The robustness of Random Forest and Support Vector Machine Algorithms to a Faulty Heart Sound Segmentation

Area of Science:

• Cardiology
• Biomedical Engineering
• Signal Processing

Background:

• Cardiac auscultation is a vital, cost-effective screening tool for heart diseases.
• Automatic recommendation systems enhance the accuracy and accessibility of cardiac screening.
• Accurate segmentation of heart sounds is essential for reliable diagnosis.

Purpose of the Study:

• To analyze the impact of common heart sound segmentation errors on diagnostic accuracy.
• To quantify the performance degradation caused by incorrect heart rate estimation versus state sequence misclassification.
• To evaluate the sensitivity of machine learning algorithms to specific segmentation errors.

Main Methods:

• Simulated common segmentation errors: inaccurate heart rate estimation and swapped S1/S2 or Systolic/Diastolic states.
• Assessed the performance impact on heart sound classifiers using Support Vector Machines (SVMs) and Random Forests (RFs).
• Compared the performance drop resulting from different types of segmentation errors.

Main Results:

• Support Vector Machines (SVMs) showed a 6% performance drop with incorrect heart rate estimation.
• Random Forests (RFs) experienced an 8% performance decrease due to inaccurate heart rate estimation.
• Both algorithms were less affected by state sequence swaps, with SVMs dropping 1.9% and RFs 4.6%.

Conclusions:

• Heart rate estimation accuracy is a critical factor in the robustness of heart sound analysis systems.
• Machine learning models like SVMs and RFs are more sensitive to heart rate errors than to state sequence mislabeling.
• Improving heart rate estimation in segmentation algorithms is key to enhancing the reliability of automated cardiac auscultation tools.