Poly-cystic Ovarian Syndrome (PCOS) is a common, yet often undiagnosed, health condition affecting 8-13% of women globally. Its effects are primarily centered around hormonal imbalances and metabolism causing problems with the ovaries. The exact cause remains unknown, but PCOS is associated with an increased risk of diabetes, heart disease, and other complications. Early diagnosis is crucial for effective management and prevention of these issues. Leveraging machine learning (ML) and data science, our study focuses on developing a robust diagnostic model for PCOS, excluding the need for ultrasonography. Statistical analysis models such as Recursive Feature Elimination (RFE), Logistic Regression, and Random Forest were used to identify key predictors of PCOS diagnosis. Notably, results revealed women with less than 5 cycle days per month were more likely to develop PCOS, contradicting the assumption that PCOS causes excessive bleeding. Cystic acne, skin discoloration, and excess hair growth were identified as notable precursors to PCOS. Anti-Müllerian hormone was a significant biomarker for PCOS development. To address disparities in access to diagnostic tools, we propose integrating Anti-Müllerian hormone testing into routine blood work for all women to enable earlier PCOS detection. Implementing these recommendations could revolutionize PCOS management by facilitating early intervention and mitigating downstream health complications. Further research is needed to fully understand the mechanisms underlying PCOS development.

Bio:

Jessica Owens is a data analyst and graduate student at Meharry Medical College, where she is pursuing a Master of Science in Data Science. She holds an undergraduate degree in Data Analysis from Lipscomb University, where she developed skills in data visualization, analytical problem solving, and working with data tools to interpret complex information. Her work focuses on analyzing large datasets to identify insights that support decision making, and she is actively exploring machine learning and neural networks through her graduate studies. As a United States Air Force veteran, Jessica brings discipline, adaptability, and a results-driven mindset to her academic and professional pursuits, and she is currently completing her capstone project on a multi-source recommendation system.

Abstract:

This capstone project focuses on improving traditional collaborative filtering models, which often rely only on user-item interactions and miss important contextual information. The project develops a hybrid recommendation system that combines collaborative filtering with content-based features from MovieLens, IMDb, and Movie Plot Synopses (MPST). User ratings are integrated with additional data such as genre, personnel, and text features to improve prediction accuracy. The study compares a baseline collaborative filtering model with hybrid models that incorporate additional data sources, using Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) to evaluate improvements. This approach is expected to show that incorporating multiple data sources can improve recommendation performance and support more informed, data-driven decision making.

Advisor: Graham West, Ph.D.

Bio

Mian Pan is a graduate student in Data Science at Meharry Medical College’s School of Applied Computational Sciences. She holds advanced degrees in Communication and Information Systems and Computer Science and is transitioning into a career in data science with a focus on healthcare applications. Her academic interests include machine learning, statistical modeling, and health data analytics, particularly in addressing disparities in underserved populations. Mian has conducted research in protein function prediction and has been involved in the AIM-AHEAD health data science training program, where she applies AI and machine learning techniques to real-world clinical and population health data. Her current work includes projects involving electronic health records and autism-related healthcare access. She is passionate about using data-driven approaches to improve clinical decision-making and public health outcomes.

Abstract

This capstone project investigates unmet healthcare needs among children with autism using data from the National Survey of Children’s Health (NSCH). The study applies machine learning models, including logistic regression, Random Forest, Gradient Boosting (XGBoost), and Support Vector Machines (SVM), to model and identify key predictors of disparities in healthcare access. Model performance is evaluated through cross-validation and comparative analysis. To enhance interpretability, SHAP (SHapley Additive exPlanations) is employed to quantify feature contributions and identify key socioeconomic and demographic determinants. The project also incorporates state-level choropleth maps to describe geographic disparities. By combining predictive modeling with interpretable analytics, this work aims to provide evidence on targeted healthcare needs for children with autism.

Advisor: Aize Cao, Ph.D.

Understanding complex biological processes, diseases, and drug discovery necessitates deciphering the intricate interactions among diverse biological entities such as proteins, drugs, metabolites, and enzymes. Knowledge graphs (KGs), representing interconnected multi-relational entities through nodes and edges, offer a promising approach to model heterogeneous biological data comprehensibly. However, many crucial links within KGs remain hidden, presenting a challenge for comprehensive understanding and analysis. In this capstone project, we address the task of link prediction within a biomedical knowledge graph to facilitate biomedical knowledge discovery. Our objectives encompass curating a comprehensive biomedical knowledge graph, constructing a general link prediction pipeline employing various knowledge graph embedding (KGE) models, and applying link prediction techniques to tackle two challenging biomedical problems: drug repurposing and protein function prediction. This project not only contributes a reusable pipeline for biomedical knowledge discovery but also lays the groundwork for future advancements in link discovery using knowledge graphs, applicable to a wide range of biomedical research tasks.

Proteins are the biomolecules that form the building blocks of life. The genetic information encoded in Deoxyribonucleic Acid (DNA) is translated into proteins from Ribonucleic Acid (RNA) transcription. Proteins perform multitudes of functions that include but are not limited to catalyzing reactions as enzymes, participating in the body’s defense mechanism as antibodies, forming structures and transporting important chemicals. Therefore, functional characterization of proteins is crucial to understanding life, diseases, and developing novel therapeutics. In this research, we propose a protein function prediction (PFP) pipeline that utilizes a new way of tokenization called SAT (structurally aware Tokenization). SAT works by decomposing a protein sequence into its constituent domains which are the evolutionarily conserved structural parts of a protein sequence. Typically, given a protein sequence as input, the objective of protein function prediction or annotation is to assign appropriate Gene Ontology (GO) terms to describe the functional characteristics of the query protein. Tokenization – the split of the input sequence into smaller subunits (tokens), is crucial in language models. Typically, these models take protein sequences of amino acids and use common tokenization techniques such as individual amino acids (AA), K-Mers, or Byte Pair Encoding (BPE) to build protein language models. However, we hypothesize that including structural constructs of protein such as ordered list of motifs and domains as tokens will provide additional valuable structural information into the model as they directly link to functions. In this work, we have explored SAT for protein function prediction in comparison with AA, k-Mers, and BPE under various machine learning and natural language processing techniques including ‘Word2Vec’ and large language models. Additionally, we have implemented a GPT2 style transformer architecture to train a PFP model using proposed SAT tokenization. The experimental outcome presents promising performance although it demands further exploration to confirm its efficacy.