FADS RESEARCH SHOWCASE

AUGUST 3-4, 2026

Faculty Assistance in Data Science

INDIANA UNIVERSITY

THE LUDDY SCHOOL

Reminders

FADS showcase

• Recognitions will be emailed to you by August 10th
• Judges feedback will be anonymized and emailed to you
• The evaluation criteria are: 
• Best Communication 
• Best Innovation 
• Overall Best 
• Honorable mention 

​

​

Day 1: August 3

Judges

• Yavuz Kadioglu (Director of Technology Development & Commercialization)
• Rahul Devajji (Info PhD Student)
• Lalit Pandey (Info PhD Student)

​

From Indiana EHRs to Alzheimer's Progression Pathways A Graph-Based Subphenotyping Framework

​

Faculty/Advisor: Dr. Selena Wang; Indiana University School of Medicine

Presenter: Uthra Sridhar

​

Approach & Technical Build

GNN + clustering framework to real-world Indiana patient records (INPC).

Technical Pipeline

1. Cohort & Data Engineering

Built the study cohort from 163,346 raw INPC records down to a 1,547-patient final cohort; encoded demographics, diagnoses (Phecodes), and medications per encounter.

2. Graph Construction

Built a directed patient graph — 128,073 encounter nodes, linked via Jaccard-similarity (cross-patient) and chronological (within-patient) edges.

3. GNN Training

Trained and directly compared two architectures, GraphSAGE and MagNet, on the full patient graph to learn outcome-oriented encounter embeddings.

4. Clustering & Subphenotyping

Consensus clustering + UMAP visualization on embedding trajectories to discover and validate distinct progression subphenotypes.

Cohort Funnel

163,346 patients screened (INPC)

↓ 31,195 with an MCI diagnosis

↓ 3,219 with MCI + later AD diagnosis

↓ 2,319 in correct temporal order

↓ 1,803 with a ≥180-day MCI→AD gap

↓ 1,547 final study cohort

Scale of Effort

✓ 128,073 clinical encounters modeled as graph nodes

✓ 2 independent GNN architectures trained and compared

✓ Consensus clustering + UMAP validation across multiple runs

Tools & Stack

Python · PyTorch Geometric · GraphSAGE / MagNet GNNs · OMOP CDM · INPC (Indiana Network for Patient Care) · UMAP · Consensus clustering

Results: Distinct Progression Pathways Across Two GNN Architectures

GraphSAGE UMAP — 4 subphenotypes

MagNet UMAP — 5 subphenotypes

GraphSAGE (K=4): top comorbidities & dementia medication use by subphenotype

MagNet (K=5): top comorbidities & dementia medication use by subphenotype

Results

✓ Encounter-level Healthy-vs-MCI classification accuracy: 97.85% (GraphSAGE) / 98.78% (MagNet)

✓ 4 distinct Healthy-to-MCI progression subphenotypes identified via GraphSAGE — trajectories diverging in speed and comorbidity pattern

✓ A deeper 5-subphenotype structure surfaced under MagNet, including one high-severity subgroup with a striking demographic concentration

Next: sharpening these subphenotype boundaries with contrastive representation learning.

Archival Data Sourcing and Analysis for Behavioural Science / Management Research (Performance Consistency and Salary Growth in Major League Baseball)

​

Faculty/Advisor: Dr. Jessica Paek

Presenter: Khushi Jhanwar

​

Does Consistent Play Earn steady Raises?

Performance Consistency and Next-Season Salary Growth in MLB

D

The Data

Lahman Baseball Database (1871–2025) — batting, pitching, salary, and biographical records

6,364

Hitter player-seasons

5,283

Pitcher player-seasons

35

MLB teams represented

31

Seasons of salary history

M

The Method

1

Measure consistency

Rolling, playing-time-weighted coefficient of variation of OPS (hitters) or ERA (pitchers), computed cumulatively across each player's career

2

Require a track record

Only players with 3+ qualifying seasons; 200+ AB or 50+ IP per season

3

Predict next year's pay

Regress next-season log salary growth on performance inconsistency, with robust (HC3) standard errors

Results

More erratic performance predicts smaller raises the following season, for both hitters and pitchers

*** p < .001 — both effects statistically significant

What it means

A one-SD rise in inconsistency is linked to an ~8.0% smaller raise for hitters, and a smaller but still significant effect for pitchers

The pattern holds after adding year and team fixed effects

Model fit is modest (R² = 0.015 hitters, 0.005 pitchers) — expected, since many factors beyond consistency shape pay — but both models are jointly significant

Bottom line: Teams appear to reward steadiness — players who perform consistently season to season see stronger salary growth than volatile players of similar overall ability.

Day 2: August 4

Judges

• Dan Melnick (Internal Audience Engagement Manager)
• Scott Michaels (Associate Vice President for Centers and Core Facilities, IU Research)
• Belen Saavedra (PhD student)

​

Harnessing Social Media to Monitor Emerging Drug Trends with ontology-enhanced large language models

​

Faculty Advisor: Dr. Maria Parker

Presenter: Nithish Sriram Srinivasan

​

Project Idea

• Current surveillance gap: Traditional drug-monitoring systems often lag by months or years, while emerging drug trends and coded language spread rapidly across social media.

​

• Core challenge: Drug slang, abbreviations, emojis, and ambiguous terms create semantic drift, weak signals, and high noise that can reduce the accuracy of conventional NLP and standalone LLMs.

​

• Proposed solution: Convert trusted drug inventories such as NIDA and DEA resources into dynamic ontologies and knowledge graphs containing drug classes, brand names, effects, routes, and slang relationships.

​

• Ontology-grounded LLM workflow: The oRAG system retrieves relevant knowledge-graph relationships, classifies known or ambiguous terms, and flags unknown referents for expert review and future ontology updates.

​

• Expected impact: Combining ontologies, LLMs, and human validation could enable faster, more accurate, adaptive, and sustainable detection of emerging drug discourse for public-health surveillance and intervention.

Contribution

• Over the summer, I worked alongside the PIs to refine a computational pipeline that combines machine learning, RDF/SPARQL, and large language models to identify drug-related content on social media and connect it to standardized concepts from the NIDA ontology. One of my key contributions was building a binary BERT classifier that acts as an initial screening step, flagging whether a given post is drug-related before it moves further through the pipeline.

​

• Ran experiments on different social media posts and validated the results of the pipeline. This additional layer reduces cost by bypassing a need for LLM tokens to perform the same task.

​

• We tested the BERT classifier with collected Reddit data and are working to refine it a little further

​

AI-Based Precipitation Nowcasting with GOES-16 Satellite Data

Presenting Team: Shubham Ganpat Khupase, Sahana Madhugiri Shankar

​

Faculty Advisor: Dr. Yaqian He

Presenter: Shubham Khupase

​

FADS Research Project · DSCI-D 590 · Indiana University Bloomington

AI-Based Precipitation

Nowcasting with GOES-16 Satellite Data

Developing and evaluating AI-based nowcasting methods using GOES-16 satellite imagery benchmarked against traditional Simple Extrapolation, S-PROG, and STEPS

Project Overview

Dataset: GOES-16 RRQPE satellite rainfall estimates over Arkansas (30 time-steps, 10-min intervals)

Traditional Baselines: Simple Extrapolation · S-PROG · STEPS (via pySTEPS)

ML Methods: ConvLSTM · XGBoost · XGBoost Quantile

Evaluation Metrics: RMSE · Fractions Skill Score (FSS)

Sahana Madhugiri Shankar

Re M.S. Data Science

Shubham Ganpat Khupase

M.S. Data Science

Advisor: Prof. Yaqian He · Indiana University Bloomington

Traditional

Extrapolation

Traditional

S-PROG

Traditional

STEPS

ML Method

XGBoost

ML Method

XGBoost Quantile

ML Method

Vision Transformer

MOTIVATION

Accurate short-term forecasting (0–6 hrs) is essential for extreme precipitation disaster preparedness.

Performance Findings

• Short-Term Advection: Skill peaks in the first 10–20 min. Adding optical flow lifts XGBoost FSS to 0.53 for heavy rain, bridging the physics gap.
• Long-Range Context: ViT outperforms both XGBoost variants on RMSE (1.18 vs 1.32 at 10m) and FSS at lead times > 10 min.
• The Core Challenge: Forecast skill collapses past 30 min (FSS < 0.2) across models due to high-frequency echo smoothing.

Machine Learning Models

Quantile XGBoost — No Optical Flow

Captures nonlinear pixel-wise relationships but lacks explicit spatial-motion context.

Quantile XGBoost — With Optical Flow

Integrates kinematic motion vectors (advection) to boost spatial tracking of severe storms.

Vision Transformer (ViT)

Uses global self-attention to map storm morphology without strict physical priors.

Metrics

• FSS (Fractions Skill Score): Evaluates spatial scale accuracy of precipitation thresholds.

• RMSE (Root Mean Square Error): Measures point-wise forecast error.

Model Trade-offs

ViT: over-smooths high-frequency extremes.

XGBoost: underestimates peak rainfall intensities.

XGBoost Quantile: captures upper-bound risk percentiles.

Analyzing the #MeToo Movement in the World’s Press

​

Faculty Advisor: Dr. Julia Dessauer

Presenter: Cindy Tantowibowo

​

Analyzing the #MeToo Movement in the World’s Press

How did news media across the globe frame #MeToo and similar feminist movements and their backlash?

WHY THIS MATTERS

Media coverage shapes public understanding of feminist movements

Feminism seeks equal human rights and holds those who brought injustice against women accountable. Movements such as #MeToo brought that reckoning into public view, yet most people encounter it through the news, and outlets differ in what they emphasize. This research analyzes how coverage of such movements varies across the world, and why.

WHAT WE WANT TO EXPLORE

▸ Where coverage came from, by country

▸ Support vs. backlash framing

▸ How coverage rose and fell over time

▸ News reporting vs. opinion / editorial

METHOD

1. Collect: Gather worldwide news using a multilingual set of movement terms
2. Classify: Code article stance and validate
3. Analyze: Compare framing across countries and examine what explains the differences

TOOLS & PIPELINE

GDELT

Global news knowledge graph, via Google BigQuery

Python

Codebook, local keyword matching, enrichment

LLM coding

Stance labels with reasoning using Gemini

Dashboard

Interactive world map, timeline, country table

FACULTY Dr. Julia Dessauer · Dr. Nilesh Shinde

PRESENTED BY Cindy Tantowibowo

212,957 news articles were retrieved, half from five countries

WHERE THE COVERAGE CAME FROM *

* Darker fill = more coverage. The United States alone accounts for 1 in 4 articles.

COVERAGE MONTH BY MONTH

PRELIMINARY INSIGHTS

Coverage is heavily concentrated

The US, France, India, Mexico and Argentina accounts for half of the coverage.

The US alone produced 60,000 articles, more than the next four combined. 56% of the articles were #MeToo-related, explaining the US’ prevalence.

It follows the news cycle

Coverage spiked around major moments, the largest in October 2018.

The Kavanaugh hearings drove the single biggest spike, at ~19,000 articles.

An interactive dashboard

Live world map, monthly timeline, and country-by-country breakdown in HTML format.

NEXT Advanced data cleaning, article classification (support/backlash), pattern clarification, final analysis

Visualizing Vulnerability Exposure in the AI Open-Source Software Ecosystem: �An Interactive Monitoring Dashboard

​

Faculty Advisor: Dr. Sagar Samtani

Presenter: Princy Reshma Ramaseshan

​

Visualizing Vulnerability Exposure in the AI Open-Source Software Ecosystem: �An Interactive Monitoring Dashboard

The Problem

• Millions of AI models on Hugging Face reference GitHub repositories carrying undetected security vulnerabilities
• No existing tool monitored this risk at ecosystem scale across models and downstream dependencies

​

System & Dashboard

• Resolves GitHub links from 2.6M+ model cards, clones repos, and runs Semgrep static analysis capturing severity labels and CWE mappings
• 3-page app (Python, Dash, Plotly) on IU Jetstream2 covering Overview, Explore, and Repo Detail with sub-second queries

​

Real-World Impact

• Helps security analysts, AI/ML engineers, and HPC stakeholders identify high-risk repositories with disproportionate downstream exposure
• Links repo-level findings to specific HF models enabling blast-radius-aware prioritization

Faculty Mentor: Dr. Sagar Samtani · Team: Princy Reshma Ramaseshan · Indiana University DSAIL · FADS Research Showcase 2026

One vulnerable repository like huggingface/transformers exposes up to 7,259 HF models, demonstrating why blast-radius-aware prioritization is critical

2.6M+

HF Models Scanned�across model cards analyzed

​

23,725

Repos Monitored�AI/ML GitHub repositories

​

​

750K+

Vulnerabilities�across all severity levels

​

​

132,317

High Severity�17.6% of all findings

​

​

​

495,529

Models at Risk�via one vulnerable repo

​

​

47.2

Avg / Repo�findings per repository

Dashboard

Learnings:

• Designed scalable pipeline integrating HF model cards, GitHub repos, and Semgrep on IU HPC�• Built interactive multi-page Dash dashboard supporting ecosystem-to-repo-level drill-down

Future Plans:

• Integrate dynamic scanners and longitudinal vulnerability trend tracking�• Add AI-enabled summarization and real-time monitoring capabilities

• 23,725 repos monitored�• 750,281 vulnerabilities detected�• Risk gauge, monthly scan trend

Page 1: Overview

• 23,725 repos · 750,281 vulnerabilities · 132,317 high severity
• Monthly trend · Risk gauge (score 83)
• Recently active high-risk repos

Overview (cont.):

• 495,529 HF models at risk
• Treemap: size = exposure, color = severity
• One repo exposes up to 86,848 models

Page 2: Explore

• Filter by severity, CWE, language, risk tier
• Top CWE categories · Sortable repo table
• Drill down to specific vulnerability rules per repo

Page 3: Repo Detail

• Per-repo findings: severity, CWE, file path, line number
• Vulnerability details with remediation guidance
• Blast radius: 7,259 HF models linked to this repo

Measuring AI Adoption in State Public Management

​

Faculty Advisor: Dr. Sanghee Park

Presenter: Virija Nandamudi

​

AI ADOPTION INTENSITY INDEX (v2)

Virija Nandamudi

WHICH AI STATE AGENCIES ARE USING AI AND HOW MUCH?

States

Sectors

50

4

State × Sector Cells

200

I(s,y) = Σd [ wd · Σt ( τt · Σr εr ) ]

domain weight × tier weight × evidence weight per record

1 · Crawl & fetch

2 · Coverage gate

3 · Text extraction

4 · Keyword matching

5 · Deduplication

6 · Evidence gating

7 · Jurisdiction resolution

8 · LLM classification

9 · Scoring & export

Intensity index · need score · AI gap

Technology Policy

Social Services

Procurement

Legislative

Prof. Sanghee Park · FADS Research Showcase 2026

BUILT FOR REPRODUCIBILITY. DESIGNED AGAINST SILENT FAILURE.

Rule matches with confidence ≥ 0.70 bypass the LLM — preserving Groq free-tier quota and model consistency across the run. Every record carries classification_source for audit traceability.

MinHash + sentence-transformers handle deduplication. LLM handles evidence class assignment, because "Indiana uses risk scoring" vs. "Indiana faces pressure to adopt risk scoring" is a grammatical distinction, not a cosine distance.

Gater-Confidence Routing

Embeddings ≠ Classification

Year attribution, seed mismatches, SharePoint/ASPX failures, and cross-state jurisdiction conflicts are flagged explicitly, never silently imputed, keeping intensity scores confound-free.

Silent Failure Controls

Virija Nandamudi

Thanks for attending!

Recognitions will be emailed to you by August 10th