Cardiovascular disease is a leading cause of mortality, placing a significant burden on healthcare systems. Transthoracic echocardiography (TTE) is the gold standard for LVEF assessment, but its accessibility is limited. Point-of-care ultrasound (PoCUS), particularly focused cardiac ultrasound (FoCUS), offers a potential solution for rapid bedside evaluation. However, the reliability of FoCUS, especially when performed by clinicians with limited training, is a concern. Artificial intelligence (AI) integrated into FoCUS devices could improve accuracy and overcome user experience limitations. This study aimed to determine the diagnostic accuracy of AI-assisted FoCUS for LVEF assessment compared to TTE, and to compare accuracy between novice and experienced users. The increasing prevalence of cardiac disease necessitates efficient and accessible diagnostic tools. The high cost and limited availability of TTE highlight the need for alternative methods. FoCUS has emerged as a promising approach due to its portability and ease of use. However, its accuracy, particularly when used by non-experts, remains a subject of ongoing research and validation. AI-assisted FoCUS has the potential to address the limitations of traditional FoCUS by providing automated image analysis and LVEF calculation. This would potentially improve both accuracy and efficiency, making LVEF assessment more widely accessible. Previous studies have shown promise for AI in echocardiography, but further validation in real-world settings with a diverse range of users is needed. The integration of AI into FoCUS offers a potential solution by improving the accuracy and consistency of LVEF assessments, regardless of the sonographer's experience level. This would potentially make rapid and accurate LVEF assessment accessible in a broader range of clinical settings.

Existing literature demonstrates the increasing importance of point-of-care ultrasound (POCUS) in various clinical settings. Studies have shown PoCUS's effectiveness in improving diagnostic accuracy compared to physical examination alone. However, the majority of previous research on FoCUS assessment of LVEF has been limited to trained sonographers. Real-world clinical settings often involve providers with limited PoCUS training, raising concerns about the accuracy and reliability of LVEF assessments. The potential impact of AI in improving the accuracy of PoCUS for LVEF assessment has been explored, with early studies showing promising results. These studies, however, often involved experienced echocardiographers and carefully controlled settings. There's a critical gap in understanding how AI-assisted FoCUS performs in real-world conditions with a diverse range of user experience levels. The integration of AI in FoCUS is a relatively new area, with limited data on its real-world applicability and effectiveness for clinicians of different experience levels. Existing studies often focus on the performance of AI in formal echocardiography settings rather than in the more challenging environment of point-of-care settings. The lack of extensive real-world studies necessitates this current investigation into the accuracy and reliability of AI-assisted FoCUS for LVEF assessment across different user experience levels.

This prospective, multicenter, observational cohort study enrolled adult patients undergoing TTE. Participants underwent FoCUS within 48 hours of TTE. The study involved two groups: novice scanners (NS) with less than 100 FoCUS assessments and experienced scanners (ES) with at least 100 assessments and 10 years of experience. The EchoNous KOSMOS device was used for AI-assisted LVEF assessment. The AI algorithm, using the modified Simpson's biplane method of discs, calculated LVEF from apical four-chamber and two-chamber views. The AI was used for image interpretation, not acquisition. A comprehensive TTE, performed by trained sonographers and interpreted by an experienced echocardiographer using the biplane method, served as the gold standard. Data analysis included simple linear regression, intraclass correlation coefficient (ICC), Bland-Altman plots, Cohen's weighted kappa, and receiver operating characteristic (ROC) analysis to assess agreement and diagnostic performance. Statistical significance was set at p<0.05. The study adhered to STARD reporting guidelines. The study population was diverse, with participants recruited from both inpatient and outpatient settings. The diverse clinical settings and inclusion criteria aimed to reflect real-world conditions. Specific inclusion/exclusion criteria were age 18 years and up and the undergoing of a TTE. Exclusions were the studies that resulted in non-diagnostic image quality. This resulted in a final study cohort of 424 participants, a substantial sample size to draw confident conclusions. The selection of the EchoNous KOSMOS device is justified by the fact that the device has received 510(k) U.S. Food and Drug Administration (FDA) clearance. The methodology section also discusses the statistical analysis methods which were simple linear regression and intraclass correlation (ICC) for agreement analysis, Bland-Altman plots to visualize agreement, Cohen's weighted Kappa for categorical agreement, and receiver operating characteristic (ROC) analysis for diagnostic performance. The choice of statistical methods is appropriate for the type of data collected and the research question.

A total of 449 participants were initially enrolled, with 424 studies included in the final analysis after excluding non-diagnostic images. The overall intraclass correlation coefficient (ICC) for LVEF assessment was excellent at 0.904, with ICCs of 0.921 for novice scanners and 0.845 for experienced scanners. A Bland-Altman analysis revealed a small bias (0.73%) towards TTE. Categorical agreement between AI-assisted FoCUS and TTE for LVEF severity was excellent (weighted kappa = 0.83). ROC analysis showed that AI-assisted FoCUS had an area under the curve (AUC) of 0.98 for detecting abnormal LVEF (<50%), with a sensitivity of 92.3%, specificity of 92.3%, negative predictive value (NPV) of 0.97, and positive predictive value (PPV) of 0.83. For severe LV dysfunction (<30%), the AUC was 0.99, with a sensitivity of 78.1%, specificity of 98.0%, NPV of 0.98, and PPV of 0.76. The results showed high diagnostic accuracy regardless of the user's experience level. The high ICC values suggest excellent reliability of the AI-assisted FoCUS LVEF assessment. The high AUC values and sensitivity/specificity values from the ROC analysis demonstrate a strong ability to detect both abnormal and severely reduced LVEF values. The high levels of agreement between the AI-assisted FoCUS and the gold standard TTE support the clinical applicability of the AI-assisted FoCUS, especially given the comparable performance between novice and experienced users.

This study demonstrates that AI-assisted FoCUS provides highly accurate and reliable LVEF assessments comparable to TTE, regardless of the user's experience. This finding is significant because it expands the accessibility of accurate LVEF assessment beyond trained echocardiographers. The high accuracy, even with novice users, suggests the potential for AI-assisted FoCUS to improve diagnostic efficiency and reduce healthcare costs, especially in resource-limited settings. The results support the use of AI-assisted FoCUS as a valuable tool for rapid and accurate LVEF assessment in various clinical settings, potentially improving patient management and outcomes. The observed similar performance between novice and experienced users emphasizes the potential for broader implementation and reduces the barrier to entry for clinicians using this technology. The study's pragmatic design, using real-world patient data and diverse clinical settings, strengthens the generalizability of its findings. This work significantly advances the field by demonstrating the real-world applicability of AI-assisted FoCUS for accurate LVEF assessment.

AI-assisted FoCUS provides a reliable and accurate method for LVEF assessment, comparable to TTE. This technology's accuracy is consistent across novice and experienced users, expanding accessibility in various clinical settings. Future research could explore the integration of AI-assisted image acquisition to further streamline the workflow and investigate the cost-effectiveness of this technology compared to traditional methods.

The study's convenience sample might introduce selection bias, potentially excluding critically ill patients. The findings are specific to the EchoNous KOSMOS device, limiting generalizability to other AI-assisted ultrasound platforms. The AI assistance was limited to image interpretation, not image acquisition, potentially influencing the results. The heterogeneity between novice and experienced user groups, particularly in recruitment settings, could have impacted the observed differences in diagnostic quality studies. While TTE served as the gold standard, inter-observer variability might exist. Future studies could address these limitations by employing randomized controlled trials with larger sample sizes and diverse AI platforms.