Speech Perception In Noise Enhancement (SPINE) via EEG-based Neurofeedback

I am awarded a fully funded research grant for a period of 2 years (of which 10 months will be spent as a residential fellowship) to work on this project. The goal is to determine if neurofeedback on auditory selective attention enhances the perception of speech in noisy environments in everyday life. The hypothesis is that providing users with feedback on their auditory selective attention skills (assessed through the classical competing speech paradigm) enables them to adjust their perceptual abilities, thereby improving their performance in speech-in-noise tasks. I collaborate with Prof. Dr. Stefan Debener and Dr. Martin Bleichner from the Department of Psychology (Neuropsychology Lab), University of Oldenburg, Germany.

Newborn EEG Artifact Removal (NEAR) Pipeline

One of the challenges in studying human cognitive development is the lack of reliable neural feature extraction from noisy neuroimaging data. My (primary) goal within my PhD research was to develop signal processing methodologies that can overcome such limitations. As a result, I developed a fully automated preprocessing pipeline to remove/filter out artifacts from EEG collected from newborns of 0-4 days old and infants of 9-months old. The pipeline is based on the EEGLAB software and is freely available here.

Automatic Noisy/Bad EEG Sensors Detection using the Local Outlier Factor (LOF) algorithm

Detecting and removing noisy sensors (such as flat-line sensors) is one of the common preprocessing steps in EEG and MEG data analysis pipeline. I introduced a novel, density-based bad channel detection using the Local Outlier Factor (LOF) algorithm. The novelty is that the method does not assume any distribution for the data and employs nearest neighbors approach to find the sensors that ‘locally’ deviated from its neighbors. More information can be found here. The algorithm is available as EEGLAB plugin (download here), MEEG-kit (download here) and MNE-Python (download here) .

Performance-efficient EEG Channel Reduction using Knowledge Distillation framework

Reducing the number of EEG channels can enhance the ecological validity of studies but often leads to decreased performance in machine learning algorithms used for downstream analysis. In this project, in collaboration with researchers from EPFL (Switzerland), we developed a Knowledge Distillation-based framework for EEG channel reduction that preserves the performance achieved with all EEG channels.

Deep-Learning based projects

1. Attempted to train a DL classifier to detect EEG bad channels (together with a Master thesis student) - see here for the paper.
2. Interpreted a DL classifier why/how models find noisy EEG channels (together with Mr. Francesco Paissan, FBK, Italy) - see here for the paper.

Non-EEG (Personal) Projects

1. ECG data cleaning using the state-of-the-art algorithms for reliable estimation of Heart Rate Variability.
2. EDA (Electrodermal Activity) data cleaning and processing to extract tonic components.
3. Interpreting the Deep Learning (DL) model that classified vocalization of chimpanzees by using signal processing methods such as Fast Fourier Transform (FFT).