I am an independent ML researcher working on applied machine learning in the field of medicine. With a B.S. in Computer Science and a strong passion for solving real-world problems, I have explored diverse domains to create impactful solutions.
Currently a graduate student at the University of Kansas, where I contribute to various research labs and serve as a graduate teaching assistant. My work is driven by a commitment to advancing healthcare through innovative ML applications.
Outside of research, I spend my free time doing combat sports 🤼, such as jiu-jitsu 🥋, muay thai 👊🏼, and boxing 🥊.
![[ project image ]](assets/research1.png)
Development of Alzheimer's Disease Risk Score for Future Integrated Primary Care: A White-Box Approach
Early diagnosis of Alzheimer’s disease is often delayed by limited access to specialists. We developed an interpretable machine learning scorecard (FLAME) using simple clinical measures—cognition, daily functioning, and demographics—that can be collected in primary care settings. Trained and validated on large, independent datasets, the model shows strong and consistent performance in predicting Alzheimer’s risk. By translating complex predictions into an easy-to-use scoring system, this approach enables earlier detection and intervention, with potential applications across other forms of dementia.
![[ project image ]](assets/research2.png)
Individualized Machine-learning Based Clinical Assessments Recommendation System (iCARE)
The iCARE framework is a specialized machine-learning system designed to personalize clinical assessments by tailoring feature selection to individual patient profiles. By utilizing locally weighted logistic regression and SHAP value analysis, the framework identifies high-impact diagnostic features that traditional standardized procedures often overlook. Evaluations on real-world datasets for diabetes and heart disease show accuracy and AUC improvements of up to 12% over global approaches. Ultimately, iCARE enhances diagnostic precision by providing data-driven, individualized recommendations for more effective medical decision-making.
-
Heterogeneous ensemble learning: modified ConvNextTiny for detecting molecular expression of breast cancer on standard biomarkers
-
A Concept-Driven Disentanglement Framework for Interpretable Graph Neural Networks in Structure-Function Coupling
-
Unsupervised Deep Learning Framework for Quantifying Atypical Motor Signatures in ASD
-
Decoding Motor Signatures in Autism from Markerless Video: Interpretable Skeleton-Based Ensembles
-
ConvNeXt Model for Breast Cancer Image Classification
-
Image Segmentation for Teeth Classification
-
In-silico Analysis for Aptamer Selection
![[ project image ]](assets/project1.png)
PROSE x Llama: Program Synthesis for Semantically Wrong Code Fixing
This project presents a detailed account of the implementation and evaluation of targeted AST-to-AST transformations aimed at correcting semantic inaccuracies in code generated by Pre-trained Language Models (PTLMs). Our approach utilizes a customized Domain-Specific Language (DSL) developed for structuring abstract syntax trees (ASTs) and a series of predefined transformation rules facilitated by the PROSE synthesis framework. We evaluated the effectiveness of these transformations on common semantic errors identified in PTLM outputs, such as incorrect indexing methods and unnecessary arguments.
![[ project image ]](assets/project2.png)
MCUNet Reproduction: Tiny Deep Learning on IoT Devices
This project began as a reproduction of the MCUNet paper: "MCUNet: Tiny Deep Learning on IoT Devices" (NeurIPS 2020) by Ji Lin et al. Our original focus was on deploying deep learning models on resource-constrained microcontrollers, specifically targeting the Visual Wake Words (VWW) task (i.e., person detection). Midway through, we pivoted based on practical challenges and resource constraints. The revised project focuses on re-implementing the TinyNAS component and demonstrating its adaptability to alternate datasets and toolchains.
-
Deep Learning for IMU Classification
-
Interpretable Scorecard Model for Early Diabetes Classification
-
Cuadrado