Abstract
Background: Aortic stenosis (AS) is frequently identified at an advanced stage after clinical symptoms appear. The aim of this systematic review and meta-analysis is to evaluate the diagnostic accuracy of artificial intelligence (AI) algorithms for AS screening.
Methods: We conducted a thorough search of six databases. Several evaluation parameters, such as sensitivity, specificity, diagnostic odds ratio (DOR), negative likelihood ratio (NLR), positive likelihood ratio (PLR), and area under the curve (AUC) value were employed in the diagnostic meta-analysis of AI-based algorithms for AS screening. The AI algorithms utilized diverse data sources including electrocardiograms (ECG), chest radiographs, auscultation audio files, electronic stethoscope recordings, and cardio-mechanical signals from non-invasive wearable inertial sensors.
Results: Of the 295 articles identified, 10 studies met the inclusion criteria. The pooled estimates for AI-based algorithms in diagnosing AS were as follows: sensitivity 0.83 (95% CI: 0.81-0.85), specificity 0.81 (95% CI: 0.79-0.84), PLR 4.78 (95% CI: 3.12-7.32), NLR 0.20 (95% CI: 0.13-0.28), and DOR 27.11 (95% CI: 14.40-51.05). The AUC value was 0.909 (95% CI: 0.889-0.929), indicating outstanding diagnostic accuracy. Subgroup and meta-regression analyses showed that continent, type of AS, data source, and type of AI-based method constituted sources of heterogeneity. Furthermore, we demonstrated proof of publication bias for DOR values analyzed using Egger’s regression test (P = 0.002) and a funnel plot.
Conclusion: Deep learning approaches represent highly sensitive, feasible, and scalable strategies to identify patients with moderate or severe AS.
Aortic stenosis (AS) is the most prevalent valvular heart disease in the developed world.1 Many individuals with severe AS have prolonged periods of asymptomatic living, which is characteristic of the disease’s history.2 The prevalence of AS increases with age, affecting 10%–15% of individuals over age 70.3 The clinical course of AS includes a decline in exercise capacity, worsening left ventricular outflow tract obstruction, and an elevated risk of exertional angina, syncope, heart failure, and sudden cardiac death.4 Mortality rates rise significantly after the onset of symptoms, with 40%–50% of symptomatic individuals dying within a year without surgery.5,6 Careful follow-up in asymptomatic individuals and prompt aortic valve replacement in symptomatic patients typically lead to positive outcomes.7
After clinical symptoms manifest, AS is frequently diagnosed at an advanced stage of the disease.8 The scalability of AS diagnosis is limited due to its reliance on echocardiographic Doppler imaging, which requires specialist imaging equipment and skilled personnel. Echocardiography, chest radiography, and electrocardiogram (ECG) are some of the diagnostic techniques used for AS, although chest radiography and ECG are not sensitive or specific enough.9 Echocardiography, the gold standard, uses ultrasound waves to create images of the heart, allowing for the assessment of valve structure and function. Despite its accuracy, echocardiography is costly, time-consuming, and requires expertise, thus it is often reserved for symptomatic patients. Chest radiography and ECG are supplementary tools; however, they lack the sensitivity and specificity needed for definitive diagnosis.10
For nearly 200 years, auscultation with a stethoscope has been used to detect AS. However, diagnostic accuracy is limited by factors such as obesity, emphysema, background noise, and reliance on clinical judgment and the human ear’s acoustic range.11 Due to the increased reliance on echocardiography in recent years, primary care physicians have found it less useful to auscultate patients. This, together with a lack of public awareness of AS and its symptoms, has resulted in a marked underdiagnosis and undertreatment of AS.12
Recent advancements in artificial intelligence (AI) offer promising improvements in the diagnosis of AS. AI training programs, including machine learning algorithms, neural networks, and deep learning models, have been developed to analyze echocardiographic and other imaging data. These AI systems are trained on large datasets to recognize patterns indicative of AS, enabling automated, rapid, and accurate diagnostics.13,14 Detection of left ventricular hypertrophy and dysfunction has been demonstrated by the successful application of AI algorithms in echocardiograms.15 Studies have demonstrated the use of Convolutional Neural Networks (CNNs) on noisy single-lead ECGs and single-view 2D echocardiography, respectively, achieving high sensitivity and specificity.16,17 Other methods include real-time heart sound recordings, computed tomography (CT) radiomics features, and chest radiographs using various deep learning architectures like InceptionV3 and DenseNet121, highlighting the diverse approaches in AI applications.18-21
The aim of this systematic review and meta-analysis was to evaluate the accuracy of AI-based approaches for AS diagnosis.
Methods
This systematic review and meta-analysis study was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines22 and was registered at PROSPERO (CRD42024514090).
Search Strategy
The following databases were searched from database inception until January 2024: PubMed/Medline, Web of Science, the Cochrane Library, Embase, Google Scholar, and CINAHL. The search strategy was based on the following key search terms: “aortic stenosis” AND “deep learning” OR “machine learning” OR “artificial intelligence”. We also manually searched the references mentioned in narrative reviews and pertinent non-systematic papers to find further relevant studies that our search approach could have overlooked. All retrieval processes were performed independently by two authors (Apurva Popat and Sweta Yadav).
Selection Criteria
Relevant articles were screened by title and abstract after removing duplicates. Studies were eligible for inclusion if they addressed the diagnosis of AS using AI. The remaining studies were then examined in full text to confirm eligibility.
Inclusion criteria for articles were: (1) observational studies reporting the diagnostic performance of AI for diagnosis of AS; (2) patients with moderate or severe AS; (3) publications reporting sensitivity and specificity data; (4) studies with sufficient epidemiological data; and (5) studies published as original articles. Exclusion criteria were: (1) no full text electronically available; (2) publication in a language other than English; (3) comments, letters, editorials, protocols, guidelines, and review papers; and (4) studies with insufficient outcome data.
Data Extraction
Two independent authors (Babita Saini and Niran Seby) retrieved information from the eligible articles following the inclusion and exclusion criteria, and information was collected on a standardized data sheet that included: (1) study ID (name of first author, year of publication), (2) country, (3) study period, (4) study population, (5) type of AS, (6) age, (7) gender, (8) data source, (9) deep learning method, (10) sensitivity, (11) specificity, (12) accuracy, and (13) AUC.
Quality Assessment of the Studies
The methodologic quality of the included studies was evaluated independently by two authors using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool, which includes four criteria: “patient selection”, “index test”, “reference standard”, and “flow and timing” to judge bias and applicability.23 Each is assessed in terms of risk of bias, and the first three domains were assessed with respect to applicability. Each item is answered with “yes,” “no,” or “unclear.” The answer of “yes” means low risk of bias, whereas “no” or “unclear” means the opposite. Any disagreements were resolved by inviting a third reviewer (Mitkumar Patel) to participate in the discussion.
Statistical Analysis
This diagnostic meta-analysis was conducted on the analytical software Meta-disc 1.4 and the statistical software Comprehensive Meta-Analysis version 3 (Biostat Inc. USA) to analyze the pooled sensitivity, specificity, Positive Likelihood Ratio (PLR), Negative Likelihood Ratio (NLR), Diagnostic Odds Ratio (DOR), and the area under the curve (AUC) values with 95% confidence intervals (CIs) across studies. Data were considered statistically significant when two-sided P < 0.05. The summary receiver operating characteristic (SROC) curve was also used based on the sensitivity and specificity of each study to assess the diagnostic performance. Because of the differences in the basic features of the included articles, their diverging results may have been caused by heterogeneity or random error. Therefore, the Cochrane chi-squared test was used to evaluate heterogeneity among articles, with P < 0.05 indicating the existence of heterogeneity. To estimate the impact of heterogeneity on the meta-analysis, the I2 value was also calculated. If P <0.05 and I2 >50%, heterogeneity was defined as significant. To explore heterogeneity, the threshold effect was assessed using the Spearman correlation coefficient. A strong positive correlation would suggest a threshold effect. Subgroup and meta-regression analyses were also performed to identify potential sources of heterogeneity according to the characteristics of the included studies and using several common evaluation indicators, including sensitivity, specificity, PLR, NLR, DOR, and AUC values. Finally, Egger’s test was conducted to evaluate publication bias. This was further assessed by the visual inspection of the symmetry in funnel plots.
Results
Identification of Studies
The database search identified 295 studies to be screened, of which 109 abstracts were identified as potentially eligible and retrieved for full-text review. Eligibility criteria were met by 10 articles, which were included in this systematic review and meta-analysis study. The PRISMA flowchart is shown in Figure 1.
PRISMA flow diagram of the literature study process and selection
Characteristics of Included Studies
The ten included articles were published between 2020 and 2023 and distributed among six countries (USA=4, Republic of Korea=2, Israel=1, Australia=1, Japan=1, Germany=1). The sample size of the included articles varied from 100 to 329,975 participants, with a total of 941,702. Six studies included patients with moderate to severe AS, while four studies reported patients with severe AS. The mean age of participants ranged from to 61.5 ± 17.6 to 77.77 ± 7.52 years, and the majority of them were men. Half of the studies used ECG as the data source and Convolutional Neural Network (CNN) as the deep learning method. The outcomes retrieved from the included articles were sensitivity, specificity, accuracy, and AUC. The characteristics of the included studies are summarized in Table 1.
Characteristics of included studies
Quality Assessment
The quality of the ten studies was methodologically assessed using the QUADAS-2 tool. Overall, the qualities of the included studies were satisfying and eligible. With respect to the domain of patient selection, the majority of studies were identified to have a low risk of bias (6/10). Additionally, only three studies revealed a high-risk bias concentrated in the field of the index test due to presetting the threshold. The domain of the reference standard was slightly affected by the risk of bias, with only one study showing a high risk of bias. Similarly, there were not too many concerns regarding the applicability of the majority of studies included in this meta-analysis. Specifically, high applicability concerns were shown in one study in patient selection, three studies in the index test, and one study in the reference standard, respectively. Figure 2 shows the details of the quality assessment form.
Risk of bias and applicability concerns graph: review authors’ judgments about each domain presented as percentages across included studies.
Data Analysis
From forest plots of pooled data from ten studies, we found significant heterogeneity in sensitivity (Chi2=57.66, P=0.0000, I2= 84.4%, Figure 3), specificity (Chi2=126.67, P=0.0000, I2= 92.9%, Figure 4), PLR (Chi2=85.82, P=0.0000, I2= 89.5%, Figure 5), NLR (Chi2=61.68, P=0.0000, I2= 85.4%, Figure 6), and DOR (Chi2=49.70, P=0.0000, I2= 81.90%, Figure 7) outcomes. Consequently, the random-effect model was used to calculate the pooled estimates of these evaluation indicators.
Forest plot for the sensitivity of AI-based methods
Forest plot for the specificity of AI-based methods
Forest plot for the PLR of AI-based methods
Forest plot for the NLR of AI-based methods
Forest plot for the DOR of AI-based methods
The pooled estimates of overall AI-based methods for the diagnosis of AS were 0.83 (95%CI: 0.81-0.85) for sensitivity (Figure 3), 0.81 (95%CI: 0.79-0.84) for specificity (Figure 4), 4.78 (95%CI: 3.12-7.32) for PLR (Figure 5), 0.20 (95%CI: 0.13-0.28) for NLR (Figure 6), and 27.11 (95%CI: 14.40-51.06) for DOR (Figure 7). Moreover, we plotted the SROC curve to evaluate the diagnostic accuracy of AI-based methods (Figure 8). AUC was 0.909 (95% CI: 0.889-0.929), suggesting an outstanding diagnostic accuracy of AI-based methods.
SROC curve for the diagnostic accuracy of AI-based methods
Analysis of Diagnostic Threshold
The threshold effect is the primary source of heterogeneity in a diagnostic test. The Moses model was weighted by inverse variance, and the Spearman correlation coefficient was used to assess the threshold impact. The findings revealed that the Spearman correlation coefficient was −0.036 (P=0.920). Thus, the threshold effect was not found to cause variations in the accuracy estimates among individual studies.
Subgroup and Meta-Regression Analyses
The forest plots demonstrated all ten studies were heterogeneous. As heterogeneity cannot be completely avoided in a meta-analysis, its source and level were further investigated. It was suggested many factors, other than the threshold effect, may have contributed to the variation in accuracy estimates. Hence, subgroup and meta-regression analyses based on continent, type of AS, data source, and type of AI-based method were also carried out (Table 2).
Subgroup and meta-regression analysis of diagnostic accuracy of AI-based methods for diagnosis of AS.
The sensitivity, specificity, DOR, and AUC values were significantly influenced by the continent, type of AS, data source, and type of AI-based method, which indicates they were a source of heterogeneity (P < 0.05). Regarding the continent, AI-based method showed the highest overall diagnostic accuracy in South America with a sensitivity of 0.84, specificity of 0.75, DOR of 18.53 and AUC of 0.892.
Regarding the type of AS, AI-based method showed the highest overall diagnostic accuracy in detecting severe AS with a sensitivity of 0.79, specificity of 0.89, DOR of 46.50, and AUC of 0.936. Regarding data source, AI-based method showed the highest overall diagnostic accuracy when ECG data sources were used with a sensitivity, specificity, DOR, and AUC of 0.84, 0.87, 48.91, and 0.924, respectively. Regarding the type of AI-based method, the highest overall diagnostic accuracy was detected in studies that used a method other than CNN with a sensitivity, specificity, DOR, and AUC of 0.82, 0.88, 59.38, and 0.932, respectively.
Publication Bias
To assess the potential publication bias of the included studies, the funnel plot asymmetry test was conducted. We revealed proof of publication bias for DOR values analyzed using Egger’s regression test (P=0.002). Moreover, a visual inspection of the funnel plot showed an asymmetrical funnel (Figure 9).
Funnel plot showing publication bias in terms of DOR values among the included studies.
Discussion
The concept of artificial intelligence (AI) was first articulated by John McCarthy in 1956 during the Dartmouth Conference, where the term “artificial intelligence” was coined.29 The 1980s and 1990s saw the growth of machine learning, enhancing diagnostic tools through pattern recognition in medical data.30 The 2000s introduced natural language processing (NLP) and integration with electronic health records (EHRs).31 Recently, deep learning and neural networks have revolutionized medical imaging, significantly improving the accuracy of disease detection, including cancers and cardiovascular conditions. AI continues to advance, driving innovations in personalized medicine, predictive analytics, and clinical decision support systems.32,33
With recent developments in transcatheter aortic valve replacement, the morbidity and mortality associated with AS can be treated, making AS care safer and more practical.34 Only 10% to 20% of patients with moderate-to-severe AS are detected, meaning that the diagnosis of AS is commonly overlooked in regular care.35 The broad use of diagnostic tools does not constitute a workable screening strategy for AS, despite the emphasis on the significance of AS screening.36 Therefore, the use of deep learning algorithms has been suggested as a modality for AS detection.17 However, there are limited data about the effectiveness of AI-based approaches on the diagnosis of AS. To the best of our knowledge, this is the first systematic review and meta-analysis that assesses the diagnostic performance of AI-based approaches for AS screening. According to our results, AI-based algorithms for the diagnosis of AS achieved a high sensitivity and specificity of greater than 80%. In comparison to currently used medical tests, such as pap smears, which can detect cervical uterine cancer (AUC 0.71)37 and AI-interpreted mammography, which can detect breast cancer (AUC 0.76–0.89),38 our study showed the AI algorithms can successfully identify patients with moderate to severe AS with high performance (AUC 0.909). AI applications in echocardiography have been reported in the last few years.39 These include automated video-based beat-to-beat evaluation of left ventricular systolic dysfunction,40 detection of diastolic dysfunction, left ventricular hypertrophy and its subtypes,15 automated classification of echocardiographic views,41 and expert-level prenatal detection of complex congenital heart disease.42 While echocardiography is still necessary for the diagnosis and grading of AS,28 the majority of AI solutions for timely AS screening have concentrated on alternative data types, including cardiac auscultation audio files,25 12-lead ECGs,24 cardio-mechanical signals using non-invasive wearable inertial sensors,43 and chest radiographs.20 Interestingly, subgroup analysis showed that AI-based method has the highest overall diagnostic accuracy when ECG data sources were used with a sensitivity, specificity, DOR, and AUC of 0.84, 0.87, 48.91, and 0.924, respectively; whereas for alternative data types, analyses were limited to small datasets without external validation.20,25 Consequently, since these devices are widely accessible and simple to install, AI-ECG constitutes an approachable method for detecting moderate and severe AS in the population. But a cautious reading is required, particularly if the model suggests the existence of AS. It was shown that the positive predictive value was minimal, with a prevalence of moderate to severe AS of roughly 4%, and there might be a risk of performing unnecessary echocardiography examinations on individuals who tested positive on the AI-ECG.24 Therefore, the AI-ECG result must be incorporated into the clinical evaluation together with the comorbidities related to AS, the patient’s symptomatic state, and a more cautious auscultation prior to conducting additional testing.
In the present review, we performed the literature search in six different databases. The principal strength of this article is the considerably high number of participants analyzed, which provides a realistic view of the performance of the AI methods. Furthermore, our meta-analysis demonstrated an acceptable methodological quality of the included studies. The main findings were validated by conducting a subgroup analysis to investigate any potential source of heterogeneity that might have affected the final results.
However, the included papers presented variable factors that may have a considerable impact on the clinical performance of AI approaches, such as data source (ECG, chest radiographs, electronic stethoscope, and auscultation audio files), type of AI algorithm (CNN, InceptionV3, ResNet50, DenseNet121, and RF and XGBoost), and type of AS (moderate and/or severe). These differences were one of the main causes of the inconsistencies in our meta-analysis, which made it difficult to compare the results of different studies and complicated the pooled analysis. As anticipated in meta-analysis research,44 the included studies showed a high degree of heterogeneity. We demonstrated this heterogeneity between the studies was not caused by the threshold effect. Nonetheless, our meta-regression analysis demonstrated a number of clinical parameters, including continent, AS type, data source, and AI-based approach type, were responsible for heterogeneity. For instance, variations in healthcare infrastructure, population genetics, and diagnostic practices across continents likely contributed to these differences. The diagnostic accuracy was notably higher for severe AS cases, indicating disease severity impacts AI performance. Additionally, the type of data source, such as ECG versus other imaging techniques, and the specific AI algorithms used, such as CNNs versus other methods, also influenced the results. Consequently, the results of the current investigation need to be carefully considered.
The future potential of AI in screening for aortic stenosis is promising, with advancements likely to enable earlier and more accurate detection of the disease. Researchers should focus on standardizing AI training data and methodologies to minimize heterogeneity in diagnostic accuracy across different regions and populations. Further studies are needed to compare the performance of various AI algorithms in real-world clinical settings, particularly for different types of AS and data sources. Additionally, exploring the integration of AI tools into routine clinical workflows and evaluating their cost-effectiveness and impact on patient outcomes will be crucial. Collaboration between AI developers and healthcare providers can enhance the practical utility and acceptance of AI-based diagnostic tools.
Conclusion
The present meta-analysis revealed that AI algorithms serve as powerful screening tools for the detection of patients with moderate to severe AS. Leveraging diverse data sources such as ECG, chest radiographs, and auscultation audio files, AI offers a comprehensive and practical approach to AS diagnosis. These methods represent highly sensitive, feasible, and scalable strategies for community-based AS screening. However, using them directly in clinical practice is still challenging. Hence, more AI algorithms should be developed, improved, and validated for more efficacious and rapid AS diagnosis.
- Received May 19, 2024.
- Revision received August 1, 2024.
- Accepted August 1, 2024.
References
- 1.↵Aluru JS, Barsouk A, Saginala K, Rawla P, Barsouk A. Valvular Heart Disease Epidemiology. Med Sci (Basel). 2022;10(2):32. doi:10.3390/medsci10020032.
- 2.↵Lancellotti P, Magne J, Dulgheru R, Outcomes of Patients With Asymptomatic Aortic Stenosis Followed Up in Heart Valve Clinics. JAMA Cardiol. 2018;3(11):1060-1068. doi:10.1001/jamacardio.2018.3152.
- 3.↵Danielsen R, Aspelund T, Harris TB, Gudnason V. The prevalence of aortic stenosis in the elderly in Iceland and predictions for the coming decades: The AGES–Reykjavík study. Int J Cardiol. 2014;176(3):916-922. doi:10.1016/j.ijcard.2014.08.053.
- 4.↵Otto CM, Prendergast B. Aortic-valve stenosis—from patients at risk to severe valve obstruction. N Engl J Med. 2014;371(8):744-756. doi:10.1056/NEJMra1313875.
- 5.↵Ben-Dor I, Pichard AD, Gonzalez MA, Correlates and causes of death in patients with severe symptomatic aortic stenosis who are not eligible to participate in a clinical trial of transcatheter aortic valve implantation. Circulation. 2010;122(11_suppl_1)(Suppl):S37-S42. doi:10.1161/CIRCULATIONAHA.109.926873.
- 6.↵Leon MB, Smith CR, Mack M, PARTNER Trial Investigators. Transcatheter aortic-valve implantation for aortic stenosis in patients who cannot undergo surgery. N Engl J Med. 2010;363(17):1597-1607. doi:10.1056/NEJMoa1008232.Medline
- 7.↵Maes F, Lerakis S, Barbosa Ribeiro H, Outcomes From Transcatheter Aortic Valve Replacement in Patients With Low-Flow, Low-Gradient Aortic Stenosis and Left Ventricular Ejection Fraction Less Than 30%: A Substudy From the TOPAS-TAVI Registry. JAMA Cardiol. 2019;4(1):64-70. doi:10.1001/jamacardio.2018.4320.
- 8.↵Wheeler MT, Heidenreich PA, Froelicher VF, Hlatky MA, Ashley EA. Cost-effectiveness of preparticipation screening for prevention of sudden cardiac death in young athletes. Ann Intern Med. 2010;152(5):276-286. doi:10.7326/0003-4819-152-5-201003020-00005.
- 9.↵Maganti K, Rigolin VH, Sarano ME, Bonow RO. Valvular heart disease: diagnosis and management. Mayo Clin Proc. 2010;85(5):483-500. doi:10.4065/mcp.2009.0706.Medline
- 10.↵Saikrishnan N, Kumar G, Sawaya FJ, Lerakis S, Yoganathan AP. Accurate assessment of aortic stenosis: a review of diagnostic modalities and hemodynamics. Circulation. 2014;129(2):244-253. doi:10.1161/CIRCULATIONAHA.113.002310.
- 11.↵Montinari MR, Minelli S. The first 200 years of cardiac auscultation and future perspectives. J Multidiscip Healthc. 2019;12:183-189. doi:10.2147/JMDH.S193904.
- 12.↵Thoenes M, Bramlage P, Zamorano P, Patient screening for early detection of aortic stenosis (AS)— review of current practice and future perspectives. J Thorac Dis. 2018;10(9):5584-5594. doi:10.21037/jtd.2018.09.02.
- 13.↵Elvas LB, Águas P, Ferreira JC, Oliveira JP, Dias MS, Rosário LB. AI-Based Aortic Stenosis Classification in MRI Scans. Electronics (Basel). 2023;12(23):4835. doi:10.3390/electronics12234835.
- 14.↵Harmon DM, Malik A, Nishimura R. Progression of Calcific Aortic Stenosis Detected by Artificial Intelligence Electrocardiogram. Mayo Clin Proc. 2022;97(6):1211-1212. doi:10.1016/j.mayocp.2022.04.004.
- 15.↵Duffy G, Cheng PP, Yuan N, High-Throughput Precision Phenotyping of Left Ventricular Hypertrophy With Cardiovascular Deep Learning. JAMA Cardiol. 2022;7(4):386-395. doi:10.1001/jamacardio.2021.6059.
- 16.↵Aminorroaya A, Dhingra LS, Sangha V, Deep Learning-enabled Detection of Aortic Stenosis from Noisy Single Lead Electrocardiograms. medRxiv. 2023:2023.2009.2029.23296310. doi:10.1101/2023.09.29.23296310
- 17.↵Holste G, Oikonomou EK, Mortazavi BJ, Severe aortic stenosis detection by deep learning applied to echocardiography. Eur Heart J. 2023;44(43):4592-4604. doi:10.1093/eurheartj/ehad456.
- 18.↵Ghanayim T, Lupu L, Naveh S, Artificial Intelligence-Based Stethoscope for the Diagnosis of Aortic Stenosis. Am J Med. 2022;135(9):1124-1133. doi:10.1016/j.amjmed.2022.04.032.
- 19.↵Kang N, Suh YJ, Han K, Kim YJ, Choi BW. Performance of Prediction Models for Diagnosing Severe Aortic Stenosis Based on Aortic Valve Calcium on Cardiac Computed Tomography: Incorporation of Radiomics and Machine Learning. Korean J Radiol. 2021;22(3):334-343. doi:10.3348/kjr.2020.0099.
- 20.↵Ueda D, Yamamoto A, Ehara S, Artificial intelligence-based detection of aortic stenosis from chest radiographs. Eur Heart J Digit Health. 2021;3(1):20-28. doi:10.1093/ehjdh/ztab102
- 21.↵Popat A, Yadav S, Patel SK, Artificial Intelligence in the Early Prediction of Cardiogenic Shock in Acute Heart Failure or Myocardial Infarction Patients: A Systematic Review and Meta-Analysis. Cureus. 2023;15(12):e50395. doi:10.7759/cureus.50395.
- 22.↵Liberati A, Altman DG, Tetzlaff J, The PRISMA statement for reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaboration. J Clin Epidemiol. 2009;62(10):e1-e34. doi:10.1016/j.jclinepi.2009.06.006.
- 23.↵Whiting PF, Rutjes AW, Westwood ME, QUADAS-2 Group. QUADAS-2: a revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155(8):529-536. doi:10.7326/0003-4819-155-8-201110180-00009.
- 24.↵Cohen-Shelly M, Attia ZI, Friedman PA, Electrocardiogram screening for aortic valve stenosis using artificial intelligence. Eur Heart J. 2021;42(30):2885-2896. doi:10.1093/eurheartj/ehab153.
- 25.↵Kwon JM, Lee SY, Jeon KH, Deep Learning–Based Algorithm for Detecting Aortic Stenosis Using Electrocardiography. J Am Heart Assoc. 2020;9(7):e014717. doi:10.1161/JAHA.119.014717.
- 26.↵Playford D, Bordin E, Mohamad R, Stewart S, Strange G. Enhanced Diagnosis of Severe Aortic Stenosis Using Artificial Intelligence: A Proof-of-Concept Study of 530,871 Echocardiograms. JACC Cardiovasc Imaging. 2020;13(4):1087-1090. doi:10.1016/j.jcmg.2019.10.013.
- 27.↵Tseng AS, Shelly-Cohen M, Attia IZ, Spectrum bias in algorithms derived by artificial intelligence: a case study in detecting aortic stenosis using electrocardiograms. Eur Heart J Digit Health. 2021;2(4):561-567. doi:10.1093/ehjdh/ztab061.
- 28.↵Voigt I, Boeckmann M, Bruder O, Wolf A, Schmitz T, Wieneke H. A deep neural network using audio files for detection of aortic stenosis. Clin Cardiol. 2022;45(6):657-663. doi:10.1002/clc.23826.
- 29.↵McCarthy J, Minsky ML, Rochester N, Shannon CE. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, August 31, 1955. AI Mag. 2006;27(4):12. doi:10.1609/aimag.v27i4.1904. https://ojs.aaai.org/aimagazine/index.php/aimagazine/article/view/1904
- 30.↵Kononenko I. Machine learning for medical diagnosis: history, state of the art and perspective. Artif Intell Med. 2001;23(1):89-109. doi:10.1016/S0933-3657(01)00077-X.
- 31.↵Olivera P, Danese S, Jay N, Natoli G, Peyrin-Biroulet L. Big data in IBD: a look into the future. Nat Rev Gastroenterol Hepatol. 2019;16(5):312-321. doi:10.1038/s41575-019-0102-5.
- 32.↵Reddy S, Fox J, Purohit MP. Artificial intelligence-enabled healthcare delivery. J R Soc Med. 2019;112(1):22-28. doi:10.1177/0141076818815510.
- 33.↵Litjens G, Kooi T, Bejnordi BE, A survey on deep learning in medical image analysis. Med Image Anal. 2017;42:60-88. doi:10.1016/j.media.2017.07.005.
- 34.↵Otto CM, Nishimura RA, Bonow RO, 2020 ACC/AHA Guideline for the Management of Patients With Valvular Heart Disease: Executive Summary: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines [published correction appears in Circulation. 2021 Feb 2;143(5):e228. doi: 10.1161/CIR.0000000000000960] [published correction appears in Circulation. 2021 Mar 9;143(10):e784. doi: 10.1161/CIR.0000000000000966]. Circulation. 2021;143(5):e35-e71. doi:10.1161/CIR.0000000000000932
- 35.↵Perego S, Zambon A, Nistri S, Prevalence, clinical correlates, and burden of undiagnosed aortic stenosis in older patients: a prospective study in a non-cardiologic acute hospital ward. Aging Clin Exp Res. 2020;32(8):1533-1540. doi:10.1007/s40520-020-01471-w.
- 36.↵Sangha V, Mortazavi BJ, Haimovich AD, Automated multilabel diagnosis on electrocardiographic images and signals. Nat Commun. 2022;13(1):1583. doi:10.1038/s41467-022-29153-3.
- 37.↵Hu L, Bell D, Antani S, An Observational Study of Deep Learning and Automated Evaluation of Cervical Images for Cancer Screening. J Natl Cancer Inst. 2019;111(9):923-932. doi:10.1093/jnci/djy225.
- 38.↵McKinney SM, Sieniek M, Godbole V, International evaluation of an AI system for breast cancer screening [published correction appears in Nature. 2020 Oct;586(7829):E19. doi: 10.1038/s41586-020-2679-9]. Nature. 2020;577(7788):89-94. doi:10.1038/s41586-019-1799-6
- 39.↵Madani A, Arnaout R, Mofrad M, Arnaout R. Fast and accurate view classification of echocardiograms using deep learning. NPJ Digit Med. 2018;1(1):6. doi:10.1038/s41746-017-0013-1.
- 40.↵Ouyang D, He B, Ghorbani A, Video-based AI for beat-to-beat assessment of cardiac function. Nature. 2020;580(7802):252-256. doi:10.1038/s41586-020-2145-8.
- 41.↵Chiou YA, Hung CL, Lin SF. AI-Assisted Echocardiographic Prescreening of Heart Failure With Preserved Ejection Fraction on the Basis of Intrabeat Dynamics. JACC Cardiovasc Imaging. 2021;14(11):2091-2104. doi:10.1016/j.jcmg.2021.05.005.
- 42.↵Narang A, Bae R, Hong H, Utility of a Deep-Learning Algorithm to Guide Novices to Acquire Echocardiograms for Limited Diagnostic Use. JAMA Cardiol. 2021;6(6):624-632. doi:10.1001/jamacardio.2021.0185.
- 43.↵Sengupta PP, Shrestha S, Kagiyama N, Artificial Intelligence for Aortic Stenosis at Risk International Consortium. A Machine-Learning Framework to Identify Distinct Phenotypes of Aortic Stenosis Severity. JACC Cardiovasc Imaging. 2021;14(9):1707-1720. doi:10.1016/j.jcmg.2021.03.020.
- 44.↵Imrey PB. Limitations of Meta-analyses of Studies With High Heterogeneity. JAMA Netw Open. 2020;3(1):e1919325-e1919325. doi:10.1001/jamanetworkopen.2019.19325.
