We interviewed the volunteers and employees of the immigrant-run non-profit organizations in North America for this research. The goal was to better understand the particular qualities and problems they face as they traverse the philanthropic ecosystems in both their home and host countries; and finally design a better user experience for the nonprofit groups, their donor partners, and the end beneficiaries. We used thematic analysis to analyze the data in this project. A publication is under preparation.

Since the advent of the COVID19 pandemic, individuals of Asian descent have been the subject of stigma and hate speech in both offline and online communities. In this study, we introduce a manually labeled and curated dataset of tweets containing sensitive opinions/news relevant to anti-Asian stigma and hate speech.

We sampled over 668M Tweets using an iterative data construction approach that includes three different stages of algorithm-driven data selection and manual labeling. Using our iterative process, we present 9,289 Tweets that are potentially stigmatizing towards the Asian community and an equal number of neutral Tweets.

This dataset can be used as a benchmark for further qualitative and quantitative research and analysis around the issue of unfair stigma and discrimination against Asian people. We have also developed machine learning models which detect tweets containing COVID-related stigma and hate speech with the best accuracy of 76%, and therefore, can be used for automatic elimination of such behavior in online communities.

Since the advent of the COVID-19 pandemic in early 2020, many artists express the effect of coronavirus on life in creative artworks and share a picture of them on social media. This collection of art is very valuable as it shows the day-to-day life of people in COVID-times. If collected and managed properly, it can provide a valuable source of data for analyzing the sociological aspect of the pandemic, and preserve a collection for future generations. In this line, we produced a sample COVID19-related art database in the form of a website and visualized the database based on the geographic and time information of the art data. We also ran a user study to inquire about the expectations of the potential users of the system, evaluate the website prototype, and gather feedback for further improvement of the database and visualization.
You can read more about this project in this Github repository.

Brain Imaging Data Structure (BIDS) is a new standard for organizing neuroimaging and behavioral data. Prior to the definition of BIDS made standard by neuroimaging world-leading scientists, researchers had different layouts for arranging the brain imaging data obtained in different experiments. Inconsistency in data structures was leading to misunderstanding and extra work on adjusting the scripts for a certain data type. Therefore, most neuroimaging organizations are moving toward adopting BIDS as their standard of data sharing and storage. Ontario Brain Institute is also working on converting the current Brain-CODE neuroimaging database into BIDS format. As a member of the Strother lab (the informatic branch of the Brain-CODE ), I developed a pipeline for their primary data-management platform, namely XNAT, to automate the conversion of data directly from XNAT to BIDS.

In this study, we investigated the short-term effects of mindfulness meditation on brain-computer interface (BCI) performance in a pediatric context. In the time of this research, the vast majority of previous BCI research had focused on creating more effective machine-learning algorithms. However, further improvements in BCI-user training was needed. On the other hand, mindfulness meditation was a promising methodology owing to its potential to increase an individual's mental control, which we used as the candidate user training method.

We recruited twelve able-bodied children between the ages of 10 and 18. An auditory-tactile P300-BCI was designed for each individual participant. Each participant completed two meditation sessions, one with 15 minutes of meditation preceding BCI operation and the other without. BCI control entailed an auditory-tactile oddball task. The average accuracy of 80.66% ± 11.68% was achieved across participants without meditation, while five participants exhibited significantly improved accuracy following meditation training. In a separate analysis using time-frequency features and a Gaussian Naive Bayes classifier, meditative and rest states were also differentiated with an accuracy of 83.94% ± 13.00%.

You can read the complete report in my thesis.

AS my summer Internship project at the Bloorview Research Institute, I designed a cough detection module to be embedded in a dysphasia detector device. Using machine learning, a set of recorded accelerometer signals for healthy participants was classified into a cough vs. non-cough group with the best classifier accuracy of 99 %. The data-set consisted of cough, swallow, speech, head movement, tongue out, tongue to left, tongue to right, and rest signals. The applied classifiers were artificial neural networks (ANN), Support Vector Machines (SVM), and Naïve-Bayes.