Algorithm for predicting valvular heart disease from heart sounds in an unselected cohort

Figure 1. Data flowchart. Flowchart overview of how the final study dataset was formed and how many samples were excluded due to noisy or incomplete data. The values at the end show the number of usable (absence of considerable noise) recordings annotated with murmur grades and grades of AR, MR, AS, and MS for each auscultation position. VHD, valvular heart disease; AR, aortic regurgitation; MR, mitral regurgitation; AS, aortic stenosis; MS, mitral stenosis.

Figure 2. Murmur detection ROC curves. Receiver operator characteristics curves illustrate the algorithm's ability to predict murmur grades ≥1 (blue) and ≥2 (red). Recordings from all four auscultation positions were used in producing the figure.

Figure 3. ROC curves for prediction of clinically relevant cases. Each receiver operating characteristics curve corresponds to prediction of a different definition of clinically relevant VHD, with definitions shown in the legend. For the yellow curve, asymptomatic cases of regurgitation were excluded from the positive class. The circles show sensitivities and specificities for murmur grade decision thresholds of 0.5, 1.0, and 2.0. The position with the highest algorithm-predicted murmur grade was used as the predictor. VHD, valvular heart disease; AR, aortic regurgitation; MR, mitral regurgitation; AS, aortic stenosis; MS, mitral stenosis.

Figure 4. Joint screening for clinically relevant cases. The diagram that shows numbers and proportions of clinically significant cases detected in the process of screening jointly for all clinical cases using the murmur detection algorithm and the prevalence of false positive predictions. The highest algorithm-predicted murmur grade was used as the predictor of clinical cases. A positive prediction was counted as a true positive if the participant had a significant grade for at least one VHD. Included in the pathological VHD categories are grade ≥ 1 stenosis and AS ≥ 1 regurgitation. The red boxes represent subgroups of participants who had or used to have symptoms (shortness of breath while walking calmly on a flat surface or resting). VHD, valvular heart disease; AR, aortic regurgitation; MR, mitral regurgitation; AS, aortic stenosis; MS, mitral stenosis.

Figure 5. Multiposition model vs. single-position AUC for AVPGmean prediction. Each panel shows AUC for the prediction of cases where AVPGmean > u for thresholds u ranging from 7 to 30 mm Hg for the single-position (dashed line) and multiposition (solid line) model. Each single-position model uses only data from the index position to make predictions, whereas the multiposition model uses data from all four positions. Significant differences are marked with *. AUC, area under the curve; AVPGmean, aortic valve mean pressure gradient.

Figure 6. AUC as a function of the aortic stenosis cutoff threshold. AUC for the prediction of AVPGmean > u across a range of thresholds u using the multiposition model on the subset with noise in up to three out of four auscultation positions. Confidence intervals were obtained using bootstrapping. Recordings annotated as noisy in all four positions were excluded from the analysis. AVPGmean, aortic valve mean pressure gradient; AUC, area under the curve.