more widely available and usable clinical tool. In the past, deep learning approaches have typically used specific views to develop and/or validate deep learning or machine learning models for LVEF estimation or classification 2 , 12 – 26 , with some studies also segmenting the left ventricle or estimating intermediate values prior to computation of LVEF 2 , 5 , 12 – 25 , 27 – 29 . However, more recent models, such as PanEcho 30 and DROID 31 , estimate measurements such as LVEF without requiring segmentation or computing intermediate values, thereby eliminating potential vulnerability to errors that may arise during this process, and furthermore do not have specific view requirements, thereby allowing for greater flexibility of model input. Like PanEcho and DROID, the current set of models are also view-agnostic and do not require segmentation, but go beyond these models in their demonstrated success with HCU, as well as in their successful estimation of patient age and classification of patient sex in a view-agnostic manner. These results have far-reaching implications. For example, rapid screening of at-risk cohorts and symptomatic patients in the clinic, emergency department, or hospital for the presence or absence of LV dysfunction would provide rapid and accurate cardiovascular triage capabilities. Additionally, these tools placed in the hands of Emergency Medical Services providers, who are typically non-expert sonographers, could allow for rapid cardiac assessment of patients before they reach the emergency department in order to better triage severe or critically ill patients prior to arrival to the hospital. Finally, although age estimation and sex classification may not have immediate clinical relevance, these tasks could potentially serve as a useful quality control measure given that ground truth values for these parameters are readily available. Furthermore, examining discrepancies between model estimates and ground truth values may allow for a better understanding of the role these factors play in cardiovascular risk prediction 32 – 36 . With that said, our work is best understood in the context of its limitations. Firstly, while the entire patient cohort was highly technically and geographically diverse, the patient cohort used for model training was not diverse in terms of race and ethnicity. This could affect the model’s accuracy and applicability across various global populations, highlighting the need for future studies to incorporate a broader patient demographic to ensure the models’ performance is universally reliable. Second, the handheld cardiac ultrasound data were collected in controlled environments, and therefore may not accurately reflect the variability typical of POCUS scenarios in diverse clinical environments. Finally, the use of a view classifier to identify valid-class views introduces the possibility of downstream errors in LVEF, age, or sex estimation. Future work should examine model performance across a broader set of view categories than just those containing the LV, and should also consider clips collected from viewing angles outside the canonical set of views. Data availability All requests for raw and analysed data and related materials, excluding programming code, will be reviewed by the Mayo Clinic legal department and Mayo Clinic Ventures to verify whether the request is subject to any intellectual property or confidentiality obligations. Requests for patient-related data not included in the paper will not be considered. Any data and materials that can be shared will be released via a Material Transfer Agreement. Code availability Programming code related to the PyTorch model specification will be made available under the GNU General Public License version 3 upon request to Z.I.A. ( attia.itzhak{at}mayo.edu ). Data Availability All requests for raw and analysed data and related materials, excluding programming code, will be reviewed by the Mayo Clinic legal department and Mayo Clinic Ventures to verify whether the request is subject to any intellectual property or confidentiality obligations. Requests for patient-related data not included in the paper will not be considered. Any data and materials that can be shared will be released via a Material Transfer Agreement. Footnotes Disclosures: D.A., J.G.M., J.I.J., E.L., J.A.N., S.V.P., G.C.K., and Z.I.A. have invented algorithms licensed to UltraSight Ltd. and may benefit from algorithm commercialization via Mayo Clinic. D.S. and T.A. are both employees of UltraSight Ltd. F.L.-J. is a member of the Scientific Advisory Board for UltraSight Ltd. P.A.P. received research support from the American Society of Echocardiography Foundation, Ultromics Ltd., GE Healthcare, and Edwards Lifesciences (all funds paid to Mayo Clinic). P.A.P. is also a consultant/advisor for Astellas Pharma. T.J.P. received research support from the American Heart Association, the Amyloidosis Foundation, Eidos Therapeutics, Pfizer, Janssen, and Edwards Lifesciences. T.J.P. also owns stock in Abbott Laboratories and Baxter International. P.A.F., F.L.-J., and Z.I.A. are members of the Scientific Advisory Board for Anumana, an AI company commercializing AI-ECG. F.L.-J. is a co-inventor of several algorithms using AI-ECG licensed to Anumana and a co-inventor of an artificial heart valve that has been licensed to Colibri Co. The remaining authors have nothing to disclose Abbreviations TTE transthoracic echocardiography HCU handheld cardiac ultrasound POCUS point-of-care ultrasound LVEF left ventricular ejection fraction ROC receiver-operator curve RMSE Root Mean Square Error MAE Mean Absolute Error PC Pearson Correlation AUC area under the curve CI confidence interval(s) PPV positive predictive value NPV negative predictive value References 1. ↵ Mor-Avi V , Khandheria B , Klempfner R , Cotella JI , Moreno M , Ignatowski D , Guile B , Hayes HJ , Hipke K , Kaminski A , et al. Real-Time Artificial Intelligence–Based Guidance of Echocardiographic Imaging by Novices: Image Quality and Suitability for Diagnostic Interpretation and Quantitative Analysis . Circulation: Cardiovascular Imaging . 2023 ; 16 . doi: 10.1161/circimaging.123.015569 OpenUrl CrossRef 2. ↵ Zhang J , Gajjala S , Agrawal P , Tison GH , Hallock LA , Beussink-Nelson L , Lassen MH , Fan E , Aras MA , Jordan C , et al. Fully Automated Echocardiogram Interpretation in Clinical Practice . Circulation . 2018 ; 138 : 1623 – 1635 . doi: 10.1161/CIRCULATIONAHA.118.034338 OpenUrl CrossRef PubMed 3. ↵ Collins K , Little SH. Sonographer Challenges Today . J Am Soc Echocardiogr . 2023 ; 36 :A7-A8. doi: 10.1016/j.echo.2023.01.008 OpenUrl CrossRef 4. ↵ Lang RM , Badano LP , Mor-Avi V , Afilalo J , Armstrong A , Ernande L , Flachskampf FA , Foster E , Goldstein SA , Kuznetsova T , et al. Recommendations for cardiac chamber quantification by echocardiography in adults: an update from the American Society of Echocardiography and the European Association of Cardiovascular Imaging . Eur Heart J Cardiovasc Imaging . 2015 ; 16 : 233 – 270 . doi: 10.1093/ehjci/jev014 OpenUrl CrossRef PubMed 5. ↵ Anisuzzaman D , Malins JG , Jackson JI , Lee E , Naser JA , Rostami B , Greason G , Bird JG , Friedman PA , Oh JK , et al. Leveraging comprehensive echo data to power AI models for handheld cardiac ultrasound . Mayo Clinic Proceedings: Digital Health . 2025 ; 3 : 100194 . 6. ↵ Bradski G . The OpenCV Library . Dr Dobb’s Journal of Software Tools . 2000 . 7. ↵ Xie S , Sun C , Huang J , Tu Z , Murphy K . Rethinking spatiotemporal feature learning: Speed-accuracy trade-offs in video classification . Paper/Poster presented at: Proceedings of the European conference on computer vision (ECCV ); 2018 ; 8. ↵ Torch Contributors. PyTorch: Transforming and Augmenting Images. https://pytorch.org/vision/stable/transforms.html . 2017. Accessed May 11 . 9. ↵ Naser JA , Lee E , Pislaru SV , Tsaban G , Malins JG , Jackson JI , Anisuzzaman D , Rostami B , Lopez-Jimenez F , Friedman PA , et al. Artificial intelligence-based classification of echocardiographic views . European Heart Journal - Digital Health . 2024 ; 5 : 260 – 269 . doi: 10.1093/ehjdh/ztae015 OpenUrl CrossRef PubMed 10. ↵ Galema T , Geleijnse M , Yap S , Vandomburg R , Biagini E , Vletter W , Tencate F . Assessment of left ventricular ejection fraction after myocardial infarction using contrast echocardiography . European Journal of Echocardiography . 2007 . doi: 10.1016/j.euje.2007.03.025 OpenUrl CrossRef PubMed Web of Science 11. ↵ Hoffmann R , Barletta G , von Bardeleben S , Vanoverschelde JL , Kasprzak J , Greis C , Becher H . Analysis of Left Ventricular Volumes and Function: A Multicenter Comparison of Cardiac Magnetic Resonance Imaging, Cine Ventriculography, and Unenhanced and Contrast-Enhanced Two-Dimensional and Three-Dimensional Echocardiography . Journal of the American Society of Echocardiography . 2014 ; 27 : 292 – 301 . doi: 10.1016/j.echo.2013.12.005 OpenUrl CrossRef PubMed 12. ↵ Asch FM , Abraham T , Jankowski M , Cleve J , Adams M , Romano N , Polivert N , Hong H , Lang R . Accuracy and reproducibility of a novel artificial intelligence deep learning-based algorithm for automated calculation of ejection fraction in echocardiography . Journal of the American College of Cardiology . 2019 ; 73 : 1447 – 1447 . doi : doi: 10.1016/S0735-1097(19)32053-4 OpenUrl CrossRef PubMed 13. Cannesson M , Tanabe M , Suffoletto MS , McNamara DM , Madan S , Lacomis JM , Gorcsan J , 3rd. A novel two-dimensional echocardiographic image analysis system using artificial intelligence-learned pattern recognition for rapid automated ejection fraction . J Am Coll Cardiol . 2007 ; 49 : 217 – 226 . doi: 10.1016/j.jacc.2006.08.045 OpenUrl FREE Full Text 14. Knackstedt C , Bekkers SC , Schummers G , Schreckenberg M , Muraru D , Badano LP , Franke A , Bavishi C , Omar AM , Sengupta PP . Fully Automated Versus Standard Tracking of Left Ventricular Ejection Fraction and Longitudinal Strain: The FAST-EFs Multicenter Study . J Am Coll Cardiol . 2015 ; 66 : 1456 – 1466 . doi: 10.1016/j.jacc.2015.07.052 OpenUrl FREE Full Text 15. Luong CL , Jafari MH , Behnami D , Shah YR , Straatman L , Van Woudenberg N , Christoff L , Gwadry N , Hawkins NM , Sayre EC , et al. Validation of machine learning models for estimation of left ventricular ejection fraction on point-of-care ultrasound: insights on features that impact performance . Echo Research & Practice . 2024 ; 11 . doi: 10.1186/s44156-024-00043-2 OpenUrl CrossRef 16. Papadopoulou SL , Sachpekidis V , Kantartzi V , Styliadis I , Nihoyannopoulos P . Clinical validation of an artificial intelligence-assisted algorithm for automated quantification of left ventricular ejection fraction in real time by a novel handheld ultrasound device . Eur Heart J Digit Health . 2022 ; 3 : 29 – 37 . doi: 10.1093/ehjdh/ztac001 OpenUrl CrossRef PubMed 17. Rahmouni HW , Ky B , Plappert T , Duffy K , Wiegers SE , Ferrari VA , Keane MG , Kirkpatrick JN , Silvestry FE , St John Sutton M. Clinical utility of automated assessment of left ventricular ejection fraction using artificial intelligence-assisted border detection . Am Heart J . 2008 ; 155 : 562 – 570 . doi: 10.1016/j.ahj.2007.11.002 OpenUrl CrossRef PubMed Web of Science 18. Liu X , Fan Y , Li S , Chen M , Li M , Hau WK , Zhang H , Xu L , Lee AP . Deep learning-based automated left ventricular ejection fraction assessment using 2-D echocardiography . Am J Physiol Heart Circ Physiol . 2021 ; 321 : H390 – H399 . doi: 10.1152/ajpheart.00416.2020 OpenUrl CrossRef PubMed 19. Tromp J , Seekings PJ , Hung CL , Iversen MB , Frost MJ , Ouwerkerk W , Jiang Z , Eisenhaber F , Goh RSM , Zhao H , et al. Automated interpretation of systolic and diastolic function on the echocardiogram: a multicohort study . Lancet Digit Health . 2022 ; 4 : e46 – e54 . doi: 10.1016/S2589-7500(21)00235-1 OpenUrl CrossRef 20. Asch FM , Mor-Avi V , Rubenson D , Goldstein S , Saric M , Mikati I , Surette S , Chaudhry A , Poilvert N , Hong H , et al. Deep Learning-Based Automated Echocardiographic Quantification of Left Ventricular Ejection Fraction: A Point-of-Care Solution . Circ Cardiovasc Imaging . 2021 ; 14 : e012293 . doi: 10.1161/CIRCIMAGING.120.012293 OpenUrl CrossRef PubMed 21. Filipiak-Strzecka D , Kasprzak JD , Wejner-Mik P , Szymczyk E , Wdowiak-Okrojek K , Lipiec P . Artificial Intelligence-Powered Measurement of Left Ventricular Ejection Fraction Using a Handheld Ultrasound Device . Ultrasound Med Biol . 2021 ; 47 : 1120 – 1125 . doi: 10.1016/j.ultrasmedbio.2020.12.003 OpenUrl CrossRef PubMed 22. Ouyang D , He B , Ghorbani A , Lungren MP , Ashley EA , Liang DH , Zou JY . Echonet-dynamic: a large new cardiac motion video data resource for medical machine learning. Paper/Poster presented at: NeurIPS ML4H Workshop ; 2019 ; Vancouver, BC, Canada . 23. Ouyang D , He B , Ghorbani A , Yuan N , Ebinger J , Langlotz CP , Heidenreich PA , Harrington RA , Liang DH , Ashley EA , et al. Video-based AI for beat-to-beat assessment of cardiac function . Nature . 2020 ; 580 : 252 – 256 . doi: 10.1038/s41586-020-2145-8 OpenUrl CrossRef PubMed 24. Samtani R , Bienstock S , Lai AC , Liao S , Baber U , Croft L , Stern E , Beerkens F , Ting P , Goldman ME . Assessment and validation of a novel fast fully automated artificial intelligence left ventricular ejection fraction quantification software . Echocardiography . 2022 ; 39 : 473 – 482 . doi: 10.1111/echo.15318 OpenUrl CrossRef PubMed 25. ↵ Reddy CD , Lopez L , Ouyang D , Zou JY , He B . Video-Based Deep Learning for Automated Assessment of Left Ventricular Ejection Fraction in Pediatric Patients . J Am Soc Echocardiogr . 2023 ; 36 : 482 – 489 . doi: 10.1016/j.echo.2023.01.015 OpenUrl CrossRef PubMed 26. ↵ Christensen M , Vukadinovic M , Yuan N , Ouyang D . Vision-language foundation model for echocardiogram interpretation . Nat Med . 2024 . doi: 10.1038/s41591-024-02959-y OpenUrl CrossRef PubMed 27. ↵ Blaivas M , Blaivas LN , Campbell K , Thomas J , Shah S , Yadav K , Liu YT . Making Artificial Intelligence Lemonade Out of Data Lemons: Adaptation of a Public Apical Echo Database for Creation of a Subxiphoid Visual Estimation Automatic Ejection Fraction Machine Learning Algorithm . J Ultrasound Med . 2022 ; 41 : 2059 – 2069 . doi: 10.1002/jum.15889 OpenUrl CrossRef PubMed 28. Ghorbani A , Ouyang D , Abid A , He B , Chen JH , Harrington RA , Liang DH , Ashley EA , Zou JY . Deep learning interpretation of echocardiograms . NPJ Digit Med . 2020 ; 3 : 10 . doi: 10.1038/s41746-019-0216-8 OpenUrl CrossRef PubMed 29. ↵ He B , Dash D , Duanmu Y , Tan TX , Ouyang D , Zou J . AI-Enabled Assessment of Cardiac Function and Video Quality in Emergency Department Point-of-Care Echocardiograms . The Journal of Emergency Medicine . 2023 . doi: 10.1016/j.jemermed.2023.02.005 OpenUrl CrossRef 30. ↵ Holste G , Oikonomou EK , Wang Z , Khera R . PanEcho: Complete AI-enabled echocardiography interpretation with multi-task deep learning . medRxiv . 2024 . doi: 10.1101/2024.11.16.24317431 OpenUrl Abstract / FREE Full Text 31. ↵ Lau ES , Di Achille P , Kopparapu K , Andrews CT , Singh P , Reeder C , Al-Alusi M , Khurshid S , Haimovich JS , Ellinor PT , et al. Deep Learning-Enabled Assessment of Left Heart Structure and Function Predicts Cardiovascular Outcomes . J Am Coll Cardiol . 2023 ; 82 : 1936 – 1948 . doi: 10.1016/j.jacc.2023.09.800 OpenUrl CrossRef PubMed 32. ↵ Kalyakulina A , Yusipov I , Moskalev A , Franceschi C , Ivanchenko M. eXplainable Artificial Intelligence (XAI) in aging clock models . arXiv . 2023 . doi: doi.org/ 10.48550/arXiv.2307.13704 OpenUrl CrossRef 33. Attia ZI , Friedman PA , Noseworthy PA , Lopez-Jimenez F , Ladewig DJ , Satam G , Pellikka PA , Munger TM , Asirvatham SJ , Scott CG , et al. Age and Sex Estimation Using Artificial Intelligence From Standard 12-Lead ECGs . Circ Arrhythm Electrophysiol . 2019 ; 12 : e007284 . doi: 10.1161/CIRCEP.119.007284 OpenUrl CrossRef PubMed 34. Baek YS , Lee DH , Jo Y , Lee SC , Choi W , Kim DH . Artificial intelligence-estimated biological heart age using a 12-lead electrocardiogram predicts mortality and cardiovascular outcomes . Front Cardiovasc Med . 2023 ; 10 : 1137892 . doi: 10.3389/fcvm.2023.1137892 OpenUrl CrossRef PubMed 35. Naser JA , Lopez-Jimenez F , Chang AY , Baez-Suarez A , Attia ZI , Pislaru SV , Pellikka PA , Lin G , Kapa S , Friedman PA , et al. Artificial Intelligence-Augmented Electrocardiogram in Determining Sex: Correlation with Sex Hormone Levels . Mayo Clin Proc . 2023 ; 98 : 541 – 548 . doi: 10.1016/j.mayocp.2022.08.019 OpenUrl CrossRef PubMed 36. ↵ Faierstein K , Fiman M , Loutati R , Rubin N , Manor U , Am-Shalom A , Cohen-Shelly M , Blank N , Lotan D , Zhao Q , et al. Artificial Intelligence Assessment of Biological Age From Transthoracic Echocardiography: Discrepancies with Chronologic Age Predict Significant Excess Mortality . Journal of the American Society of Echocardiography . 2024 ; 37 : 725 – 735 . doi: 10.1016/j.echo.2024.04.017 OpenUrl CrossRef PubMed 37. Gasparini A. comorbidity: An R package for computing comorbidity scores . Journal of Open Source Software . 2018 ; 3 . doi: 10.21105/joss.00648 OpenUrl CrossRef 38. ↵ Lex A , Gehlenborg N , Strobelt H , Vuillemot R , Pfister H . UpSet: Visualization of intersecting sets . IEEE Trans Vis Comput Graph . 2014 ; 20 : 1983 – 1992 . doi: 10.1109/TVCG.2014.2346248 OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted July 11, 2025. Download PDF Supplementary Material Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography D M Anisuzzaman , Jeffrey G. Malins , John I. Jackson , Eunjung Lee , Jwan A. Naser , Behrouz Rostami , Jared G. Bird , Dan Spiegelstein , Talia Amar , Christie C. Ngo , Jae K. Oh , Patricia A. Pellikka , Jeremy J. Thaden , Francisco Lopez-Jimenez , Timothy J. Poterucha , Paul A. Friedman , Sorin V. Pislaru , Garvan C. Kane , Zachi I. Attia medRxiv 2025.07.10.25331304; doi: https://doi.org/10.1101/2025.07.10.25331304 Share This Article: Copy Citation Tools A View-Agnostic Deep Learning Framework for Comprehensive Analysis of 2D-Echocardiography D M Anisuzzaman , Jeffrey G. Malins , John I. Jackson , Eunjung Lee , Jwan A. Naser , Behrouz Rostami , Jared G. Bird , Dan Spiegelstein , Talia Amar , Christie C. Ngo , Jae K. Oh , Patricia A. Pellikka , Jeremy J. Thaden , Francisco Lopez-Jimenez , Timothy J. Poterucha , Paul A. Friedman , Sorin V. Pislaru , Garvan C. Kane , Zachi I. Attia medRxiv 2025.07.10.25331304; doi: https://doi.org/10.1101/2025.07.10.25331304 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Cardiovascular Medicine Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4435) Dentistry and Oral Medicine (444) Dermatology (382) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1509) Epidemiology (15228) Forensic Medicine (30) Gastroenterology (1124) Genetic and Genomic Medicine (6597) Geriatric Medicine (668) Health Economics (997) Health Informatics (4534) Health Policy (1368) Health Systems and Quality Improvement (1613) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15916) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (146) Nephrology (667) Neurology (6599) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1144) Occupational and Environmental Health (957) Oncology (3332) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (691) Primary Care Research (711) Psychiatry and Clinical Psychology (5447) Public and Global Health (9230) Radiology and Imaging (2198) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0047da73e99dfa9',t:'MTc3OTU0MzgxOQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();