Figure 1.
An example of the processing pipeline for the DETECT toolbox using the default settings.
Step 1 creates a set of trials from training data using a window size of 0.5 s at data sampling frequency of 256 Hz. Step 2 builds and validates the model using the default settings. Step 3 labels test data using a slide width of 0.25 s. Step 4 applies an optional certainty policy based on a specified certainty threshold to relabel. Step 5 visualizes the results and produces an event log. Step 6 calculates the amount of agreement between the current labeling (events1) and another labeling (events2). Additional resources for optional parameters can be found in the DETECT Users Guide bundled together with the toolbox. HTML documentation for each function can be found at http://visual.cs.utsa.edu/detect/documentation/help/.
Figure 2.
The getLabels function displays a graphical user interface (GUI) that allows users to label different events in a continuous dataset.
The toolbar buttons at the top (arrow) allow the user to choose the event label. In this figure two events have been identified – a “None” event shown in yellow and a “Blink” event shown in blue.
Figure 3.
An example of the continuous detection for a window size of 0.5 s and slide width of 0.25 s.
In Step 1, the analysis window W1 is [0, 0.5) seconds, and the prediction for this data is made at [M1 − 0.5 S, M1+0.5 S) where M1 = the midpoint of W1 and S = the slide width. The blue color denotes the time interval of applicability of the prediction based on the first window, W1. In Step 2, the window is shifted by 0.25 s, so the analysis window W2 is [0.25, 0.75) seconds, and the prediction for the data is made at [M2 − 0.5 S, M2+0.5 S). The green color denotes the time interval of applicability of the prediction based on the first window, W2.
Figure 4.
A comparison of two labeled datasets using the fuzzy window approach for allowable timing errors for two slightly different window sizes using the compareLabels function.
There are two events, colored blue and orange. Blank spaces denote the absence of an event in the time window. The Decision row indicates the decision made by compareLabels. (B) The comparison with using a fuzzy window size of 100 ms. The fuzzy window extends regions of agreement by 100 ms on each end. The extension is shown in Green. The decision codes are: NullAgreement (NA), False Negative (FN), Agreement (Agree), False Positive (FP) and TypeError (TE).
Table 1.
DETECT Function List and Summary.
Table 2.
Summary of the comparison between different user labelings and automated labelings by DETECT for Dataset D1.
Figure 5.
Fraction of the data in different groups when using different threshold values in the thresholdPolicy certainty policy (Agreement and NullAgreement are combined).
Figure 6.
Example of disagreement between the labeling from User 1 (top graph) and DETECT (bottom graph) for dataset D2.
Table 3.
Summary of the comparison between user labelings and automated labelings by DETECT for Dataset D2.
Figure 7.
An example of a FalseNegative error in dataset D2 when using DETECT.
Increased periods of alpha spindling are found in the highlighted regions of the data. These regions are being misclassified as Eye and Blink activity.
Figure 8.
A segment of dataset D2 with regions labeled from a user (A) and the automated labeling from DETECT (B).
Table 4.
Summary of the comparison between user labelings and automated labelings by DETECT for Dataset D3.
Figure 9.
DETECT labeling of ECG data obtained from the PhysioNet online database.
Aqua regions denote predictions of normal heart waveforms while yellow regions denote predictions of PVC waveforms. Periods of no color shading denote the absence of a heart waveform. The event codes N and V denote a Normal and PVC waveform based on expert labeling, respectively.