Automated Analysis of Heart Sound Signals in Screening for Structural Heart Disease in Children

Automated Analysis of Heart Sound Signals in Screening for Structural Heart Disease in Children | 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 Research Article Automated Analysis of Heart Sound Signals in Screening for Structural Heart Disease in Children Ida Papunen, Kaisa Ylänen, Oliver Lundqvist, Martin Porkholm, and 8 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4697876/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 21 Sep, 2024 Read the published version in European Journal of Pediatrics → Version 1 posted 9 You are reading this latest preprint version Abstract Purpose Our aim was to investigate the ability of an artificial intelligence (AI)-based algorithm to differentiate innocent murmurs from pathologic ones. Methods An AI-based algorithm was developed using heart sound recordings collected from 1413 patients at the five university hospitals in Finland. The corresponding heart condition was verified using echocardiography. In the second phase of the study, patients referred to Helsinki New Children’s Hospital due to a heart murmur were prospectively assessed with the algorithm, and then the results were compared with echocardiography findings. Results Ninety-eight children were included in this prospective study. The algorithm classified 72 (73%) of the heart sounds as normal and 26 (27%) as abnormal. Echocardiography was normal in 63 (64%) children and abnormal in 35 (36%). The algorithm recognised abnormal heart sounds in 24 of 35 children with abnormal echocardiography and normal heart sounds with normal echocardiography in 61 of 63 children. When the murmur was audible, the sensitivity and specificity of the algorithm were 83% (24/29) (confidence interval (CI) 64–94%) and 97% (59/61) (CI 89–100%), respectively. Conclusions The algorithm was able to distinguish murmurs associated with structural cardiac anomalies from innocent murmurs with good sensitivity and specificity. The algorithm was unable to identify heart defects that did not cause a murmur. Further research is needed on the use of the algorithm in screening for heart murmurs in primary health care. artificial intelligence congenital heart defect electronic stethoscope heart murmur screening Figures Figure 1 Figure 2 Figure 3 What is known Innocent murmurs are common in children, while the incidence of moderate or severe congenital heart defects is low. Auscultation plays a significant role in assessing the need for further examinations of the murmur. The ability to differentiate innocent murmurs from those related to congenital heart defects requires clinical experience on the part of general practitioners. No AI-based auscultation algorithms have been systematically implemented in primary health care. What is new: We developed an AI-based algorithm using a large dataset of sound samples validated by echocardiography The algorithm performed well in recognising pathological and innocent murmurs in children from different age groups. Introduction A heart murmur can be a sign of a congenital heart defect (CHD), but innocent murmurs are common in children and can be heard in 50% of school-aged children 1,2 . CHDs are found in about 75 per 1000 live-born infants. However, the incidence of moderate or severe CHD is only six cases per 1000 live births 3 . About 10% of children with significant CHDs are discharged from the maternity hospital without receiving a diagnosis 4 , and CHDs are rarely found in children over six months of age 5 6 . Although a murmur is the most common finding that raises suspicion of CHD, most of the children referred to paediatric cardiologists have an innocent murmur. According to some estimates, up to 90% of initial visits to paediatric cardiology clinics were due to an innocent murmur and thus unnecessary 5,7 . Despite the development of medical technology, auscultation still plays a significant role in assessing the need for further examinations of the murmur. The interpretation of cardiac auscultation is subject to uncertainty because it depends on the subjective perceptions and experience of the clinician. Computer decision algorithms based on artificial intelligence (AI) have been developed in recent years to identify murmurs collected by electronic stethoscopes 8 . The sensitivity and specificity of an AI-assisted murmur detection algorithm were shown to be good in differentiating innocent murmurs from pathologic ones in a validation study involving both adults and children 9 . To the best of our knowledge, only a few studies have been published on paediatric populations, with the algorithms differentiating pathologic murmurs from normal heart sounds with good sensitivity 9,10 . While the use of AI has potential for cost-effective CHD screening, it has not yet been utilised in clinical practice. Our aim was to evaluate the ability of an AI-based algorithm to identify innocent and pathologic murmurs in children. Materials and Methods The development of the algorithm An AI-based algorithm was developed by fine-tuning a state-of-the-art speech model using heart sound recordings collected at the five university hospitals in Finland. The heart sound recordings, selected for the training of the algorithm, were collected from 1413 patients, including 1061 (75%) children (<18 years) and 352 (25%) adults (Table 1). The recordings from CHDs without murmurs were not used in the development of the algorithm. The recordings were made with the Thinklabs One® electronic stethoscope, which is commercially available and is licensed for medical use by the manufacturer (US FDA Class 2 diagnostic device 2004 and CE-marking 2017). The data were collected and saved on a Samsung Galaxy Tab A10 tablet (CE) (Figure 1). Recordings were made from four different locations (aortic, pulmonary, tricuspid, and mitral valve areas), with a sample rate of 44 kHz and a duration of 10 seconds. If patient co-operation was not adequate to complete all four recordings with good signal quality, only the samples with adequate signal were included in training the algorithm. Table 1 Basic information of the patients used to train the algorithm (N=1413). Characteristic N= 1413 (%) Sex Female 671 (47 %) Age 0–1 month 1–6 months 6–12 months 1–4 years 4–12 years 12–18 years > 18 years 83 (6 %) 155 (11 %) 66 (5 %) 143 (10 %) 375 (27 %) 239 (17 %) 352 (25 %) Heart murmur No murmur and no CHD Innocent murmur Pathological murmur 514 (36 %) 339 (24 %) 560 (40 %) Result of echocardiography Normal Abnormal 853 (60 %) 560 (40 %) The trained algorithm was divided into two leading AI algorithms. A quality algorithm was used to pre-screen the raw phonocardiogram signal and remove noise and other artifacts, and a binary classifier input the high-quality phonocardiogram segments and used this information to predict the presence or absence of a heart defect causing an audible heart murmur (Figure 2). The algorithm formed a floating-point value between 0 and 1. A threshold of 0.5 was set as the divider for the two categories of normal and abnormal heart sounds. If the output value (probability of defect) was less than 0.5, the result was classified as normal, and an output value above 0.5 was classified as abnormal. The normal category included heart sounds with no murmur and innocent murmurs, while the abnormal heart sound included pathological murmurs. Both algorithms were trained using the same training dataset and utilised the same separate development set for hyperparameter tuning and establishing performance metrics on data not used during training to understand how well each model generalised on unseen data. The development set was constructed using a 90–10 split (90% of the data was used for training, and 10% was placed in the development set). The development set was not the same dataset as the result set of the 98 patients, which was used in this research. Study patients A total of 98 children were prospectively recruited for the study during their outpatient visits at the Paediatric Cardiology unit at the New Children’s Hospital in Helsinki, Finland, between 29 Mar 2022 and 9 Aug 2023. The patients were examined by a paediatric cardiologist due to a heart murmur or CHD. Children with previous heart surgery or insufficient co-operation were excluded. The patients were first analysed with the algorithm, and then the results were compared with clinical evaluations and echocardiography. The paediatric cardiologists performed auscultation with a conventional stethoscope before the echocardiography and assessed whether a murmur was audible. The cardiologist was not aware of the analysis result of the AI. Finally, echocardiography was performed, and the findings were classified as normal or abnormal. The algorithm’s analysis of the murmur (normal or abnormal) was compared with the diagnosis made by echocardiography. The algorithm result “abnormal” included all pathologic murmurs identified by the algorithm. The result “normal” included innocent murmurs and normal heart sounds without murmurs. All the study patients and/or their legal guardians provided written informed consent. This study complied with the Declaration of Helsinki and was approved by the ethical committee (HUS/1630/2019 23 August 2019) of Helsinki University Hospital (HUS). The statistical analyses were performed using IBM SPSS Statistics for Macintosh (version 29.0.1.0). Categorical values were expressed as frequencies and percentages. The median and interquartile range (IQR) (Q1–Q3) were expressed for the non-normally distributed variables. The sensitivity, specificity, and accuracy of the algorithm to differentiate an innocent murmur from a pathologic one were calculated. Categorical variables were compared with chi-square tests. Sensitivity, specificity, and accuracy were calculated with 95% confidence intervals. Results Study population Heart sounds of 98 paediatric patients were analysed in this study. The patient characteristics and echocardiography findings are presented in Table 2. Table 2 Basic patient data of the patients examined with the algorithm. Characteristic N=98 (%) Sex Female 50 (51 %) Age 0–1 month 1–6 months 6–12 months 1–4 years 4–12 years >12 years 3 (3 %) 14 (14 %) 7 (7 %) 20 (20 %) 31 (32 %) 23 (24 %) Heart murmur Yes 90 (92 %) Result of echocardiography Normal Abnormal * Shunt lesions Valvular lesions Other lesions 63 (64 %) 35 (36 %) 22 (22 %) 17 (17 %) 3 (3 %) *One patient can have more than one diagnosis Echocardiography was normal in 63 (64%) children and abnormal in 35 (36%). The algorithm identified 72 (73%) of the heart sounds as normal and 26 (27%) as abnormal. The sensitivity of the algorithm was 69% (CI 51–83%), and the specificity was 97% (CI 89–100%) (Table 3). Table 3 Performance of the AI-based algorithm in detecting pathologic murmurs verified by echocardiography. All patients (N = 98) Echocardiography Abnormal Normal Algorithm Abnormal True positive 24 False positive 2 Positive predictive value 0.92 Normal False negative 11 True negative 61 Negative predictive value 0.85 Sensitivity 0.69 Specificity 0.97 Accuracy 0.87 The algorithm recognised pathologic murmurs in 24 of 35 cases with abnormal echocardiography and misdiagnosed 11 cases. Of the 11 misdiagnosed cases, five had an audible murmur and six had no murmur. The algorithm did not identify murmurs caused by aortic valve insufficiency (diastolic grade I), patent ductus arteriosus (PDA) (grade 1 systo-diastolic), atrial septal defect (ASD), or a small ventricular septal defect (VSD). In total, 61 of 63 children with normal echocardiography findings were identified as having normal heart sounds by the algorithm. The level of confidence of the algorithm was unsure in six (6%) analyses and confident in 92 (94%). The details of the algorithm analyses and the findings of the patients with abnormal echocardiography are presented in Table 4. In the group of patients with abnormal echocardiography findings, the median (IQR) probability set by the algorithm of a CHD was 0.66 (0.27–0.98) (Table 4). When comparing age groups, the algorithm had a sensitivity of 75% (12/16) (CI 48–93%) and a specificity of 96% (27/28) (CI 82–100%) in the group of children aged 0–4 years (n=44). In children over four years of age (n=54), the sensitivity was 63% (12/19) (CI 38–84%), while the specificity was 97% (34/35) (CI 85–100%). There was one CHD without a murmur in children aged 0–4 years (VSD) and five in children over four years of age (VSD, bicuspid aortic valve, and ASD) (Table 4). Table 4 The diagnoses found in the echocardiography (n = 35) classified according to the interpretation of the algorithm. Algorithm analysis Diagnoses Murmur n Probability of defect (range) Abnormal n = 24 Ventricular septal defect Atrial septal defect Aortic stenosis Pulmonary stenosis Coarctation of aorta Patent ductus arteriosus Aortic and pulmonary stenosis Others Yes Yes Yes Yes Yes Yes Yes Yes 7 4 4 3 2 1 1 2 0.54–0.99 0.58–0.99 0.87–0.97 0.99 0.78–0.87 0.92 0.98 0.94–0.95 Normal n = 11 Bicuspid aortic valve Ventricular septal defect Atrial septal defect Atrial septal defect Ventricular septal defect Aortic valve insufficiency Patent ductus arteriosus No No No Yes Yes Yes Yes 2 2 2 2 1 1 1 0.03 0.09–0.25 0.18–0.33 0.11–0.49 0.41 0.04 0.07 Patients with a murmur Paediatric cardiologists heard a murmur from 90 of 98 patients (92%). In these 90 patients, the sensitivity of the algorithm was 83% (24/29) (CI 64–94%) and the specificity was 97% (59/61) (CI 89–100%) (Table 5). The algorithm was confident in 83 (93%) and unsure in six (7%) analyses. Table 5 Performance of the AI based algorithm in detecting pathologic murmurs verified by echocardiography among patients with audible murmur (N = 90). Echocardiography Abnormal Normal Algorithm Abnormal True positive 24 False positive 2 Positive predictive value 0.92 Normal False negative 5 True negative 59 Negative predictive value 0.94 Sensitivity 0.86 Specificity 0.97 Accuracy 0.93 The median (IQR) probability of CHD was 0.15 (0.03–0.27) in children with normal echocardiography findings and an audible murmur and 0.81 (0.79–0.99) in children with abnormal echocardiography findings and an audible murmur (Figure 3). Comparing age groups, the sensitivity of the algorithm was 80% (12/15) (CI 52–96%) and its specificity was 96% (26/27) (CI 81–100%) in the group of children aged 0–4 years (n=42). In children over four years of age (n=48), the algorithm’s sensitivity was 86% (12/14, CI 57–98%) and its specificity was 97% (33/34, CI 85–100%). The vertical black line marks the threshold probability of the defect (0.5). The black dots represent the abnormalities that were not detected by the algorithm. The dots with “X” are the normal cases interpreted incorrectly as abnormal. The white dots represent the cases that the algorithm identified correctly. Discussion This is the first clinical study to investigate the ability of AI to differentiate between benign and pathological heart murmurs in children to this extent using echocardiography as the gold standard. The algorithm was able to distinguish a murmur caused by a CHD from innocent murmurs with good sensitivity and specificity when echocardiography was used as the reference. Since the method is based on the identification of different heart sound signals, it is not suitable for screening CHDs without a murmur. Despite the routine use of pulse oximetry screening, which is highly sensitive in detecting critical cyanotic CHDs, some newborns with acyanotic CHDs may still go undiagnosed before discharge from the maternity ward 11 . Indeed, many CHDs do not cause symptoms during the first week of life 4 , and some are only diagnosed after normal postnatal adaptation has taken place. For example, VSD causes murmur after a reduction in pulmonary vascular resistance leads to a pressure difference between ventricles and coarctation when the arterial duct has closed and caused narrowing of the aorta. CHDs diagnosed in children over six months of age are usually mild and asymptomatic, including ASD, PDA, and valvular defects 6 . Some valvular defects and hypertrophic cardiomyopathy may manifest later in adolescence. Auscultation during childhood plays a vital role in identifying these CHDs, but the accuracy of interpretation depends on the experience and skills of the physician. In previous studies, the sensitivity and specificity of clinical assessments of CHD have varied widely depending on the clinical experience of the physician, while specificity generally increased with experience. For example, medical students and paediatric residents had a sensitivity of 82% but a specificity of only 56% in assessing CHD 12 . Among paediatricians, sensitivity was found to be better (93%), but specificity remained low (59%) 6 . Meanwhile, clinical assessment by paediatric cardiologists had a sensitivity of 81% and specificity of 91% in identifying neonates with CHD 13 . Auscultation using a conventional stethoscope or AI requires optimal conditions, including good patient co-operation and quiet environment. If a child is crying, both the physician and the algorithm may struggle to recognise murmurs accurately. In our study, all children were co-operative, ensuring a reliable evaluation of the algorithm’s performance. The AI used in this study includes a quality algorithm that screens raw phonocardiogram signals and removes noise and other artifacts, allowing high-quality phonocardiogram segments to be used for analysis of heart sounds. A prerequisite for the development and utilisation of AI in health care is the ability to reliably listen to and record heart sounds. In recent decades, electronic stethoscopes have undergone significant development, resulting in enhanced capabilities for analysing murmurs. These devices improve sound signals and reduce background noise, facilitating auscultation. Most available models not only aid in listening but also can record sounds and store data on murmurs for future reference 14 . The accuracy of algorithms primarily hinges on the quantity and quality of the heart sound samples used in their development. The training data must encompass a sufficient variety of normal heart sounds, innocent murmurs, and abnormal murmurs associated with different CHDs. AI algorithms trained with large high-quality datasets outperform interpretations made by inexperienced listeners when assessing murmurs. In this study, the AI was trained on a dataset comprising voice samples from 1413 patients, after which it was prospectively tested with 98 new cases. Compared to other studies that have used AI in the analysis of heart sounds in children, the algorithm used in our study was trained with a larger number of samples 9 , 15 . In our study, electronic stethoscope recording was performed with four standard anterior auscultation points. This technique differs from that used in a small pilot study, in which the recording was made at the loudest location of the murmur 15 . Murmurs caused by some CHDs are best heard in areas not covered by standard auscultation points. For example, a PDA murmur is often heard just below the left clavicle, and small VSD murmurs are only heard in very small areas. A coarctation of aorta (CoA) murmur is usually heard most clearly from the back near the left scapula. In our study, the algorithm identified all CoAs as abnormal even though the back was not included in the recording areas. However, the possibility of a false negative result increases if heart sounds are analysed in an area where the murmur is not heard best. The ability of an algorithm to identify CHD is also weakened if the murmur associated with it does not clearly differ from innocent murmurs. ASD can occur without a murmur, or the murmur of ASD can mimic an innocent murmur from the pulmonary artery area, which was also observed in this study as a false negative finding for ASD. The ability of AI to recognise murmurs outside standard areas can be improved by directing the recording to the point where the physician hears the murmur best. To improve the reliability of murmur examinations in children, a promising approach would be to combine the results of an AI algorithm with findings from a clinical examination. AI algorithms based on the recognition of murmurs and normal heart sounds are unable to recognise CHDs without an audible murmur. Therefore, heart defects without murmurs could not be used to train our algorithm and were also excluded from our training dataset. In this study, AI failed to recognise ASDs, small VSDs, and bicuspid aortic valves with normal function (no stenosis or insufficiency) without a murmur. These defects represent “false negatives” and explain the decrease in sensitivity in the entire study population, which also included patients without a murmur. The age group over four years had more CHDs without a murmur, which explains the lower sensitivity in the older age group in our study. Breathing sounds and heart rate can affect the quality of heart sound recordings 10 . Both are faster in children than in adults, and both decrease as the child ages. The effect of breathing sounds on the quality of murmur recordings can be mitigated by performing the recordings during breath holding, as reported in a small pilot study 15 . However, breath holding requires good co-operation and is not possible for small children. Algorithms based on adult heart sound samples cannot be used for screening children, as heart diseases and murmurs differ between children and adults. Our algorithm was developed with samples from different age groups of children and adolescents (from 0 to 18 years), so it could be a promising method for broader use in the field of paediatrics and adolescent medicine. Previous validation studies assessing AI algorithms in the identification of pathologic murmurs in children have reported similar results to ours. In a virtual clinical study (n = 120, age 2-17 years) based on a database of recordings of children`s heart sounds and murmurs, AI identified Still’s innocent murmur with sensitivity of 90% and specificity of 98%. This selected patient sample had no other innocent murmurs or normal heart sounds and only sound samples recorded at the lower left sternal border were used in the analysis, distinguishing it from our study. The performance of the algorithm worsened when also the sound samples without a murmur and all auscultation areas were included, resulting in the sensitivity of 83% and specificity of 89%. 16 A small ( n = 34) AI pilot study in children over 3.5 years of age reported a sensitivity of 87% and a specificity of 100% in identifying pathologic murmurs 15 . A virtual clinical trial ( n = 603) using AI identified pathologic murmurs with a sensitivity of 93% and a specificity of 81%. In that study, previously recorded patient heart sounds were analysed from a sound bank with the help of AI. Pathologic cases had at least one pathologic diagnosis by echocardiogram and at least one murmur considered to be caused by the pathology. CHD patients without a murmur or with innocent murmurs were excluded, which increased the accuracy of the algorithm 10 . However, due to the differences in patient selection, these results cannot be directly compared with those of our study. The virtual clinical trial by Thompson et al. included also adults, and the age range was wide (0.3–80.9 years), with a median age of 8.8 years and 34% of the patients being over 12 years of age 10 . The strength of our clinical study is the use of echocardiography combined with AI analysis and clinical examination. Another strength of our study is that versatile data were collected from different age groups, covering normal heart sounds as well as innocent and pathologic murmurs related to CHDs. In addition, a large dataset of over 1400 sound samples, validated with echocardiography, was used in developing the algorithm. One of the limitations of this study was the small number of children under one month of age with fast heart and respiratory rates, raising questions about the algorithm’s utility in that age group, which warrants further evaluation. In addition, the exclusion of children with prior heart surgeries means it was not possible to assess the algorithm’s effectiveness in identifying murmurs in this specific paediatric population. The inclusion of children with innocent murmurs makes it difficult to compare the results to those of previous studies 10 ,15 . In Finland, most referrals due to murmurs or suspicion of CHD come from primary health care, so our algorithm could be most useful in screening for murmurs in that context 7 . The high prevalence of innocent murmurs detected in primary health care strains the limited resources of specialised care. Evaluations of auscultatory findings by inexperienced listeners leads to increased numbers of referrals to specialised medical services. Thus, if innocent murmurs could be reliably diagnosed in primary care settings using AI as an aid to clinical examination, the costs of specialised care could be reduced. Then, direct specialised health care resources could be targeted to those patients who need them most. In conclusion, the AI algorithm developed in this study showed promising results among paediatric cardiology outpatients in distinguishing between innocent and pathologic murmurs, exhibiting good sensitivity and specificity. It could be used as an aid to identify murmurs that require further analysis by echocardiography. In addition, when combined with clinical examination, the use of this AI algorithm could increase the number of accurate diagnoses of benign murmurs without a need for echocardiography, thus decreasing health care expenses. Additional research is needed to investigate the potential application of AI algorithms in primary health care settings for screening murmurs in children. A working algorithm could be most useful in developing countries, in which the availability of echocardiography can be limited 17 . Conclusions The algorithm was able to distinguish murmurs associated with structural cardiac anomalies from innocent murmurs with good sensitivity and specificity. The algorithm was unable to identify heart defects that did not cause a murmur. Further research is needed on the use of the algorithm in screening for heart murmurs in primary health care. Abbreviations AI = artificial intelligence ASD = atrial septal defect CHD = congenital heart defect CI= confidence interval CoA = coarctation of aorta HUS = Helsinki University Hospital PDA = patent ductus arteriosus PS = pulmonary stenosis VSD = ventricular septal defect Declarations S tatements of conflict of interest and of funding: Ida Papunen received a 2500 € grant from the Finnish Medical Foundation in October 2021. Martin Porkholm, Oliver Lundqvist, Ilkka Jaakkola, and Otto Rahkonen are shareholders of the company AusculThing Oy. Tuija Poutanen, Kaisa YIänen, Merja Kallio, Minna Mecklin, Anneli Eerola, Anita Arola, and Jussi Niemelä do not have conflicts of interest to declare. References Ip HL, Menahem S. Does echocardiography have a role in the cardiologist’s diagnosis of innocent murmurs in childhood? Heart, Lung and Circulation . 2020;29(2):242-245. doi:10.1016/j.hlc.2019.02.003 Van Oort A, Hopman J, De Boo T, Van Der Werf T, Rohmer J, Daniëls O. The vibratory innocent heart murmur in schoolchildren: A case-control Doppler echocardiographic study. Pediatr Cardiol . 1994;15(6):275-281. doi:10.1007/BF00798120 Hoffman JIE, Kaplan S. The incidence of congenital heart disease. Journal of the American College of Cardiology . 2002;39(12):1890-1900. doi:10.1016/S0735-1097(02)01886-7 Liberman RF, Getz KD, Lin AE, et al. Delayed diagnosis of critical congenital heart defects: Trends and associated factors. Pediatrics . 2014;134(2):e373-e381. doi:10.1542/peds.2013-3949 Kwiatkowski D, Wang Y, Cnota J. The utility of outpatient echocardiography for evaluation of asymptomatic murmurs in children: Outpatient echocardiography for asymptomatic murmurs. Congenital Heart Disease . 2012;7(3):283-288. doi:10.1111/j.1747-0803.2012.00637.x Sackey AH. Prevalence and diagnostic accuracy of heart disease in children with asymptomatic murmurs. Cardiol Young . 2016;26(3):446-450. doi:10.1017/S1047951115000396 Papunen I, Poutanen T, Ylänen K. Major congenital heart defects are rarely diagnosed after newborns’ hospital discharge with modern screening. Acta Paediatrica . 2024;113(1):143-149. doi:10.1111/apa.16928 Chowdhury MEH, Khandakar A, Alzoubi K, et al. Real-time smart-digital stethoscope system for heart diseases monitoring. Sensors . 2019;19(12):2781. doi:10.3390/s19122781 Prince J, Maidens J, Kieu S, et al. Deep learning algorithms to detect murmurs associated with structural heart disease. JAHA . 2023;12(20):e030377. doi:10.1161/JAHA.123.030377 Thompson WR, Reinisch AJ, Unterberger MJ, Schriefl AJ. Artificial intelligence-assisted auscultation of heart murmurs: Validation by virtual clinical trial. Pediatr Cardiol . 2019;40(3):623-629. doi:10.1007/s00246-018-2036-z Singh Y, Chen SE. Impact of pulse oximetry screening to detect congenital heart defects: 5 years’ experience in a UK regional neonatal unit. Eur J Pediatr . 2022;181(2):813-821. doi:10.1007/s00431-021-04275-w Kumar K, Thompson WR. Evaluation of cardiac auscultation skills in pediatric residents. Clin Pediatr (Phila) . 2013;52(1):66-73. doi:10.1177/0009922812466584 Mackie AS, Jutras LC, Dancea AB, Rohlicek CV, Platt R, Béland MJ. Can cardiologists distinguish innocent from pathologic murmurs in neonates? The Journal of Pediatrics . 2009;154(1):50-54.e1. doi:10.1016/j.jpeds.2008.06.017 Leng S, Tan RS, Chai KTC, Wang C, Ghista D, Zhong L. The electronic stethoscope. BioMed Eng OnLine . 2015;14(1):66. doi:10.1186/s12938-015-0056-y Lai LSW, Redington AN, Reinisch AJ, Unterberger MJ, Schriefl AJ. Computerized automatic diagnosis of innocent and pathologic murmurs in pediatrics: A pilot study: Computerized diagnosis of murmurs. Congenital Heart Disease . 2016;11(5):386-395. doi:10.1111/chd.12328 Shekhar R, Vanama G, John T, Issac J, Arjoune Y, Doroshow RW. Automated identification of innocent Still's murmur using a convolutional neural network. Front Pediatr . 2022;10:923956. doi:10.3389/fped.2022.923956 Lv J, Dong B, Lei H, et al. Artificial intelligence-assisted auscultation in detecting congenital heart disease. European Heart Journal – Digital Health . 2021;2(1):119-124. doi:10.1093/ehjdh/ztaa017 Additional Declarations Competing interest reported. Martin Porkholm, Oliver Lundqvist, Ilkka Jaakkola, and Otto Rahkonen are shareholders of the company AusculThing Oy. Cite Share Download PDF Status: Published Journal Publication published 21 Sep, 2024 Read the published version in European Journal of Pediatrics → Version 1 posted Editorial decision: Revision requested 25 Jul, 2024 Reviews received at journal 24 Jul, 2024 Reviews received at journal 15 Jul, 2024 Reviewers agreed at journal 11 Jul, 2024 Reviewers agreed at journal 09 Jul, 2024 Reviewers invited by journal 09 Jul, 2024 Editor assigned by journal 08 Jul, 2024 Submission checks completed at journal 08 Jul, 2024 First submitted to journal 06 Jul, 2024 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-4697876","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":331638780,"identity":"aebc4162-6210-47d8-a0ca-34680c2221de","order_by":0,"name":"Ida Papunen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3klEQVRIiWNgGAWjYBACNoYEOJvxAZDgYWA+2EC0FmaDAyAtbIn4tTAgaWGTOACmErArhAE+9uQHzAU1dxK3s/ceq/7YdkeGgY2ZgMN4nhkwzzj2LHFnz7m0GwfbngEdxkhAi0SCATMP2+HEDTdyzIBaDvMwyDcS0pL+gZnnH1DL/TdmBWAthG3JMWDmbQPZwmPGQJwWnjcFh2f2PTPecCbHWOLMucM8bIS0yLenb3xc8O2O7IbjZww/VJQdtudnY3+AVwsIHGZgOIBkL0H1QMCMomUUjIJRMApGAToAADgBSQU4IX75AAAAAElFTkSuQmCC","orcid":"","institution":"Tampere Center for Child, Adolescent and Maternal Health Research, Faculty of Medicine and Health Technology, Tampere University","correspondingAuthor":true,"prefix":"","firstName":"Ida","middleName":"","lastName":"Papunen","suffix":""},{"id":331638781,"identity":"9f476bc8-be27-4f59-ba71-4d6e909677c3","order_by":1,"name":"Kaisa Ylänen","email":"","orcid":"","institution":"Department of Pediatrics, Tampere University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Kaisa","middleName":"","lastName":"Ylänen","suffix":""},{"id":331638782,"identity":"2d8e7450-22ed-4a85-84f9-758ce01581d2","order_by":2,"name":"Oliver Lundqvist","email":"","orcid":"","institution":"AusculThing Oy","correspondingAuthor":false,"prefix":"","firstName":"Oliver","middleName":"","lastName":"Lundqvist","suffix":""},{"id":331638783,"identity":"f344a7f9-0ef2-41eb-bdcf-ec39fa0aedcc","order_by":3,"name":"Martin Porkholm","email":"","orcid":"","institution":"AusculThing Oy","correspondingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"Porkholm","suffix":""},{"id":331638785,"identity":"827cd8bc-8364-46ce-8e6a-4d436d05f0d1","order_by":4,"name":"Otto Rahkonen","email":"","orcid":"","institution":"Department of Pediatric Cardiology, New Children’s Hospital, University of Helsinki and Helsinki University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Otto","middleName":"","lastName":"Rahkonen","suffix":""},{"id":331638786,"identity":"c4f29a0c-fdb9-40a2-bd00-de841baa8c06","order_by":5,"name":"Minna Mecklin","email":"","orcid":"","institution":"Department of Pediatrics, Tampere University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Minna","middleName":"","lastName":"Mecklin","suffix":""},{"id":331638789,"identity":"284e450e-31ab-4e22-bf84-76fa06cc725e","order_by":6,"name":"Anneli Eerola","email":"","orcid":"","institution":"Department of Pediatrics, Tampere University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Anneli","middleName":"","lastName":"Eerola","suffix":""},{"id":331638792,"identity":"238e5cc0-2ed7-4648-80e0-6b4e5a4eb27a","order_by":7,"name":"Merja Kallio","email":"","orcid":"","institution":"Department of Pediatric Cardiology, New Children’s Hospital, University of Helsinki and Helsinki University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Merja","middleName":"","lastName":"Kallio","suffix":""},{"id":331638794,"identity":"f08547af-f818-4e5d-9f18-56cd67625041","order_by":8,"name":"Anita Arola","email":"","orcid":"","institution":"Department of Pediatrics and Adolescent Medicine Turku University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Anita","middleName":"","lastName":"Arola","suffix":""},{"id":331638795,"identity":"24819667-9847-41e1-bd97-31db7dce5de4","order_by":9,"name":"Jussi Niemelä","email":"","orcid":"","institution":"Department of Pediatrics and Adolescent Medicine Turku University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jussi","middleName":"","lastName":"Niemelä","suffix":""},{"id":331638797,"identity":"5401a4c8-3948-4f95-8d53-681b67d78a5c","order_by":10,"name":"Ilkka Jaakkola","email":"","orcid":"","institution":"Department of Pediatric Cardiology, New Children’s Hospital, University of Helsinki and Helsinki University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ilkka","middleName":"","lastName":"Jaakkola","suffix":""},{"id":331638798,"identity":"23d8d946-1a06-4ff2-9ed3-8fda086c6bce","order_by":11,"name":"Tuija Poutanen","email":"","orcid":"","institution":"Department of Pediatrics, Tampere University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Tuija","middleName":"","lastName":"Poutanen","suffix":""}],"badges":[],"createdAt":"2024-07-06 18:03:49","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4697876/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4697876/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s00431-024-05773-3","type":"published","date":"2024-09-21T15:58:01+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":62153778,"identity":"664f81f7-dc22-43fe-aad1-97adebea8c45","added_by":"auto","created_at":"2024-08-09 20:54:24","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":54236,"visible":true,"origin":"","legend":"\u003cp\u003eEquipment for analysis of the heart sounds.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4697876/v1/3f286027069c7d6244832582.jpg"},{"id":62153780,"identity":"019f05cc-826f-46c1-ba52-bc26118e9c2e","added_by":"auto","created_at":"2024-08-09 20:54:24","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":82190,"visible":true,"origin":"","legend":"\u003cp\u003ePhonocardiograms of normal heart sounds (top), innocent murmur (middle) and ventricular septal defect (bottom).\u003c/p\u003e","description":"","filename":"Figure2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4697876/v1/7ca4df5f08320e5524e8cbbe.jpg"},{"id":62153779,"identity":"5bacdb9e-fdb0-4a87-bf28-761b2260aec0","added_by":"auto","created_at":"2024-08-09 20:54:24","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":33243,"visible":true,"origin":"","legend":"\u003cp\u003eThe probability of congenital heart defect as analyzed by the algorithm compared to echocardiography results in the patients with a murmur (n=90)\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-4697876/v1/2d3610d97bd93f682ea4b4ec.jpg"},{"id":65104158,"identity":"67866f41-393c-4b41-b699-e674fb86006d","added_by":"auto","created_at":"2024-09-23 16:12:23","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":699227,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4697876/v1/9292a9b3-d048-4ba3-84e3-bcbb3ecc7478.pdf"}],"financialInterests":"Competing interest reported. Martin Porkholm, Oliver Lundqvist, Ilkka Jaakkola, and Otto Rahkonen are shareholders of the company AusculThing Oy.","formattedTitle":"Automated Analysis of Heart Sound Signals in Screening for Structural Heart Disease in Children","fulltext":[{"header":"What is known","content":"\u003cp\u003eInnocent murmurs are common in children, while the incidence of moderate or severe congenital heart defects is low.\u0026nbsp;Auscultation plays a significant role in assessing the need for further examinations of the murmur. The ability to differentiate innocent murmurs from those related to congenital heart defects requires clinical experience on the part of general practitioners. No AI-based auscultation algorithms have been systematically implemented in primary health care.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eWhat is new:\u0026nbsp;\u003c/strong\u003eWe developed an AI-based algorithm using a large dataset of sound samples validated by echocardiography The algorithm performed well in recognising pathological and innocent murmurs in children from different age groups.\u003c/p\u003e"},{"header":"Introduction","content":"\u003cp\u003eA heart murmur can be a sign of a congenital heart defect (CHD), but\u0026nbsp;innocent\u0026nbsp;murmurs are common in children and can be heard in 50% of school-aged children\u0026nbsp;\u003csup\u003e1,2\u003c/sup\u003e. CHDs are found in about 75 per 1000 live-born infants. However, the incidence of moderate or severe CHD is only six cases per 1000 live births\u0026nbsp;\u003csup\u003e3\u003c/sup\u003e. About 10% of children with significant CHDs are discharged from the maternity hospital without receiving a diagnosis\u0026nbsp;\u003csup\u003e4\u003c/sup\u003e, and CHDs are rarely found in children over six months of age\u0026nbsp;\u003csup\u003e5\u003c/sup\u003e \u003csup\u003e6\u003c/sup\u003e. Although a murmur is the most common finding that raises suspicion of CHD, most of the children referred to paediatric cardiologists have an\u0026nbsp;innocent\u0026nbsp;murmur. According to some estimates, up to 90% of initial visits to paediatric cardiology clinics were due to an\u0026nbsp;innocent\u0026nbsp;murmur and thus unnecessary\u0026nbsp;\u003csup\u003e5,7\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDespite the development of medical technology, auscultation still plays a significant role in assessing the need for further examinations of the murmur. The interpretation of cardiac auscultation is subject to uncertainty because it depends on the subjective perceptions and experience of the clinician. Computer decision algorithms based on artificial intelligence (AI) have been developed in recent years to identify murmurs collected by electronic stethoscopes\u0026nbsp;\u003csup\u003e8\u003c/sup\u003e. The sensitivity and specificity of an AI-assisted murmur detection algorithm were shown to be good in differentiating\u0026nbsp;innocent\u0026nbsp;murmurs from pathologic ones in a validation study involving both adults and children\u0026nbsp;\u003csup\u003e9\u003c/sup\u003e. To the best of our knowledge, only a few studies have been published on paediatric populations, with the algorithms differentiating pathologic murmurs from normal heart sounds with good sensitivity\u0026nbsp;\u003csup\u003e9,10\u003c/sup\u003e. While the use of AI has potential for cost-effective CHD screening, it has not yet been utilised in clinical practice. Our aim was to evaluate the ability of an AI-based algorithm to identify innocent and pathologic murmurs in children.\u0026nbsp;\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eThe development of the algorithm\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn AI-based algorithm was developed by fine-tuning a state-of-the-art speech model using heart sound recordings collected at the five university hospitals in Finland. The heart sound recordings, selected for the training of the algorithm, were collected from 1413 patients, including 1061 (75%) children (\u0026lt;18 years) and 352 (25%) adults (Table 1).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe recordings from CHDs without murmurs were not used in the development of the algorithm. The recordings were made with the Thinklabs One\u0026reg; electronic stethoscope, which is commercially available and is licensed for medical use by the manufacturer (US FDA Class 2 diagnostic device 2004 and CE-marking 2017). The data were collected and saved on a Samsung Galaxy Tab A10 tablet (CE) (Figure 1). Recordings were made from four different locations (aortic, pulmonary, tricuspid, and mitral valve areas), with a sample rate of 44 kHz and a duration of 10 seconds. If patient co-operation was not adequate to complete all four recordings with good signal quality, only the samples with adequate signal were included in training the algorithm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1\u0026nbsp;\u003c/strong\u003eBasic information of the patients used to train the algorithm (N=1413).\u003cstrong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"63.872832369942195%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.127167630057805%\" valign=\"top\"\u003e\n \u003cp\u003eN= 1413 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"63.872832369942195%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex\u0026nbsp;\u003c/strong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eFemale\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.127167630057805%\" valign=\"top\"\u003e\n \u003cp\u003e671 (47 %)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"63.872832369942195%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0\u0026ndash;1 month\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e1\u0026ndash;6 months\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e6\u0026ndash;12 months\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e1\u0026ndash;4 years\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e4\u0026ndash;12 years\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e12\u0026ndash;18 years\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026gt; 18 years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.127167630057805%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e83 (6 %)\u003c/p\u003e\n \u003cp\u003e155 (11 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e66 (5 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e143 (10 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e375 (27 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e239 (17 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e352 (25 %)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"63.872832369942195%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHeart murmur\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eNo murmur and no CHD\u003c/p\u003e\n \u003cp\u003eInnocent murmur\u003c/p\u003e\n \u003cp\u003ePathological murmur\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.127167630057805%\" valign=\"top\"\u003e\n \u003cp\u003e514 (36 %)\u003c/p\u003e\n \u003cp\u003e339 (24 %)\u003c/p\u003e\n \u003cp\u003e560 (40 %)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"63.872832369942195%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eResult of echocardiography\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eNormal\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eAbnormal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"36.127167630057805%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e853 (60 %) \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e560 (40 %)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe trained algorithm was divided into two leading AI algorithms. A quality algorithm was used to pre-screen the raw phonocardiogram signal and remove noise and other artifacts, and a binary classifier input the high-quality phonocardiogram segments and used this information to predict the presence or absence of a heart defect causing an audible heart murmur (Figure 2). The algorithm formed a floating-point value between 0 and 1. A threshold of 0.5 was set as the divider for the two categories of normal and abnormal heart sounds. If the output value (probability of defect) was less than 0.5, the result was classified as normal, and an output value above 0.5 was classified as abnormal. The normal category included heart sounds with no murmur and innocent murmurs, while the abnormal heart sound included pathological murmurs. Both algorithms were trained using the same training dataset and utilised the same separate development set for hyperparameter tuning and establishing performance metrics on data not used during training to understand how well each model generalised on unseen data. The development set was constructed using a 90\u0026ndash;10 split (90% of the data was used for training, and 10% was placed in the development set). The development set was not the same dataset as the result set of the 98 patients, which was used in this research.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStudy patients\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u003c/strong\u003eA total of 98 children were prospectively recruited for the study during their outpatient visits at the Paediatric Cardiology unit at the New Children\u0026rsquo;s Hospital in Helsinki, Finland, between 29 Mar 2022 and 9 Aug 2023. The patients were examined by a paediatric cardiologist due to a heart murmur or CHD. Children with previous heart surgery or insufficient co-operation were excluded. The patients were first analysed with the algorithm, and then the results were compared with clinical evaluations and echocardiography. The paediatric cardiologists performed auscultation with a conventional stethoscope before the echocardiography and assessed whether a murmur was audible. The cardiologist was not aware of the analysis result of the AI. Finally, echocardiography was\u0026nbsp;performed, and the findings were\u0026nbsp;classified\u0026nbsp;as normal or abnormal.\u0026nbsp;The algorithm\u0026rsquo;s analysis of the murmur (normal or abnormal) was compared with the diagnosis made by echocardiography.\u0026nbsp;The algorithm result \u0026ldquo;abnormal\u0026rdquo; included all pathologic murmurs identified by the algorithm. The result \u0026ldquo;normal\u0026rdquo; included\u0026nbsp;innocent\u0026nbsp;murmurs and normal heart sounds without murmurs.\u003c/p\u003e\n\u003cp\u003eAll the study patients and/or their legal guardians provided written informed consent. This study complied with the Declaration of Helsinki and was approved by the ethical committee (HUS/1630/2019 23 August 2019) of Helsinki University Hospital (HUS).\u003c/p\u003e\n\u003cp\u003eThe statistical analyses were performed using IBM SPSS Statistics for Macintosh (version 29.0.1.0). Categorical values were expressed as frequencies and percentages.\u0026nbsp;The median and interquartile range (IQR) (Q1\u0026ndash;Q3) were expressed for the non-normally distributed variables. The sensitivity, specificity, and accuracy of the algorithm to differentiate an innocent murmur from a pathologic one were calculated.\u0026nbsp;Categorical variables were compared with chi-square tests. Sensitivity, specificity, and accuracy were calculated with 95% confidence intervals.\u0026nbsp;\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003e\u003cem\u003eStudy population\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHeart sounds of 98 paediatric patients were analysed in this study. The patient characteristics and echocardiography findings are presented in Table 2.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2\u003c/strong\u003e Basic patient data of the\u0026nbsp;patients examined with the algorithm.\u003cstrong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"71.8543046357616%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.14569536423841%\" valign=\"top\"\u003e\n \u003cp\u003eN=98 (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"71.8543046357616%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex\u0026nbsp;\u003c/strong\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp; \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.14569536423841%\" valign=\"top\"\u003e\n \u003cp\u003e50 (51 %)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"71.8543046357616%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0\u0026ndash;1 month\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e1\u0026ndash;6 months\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e6\u0026ndash;12 months\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e1\u0026ndash;4 years\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e4\u0026ndash;12 years\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e\u0026gt;12 years\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.14569536423841%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e3 (3 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e14 (14 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e7 (7 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e20 (20 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e31 (32 %)\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e23 (24 %)\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"71.8543046357616%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eHeart murmur\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.14569536423841%\" valign=\"top\"\u003e\n \u003cp\u003e90 (92 %)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"71.8543046357616%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eResult of echocardiography\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eNormal\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eAbnormal\u003cstrong\u003e*\u003c/strong\u003e\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eShunt lesions\u003c/p\u003e\n \u003cp\u003eValvular lesions\u003c/p\u003e\n \u003cp\u003eOther lesions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.14569536423841%\" valign=\"top\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e63 (64 %) \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e35 (36 %)\u003c/p\u003e\n \u003cp\u003e22 (22 %) \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e17 (17 %) \u0026nbsp;\u003c/p\u003e\n \u003cp\u003e3 (3 %) \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*One patient can have more than one diagnosis\u003c/p\u003e\n\u003cp\u003eEchocardiography was normal in\u0026nbsp;63 (64%) children and abnormal in 35 (36%).\u003c/p\u003e\n\u003cp\u003eThe algorithm identified 72 (73%) of the heart sounds as normal and 26 (27%) as abnormal. The sensitivity of the algorithm was 69% (CI 51\u0026ndash;83%), and the specificity was 97% (CI 89\u0026ndash;100%) (Table 3).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003ePerformance of the AI-based algorithm in detecting pathologic murmurs verified by echocardiography.\u0026nbsp;All patients (N = 98)\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.889816360601%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"42.07011686143573%\" colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEchocardiography\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.040066777963272%\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.889816360601%\" colspan=\"2\"\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.535893155258766%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbnormal\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.534223706176963%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.040066777963272%\"\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"16.52754590984975%\" rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"17.362270450751254%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbnormal\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.535893155258766%\" valign=\"top\"\u003e\n \u003cp\u003eTrue positive\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e24\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.534223706176963%\" valign=\"top\"\u003e\n \u003cp\u003eFalse positive\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.040066777963272%\" valign=\"top\"\u003e\n \u003cp\u003ePositive predictive value\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.92\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"20.8%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.8%\" valign=\"top\"\u003e\n \u003cp\u003eFalse negative\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e11\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.6%\" valign=\"top\"\u003e\n \u003cp\u003eTrue negative\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e61\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.8%\" valign=\"top\"\u003e\n \u003cp\u003eNegative predictive value\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.85\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"33.889816360601%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.535893155258766%\" valign=\"top\"\u003e\n \u003cp\u003eSensitivity\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.69\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"20.534223706176963%\" valign=\"top\"\u003e\n \u003cp\u003eSpecificity\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.97\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.040066777963272%\" valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.87\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe algorithm recognised pathologic murmurs in 24 of 35 cases with abnormal echocardiography and misdiagnosed 11 cases. Of the 11 misdiagnosed cases, five had an audible murmur and six had no murmur. The algorithm did not identify murmurs caused by\u0026nbsp;aortic valve insufficiency\u0026nbsp;(diastolic grade I), patent ductus arteriosus (PDA) (grade 1 systo-diastolic), atrial septal defect (ASD), or a small ventricular septal defect (VSD). In total, 61 of 63 children with normal echocardiography findings were identified as having normal heart sounds by the algorithm. The level of confidence of the algorithm was unsure in six (6%) analyses and confident in 92 (94%). The details of the algorithm analyses and the findings of the patients with abnormal echocardiography are presented in Table 4.\u0026nbsp;In the group of patients with abnormal echocardiography findings, the median (IQR) probability set by the algorithm of a CHD was 0.66 (0.27\u0026ndash;0.98) (Table 4).\u003c/p\u003e\n\u003cp\u003eWhen comparing age groups, the algorithm had a sensitivity of 75% (12/16) (CI 48\u0026ndash;93%) and a specificity of 96% (27/28) (CI 82\u0026ndash;100%) in the group of children aged 0\u0026ndash;4 years\u0026nbsp;(n=44). In children over four years of age (n=54), the sensitivity was 63% (12/19) (CI 38\u0026ndash;84%), while the specificity was 97% (34/35) (CI 85\u0026ndash;100%). There was one CHD without a murmur in children aged 0\u0026ndash;4 years (VSD) and five in children over four years of age (VSD, bicuspid aortic valve, and ASD) (Table 4).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4\u0026nbsp;\u003c/strong\u003eThe diagnoses\u0026nbsp;found in the\u0026nbsp;echocardiography (n = 35)\u0026nbsp;classified according to\u0026nbsp;the interpretation of the algorithm.\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"19.781931464174455%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm analysis\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.81619937694704%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eDiagnoses\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.2398753894081%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eMurmur\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.878504672897197%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003en\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.28348909657321%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eProbability of defect\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e(range)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"19.781931464174455%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbnormal\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003en = 24\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.81619937694704%\" valign=\"top\"\u003e\n \u003cp\u003eVentricular septal defect\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eAtrial septal defect\u003c/p\u003e\n \u003cp\u003eAortic stenosis\u0026nbsp;\u003c/p\u003e\n \u003cp\u003ePulmonary stenosis\u003c/p\u003e\n \u003cp\u003eCoarctation of aorta\u003c/p\u003e\n \u003cp\u003ePatent ductus arteriosus\u003c/p\u003e\n \u003cp\u003eAortic and pulmonary stenosis\u003c/p\u003e\n \u003cp\u003eOthers\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.2398753894081%\" valign=\"top\"\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.878504672897197%\" valign=\"top\"\u003e\n \u003cp\u003e7\u003c/p\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003cp\u003e4\u003c/p\u003e\n \u003cp\u003e3\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.28348909657321%\" valign=\"top\"\u003e\n \u003cp\u003e0.54\u0026ndash;0.99\u003c/p\u003e\n \u003cp\u003e0.58\u0026ndash;0.99\u003c/p\u003e\n \u003cp\u003e0.87\u0026ndash;0.97\u003c/p\u003e\n \u003cp\u003e0.99\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.78\u0026ndash;0.87\u003c/p\u003e\n \u003cp\u003e0.92\u003c/p\u003e\n \u003cp\u003e0.98\u003c/p\u003e\n \u003cp\u003e0.94\u0026ndash;0.95\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"19.781931464174455%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003en = 11\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.81619937694704%\" valign=\"top\"\u003e\n \u003cp\u003eBicuspid aortic valve\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eVentricular septal defect\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eAtrial septal defect\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eAtrial septal defect\u0026nbsp;\u003c/p\u003e\n \u003cp\u003eVentricular septal defect\u003c/p\u003e\n \u003cp\u003eAortic valve insufficiency\u003c/p\u003e\n \u003cp\u003ePatent ductus arteriosus\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"13.2398753894081%\" valign=\"top\"\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003cp\u003eNo\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003cp\u003eYes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"8.878504672897197%\" valign=\"top\"\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003cp\u003e2\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003cp\u003e1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"29.28348909657321%\" valign=\"top\"\u003e\n \u003cp\u003e0.03\u003c/p\u003e\n \u003cp\u003e0.09\u0026ndash;0.25\u003c/p\u003e\n \u003cp\u003e0.18\u0026ndash;0.33\u003c/p\u003e\n \u003cp\u003e0.11\u0026ndash;0.49\u003c/p\u003e\n \u003cp\u003e0.41\u003c/p\u003e\n \u003cp\u003e0.04\u003c/p\u003e\n \u003cp\u003e0.07\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003ePatients with a murmur\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePaediatric cardiologists heard a murmur from 90 of 98 patients (92%). In these 90 patients, the sensitivity of the algorithm was 83% (24/29) (CI 64\u0026ndash;94%) and the specificity was 97% (59/61) (CI 89\u0026ndash;100%) (Table 5). The algorithm was confident in 83 (93%) and unsure in six (7%) analyses.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 5\u0026nbsp;\u003c/strong\u003ePerformance of the AI based algorithm in detecting pathologic murmurs verified by echocardiography among patients with audible murmur (N = 90).\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd width=\"32.56198347107438%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"42.97520661157025%\" colspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eEchocardiography\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.462809917355372%\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"32.615894039735096%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.357615894039736%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbnormal\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.52317880794702%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.503311258278146%\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; \u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"14.072847682119205%\" rowspan=\"2\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"18.543046357615893%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eAbnormal\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.357615894039736%\" valign=\"top\"\u003e\n \u003cp\u003eTrue positive\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e24\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.52317880794702%\" valign=\"top\"\u003e\n \u003cp\u003eFalse positive\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e2\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.503311258278146%\" valign=\"top\"\u003e\n \u003cp\u003ePositive predictive value\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.92\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"21.57996146435453%\" valign=\"top\"\u003e\n \u003cp\u003e\u003cstrong\u003eNormal\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.85549132947977%\" valign=\"top\"\u003e\n \u003cp\u003eFalse negative\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"25.048169556840076%\" valign=\"top\"\u003e\n \u003cp\u003eTrue negative\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e59\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"28.516377649325626%\" valign=\"top\"\u003e\n \u003cp\u003eNegative predictive value\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.94\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd width=\"32.615894039735096%\" colspan=\"2\"\u003e\n \u003cp\u003e\u0026nbsp;\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.357615894039736%\" valign=\"top\"\u003e\n \u003cp\u003eSensitivity\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.86\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"21.52317880794702%\" valign=\"top\"\u003e\n \u003cp\u003eSpecificity\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.97\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd width=\"24.503311258278146%\" valign=\"top\"\u003e\n \u003cp\u003eAccuracy\u0026nbsp;\u003c/p\u003e\n \u003cp\u003e0.93\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe median (IQR) probability of CHD was 0.15 (0.03\u0026ndash;0.27) in children with normal echocardiography findings and an audible murmur and 0.81 (0.79\u0026ndash;0.99) in children with abnormal echocardiography findings and an audible murmur (Figure 3). Comparing age groups, the sensitivity of the algorithm was 80% (12/15) (CI 52\u0026ndash;96%) and its specificity was 96% (26/27) (CI 81\u0026ndash;100%) in the group of children aged 0\u0026ndash;4 years (n=42). In children over four years of age (n=48), the algorithm\u0026rsquo;s sensitivity was 86% (12/14, CI 57\u0026ndash;98%) and its specificity was 97% (33/34, CI 85\u0026ndash;100%).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe vertical black line marks the threshold probability of the defect (0.5). The black dots represent the abnormalities that were not detected by the algorithm. The dots with \u0026ldquo;X\u0026rdquo; are the normal cases interpreted incorrectly as abnormal. The white dots represent the cases that the algorithm identified correctly.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis is the first clinical study to investigate the ability of AI to differentiate between benign and pathological heart murmurs in children to this extent using echocardiography as the gold standard.\u0026nbsp;The algorithm was able to distinguish a murmur caused by a CHD from innocent murmurs with good sensitivity and specificity when echocardiography was used as the reference. Since the method is based on the identification of different heart sound signals, it is not suitable for screening CHDs without a murmur.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDespite the routine use of pulse oximetry screening, which is highly sensitive in detecting critical cyanotic CHDs, some newborns with acyanotic CHDs may still go undiagnosed before discharge from the maternity ward\u0026nbsp;\u003csup\u003e11\u003c/sup\u003e. Indeed, many CHDs do not cause symptoms during the first week of life\u0026nbsp;\u003csup\u003e4\u003c/sup\u003e, and some are only diagnosed after normal postnatal adaptation has taken place. For example, VSD causes murmur after a reduction in pulmonary vascular resistance leads to a pressure difference between ventricles and coarctation when the arterial duct has closed and caused narrowing of the aorta.\u0026nbsp;CHDs diagnosed in children over six months of age are usually mild and asymptomatic, including ASD, PDA, and valvular defects\u0026nbsp;\u003csup\u003e6\u003c/sup\u003e. Some valvular defects and hypertrophic cardiomyopathy may manifest later in adolescence. Auscultation during childhood plays a vital role in identifying these CHDs, but the accuracy of interpretation depends on the experience and skills of the physician. In previous studies, the sensitivity and specificity of clinical assessments of CHD have varied widely depending on the clinical experience of the physician, while specificity generally increased with experience. For example, medical students and paediatric residents had a sensitivity of 82% but a specificity of only 56% in assessing CHD\u003csup\u003e12\u003c/sup\u003e. Among paediatricians, sensitivity was found to be better (93%), but specificity remained low (59%) \u003csup\u003e6\u003c/sup\u003e. Meanwhile, clinical assessment by paediatric cardiologists had a sensitivity of 81% and specificity of 91% in identifying neonates with CHD \u003csup\u003e13\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eAuscultation using a conventional stethoscope or AI requires optimal conditions, including good patient co-operation and quiet environment. If a child is crying, both the physician and the algorithm may struggle to recognise murmurs accurately. In our study, all children were co-operative, ensuring a reliable evaluation of the algorithm\u0026rsquo;s performance.\u0026nbsp;The AI used in this study includes a quality algorithm that screens raw phonocardiogram signals and removes noise and other artifacts, allowing high-quality phonocardiogram segments to be used for analysis of heart sounds.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eA prerequisite for the development and utilisation of AI in health care is the ability to reliably listen to and record heart sounds. In recent decades, electronic stethoscopes have undergone significant development, resulting in enhanced capabilities for analysing murmurs. These devices improve sound signals and reduce background noise, facilitating auscultation. Most available models not only aid in listening but also can record sounds and store data on murmurs for future reference \u003csup\u003e14\u003c/sup\u003e.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe accuracy of algorithms primarily hinges on the quantity and quality of the heart sound samples used in their development. The training data must encompass a sufficient variety of normal heart sounds, innocent murmurs, and abnormal murmurs associated with different CHDs. AI algorithms trained\u0026nbsp;with\u0026nbsp;large high-quality datasets\u0026nbsp;outperform interpretations made by inexperienced listeners when assessing murmurs. In this study, the AI was trained on a dataset\u0026nbsp;comprising\u0026nbsp;voice samples\u0026nbsp;from\u0026nbsp;1413 patients, after which it was prospectively tested with 98 new cases. Compared to other studies that have used AI in the analysis of\u0026nbsp;heart sounds in children, the algorithm used in our study was trained with a larger number of samples\u0026nbsp;\u003csup\u003e9\u003c/sup\u003e\u003csup\u003e,\u003c/sup\u003e\u003csup\u003e15\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn our study, electronic stethoscope recording was performed with four standard anterior auscultation points. This technique differs from that used in a small pilot study, in which the recording was made at the loudest location of the murmur \u003csup\u003e15\u003c/sup\u003e. Murmurs caused by some CHDs are best heard in areas not covered by standard auscultation points. For example, a PDA murmur is often heard just below the left clavicle, and small VSD murmurs are only heard in very small areas. A coarctation of aorta (CoA) murmur is usually heard most clearly from the back near the left scapula. In our study, the algorithm identified all CoAs as abnormal even though the back was not included in the recording areas. However, the possibility of a false negative result increases if heart sounds are analysed in an area where the murmur is not heard best.\u0026nbsp;The ability of an algorithm to identify CHD is also weakened if the murmur associated with it does not clearly differ from innocent murmurs. ASD can occur without a murmur, or the murmur of ASD can mimic an innocent murmur from the pulmonary artery area, which was also observed in this study as a false negative finding for ASD. The ability of AI to recognise murmurs outside standard areas can be improved by directing the recording to the point where the physician hears the murmur best. To improve the reliability of murmur examinations in children, a promising approach would be to combine the results of an AI algorithm with findings from a clinical examination.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;AI algorithms based on the recognition of murmurs and normal heart sounds are unable to recognise CHDs without an audible murmur. Therefore, heart defects without murmurs could not be used to train our algorithm and were also excluded from our training dataset. In this study, AI failed to recognise ASDs, small VSDs, and bicuspid aortic valves with normal function (no stenosis or insufficiency) without a murmur. These defects represent \u0026ldquo;false negatives\u0026rdquo; and explain the decrease in sensitivity in the entire study population, which also included patients without a murmur. The age group over four years had more CHDs without a murmur, which explains the lower sensitivity in the older age group in our study.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eBreathing sounds and heart rate can affect the quality of heart sound recordings\u0026nbsp;\u003csup\u003e10\u003c/sup\u003e.\u0026nbsp;Both are faster in children than in adults, and both decrease as the child ages.\u0026nbsp;The effect of breathing sounds on the quality of murmur recordings can be mitigated by performing the recordings during breath holding, as reported in a small pilot study \u003csup\u003e15\u003c/sup\u003e. However, breath holding requires good co-operation and is not possible for small children.\u0026nbsp;Algorithms based on adult heart sound samples cannot be used for screening children, as heart diseases and murmurs differ between children and adults. Our algorithm was developed with samples from different age groups of children and adolescents (from 0 to 18 years), so it could be a promising method for broader use in the field of paediatrics and adolescent medicine.\u003c/p\u003e\n\u003cp\u003ePrevious validation studies assessing AI algorithms in the identification of pathologic murmurs in children have reported similar results to ours. In a virtual clinical study (n = 120, age 2-17 years) based on a database of recordings of children`s heart sounds and murmurs, AI identified Still\u0026rsquo;s innocent murmur with sensitivity of 90% and specificity of 98%. This\u0026nbsp;selected patient sample had no other innocent murmurs or normal heart sounds and only sound samples recorded at the lower left sternal border were used in the analysis, distinguishing it from our study. The performance of the algorithm worsened when also the sound samples without a murmur and all auscultation areas were included, resulting in the sensitivity of 83% and specificity of 89%.\u003csup\u003e16\u003c/sup\u003e A small (\u003cem\u003en\u003c/em\u003e = 34) AI pilot study in children over 3.5 years of age reported a sensitivity of 87% and a specificity of 100% in identifying pathologic murmurs \u003csup\u003e15\u003c/sup\u003e. A virtual clinical trial (\u003cem\u003en =\u003c/em\u003e 603) using AI identified pathologic murmurs with a sensitivity of 93% and a specificity of 81%. In that study, previously recorded patient heart sounds were analysed from a sound bank with the help of AI. Pathologic cases had at least one pathologic diagnosis by echocardiogram and at least one murmur considered to be caused by the pathology. CHD patients without a murmur or with innocent murmurs were excluded, which increased the accuracy of the algorithm\u0026nbsp;\u003csup\u003e10\u003c/sup\u003e. However, due to the differences in patient selection, these results cannot be directly compared with those of our study. The virtual clinical trial by Thompson et al. included also adults, and the age range was wide (0.3\u0026ndash;80.9 years), with a median age of 8.8 years and 34% of the patients being over 12 years of age\u0026nbsp;\u003csup\u003e10\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eThe strength of our clinical study is the use of echocardiography combined with AI analysis and clinical examination. Another strength of our study is that versatile data were collected from different age groups, covering normal heart sounds as well as innocent and pathologic murmurs related to CHDs. In addition, a large dataset of over 1400 sound samples, validated with echocardiography, was used in developing the algorithm.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eOne of the limitations of this study was the small number of children under one month of age with fast heart and respiratory rates, raising questions about the algorithm\u0026rsquo;s utility in that age group, which warrants further evaluation. In addition, the exclusion of children with prior heart surgeries means it was not possible to assess the algorithm\u0026rsquo;s effectiveness in identifying murmurs in this specific paediatric population.\u0026nbsp;The inclusion of children with innocent murmurs makes it difficult to compare the results to those of previous studies\u0026nbsp;\u003csup\u003e10\u003c/sup\u003e\u003csup\u003e,15\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn Finland, most referrals due to murmurs or suspicion of CHD come from primary health care, so our algorithm could be most useful in screening for murmurs in that context\u0026nbsp;\u003csup\u003e7\u003c/sup\u003e. The high prevalence of innocent murmurs detected in primary health care strains the limited resources of specialised care. Evaluations of auscultatory findings by inexperienced listeners leads to increased numbers of referrals to specialised medical services. Thus, if innocent murmurs could be reliably diagnosed in primary care settings using AI as an aid to clinical examination, the costs of specialised care could be reduced. Then, direct specialised health care resources could be targeted to those patients who need them most.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIn conclusion, the AI algorithm developed in this study showed promising results among paediatric cardiology outpatients in distinguishing between innocent and pathologic murmurs, exhibiting good sensitivity and specificity. It could be used as an aid to identify murmurs that require further analysis by echocardiography.\u0026nbsp;In addition, when combined with clinical examination, the use of this AI algorithm could increase the number of accurate diagnoses of benign murmurs without\u0026nbsp;a need for echocardiography, thus decreasing health care expenses. Additional research is needed to investigate the potential application of AI algorithms in primary health care settings for screening murmurs in children. A working algorithm could be most useful in developing countries, in which the availability of echocardiography can be limited \u003csup\u003e17\u003c/sup\u003e.\u003c/p\u003e"},{"header":"Conclusions","content":"\u003cp\u003eThe algorithm was able to distinguish murmurs associated with structural cardiac anomalies from innocent murmurs with good sensitivity and specificity. The algorithm was unable to identify heart defects that did not cause a murmur. Further research is needed on the use of the algorithm in screening for heart murmurs in primary health care.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAI = artificial intelligence\u003c/p\u003e\n\u003cp\u003eASD = atrial septal defect\u003c/p\u003e\n\u003cp\u003eCHD = congenital heart defect\u003c/p\u003e\n\u003cp\u003eCI= confidence interval\u003c/p\u003e\n\u003cp\u003eCoA = coarctation of aorta\u003c/p\u003e\n\u003cp\u003eHUS = Helsinki University Hospital\u003c/p\u003e\n\u003cp\u003ePDA = patent ductus arteriosus\u003c/p\u003e\n\u003cp\u003ePS = pulmonary stenosis\u003c/p\u003e\n\u003cp\u003eVSD = ventricular septal defect\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eS\u003c/strong\u003e\u003cstrong\u003etatements of conflict of interest and of funding:\u003c/strong\u003e Ida Papunen received a 2500 \u0026euro; grant from the Finnish Medical Foundation in October 2021. Martin Porkholm, Oliver Lundqvist, Ilkka Jaakkola, and Otto Rahkonen are shareholders of the company AusculThing Oy. Tuija Poutanen, Kaisa YIänen, Merja Kallio, Minna Mecklin, Anneli Eerola, Anita Arola, and Jussi Niemel\u0026auml; do not have conflicts of interest to declare.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eIp HL, Menahem S. Does echocardiography have a role in the cardiologist\u0026rsquo;s diagnosis of innocent murmurs in childhood? \u003cem\u003eHeart, Lung and Circulation\u003c/em\u003e. 2020;29(2):242-245. doi:10.1016/j.hlc.2019.02.003 \u003c/li\u003e\n\u003cli\u003eVan Oort A, Hopman J, De Boo T, Van Der Werf T, Rohmer J, Dani\u0026euml;ls O. The vibratory innocent heart murmur in schoolchildren: A case-control Doppler echocardiographic study. \u003cem\u003ePediatr Cardiol\u003c/em\u003e. 1994;15(6):275-281. doi:10.1007/BF00798120 \u003c/li\u003e\n\u003cli\u003eHoffman JIE, Kaplan S. The incidence of congenital heart disease. \u003cem\u003eJournal of the American College of Cardiology\u003c/em\u003e. 2002;39(12):1890-1900. doi:10.1016/S0735-1097(02)01886-7 \u003c/li\u003e\n\u003cli\u003eLiberman RF, Getz KD, Lin AE, et al. Delayed diagnosis of critical congenital heart defects: Trends and associated factors. \u003cem\u003ePediatrics\u003c/em\u003e. 2014;134(2):e373-e381. doi:10.1542/peds.2013-3949 \u003c/li\u003e\n\u003cli\u003eKwiatkowski D, Wang Y, Cnota J. The utility of outpatient echocardiography for evaluation of asymptomatic murmurs in children: Outpatient echocardiography for asymptomatic murmurs. \u003cem\u003eCongenital Heart Disease\u003c/em\u003e. 2012;7(3):283-288. doi:10.1111/j.1747-0803.2012.00637.x \u003c/li\u003e\n\u003cli\u003eSackey AH. Prevalence and diagnostic accuracy of heart disease in children with asymptomatic murmurs. \u003cem\u003eCardiol Young\u003c/em\u003e. 2016;26(3):446-450. doi:10.1017/S1047951115000396 \u003c/li\u003e\n\u003cli\u003ePapunen I, Poutanen T, Yl\u0026auml;nen K. Major congenital heart defects are rarely diagnosed after newborns\u0026rsquo; hospital discharge with modern screening. \u003cem\u003eActa Paediatrica\u003c/em\u003e. 2024;113(1):143-149. doi:10.1111/apa.16928 \u003c/li\u003e\n\u003cli\u003eChowdhury MEH, Khandakar A, Alzoubi K, et al. Real-time smart-digital stethoscope system for heart diseases monitoring. \u003cem\u003eSensors\u003c/em\u003e. 2019;19(12):2781. doi:10.3390/s19122781 \u003c/li\u003e\n\u003cli\u003ePrince J, Maidens J, Kieu S, et al. Deep learning algorithms to detect murmurs associated with structural heart disease. \u003cem\u003eJAHA\u003c/em\u003e. 2023;12(20):e030377. doi:10.1161/JAHA.123.030377 \u003c/li\u003e\n\u003cli\u003eThompson WR, Reinisch AJ, Unterberger MJ, Schriefl AJ. Artificial intelligence-assisted auscultation of heart murmurs: Validation by virtual clinical trial. \u003cem\u003ePediatr Cardiol\u003c/em\u003e. 2019;40(3):623-629. doi:10.1007/s00246-018-2036-z \u003c/li\u003e\n\u003cli\u003eSingh Y, Chen SE. Impact of pulse oximetry screening to detect congenital heart defects: 5 years\u0026rsquo; experience in a UK regional neonatal unit. \u003cem\u003eEur J Pediatr\u003c/em\u003e. 2022;181(2):813-821. doi:10.1007/s00431-021-04275-w \u003c/li\u003e\n\u003cli\u003eKumar K, Thompson WR. Evaluation of cardiac auscultation skills in pediatric residents. \u003cem\u003eClin Pediatr (Phila)\u003c/em\u003e. 2013;52(1):66-73. doi:10.1177/0009922812466584 \u003c/li\u003e\n\u003cli\u003eMackie AS, Jutras LC, Dancea AB, Rohlicek CV, Platt R, B\u0026eacute;land MJ. Can cardiologists distinguish innocent from pathologic murmurs in neonates? \u003cem\u003eThe Journal of Pediatrics\u003c/em\u003e. 2009;154(1):50-54.e1. doi:10.1016/j.jpeds.2008.06.017 \u003c/li\u003e\n\u003cli\u003eLeng S, Tan RS, Chai KTC, Wang C, Ghista D, Zhong L. The electronic stethoscope. \u003cem\u003eBioMed Eng OnLine\u003c/em\u003e. 2015;14(1):66. doi:10.1186/s12938-015-0056-y \u003c/li\u003e\n\u003cli\u003eLai LSW, Redington AN, Reinisch AJ, Unterberger MJ, Schriefl AJ. Computerized automatic diagnosis of innocent and pathologic murmurs in pediatrics: A pilot study: Computerized diagnosis of murmurs. \u003cem\u003eCongenital Heart Disease\u003c/em\u003e. 2016;11(5):386-395. doi:10.1111/chd.12328 \u003c/li\u003e\n\u003cli\u003eShekhar R, Vanama G, John T, Issac J, Arjoune Y, Doroshow RW. Automated identification of innocent Still\u0026apos;s murmur using a convolutional neural network. \u003cem\u003eFront Pediatr\u003c/em\u003e. 2022;10:923956. doi:10.3389/fped.2022.923956\u003c/li\u003e\n\u003cli\u003eLv J, Dong B, Lei H, et al. Artificial intelligence-assisted auscultation in detecting congenital heart disease. \u003cem\u003eEuropean Heart Journal \u0026ndash; Digital Health\u003c/em\u003e. 2021;2(1):119-124. doi:10.1093/ehjdh/ztaa017\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-pediatrics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejpe","sideBox":"Learn more about [European Journal of Pediatrics](https://www.springer.com/journal/431)","snPcode":"431","submissionUrl":"https://submission.nature.com/new-submission/431/3","title":"European Journal of Pediatrics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"artificial intelligence, congenital heart defect, electronic stethoscope, heart murmur, screening","lastPublishedDoi":"10.21203/rs.3.rs-4697876/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4697876/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003ePurpose\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOur aim was to investigate the ability of an artificial intelligence (AI)-based algorithm to differentiate innocent murmurs from pathologic ones.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn AI-based algorithm was developed using heart sound recordings collected from 1413 patients at the five university hospitals in Finland. The corresponding heart condition was verified using echocardiography. In the second phase of the study, patients referred to Helsinki New Children’s Hospital due to a heart murmur were prospectively assessed with the algorithm, and then the results were compared with echocardiography findings.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNinety-eight children were included in this prospective study. The algorithm classified 72 (73%) of the heart sounds as normal and 26 (27%) as abnormal. Echocardiography was normal in 63 (64%) children and abnormal in 35 (36%). The algorithm recognised abnormal heart sounds in 24 of 35 children with abnormal echocardiography and normal heart sounds with normal echocardiography in 61 of 63 children. When the murmur was audible, the sensitivity and specificity of the algorithm were 83% (24/29) (confidence interval (CI) 64–94%) and 97% (59/61) (CI 89–100%), respectively.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe algorithm was able to distinguish murmurs associated with structural cardiac anomalies from innocent murmurs with good sensitivity and specificity. The algorithm was unable to identify heart defects that did not cause a murmur. Further research is needed on the use of the algorithm in screening for heart murmurs in primary health care.\u003c/p\u003e","manuscriptTitle":"Automated Analysis of Heart Sound Signals in Screening for Structural Heart Disease in Children","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-08-09 20:54:19","doi":"10.21203/rs.3.rs-4697876/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2024-07-25T09:57:29+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-25T02:33:26+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2024-07-15T05:44:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"50074302823690219223219123753418592135","date":"2024-07-11T12:26:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"317322697809260561287787457585941408798","date":"2024-07-09T10:41:19+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2024-07-09T10:29:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2024-07-09T01:29:27+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-07-09T01:28:46+00:00","index":"","fulltext":""},{"type":"submitted","content":"European Journal of Pediatrics","date":"2024-07-06T18:00:36+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"european-journal-of-pediatrics","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ejpe","sideBox":"Learn more about [European Journal of Pediatrics](https://www.springer.com/journal/431)","snPcode":"431","submissionUrl":"https://submission.nature.com/new-submission/431/3","title":"European Journal of Pediatrics","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"197f285a-2d5b-43c8-a82e-abd62fc35d7a","owner":[],"postedDate":"August 9th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2024-09-23T16:05:59+00:00","versionOfRecord":{"articleIdentity":"rs-4697876","link":"https://doi.org/10.1007/s00431-024-05773-3","journal":{"identity":"european-journal-of-pediatrics","isVorOnly":false,"title":"European Journal of Pediatrics"},"publishedOn":"2024-09-21 15:58:01","publishedOnDateReadable":"September 21st, 2024"},"versionCreatedAt":"2024-08-09 20:54:19","video":"","vorDoi":"10.1007/s00431-024-05773-3","vorDoiUrl":"https://doi.org/10.1007/s00431-024-05773-3","workflowStages":[]},"version":"v1","identity":"rs-4697876","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4697876","identity":"rs-4697876","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}