Fusing vital signs with correlation features using Fisher Vector Encoding and a linear SVM achieved an overall classification accuracy of 98.46%, a 1.32% absolute improvement over baseline.
Does fusing vital signs with correlation features using Fisher Vector Encoding improve the classification performance of a multi-parameter patient monitor?
Fusing vital signs with correlation features using Fisher Vector Encoding significantly improves the classification accuracy of multi-parameter patient monitors.
Effect estimate: 1.32% absolute improvement
Multi-parameter patient monitors(MPM) are widely used for monitoring critical physiological vital signs such as heart rate, oxygen saturation, blood pressure and respiration rate. We developed a baseline MPM using Support Vector Machine(SVM) backend classifier using Radial basis Function (RBF) kernel with four vital signs as its input. A critical study in healthy people revealed the fact that there always exists an intrinsic relationship between the vital parameters. In order to capture this intrinsic association between the four parameters, the correlation features(CFs) were used which are calculated by taking the geometric mean of two pairs of vital signs. The RBF-SVM based MPM with six CFs as input obtained a sensitivity of 97.76%, specificity of 98.71%, and overall classification accuracy of 94.72%. On combining the four vital signs with six CFs, the performance was found to be deteriorating.The input feature vectors are expressed as patient independent basis vectors using Fisher Vector Encoding (FVE). In this work we focus on fusing the vital signs with CFs after encoding with an objective of enhancing the performance of the MPM significantly. We obtained an overall classification accuracy of 98.46%, sensitivity of 98.59%, and specificity of 98.01% with FVE-linear SVM which achieved a performance improvement of 1.32% absolute for classification accuracy, 3.19% absolute for sensitivity and 0.78% absolute for specificity with respect to the baseline system with four vital signs. MIMIC II database was used in this work.
Bose et al. (Fri,) conducted a other in Critical physiological vital signs monitoring. Fusing vital signs with correlation features using Fisher Vector Encoding and linear SVM vs. Baseline system with four vital signs was evaluated on Overall classification accuracy (1.32% absolute improvement). Fusing vital signs with correlation features using Fisher Vector Encoding and a linear SVM achieved an overall classification accuracy of 98.46%, a 1.32% absolute improvement over baseline.