Early Detection of ERP Indicators

Title: Early Detection of ERP Indicators for Developmental Dyslexia Using Predictive Analytics

Type: Master Thesis

Student: Peter Maier

Supervisor: Christian Wachinger

Collaboration with: Nassir Navab

Status: Finished



This thesis investigates early indicators for developmental dyslexia in children, using various predictive modeling algorithms on a longitudinal sample of event-related EEG data ranging from a preliterate stage (kindergarten) to a late stage of reading acquisition (end of second grade). The dataset was originally acquired and presented by Wachinger et al. in [1]. Basic concepts of dyslexia research and data mining are introduced for unfamiliar readers. Exploratory data analysis, including a full topographical visualization of the historical data, is conducted and used to guide design choices in the implementation of a custom-tailored classification and regression framework, excerpts of which are made available to the reader. Modeling algorithms are fine-tuned on the first session of the longitudinal dataset and validated on data from later sessions. Elastic Net classifier (ACC=77.42%, TPR=0.75, TNR=0.80, AUC=0.80, F1=0.77 with 7-fold cross-validation, ACC=69.97% for 100-times-7-fold cross-validation, and ACC=80.65% for LOOCV) and Gradient Boosting classifier (ACC=83.87%, TPR=0.88, TNR=0.80, AUC=0.83, F1=0.85 with 7-fold cross-validation, ACC=75.00% for 100-times-7-fold cross-validation, and ACC=90.32% for LOOCV) yield the most accurate models. These models are briefly examined with regards to their medical implications.

Thesis for download.