Work

Abstract: This project addresses the 2025 George B. Moody PhysioNet Challenge focusing on Chagas detection using 12-lead electrocardiogram (ECG) data. Our approach used a balanced random forest classification (BRFC) with SMOTE to address class imbalance. Feature engineering extracted metrics including R-R peak distances, QRS widths, and FFT components from standardized ECG signals. After training in the CODE-15% data set, the model achieved 60% true positive identification while maintaining fewer than 13% false positives, with an AUC of approximately 0.82. Feature importance analysis revealed that variance in FFT magnitude and average QRS distance, particularly in AVR and DI leads, were the most significant predictors. While establishing a useful baseline, we determined BRFC is insufficient for optimal Chagas detection and are developing a Kolmogorov-Arnold Network (KAN) model that leverages grid extension techniques, B-spline architecture, and interpretable results for improved medical applications with diverse patient populations.

Abstract: As artificial intelligence becomes increasingly integrated into workplace decision-making, effective training for human-AI collaboration becomes essential. This paper presents results from a series of experiments involving 127 participants engaged in collaborative problem-solving with AI assistants. We introduce a structured training protocol that emphasizes mutual understanding, appropriate trust calibration, and efficient communication patterns. Results demonstrate significant performance improvements (25% reduction in time-to-solution, p<0.01) and higher satisfaction ratings compared to traditional training approaches. We provide guidelines for implementing effective human-AI training programs across domains and discuss implications for organizational adoption of collaborative intelligence systems.

Abstract: This study investigates how small pieces of information shared online can act as "breadcrumbs" that adversaries may follow to access sensitive or critical information. Using web scraping techniques to analyze LinkedIn profiles of U.S. military personnel, the research demonstrates how Operations Security (OPSEC) can be compromised through the aggregation of publicly available data. The paper follows the OPSEC process to identify critical information and indicators, assess threats, analyze vulnerabilities, assess risk, and recommend appropriate countermeasures. Results showed that 75.5% of sampled accounts posed potential security risks, with 12% classified as high-risk. The research highlights the importance of comprehensive OPSEC awareness among military personnel and their families, who are often identified as the weakest link in information security.

Abstract: The COVID-19 pandemic has been accompanied by an "infodemic" of misleading and false information spreading across social media platforms. This paper presents a multi-modal approach to misinformation detection, combining natural language processing, network analysis, and behavioral pattern recognition to identify and classify misleading health information. Our system achieved 87% accuracy in identifying misinformation from mixed-media posts across Twitter, Facebook, and Instagram. We provide an open-access database of over 10,000 classified posts that researchers can use for further analysis. Additionally, we identify key propagation patterns that differentiate how accurate versus misleading information spreads through social networks, offering insights for intervention strategies to combat health misinformation.