Worcester Polytechnic Institute: 2016 - Present
More details to come for recent projects…
Attributed hierarchical attention
Disentangled recurrent memory curation
Mentoring one new PhD student on time series missing value imputation
Internship with UMass Medical School
- Working with Jomol Matthew on a Microsoft Word plugin for automatically extracting and recommending eligibility criteria from the US NLM clinical trials database to assist clinicians with writing new clinical trials.
Multimodal data fusion for MRSA detection
Early time series classification
Led NSF REU project - Classifying time series with missing values
Leading NSF REU project - Early Prediction of MRSA Infections Using EHRs
- Worked with Sarah Brownell, an REU Intern, to generate Early Warnings for MRSA infections using historical clinical data.
- Used time series from the MIMIC III database and aligned them into a tabular format for use by common machine learning algorithms.
- Varying the number of days used to predict MRSA along with the number of days in advance of MRSA to generate prediction allowed us to understand what portions of a patient’s stay should be used to make predictions.
- Our models detected signs of MRSA infections accurately far in advance of actual diagnosis dates.
- Our paper was published at BIOSTEC 2018 and I presented this work in January 2018 in Funchal, Madeira.
- Sarah also presented the poster version of this work at the 2017 IEEE MIT Undergraduate Research Technology Conference.
Leading NSF REU project - Detecting Clostridium Difficile Infections Using Recurrent Neural Networks
- Worked with Sean Tocci, another REU Intern, to figure out a deep learning approach to infection prediction.
- We developed a Python framework for Recurrent Neural Networks using Keras.
- Using the same patients as in CREST, we trained Long Short-Term Memory Recurrent Neural Networks to identify Clostridium Difficile infections, approaching similar results as in our prior work, CREST.
Clostridium Difficile Risk Estimation (CREST)
Clinical data offers many opportunities for impactful and technically fascinating research. Beginning in August, 2016, I worked with Cansu Sen using the MIMIC III Intensive Care Unit Database. This database is publicly available, consists of 12 years of clinical data from ~58,000 admissions. You can request access to the database here. We focused on the detection of one infection, Clostridium Difficile, and therefore extracted patients who got this infection during their stay. As is required for supervised-learning algorithms, we also extracted a set of patients who did not get this infection and trained Logistic Regression, Random Forest, and Support Vector Machine models to detect patterns that indicate C. Diff. based on the following data sources:
- On-admission demographic information
- Clinical notes written upon admissions
- 80-dimensional time series of clinical measurements
- Hand-crafted features summarizing the time series to capture their temporal nature
We trained classifiers on each source of data, then merged their outputs to create a well-informed meta-learner to detect C. Diff. far in advance of the date of infection-confirmation. This work was published at the European Conference of Machine Learning (ECML), presented by Cansu Sen on September 19, 2017 in Skopje, Macedonia. Our paper can be found here.
University of Arizona REU: Summer 2015
Image segmentation to understand the phenological development of desert shrubs through drought periods
From June-August 2015, I took part in an NSF-funded Research Experience for Undergraduates at the University of Arizona in Tucson, AZ. I was stationed at BioSphere 2 and worked with Dr. Shirley Papuga in the School of Natural Resources and the Environment.
- Prior to my REU, Dr. Papuga and her students collected thousands of image sequences of creosote bushes over eight seasons.
- We worked with this sequence of images of one creosote bush to investigate how its health changed through the seasons and across three years.
- Using MATLAB, I trained decision trees to segment these color-images into sections consisting of background, leaves, stems, and flowers.
- We asked two main questions:
- How does the image segmentation change through the seasons and over the years?
- Can we use such segmented information to approximate the normalized difference vegetation index (NDVI), an expensive but reliable method for understanding global plant health?
- By the end of the summer we had uncovered that this method of segmenting images captured seasonal trends but did not approximate NDVI.
- I presented this work at a poster-session at the 2015 Undergraduate Research Opportunities Consortium (UROC 2015).