Data processing is usually based on uniformly sampled data in time. This sampling scheme is often unnecessary for non-stationary signals because samples are also taken in inactive regions. In case of embedded system, this useless samples significantly increase the power consumption. One solution to avoid this useless power consumption is the level crossing sampling scheme (LCSS). This method is not taking samples according to the time but according to amplitudes, therefore data are not sampling during inactive region. To optimize LCSS, thresholds need to be appropriately placed along the amplitude range of the input signal. In this paper, we present an algorithm to determine automatically these thresholds. This algorithm has been tested on experimental biomedical signals. We demonstrate that we could ensure the information and in the same time drastically decrease the number of processed samples.