Preparing a data set for Deep Learning from zipped ABR raw data files

Version 1

This notebook is about pre-processing the Auditory Brainstem Response (ABR) raw data files provided by Ingham et. al to create a data set for Deep Learning models.

The unprocessed ABR data files are available at Dryad.

Since the ABR raw data are available as zip-archives, these have to be unzipped and the extracted raw data files parsed so that the time series corresponding to the ABR audiograms can be saved in a single csv file.

The final data set contains the ABR time series, an individual mouse identifier, stimulus frequency, stimulus sound pressure level (SPL) and a manually determined hearing threshold. For each mouse there are different time series corresponding to six different sound stimuli: broadband click, 6, 12, 18, 24, and 30 kHz, each of which was measured for a range of sound pressure levels. The exact range of sound levels can vary between the different mice and stimuli.

The following is done:

• The zip archives are unpacked.

• The extracted ABR raw data files are parsed and collected in one csv file per archive.

• The csv files are merged into a data set of time series. Each time series corresponds to an ABR audiogram measured for a mouse at a specific frequency and sound level.

• The mouse phenotyping data are available in Excel format. The individual data sheets are combined into one mouse phenotyping data set, maintaining the mouse pipeline and the cohort type mapping. In addition, the hearing thresholds are added to the ABR audiogram data set.

• The data sets are curated:

  • there is a single curve per mouse, stimulus frequency and sound level,
  • each sound level is included in the list of potential sound pressure levels,
  • for each mouse for which an ABR audiogram has been measured, mouse phenotyping data are also provided.