Ward Melville High School seniors Katie Duong, Daniel Liang and Sean Skinner have been named Scholars in the 2025 Regeneron Science Talent Search, a program run by the Society of Science since 1942. This year’s contest saw a record number of submissions, with close to 2,500 entrants from 796 high schools. To enter, each student completed independent, original research projects that were judged by a panel of experts in their respective fields.

As Scholars, Katie, Daniel and Sean will each receive a $2,000 scholarship, with a matching donation being made to the school. Their projects will also now have the chance to advance in the contest, with finalists being named later this month. All finalists will then compete in March for additional scholarships – the top of which is $250,000.

Below is some additional information about Katie’s, Daniel’s and Sean’s individual projects.

Katie Duong
Project Title: SARS-CoV-2 Infection Increases Long-Term Risk of Pneumonia in an Inner-City Population
A history of SARS-CoV-2 (COVID-19) infection could increase susceptibility of future pulmonary infections, such as pneumonia. Using a retrospective cohort design, we analyzed data from the Montefiore Health System, which serves a diverse urban population in the Bronx, over a period of four years from January 2020-December 2023. We identified 64,376 patients with a recorded SARS-CoV-2 infection (defined by a PCR test) and 1.2 million patients without (controls). Individuals with and without SARS-CoV-2 infection were propensity-matched to control for factors including age, sex, race, ethnicity and observation time. We found that individuals hospitalized for COVID-19 were 3.69 times (aHR) (p<0.0001) and those not hospitalized were 1.40 times (p<0.0001) more likely to develop future pneumonia compared to controls. Individuals with medical history of obesity, diabetes, liver disease, chronic diseases and asthma were significantly associated with outcomes (p<0.01). Additionally, individuals on Medicaid, Medicare (relative to private insurance) or with unmet social needs were at even higher risk of new-onset pneumonia (p<0.001).

DANIEL LIANG:
Project Title: Radiomic Feature Engineering for Automated Colorectal Polyp Classification Using Deep Learning Architectures
Computed tomography (CT) scans are three-dimensional X-rays: photons are fired with varying energies. Depending on their energy, some photons are absorbed or partially scattered by different tissues, while the rest are recorded by a detector-bin, from which the scan is made. From the scan, pseudo CT scans called virtual monoenergetic images (VMIs) are constructed by theoretically modeling the photons being fired at fixed (as opposed to varying) energy levels. This way, VMIs carry energy-specific information.

I experimented with methods for formatting this energy-specific information – called feature-engineering methods – in VMIs to see if, and why, such methods lead to more accurate results through machine-learning models for cancer diagnoses through CT scans.

I created two different ML-models, testing feature-engineering methods on each. Using a metric-of-accuracy known as AUC-ROC minimizes bias in the data (the VMIs).

I present the most accurate method with each model as breakthrough work: namely, energy-integrated BioGLCM and energy-integrated FOM methods. These can be utilized in hospitals for accurate diagnosis.

I conjecture why such methods lead to accurate results: They use “task-driven” strategies aligned with optimizing accuracy, preserve energy-specific information by reformatting information within individual VMI energy levels, then considering VMIs of multiple energy levels at a time, and incorporate physics-based information, formulae, and models not evident in the CT scans.

Sean Skinner:
Project Title: Physics-Informed Machine Learning for Many-Objective Generative Design
Artificial intelligence can be used to create designs for engineers. Such “generative design” exists currently but is expensive and generally only considers the stiffness and weight of designs. These programs are limited because it is difficult to train the AI. My project helps generative design programs learn how to create better designs with more goals by coding physics equations into their learning methods. As a demonstration, I made a program that generates car wheel designs such that they are stiff, lightweight, and able to prevent tires and brakes from overheating. I created this physics-based AI model so that the primary decisions it makes in the design, such as the curvature of each wheel spoke, are things that physics predicts would have a large impact on the performance goals. I also programmed the AI to evaluate how well its designs fit those decisions using calculations that it can look back at to efficiently improve itself. My program improves the foundation for programs creating better car wheels, which demonstrates how physics-informed AI can make generative design programs more effective. This will one day make generative design feasible for creating more optimal versions of a wide range of engineered products.

Date Added: 1/17/2025