Artifact removal for intracranial pressure monitoring signals: A robust solution with signal decomposition

Abstract

Intracranial Pressure (ICP) monitoring signal collected in Neuro Intensive Care Units often contains large amount of artifacts. The artifacts not only directly lead to false alarms in automatic Intracranial Hypertension (IH) alert systems, and they also severely contaminate the characteristics of the underlying signal, which makes accurate forecasting of impending IH impossible. Therefore, in this paper, we propose a novel solution to effectively remove artifacts from ICP monitoring signals. The proposed method effectively detects artifacts by decomposing the ICP monitoring signal with Empirical Mode Decomposition (EMD) method. An iterative filtering method is also proposed to extract artifacts from the decomposed components of ICP signals. The proposed filter is robust. That is, the parameters of the iterative filter are estimated with robust statistics, which ensures the performance of the proposed filter will not be unduly affected by artifacts. The detected artifacts are then imputed based on the Auto-Regressive Moving Average (ARMA) model to preserve the original characteristics of the ICP signal. The effectiveness of the proposed artifact removal method is experimentally justified based on the ICP monitoring signals of 59 patients.

An online approach for intracranial pressure forecasting based on signal decomposition and robust statistics

“…Simulations results show good performance, i.e. a GPER below 5 % in situations where the approach by [2] fails. Visual inspection of clinically measured ICP signals indicates stable performance, even in seriously contaminated data sets.…”

“…This high computational cost is inapplicable for online applications. We compare our work to [2], which also forecasts ICP signals at a low sampling frequency (0.1 Hz). Our work has its roots in signal decomposition [5,6,7] and robust statistics for dependent data [7,8,9,10,11].…”

“…As stated in [2], the motion artifacts in an ICP measurement sequence are significantly observable in the IMFs. This potentials an outlier detection in the EMD domain by using the model described in Eqs.…”

“…Average results based on 100 Monte Carlo runs, are shown in Tables 4.1 and 4.1. Here, NN stands for the neural networks based method described in [2], KF ARIMA represents our method using a Kalman filter forecasting (assuming perfect artifact removal, see Section 3.2), and ROB ARIMA denotes our method using robust ARIMA forecasting. Both our methods use the suggested robust updating, which in general lowered the GPER (best GPER without: ROB ARIMA 9.20% for random walk, ROB ARIMA 8.54% for velocity sensor Table 2.…”

“…The accurate prediction of an ICP signal requires dealing with a number of difficulties: First, the non-stationarity of ICP signals is too high to be cancelled by simple methods like timedifferencing or short-time analysis. Second, the inevitable artifacts, which can be caused by motion of the patients or equipment errors, contaminate the signal seriously [2]. A further challenge arises from the fact that, up to now, no convincing statistical model for ICP measurements has been established.…”