Diagnosis of coronary artery disease in nuclear medicine using generalized linear model and singular value decomposition

Objectives Our group has previously introduced a generalized linear model (GLM) to construct a logistic equation for mapping the SPECT Bulls-eye into the angiogram space for predicting the stenosis levels (SL) of the main coronary arteries using the normalized activities measured in three different anatomical areas (LAD, RCA, and LCX) specified in a feature map constructed from rest and stress maps of bulls-eyes. However, the model order selection in the GLM strongly affects the prediction accuracy. In this study, a Singular Value Decomposition (SVD) technique was recruited to estimate the actual rank of the information matrix to select the best possible order of the GLM for increasing the accuracy of the SL prediction.

Methods Thirty two patients (Mean=58.9±12.3 years old, Mean LVEF%=53.6%±15.7%, 16 for training and 16 for testing) who had both coronary Angiography and myocardial SPECT were studied. All patients had gone through the Angiography procedure no later than three months from the time of their myocardial perfusion (SPECT99m-Tc-MIBI). A 2D feature map (Feature map=[Rest-Stress]/Rest) was constructed from the rest and stress Bulls-eye maps. All GLM coefficients were estimated from the singularity matrix generated by the SVD technique. The predictor was also evaluated by the angiography results and the mean activities measured in three different zones in 16 patients.

Results Results imply that the GLM-SVD improves the accuracy of the predictor compared to GLM and also the GLM-SVD reduces the errors type-I and type-II which are associated with the False-Positive and False-Negative predictions respectively. The GLM-SVD detects the SL of the RCA (r=0.84, p<0.0001), LAD (r=0.82, p<0.0001), and LCX (r=0.74, p<0.0001) using rest and stress images.

Conclusions SVD technique was employed in the GLM to improve the prediction accuracy of the stenosis level for the three main coronary arteries. Results imply that the proposed model increases the prediction accuracies(RCA:5%-LAD:6%-LCX:3%)