Classification of auditory brainstem responses through symbolic pattern discovery.

Area of Science:

• Medical Informatics
• Biomedical Signal Processing
• Machine Learning

Background:

• Numeric time series data are prevalent in medicine, necessitating effective data mining techniques for knowledge discovery and decision support.
• Classifying time series based on frequent patterns can enhance diagnostic accuracy and interpretability in clinical settings.

Purpose of the Study:

• To develop and validate a method for classifying time series by discovering frequent symbolic patterns derived from temporal abstraction.
• To enable explainable classification of medical time series data using domain knowledge.

Main Methods:

• Transforming numeric time series into symbolic sequences using domain knowledge.
• Applying symbolic pattern discovery to identify frequently occurring subsequences representative of specific groups.
• Utilizing a classification technique based on discovered patterns for new individual classification.

Main Results:

• The method was applied to brainstem auditory evoked potentials (BAEPs) time series from 83 individuals across four classes.
• Achieved an overall accuracy of 99.4%, with 97.6% sensitivity and 100% specificity in cross-validation.
• Demonstrated high precision in classifying patients with auditory-related disorders.

Conclusions:

• The proposed method effectively reduces dimensionality and enhances data representation clarity through symbolic transformation with domain knowledge.
• The approach accurately identifies patterns in BAEPs time series, leading to precise predictions of auditory disorders.
• This method offers a valuable tool for the interpretation and classification of medical time series data by domain specialists.