Samuel Schmidt

An open access database for the evaluation of heart sound algorithms

In the past few decades, analysis of heart sound signals (i.e. the phonocardiogram or PCG), especially for automated heart sound segmentation and classification, has been widely studied and has been reported to have the potential value to detect pathology accurately in clinical applications. However, comparative analyses of algorithms in the literature have been hindered by the lack of high-quality, rigorously validated, and standardized open databases of heart sound recordings. This paper describes a public heart sound database, assembled for an international competition, the PhysioNet/Computing in Cardiology (CinC) Challenge 2016. The archive comprises nine different heart sound databases sourced from multiple research groups around the world. It includes 2435 heart sound recordings in total collected from 1297 healthy subjects and patients with a variety of conditions, including heart valve disease and coronary artery disease. The recordings were collected from a variety of clinical or nonclinical (such as in-home visits) environments and equipment. The length of recording varied from several seconds to several minutes. This article reports detailed information about the subjects/patients including demographics (number, age, gender), recordings (number, location, state and time length), associated synchronously recorded signals, sampling frequency and sensor type used. We also provide a brief summary of the commonly used heart sound segmentation and classification methods, including open source code provided concurrently for the Challenge. A description of the PhysioNet/CinC Challenge 2016, including the main aims, the training and test sets, the hand corrected annotations for different heart sound states, the scoring mechanism, and associated open source code are provided. In addition, several potential benefits from the public heart sound database are discussed.

Segmentation of heart sound recordings by a duration-dependent hidden Markov model

Digital stethoscopes offer new opportunities for computerized analysis of heart sounds. Segmentation of heart sound recordings into periods related to the first and second heart sound (S1 and S2) is fundamental in the analysis process. However, segmentation of heart sounds recorded with handheld stethoscopes in clinical environments is often complicated by background noise. A duration-dependent hidden Markov model (DHMM) is proposed for robust segmentation of heart sounds. The DHMM identifies the most likely sequence of physiological heart sounds, based on duration of the events, the amplitude of the signal envelope and a predefined model structure. The DHMM model was developed and tested with heart sounds recorded bedside with a commercially available handheld stethoscope from a population of patients referred for coronary arterioangiography. The DHMM identified 890 S1 and S2 sounds out of 901 which corresponds to 98.8% (CI: 97.8-99.3%) sensitivity in 73 test patients and 13 misplaced sounds out of 903 identified sounds which corresponds to 98.6% (CI: 97.6-99.1%) positive predictivity. These results indicate that the DHMM is an appropriate model of the heart cycle and suitable for segmentation of clinically recorded heart sounds.

Segmentation of heart sound recordings from an electronic stethoscope by a duration dependent Hidden-Markov Model

Digital stethoscopes offer new opportunities for computerized analysis of heart sounds. Segmentation of hearts sounds is a fundamental step in the analyzing process. However segmentation of heart sounds recorded with handheld stethoscopes in clinical environments is often complicated by recording and background noise. A duration-dependent hidden Markov model (DHMM) is proposed for robust segmentation of heart sounds. The DHMM model was developed and tested with heart sounds recorded at bedside with a commercially available handheld stethoscope. In a population of 60 patients, the DHMM identified 739 S1 and S2 sounds out of 744 which corresponded to a 99.3% sensitivity. There were seven incorrectly classified sounds which corresponded to a 99.1% positive predictive value. Our results suggest that DHMM could be a suitable method for segmentation of clinically recorded heart sounds.

Detection of coronary artery disease with an electronic stethoscope

A noninvasive method for detection of coronary artery disease (CAD) with an electronic stethoscope is proposed. Heart sounds recorded in clinical settings are often contaminated with background noise and noise caused by friction between the skin and the stethoscope. A method was developed to reduce the influence of the noise artifacts. The diastolic parts of the heart sounds were divided into multiple sub-segments, where noisy sub-segments were indentified as sub-segments with a low degree of stationarity or with a high energy level. The sub-segments not identified as noisy were analyzed with an autoregressive (AR) model, where the pole-magnitude of the 1st pole was used as a discriminating parameter. A test on 50 subjects showed that removal of the noisy sub-segments before analyses improved the diagnostic performance of the AR-model considerably, thereby reducing the influence of noise related to the use of a handhold stethoscope.

Acoustic Features for the Identification of Coronary Artery Disease

Goal: Earlier studies have documented that coronary artery disease (CAD) produces weak murmurs, which might be detected through analysis of heart sounds. An electronic stethoscope with a digital signal processing unit could be a low cost and easily applied method for diagnosis of CAD. The current study is a search for heart sound features which might identify CAD. Methods: Nine different types of features from five overlapping frequency bands were obtained and analyzed using 435 recordings from 133 subjects. Results: New features describing an increase in low-frequency power in CAD patients were identified. The features of the different types were relatively strongly correlated. Using a quadratic discriminant function, multiple features were combined into a CAD-score. The area under the receiving operating characteristic for the CAD score was 0.73 (95% CI: 0.69-0.78). Conclusion: The result confirms that there is a potential in heart sounds for the diagnosis of CAD, but that further improvements are necessary to gain clinical relevance.

Stenosis Detection Algorithm for Screening of Arteriovenous Fistulae

The aim of the study was to develop an algorithm that can detect stenosis formation in arteriovenous fistulae based on audio recordings. 34 patients with a mature arteriovenous fistula were examined with use of an electronic stethoscope and subsequently by ultrasound. 27 patients had a patent fistula, while the other group consisted of 5 patients with stenosis and 2 with artificial narrowing of the fistula. Feature extraction was carried out using wavelet packet decomposition at depth 4. For each recording the scale energies SE i and the percentage of scale energy versus total energy SEp i , were calculated. The two most discriminative features with low correlation were found to be SE8 and SEp8. These features were evaluated using leave-one-out cross-validation with a quadratic discriminant function. Cross-validation using SE8 and SEp8 yielded a sensitivity of 100% and a specificity of 94%. The algorithm developed using the features obtained by wavelet analysis is reliable for detecting stenosis in a vein segment of an arteriovenous fistula. Based on these results, the prospects of developing an accurate, low-cost screening method for patients undergoing hemodialysis, are promising.

Noise and the detection of coronary artery disease with an electronic stethoscope

Recent studies demonstrated that diastolic heart sounds, recorded with an electronic stethoscope, contain markers of coronary artery disease (CAD). A difficult is that the CAD-related sound is very weak and recordings are often contaminated by noise. The current study analyses the noise contamination of 633 stethoscope recordings from a clinical environment. Respiration noise, ambient noise, recording noise and abdominal noise were identified in the recordings and were classified according to duration and intensity. To monitor how noise influences the classification performance AR-pole magnitudes were extracted from both the 25–250 Hz frequency band and the 250–1000 Hz frequency band. The classification performance was quantified by the Area Under the receiver operating Characteristic (AUC). Ambient noise was present in 39.9% of the recordings and was the most common noise source. Abdominal noise was the least common noise source, present in 10.8% of the recordings. The best pole, with respect to detection of CAD, extracted from the 250–1000 Hz frequency band was sensitive to noise, since the AUC dropped from 0.70 in to 0.57 when noisy recordings were included. Contrary the best pole from the 25–250 Hz frequency band was relatively robust against noise, since the AUC dropped from 0.73 to only 0.70 when noisy recordings were included. The study demonstrated that noise contamination is a frequent problem and that features from lower frequency bands are more robust against noise than features from higher frequency bands.

No evidence of nonlinear or chaotic behavior of cardiovascular murmurs

Abstract The current work examines cardiovascular murmurs for evidence of nonlinear and chaotic characteristics. In recent studies, methods suited for analysis of nonlinear and chaotic systems have been suggested for quantification of cardiovascular murmurs, but neither nonlinear nor chaotic behavior in cardiovascular murmurs have been confirmed. We obtained seven digital recordings from subjects with known carotid artery stenosis and audible carotid murmurs with a digital stethoscope. Measures as time reversibility, the power of third order cumulants, prediction errors from a nonlinear model, sample entropy, the correlation dimension and BDS statistics were calculated and compared to measures from surrogates generated by stationary Gaussian linear processes. Additionally nonlinear measures appropriateness for murmur quantification was evaluated by comparing sample entropy between periods with and without audible murmurs. Except from the power of third order cumulants from a single recording none of the measures from the seven recordings differed significantly from the surrogates with known linear properties. The sample entropy values increased significantly in periods with audible murmurs compared to periods without. Our study confirms that a nonlinear measure as sample entropy is suited for quantification of cardiovascular murmurs, but there is no evidence of either nonlinear or chaotic behavior in the signals.

Danish study of Non-Invasive testing in Coronary Artery Disease (Dan-NICAD): study protocol for a randomised controlled trial

BackgroundCoronary computed tomography angiography (CCTA) is an established method for ruling out coronary artery disease (CAD). Most patients referred for CCTA do not have CAD and only approximately 20–30 % of patients are subsequently referred to further testing by invasive coronary angiography (ICA) or non-invasive perfusion evaluation due to suspected obstructive CAD. In cases with severe calcifications, a discrepancy between CCTA and ICA often occurs, leading to the well-described, low-diagnostic specificity of CCTA. As ICA is cost consuming and involves a risk of complications, an optimized algorithm would be valuable and could decrease the number of ICAs that do not lead to revascularization.The primary objective of the Dan-NICAD study is to determine the diagnostic accuracy of cardiac magnetic resonance imaging (CMRI) and myocardial perfusion scintigraphy (MPS) as secondary tests after a primary CCTA where CAD could not be ruled out. The secondary objective includes an evaluation of the diagnostic precision of an acoustic technology that analyses the sound of coronary blood flow. It may potentially provide better stratification prior to CCTA than clinical risk stratification scores alone.Methods/designDan-NICAD is a multi-centre, randomised, cross-sectional trial, which will include approximately 2,000 patients without known CAD, who were referred to CCTA due to a history of symptoms suggestive of CAD and a low-risk to intermediate-risk profile, as evaluated by a cardiologist. Patient interview, sound recordings, and blood samples are obtained in connection with the CCTA. All patients with suspected obstructive CAD by CCTA are randomised to either stress CMRI or stress MPS, followed by ICA with fractional flow reserve (FFR) measurements. Obstructive CAD is defined as an FFR below 0.80 or as high-grade stenosis (>90 % diameter stenosis) by visual assessment.Diagnostic performance is evaluated as sensitivity, specificity, predictive values, likelihood ratios, and C statistics. Enrolment commenced in September 2014 and is expected to be complete in May 2016.DiscussionDan-NICAD is designed to assess whether a secondary perfusion examination after CCTA could safely reduce the number of ICAs where revascularization is not required. The results are expected to add knowledge about the optimal algorithm for diagnosing CAD.Trial registrationClinicaltrials.gov identifier, NCT02264717. Registered on 26 September 2014.

Semi-automatic vessel tracking and segmentation using epicardial ultrasound in bypass surgery

The purpose of intraoperative quality assessment of coronary artery bypass graft surgery is to confirm graft patency and disclose technical errors to reduce cardiac mortality, morbidity and improve clinical outcome for the patient. Epicardial ultrasound has been suggested as an alternative approach for quality assessment of anastomoses. To make a quantitative assessment of the anastomotic quality using ultrasound images, the vessel border has to be delineated to estimate the area of the vessel lumen. A tracking and segmentation algorithm was developed consisting of an active contour modeling approach and quality control of the segmentations. Evaluation of the tracking algorithm showed 91.96% of the segmentations were segmented correct with a mean error in height and width at 5.65% and 11.50% respectively.