Interactive event recognition and analysis

DataView provides quite sophisticated facilities for recognising and analysing waveform events such as nerve spikes. These operate in a highly interactive manner, which makes it easy for the user to fine-tune and edit event recognition, while seeing the immediate changes in both the analysis and the data file.

Recognition

DataView has two primary modes of automatic event recognition: template matching and threshold crossing. Template matching has three sub-modes - using a fixed template, using an adaptive template, or using an optimally scaled template.

• Fixed Template matching
  A user-defined segment of waveform is captured as a template, and then the entire record (or a selected portion of it) is scanned for waveforms that match the template within a user-specified error range. Recognised events can be averaged and used as a new template to refine recognition criteria.

  An extracellular recording showing several different active units. Waveform events (red line; T0) and their associated errors were recognised by template match on the 2nd spike (red arrow; error 0). The allowed error was set to 40. A close match to the template (purple arrow; error 7) is probably the same unit. Two smaller units are recognised (green arrows; error > 30), while a third much smaller unit has an error > 40 and is not recognised as an event (open arrow). By reducing the allowed error, a single specific unit could easily have been be identified. [The coloured arrows in the figure have been added using a bitmap editor.]

• Adaptive Template matching
  Adaptive template matching is like fixed template matching, except that the template adapts itself to systematic changes in the waveform such as a gradual decrease in overall amplitude. The adaptive template works by replacing the original template with a running average of the waveforms of recently-recognised events. The number of events constituting the average is determined by the user (typically 8), and as each new event is recognised, the earliest event in the average drops out to be replaced by the most recent. If an average of 1 is selected, then the template is successively replaced by each identified event, so that each found event constitutes the template for the next event search.

• Optimally Scaled Template matching
  Optimally scaled template matching is used when the intention is to find data elements of a particular shape, irrespective of their amplitude or dc-offset. A user-defined segment of data showing a waveform of the desired shape is first captured as a template. The record is then scanned on a point-by-point basis, and at each point the normalised template is scaled and offset to produce the best fit to the record. If the goodness-of-fit of this scaled template exceeds the acceptance criterion, it is counted as a match. The goodness-of-fit is defined as the scale factor divided by the standard error between the scaled template and the actual data.

The algorithm used for this procedure is that described by Clements & Bekker (1997).

Clements, J. D. & Bekkers, J.M. (1997) Detection of spontaneous synaptic events with an optimally scaled template. Biophysics J. 73, 220-229.

Optimally scaled template matching used to detect a series of EPSPs (brown line) of varying amplitudes. Waveform events (red line; E0) and their associated goodness-of-fit are shown at the top.The acceptance criterion was set to 2. The lower trace (grey line) shows the goodness-of-fit metric corresponding to each point in the data waveform.

• Threshold crossing
  Onset and offset voltage thresholds are defined (either numerically or visually by cursors), and events are recognised as waveform segments that lie on one side or the other of the thresholds. Minimum on- and off-times can be specified, so that brief waveforms "glitches" are not accepted as genuine events. Limit cursors can be specified to exclude artefacts.

  The data display (below left) shows events detected by crossing the threshold set by the horizontal cursor. A minimum on-time of 1 ms and off-time of 5 ms ensures that spike bursts, rather than individual spikes, are recognised as events, . A scatter graph of event time against frequency (below right)) reveals several below-trend events. One of these was clicked with the mouse, causing that particular event (number 272) to be centered and highlighted in the data display. This showed that the low frequency was caused by an unusually brief spike burst, which did not exceed the minimum on-time and therefore was not recognised. To correct this, either the recognition criteria could be adjusted (which might increase the number of false-positive events), or an event could be entered manually at the appropriate time.

• Manual
  If it proves impossible to define criteria which completely avoid false recognition, events can be added, deleted, or merged under precise manual control.

Tuning Recognition

Finding the best recognition criteria to minimise mismatches can be very time-consuming. In Dataview, this process is highly interactive. The display can show event parameters (time of occurrence, frequency, duration, error [as above], peak amplitude, peak width, principal components) and these parameters can be plotted against each other in a scattergraph. This often allows outlier events to be easily identified. These can either be deleted by selecting them directly from the graph, or the display can automatically centre a selected event so that the user can decide what to do. The graphs update automatically as events are added or deleted. Events can be displayed as a superimposed "stack", which resembles the display of an analog storage oscilloscope in triggered mode. The stack display can be turned into a video, where consecutive events are displayed as consecutive frames of the video. Alternatively, the waveform of events can be averaged (in all displayed channels), similar to the event-triggered signal averaging capability of many data acquisition systems.

Analysis and Manipulation

• Export
  All event parameters and the underlying channel waveforms can be exported in a format suitable for entering into a spreadsheet for further analysis.

• Principal component or wavelet analysis
  The principal components or wavelet coefficients of the waveforms comprising each event can be determined.

• Graphic display
  The main univariate (interval, peak-to-peak amplitude etc) and bivariate (phase, latency) event parameters can be displayed as a histogram, or plotted one against the other as a 2D scattergraph. Clicking on a point in the scattergraph centres that point in the main display.

• Cluster cutting
  A rotatable 3D scatter graph view of event parameters such as principal components aids manual or automatic cluster cutting.

• Point process
  Normal events with variable durations can be turned into point process events with minimal durations, located at the start, end, middle or "center of gravity" of the parent event.

• Bursts
  The "Poisson surprise" method can be used to detect bursts within event lists.

• Interval prediction
  The user can check whether the phase of a cyclic oscillatory activity detected as a series of events has been reset as a result of an experimental perturbation.

• Edit waveform
  The data of channels within the events can be set to user-specified values. This can be used to remove major artefact "glitches" from waveforms.

• Averaging
  The data of channels within the events can be averaged, and saved as a new data file. Thus stimulus-evoked responses can be averaged, if some stimulus-locked data feature is used to generate events.