Projects
Research Projects
Dementia Detection
- We experiment with domain adaptation (DA) techniques on heterogeneous spoken
language data to evaluate generalizability across diverse datasets for a common
task: dementia detection. We empirically find that the feature-augmented
DA method achieves a 22% increase in accuracy adapting from a conversational
to task-specific dataset compared to a jointly trained baseline. - Annotated conversational data with Dialog-Act (DA) tags from two different cognitive
tasks of DementiaBank dataset, proposed task-agnostic Dementia detection
framework leveraging the annotations, and achieved competitive performance relative
to the models trained on a specific cognitive task. - An automated system to predict MMSE scores reflecting individuals’ cognitive
health status, based on their free speech samples from conversational interviews.
The models experiment with traditional machine learning algorithms, pre-trained
language model, and the best performing model achieved 16.5% decrease in
RMSE score from the linguistic benchmark of the baseline paper in ADReSS
Challenge2020.
Smart and Connected Family Engagement for Equitable Early Intervention Service Design
- collaborating with researchers from UIC's College of Applied Health Sciences
on designing and implementing a chatbot to customize caregiver navigation when using a
web-based application prototype known as the Participation and Environment Measure (PEM+).
In addition to making this application more tailored to a diverse set of stakeholders,
part of this process also involves developing automated techniques to facilitate solution-focused
caregiver strategy exchange for personalized pediatric rehabilitation service design. - Thus far, we have prototyped this chatbot using Google DialogFlow with an FAQ dataset
specifically collected from real-world end users (caregiver using PEM+). We are currently working on
extending the chatbot to facilitate caregiver strategy exchange using a opensource framework
known as OpenDial. - For training the dialog manager of the chatbot, we have annotated a corpus containing
early intervention strategies categorized into four broader classes based on key, clinically validated
drivers of participation. Using the corpus we plan to train the dialog manger of our chatbot to
recommend personalized early intervention strategies to parents and caregivers and in turn
conduct human evaluations on the extended prototype.