Computer-aided auscultation aimed at detecting and characterizing heart murmurs is called computer-aided heart auscultation (also known as automatic heart sound analysis).
Motivation Auscultation of the heart using a stethoscope is the standard examination method worldwide to screen for heart defects by identifying murmurs. It requires that an examining
physician have acute hearing and extensive experience. An accurate diagnosis remains challenging for various reasons including noise, high heart rates, and the ability to distinguish innocent from pathological murmurs. Properly performed, the auscultatory examination of the heart is commonly regarded as an inexpensive, widely available tool in the detection and management of heart disease. The auscultation skills of physicians, however, have been reported to be declining. This leads to missed disease diagnoses and/or excessive costs for unnecessary and expensive diagnostic testing. A study suggests that more than one third of previously undiagnosed congenital heart defects in newborns are missed by their 6-week examination. More than 60% of referrals to medical specialists for costly echocardiography are due to a misdiagnosis of an innocent murmur. • Class I: pathological murmur • Class III: innocent murmur or no murmur More sophisticated CAA systems provide additional descriptive murmur information like murmur timing, grading, or the ability to identify the positions of the S1/S2 heart sounds.
Heart sound analysis The detection of heart murmurs in CAA systems is based on the analysis of digitally recorded heart sounds. Most approaches use the following four stages: •
Heart rate detection: In the first stage, the heart rate is determined based on the audio signal of the heart. It is a crucial step for the following stages and high accuracy is required. Automated heart rate determination based on acoustic recordings is challenging because the heart rate can range from 40-200bpm, noise and murmurs can camouflage the peaks of the heart sounds (S1 and S2), and irregular heartbeats can disturb the quasi-periodic nature of the heartbeat. •
Heart sound segmentation: After the heart rate has been detected, the two main phases of the heartbeat (
systole and
diastole) are identified. This differentiation is important since most murmurs occur in specific phases during the heartbeat. External noise from the environment or internal noise from the patient (e.g. breathing) make heart sound segmentation challenging. • Feature extraction: Having identified the phases of the heartbeat, information (
features) from the heart sound is extracted that enters a further classification stage. Features can range from simple energy-based approaches to higher-order multi-dimensional quantities. • Feature classification: During classification, the features extracted in the previous stage are used to classify the signal and assess the presence and type of a murmur. The main challenge is to differentiate no-murmur recordings from low-grade innocent murmurs, and innocent murmurs from pathological murmurs. Usually machine-learning approaches are applied to construct a classifier based on training data.
Clinical evidence of CAA systems The most common types of performance measures for CAA systems are based on two approaches: retrospective (non-blinded) studies using existing data and prospective blinded clinical studies on new patients. In retrospective CAA studies, a classifier is trained with
machine learning algorithms using existing data. The performance of the classifier is then assessed using the same data. Different approaches are used to do this (e.g.,
k-Fold cross-validation,
leave-one-out cross-validation). The main shortcoming of judging the quality (sensitivity, specificity) of a CAA system based on retrospective performance data alone comes from the risk that the approaches used can overestimate the true performance of a given system. Using the same data for training and validation can itself lead to significant
overfitting of the validation set, because most classifiers can be designed to analyse known data very well, but might not be general enough to correctly classify unknown data; i.e. the results look much better than they would if tested on new, unseen patients. “The true performance of a selected network (CAA system) should be confirmed by measuring its performance on a third independent set of data called a test set”. In summary, the reliability of retrospective, non-blinded studies are usually considered to be much lower than that of prospective clinical studies because they are prone to
selection bias and retrospective bias. Published examples include Pretorius et al. Prospective clinical studies, on the other hand, are better suited to assess the true performance of a CAA system (provided that the study is blinded and well controlled). In a prospective clinical study to evaluate the performance of a CAA system, the output of the CAA system is compared to the
gold standard diagnoses. In the case of heart murmurs, a suitable gold standard diagnosis would be auscultation-based expert physician diagnosis, stratified by an
echocardiogram-based diagnosis. Published examples include Lai et al. ==See also==