Fusion of Bottleneck Features Derived from CNNs to Enhance the Performance of Multi-Parameter Patient Monitors

Multi-parameter patient monitor (MPM) measures multiple physiological parameters of a patient to generate alarms when the patient's condition is deteriorating. Present monitors function based on early warning score(EWS) system which results in low alarm accuracy. Hence, there is a need to improve the accuracy of the alarms. In this work, how bottleneck features can be used for improving the performance of multi-parameter patient monitors(MPMs) is explored. Bottleneck features are derived from convolutional neural networks(CNN) as they are good in extracting the local patterns and making decisions at a higher level. The intrinsic relationship between the parameters can be captured using kernels of different sizes. Subsequently, the bottleneck features extracted were fused and given as an input to a model. The fusion model gives a performance of 98.33% absolute sensitivity, 99.49% absolute specificity and 99.22% absolute overall accuracy; which is an improvement of 0.13%, 2.08% and 1.56% in sensitivity, specificity and overall accuracy compared to the baseline model.