Preprint
Social Determinants of Health and Fentanyl Overdose Mortality Across US Counties: An XGBoost and SHAP Analysis Identifying Silent Risk Counties and Treatment Deserts
Kabi Raj Tiruwa, Abhisan Ghimire, Anuj Kumar Shah · arXiv:2605.08230 · 2026
A county-level machine learning study combining CDC WONDER mortality data, the Social Vulnerability Index, health behavior data, and healthcare workforce files across 975 U.S. counties. An XGBoost model with SHAP interpretation identified disability rate, hypertension, smoking, and lack of vehicle access as the strongest predictors of fentanyl overdose mortality, correctly flagging 71% of high-risk counties (R² = 0.457). The analysis surfaced 143 "silent risk counties" with elevated structural vulnerability but not-yet-elevated death counts, and found treatment-desert counties had 52.6% higher mortality, pointing to where prevention investment matters most.
- XGBoost
- SHAP
- Machine learning
- Public health
- Python
Preprint
Wearable-Grade Lead Reduction Disproportionately Degrades ECG AI Performance in Elderly Patients: Evidence from PTB-XL and MIT-BIH
Kabi Raj Tiruwa, Abhisan Ghimire · medRxiv · June 2026
A 1D residual neural network trained on 21,091 PTB-XL recordings shows that reducing clinical 12-lead ECGs down to a single wearable-grade lead degrades AI diagnostic accuracy far more for older patients than younger ones: a 14.1 percentage-point drop for patients 75+ versus just 0.4 points for patients under 40, a roughly 40-fold gap (p < 0.0001), with external validation confirming the effect held across datasets. The findings argue for age-stratified performance reporting as a minimum standard in wearable AI-ECG validation and regulatory review.
- Deep learning
- ECG analysis
- Medical AI
- Health equity
- Python
Publication
Understanding Hate Group Videos on YouTube
Shuo Niu, Keegan Veazey, Phoenix Pagan, Abhisan Ghimire · ACM CSCW'22 Companion · pp. 32–36 · November 2022
A preliminary grounded-theory analysis of 96 YouTube videos from U.S. hate groups, using open-coding and affinity diagramming to identify common video styles, and annotating each video's diagnostic, prognostic, and motivational framing to understand how these groups promote collective action. Presented in the CSCW'22 companion track. My contribution centered on data collection and Python-based research tooling supporting the analysis.
- Python
- Data collection
- Online video analysis
- Computational social science
- Qualitative & quantitative analysis
Preprint
A Survey on Watching Social Issue Videos among YouTube and TikTok Users
Shuo Niu, Dilasha Shrestha, Abhisan Ghimire, Zhicong Lu · arXiv:2310.19193 [cs.HC] · October 2023
A survey of 659 YouTube and 127 TikTok users examining why people watch social-issue videos on video-sharing platforms, using the uses-and-gratifications framework to compare information-seeking, entertainment, and social motivations across the two platforms.
- Survey research
- Human-computer interaction
- Social computing
- Video-sharing platforms
Conference Paper
Integrated Library Management System Using Barcode via Android and Web
Abhisan Ghimire · National Conference on Recent Trends in Science, Engineering, and Management
Presented work on ILMS, a library management system built as an undergraduate capstone project: administrators update book records via barcode scan, and users browse the catalog from a mobile app. Built with Java, XML, and the Google Books API, and released on the Google Play Store.
- Java
- Android
- Barcode scanning
- Google Books API