Detecting dynamical changes in vital signs using switching Kalman filter

Vital signs contain valuable information about the health condition of patients during their stay in the ward, when deterioration process begins. The use of methods to predict and detect regime changes such as switching models can help to understand how vital sign dynamics are altered in health and disease. However, time series of vital signs are remarkably non-stationary in these scenarios. The objective of this study is to quantify the potential bias of the switching models in the presence of non-stationary time series, when the inputs are spectral, symbolic and entropy indices. To distinguish stationary periods from non-stationary, a stationarity test was used to verify the stability of the mean and variance over short periods. Then, we compared the results from a switching Kalman filter (SKF) model trained using only indices obtained over stationary periods, with a model trained using indices obtained solely over non-stationary periods. It was observed that the indices measured over stationary and non-stationary periods were significantly different. The results were highly dependent of what indices were used as input, being the multiscale entropy (MSE) the most efficient approach, achieving an average correlation coefficients of 38%.