Aghaeepour Lab

Apr 2024 - present

Enhance pre-existing foundation models for electronic health records through weakly supervised pretraining using proxy labels, to ultimately improve patient outcome prediction
Develop a method for creating weak labels, given the lack of gold-standard labels and the challenges that arise with predicting conditions at scale using supervised methods

Singh Lab

Sep 2019 - Jun 2021

Remodeled a Python tool that automatically selected an appropriate fixed-width bin size for a given set of Hi-C reads by calculating SSEs between original and binned reads for different resolutions
Ran the code on Hi-C pairs files from 4DN for three cell lines: K562, HFFc6 and H1, across all chromosomes
Explored open-source loop-calling scripts including OnTAD to verify the accuracy of the code in selecting a bin size that conserved genomic features like chromatin loops
Calculated the correlation between the first eigenvector (containing information about active and inactive chromatin regions) and histone modifications for different resolutions as an additional way to check that the tool was conserving non-noisy data

Adapting U-Net for Brain Tumor Classification from MRIs

BMI 260: Computational Methods for Biomedical Image Analysis @ Stanford

Apr 2024 - Jun 2024

Achieved an accuracy of 90% and sensitivity of up to 0.99 for classifying brain tumor severity as one of four classes: glioma, meningioma, pituitary tumors, and non-tumorous.
Leveraged feature extraction for U-Net segmentation by adding a classification head with a softmax activation
Improved upon base model accuracy using augmentation techniques and performed alation analysis to determine that flipping

Interactive Stem Player

MUSC 356: Music and AI @ Stanford

Feb 2024 - Mar 2024

Created an interactive tool that allows users to combine song stems in different ways using ChucK, Wekinator and Processing
Trained a regression model to play stems at different levels depending on x,y input from the mouse
More details:
- https://medium.com/@shreyadsouza/interactive-stem-player-9dde80fe16c3
- https://github.com/shreyadsouza/stem-splitter

Predicting Parkinson’s Disease Using Gait Data

CS229: Machine Learning @ Stanford

Oct 2023 - Dec 2023

Implemented feature engineering to calculate stride length, stride time, and swing time to effectively capture characteristics of ‘Parkinsonian gait’ and obtain a baseline accuracy of 74.68% using logistic regression
Conducted feature selection to assess feature importance, and used scikit-learn to create supervised, unsupervised, and deep learning models to achieve the highest accuracy of 88% with an LSTM with attention, given that the input data was temporal
Experimented with transfer learning by using this LSTM on smaller datasets for other neurodegenerative diseases like ALS

EyeWise: Endocrinologist in your Pocket

BIOE273: Biodesign for Digital Health @ Stanford

Sep 2023 - Dec 2023

Carried out the need-finding process by interviewing clinicians to identify gaps in the diabetes market, including the lack of education provided to people with prediabetes on complications like retinopathy, and patient noncompliance
Formalized a business model for an LLM chatbot that provides users with personalized recommendations for their diabetes journey and streamlines the appointment-making process, and calculated a total addressable market of $317 million

Video Game Style Transfer

CS1430: Computer Vision @ Brown

Apr 2021

Investigated, along with three other students, style transfer by transferring video game styles onto landscape, city, and animal images using a convolutional neural network
Referenced papers on style transfer from paintings onto images, and employed a version of the VGG19 model by utilizing max pooling layers to capture more extreme edge and texture information common in video game styles
Carried out hypothesis testing to determine how to best optimize output by using different input images and style weights

Exploring the Role of Computer Science in Cancer Research

CSCI1951t: Surveying VR Data Visualization Software for Research, CSCI1970: Independent Study @ Brown

Jan 2020 - May 2020; Aug 2020 - Dec 2020

Visualized glioblastomas from MRI scans in UPenn’s BraTS dataset by segmenting the tumors and brain tissue in Slicer and exporting the models to Paraview and Unity before viewing them in virtual reality in the YURT, Brown’s virtual reality lab
Documented how virtual reality can be used in the medical field and created tutorials for how to recreate my work on the class website
Developed a syllabus for an independent study, under the guidance of Professor Sorin Istrail, to continue research into how computer science, particularly machine learning, has been used in cancer detection and treatment
Read and summarized papers on topics including predictive diagnostics, tumor classification, and patient privacy

The Hidden Power of Period Planning

CSCI 1320: Creating Modern Web and Mobile Applications (Spring 2020) @ Brown

Feb 2020 - May 2020

Built a mobile application aimed to be a resource for victims of sexual assault to document any encounters while being disguised as a period-tracking app

Acted as a full-stack engineer in a group with three other students to fulfill a client’s objectives using JavaScript frameworks such as Node.js, and MongoDB to store user information