Improving the performance of multi-parameter patient monitors using feature mapping and decision fusion

Multi-parameter patient monitor (MPM) uses vital signs, heart rate, blood pressure, oxygen saturation (SpO2) and respiration rate to identify the condition of patients. In this work, we use a support vector machine (SVM) backend classifier with four vital signs as its input and experimented using different kernels. It was observed that the SVM with a radial basis function kernel (SVM-RBF) outperforms the other kernels. Compared to non-linear SVMs, the linear SVM is computationally more efficient. Therefore, in this work we explore the use of feature mapping using locality constrained linear coding (LLC) to linearize the input features and thereby enhancing the performance of MPMs with a linear SVM (LLC-linSVM). To improve the performance further, we normalized LLC features by l2-norm (nLLC-linSVM). A performance improvement of 0.53% and 0.96% absolute for overall classification accuracy and specificity respectively was obtained over the baseline SVM-RBF system. However, a deterioration in the sensitivity was noted. To take advantage of both SVM-RBF and nLLC-linSVM, we finally fused the decision scores of both the systems. The fusion weights were estimated empirically using a dataset which is used neither for training nor testing. After decision fusion, we achieved a performance improvement of 0.90% absolute for classification accuracy, 0.24% absolute for sensitivity and 1.12% absolute for specificity compared to the baseline. All the systems were compared using receiver operating characteristics (ROC) and the results show that the performance of the fused system is better than the individual systems.