Learning From (Multiple) Demonstrations
Sept 2013 to present
I am working together with three first year graduate students on improving the robustness of learning from multiple demonstrations. (Section in progress)
Predicting Diagnoses of Depression (CS Honors Thesis)
May 2012 to July 2013
Huang SH, LePendu P, Iyer SV, Tai-Seale M, Carrell D, Shah NH. Toward Personalizing Treatment for Depression: Predicting Diagnosis and Severity. Journal of the American Medical Informatics Association (JAMIA). Under review.
Presentations and Posters
Huang SH, LePendu P, Iyer SV, Tai-Seale M, Carrell D, Shah NH. Predictive Models in Mental Health: From Diagnosis to Treatment. American Medical Informatics Association 2013 Annual Symposium (AMIA). Abstract Presentation.
Huang SH, LePendu P, Iyer SV, Bauer-Mehren A, Olson C, Shah NH. Developing Computational Models for Predicting Diagnoses of Depression. 2013 AMIA Clinical Research Informatics Summit (CRI). Poster.
For my honors thesis, under Professors Nigam Shah and Jeffrey Ullman, I built machine learning models for predicting a diagosis of depression in patients, based on their electronic medical records. Depression is often under-diagnosed, so my goal was to produce a screening tool for identifying high-risk patients for follow-up. I worked with a dataset of over 10,000 depression patients and 60,000 non-depression patients from the Palo Alto Medical Foundation. My most effective model, based on LASSO, was able to predict a future diagnosis of depression up to a year in advance, with an AUC of 0.70 to 0.80.
My colleagues and I also investigated the potential of personalizing treatment for depression, through not only developing models for predicting diagnoses of depression, but also building models for predicting baseline severity (before treatment) and identifying patient characteristics that predict differential treatment response (medication vs. psychotherapy). We recently submitted a paper on our results.
June 2012 to November 2012
Suen C & Huang S & Eksombatchai P (Joint), Sosic R, Leskovec J. NIFTY: A System for Large Scale Information Flow Tracking and Clustering. Submitted to the 2013 International World Wide Web Conference (WWW).
I worked with Caroline Suen and Pong Eksombatchai, under Rok Sosic and Professor Jure Leskovec, on developing a system for tracking quote popularity in news articles and blog posts. Our final product processed a 20-terabyte dataset of over six billion news articles and blog posts, spanning the last four years. You can take a look at the results here!
Caroline, Pong, and I made design decisions and developed the underlying data structures together. My role in this project included working together with Caroline on implementing the clustering step, in which quotes are clustered together if they are likely to have been derived from the same source quote. I also created heuristics to filter out non-quotes – quoted phrases that were movie, TV show, or song titles rather than a phrase spoken by someone. In addition, I developed the first iteration of the visualization, and linked the back-end output to the visualization.