Comparison of EDHMM classification and threshold-crossing classification.

<p><b>A</b>) The sliding-window density estimate of the LF-amplitude is shown for an example LFP recording. Log probability density is depicted as the color map, with the violet and green traces showing the time-varying state-conditional means. The regions indicated by the vertical black lines show the locations within the data where the example traces in panels C and D were taken. <b>B</b>) The overall amplitude distribution, along with the UP and DOWN state component distributions fit using a (static) Gaussian mixture model. The two values of the fixed threshold used for the “threshold-crossing” algorithm, chosen using the non-parametric (Np) and static mixture model (SMM) approaches, are shown by the finely and coarsely dashed black horizontal lines respectively. <b>C</b>) An example LFP trace (blue) from the region indicated by the black lines. The Viterbi state sequence from the EDHMM is shown in brown. For comparison, the state sequence classified using the nonparametric (Np) threshold-crossing method is shown in light blue. In this example, the threshold value appears to be too high, and many of the UP states are erroneously split by falsely detected DOWN states. <b>D</b>) Another example LFP trace from later in the recording. In this case, the state sequence classified using the SMM threshold-crossing method is shown in light blue for comparison. The SMM threshold appears to be too low in this example, producing a number of falsely detected UP states while the LFP clearly remains in the DOWN state.