Support Vector Machine (SVM) Learning Tool for the Interpretation of Cardiopulmonary Exercise Test Results

Introduction: Cardiopulmonary exercise testing (CPET) is an important modality for the evaluation and management of patients with a diverse array of medical problems and symptoms. However, interpreting the results is subjective and require significant expertise and is often difficult as well as time consuming.

Aim: To develop a Computer-Aided Disease Classification System (CADCx) and an Exercise Limitation Classification Module (CAELCx) to help physicians and exercise physiologists in objectively evaluating CPET results.

Methods: CPET data results (n=239) from the Institutes of Lung and Heart Diseases, Sheba Medical Center, were used. Of the officially diagnosed CPET files, 150 were employed in the Support Vector Machine (SVM) learning stage: 50 healthy participants, 50 patients with moderate-to-severe chronic heart failure (CHF), and 50 patients with moderate-to-severe chronic obstructive pulmonary disease (COPD). The remaining 89 files from similar populations were used for SVM module validation. Performance of the SVM in the form of classification was compared with prior expert diagnosis using distribution analysis.

Results: The CADCx successfully classified healthy, CHF, and COPD patients. The positive predictive values (PPV) were 94%, 92%, and 79% for normal, CHF, and COPD patients, respectively. The negative predictive values (NPV) were correspondingly 96%, 97%, and 91%. Consequently, sensitivity, specificity and overall precision of the proposed interpretive module were 89%, 94%, and 89%, respectively. In addition, the CAELCx, also using the SVM learning tool, was capable of accurately classifying type and level of exercise limitations with 95% sensitivity, 97% specificity, and 95% precision.

Conclusions: The new CADCx and CAELCx modules developed herein for CPET interpretations are highly sensitive and accurate. Their use may reduce complexity, lack of objectivity, and time spending of the CPET interpretation process in clinical settings. Additional populations and increased number of patients in each category is needed to make this tool clinically relevant.