Trend analysis is an efficient tool for process monitoring and diagnosis. However, the performances of a trend-based diagnosis system depend on the reliability of the trends extracted from the signals. One challenge in trend analysis is to design algorithms able to adapt themselves to the varying conditions of background noise and artefacts occurring non-deterministically on a same signal. Moreover, while long term trends such as decreasing/increasing have been extensively studied other subtle changes such as slow drifts and step-like transients have received little attention. In this paper, an adaptive on-line trend-extraction method is presented. It extends a former algorithm based on a linear segmentation to filter the signal and extract trends. In this version, the tuning parameters are not set to a fixed value for a given signal but can self-adapt on-line according to an estimation of the noise variance. An increasing or decreasing trend is detected if the variations on the signal are significantly higher than the level of the background noise. An initialisation phase is proposed to automatically set the initial values of the parameters, making the algorithm a self-tuned algorithm with minimal user intervention. The method was evaluated on a set of simulated data with various levels of background noise. It was also applied on real physiological data recorded from babies hospitalised in a Neonate Intensive Care Unit. It showed improved performances compared to the non adaptive algorithm, whatever the level of noise corrupting the data. (C) 2012 Elsevier Ltd. All rights reserved.