Background

A brief overview of the project. Who is involved. What it entails. Why it important. And how it will be achieved.

Thesis Title: Generalizability of an Unsupervised Deep Learning Algorithm in Detecting Burst Suppression in Pediatric EEG Data

HSLU MSc Student: Carlos Arzaga

Supervisors: Oliver Staubli, Prof. Dr. Mirko Birbaumer

Co-Supervisors: Prof. Dr. med Georgia Ramantani, Prof. Dr. Med Emanuela Keller

Internal Support: Dr. med Andrea Rüegger, Alex Lo Biundo Santo Pietro

External Support: Jenny Schmid, Marko Seric

Assess the generalizability (geographical/domain) via an external validation of a pre-validated unsupervised ML algorithm for BS detection in pediatric patients.
Improve BS detection by pairing the ML algorithm with a DL architecture.
- Identify the most effective unsupervised DL architecture for EEG data analysis.
  - Self Organizing Maps (SOM)
  - Autoencoder (AE)
  - Artificial Neural Networks (ANN)
- Evaluate new architecture pairing(s).
Investigate the agreement between human EEG annotators and the DL algorithm.

Data: Retrospective collection of EEG data from pediatric patients in neurocritical care at the Universitäts-Kinderspital Zürich (KISPI).
Data Preparation: Anonymization and preprocessing of EEG and clinical data.
Data Labelling: Manual labeling of BS patterns by experienced EEG annotators.
- Dr. Med. Rüegger Andrea | Kinderspital Zürich
- Alex Lo Biundo Santo Pietro | Kinderspital Zürich
Model Development: Iterative process in training and validation of unsupervised ML and DL models.
Evaluation: Assessment of model performance using metrics like sensitivity, specificity, precision, AUROC, F1-score, NPV, MAE, and Cohen's kappa.