Feature Extraction with Wavelets for Plethysmography Signal Classification

The purpose of this research work is to recognize blood pressure conditions from the analysis of photoplethysmography signals of patients under anesthesia. A multiresolution analysis was applied to the signals using the discrete wavelet transform to obtain the detail and approximation coefficients that provide information on each cardiac cycle. Amplitude - time related characteristics to the morphology of the signal were also obtained. A set of signals was selected for the different blood pressure conditions, extracting their characteristics. Thus, a dataset was created with the characteristics as inputs and the blood pressure conditions as targets (hypotension, normotension, hypertension). Using this data, fully connected multilayer neural networks were trained, evaluating the performance of different architectures. The network with the best results was selected by means of statistical quality measures. The architecture selected was 80.5% accurate regarding data from new patients.