Jia Xu.

Projects

Research and applied projects organized by research area, together with the topics of associated publications.

01

Large Language Models and Conversational AI

Benchmark construction, instruction fine-tuning, and rigorous evaluation of large language models, with a focus on conversational mental-health support.

💬

MentalChat16K

A benchmark dataset of 16K counseling conversations that combines GPT-generated multi-turn dialogues across 33 mental-health topics with question-answer pairs drawn from real behavioral-intervention interviews, built for training and evaluating conversational mental-health models. Open models fine-tuned on it (LLaMA-2, Mistral, Vicuna, Zephyr) were assessed by both human experts and LLM judges across seven dimensions; it is now among the most-downloaded mental-health datasets on Hugging Face.

KDD ’25 OralBenchmarkOpen Source
Hugging Face Paper
🧠

MentalGPT

A system of LLMs fine-tuned on MentalChat16K that outperforms base models and baselines on 7 mental-health metrics, evaluated by human experts alongside LLM judges (GPT-3.5, Gemini Pro).

AMIA 2024 PodiumLLM Fine-tuning
Abstract
🎙️

MI Session Analyzer

AI-powered feedback for Motivational Interviewing: Whisper transcription with speaker diarization, and Gemini scoring against the gold-standard MITI coding system, giving counselors instant, structured feedback. Built for the global gap in addiction care, where only 20% of 39.5M affected people access treatment.

1st · MEXA HackathonSpeech AIMITI
Code
🗣️

LLM-Enhanced Dementia Talk Therapy

Fine-tuned large language models that enhance talk therapy for dementia care, trained on GPT self-generated data to expand limited clinical dialogue resources. Presented at AAIC 2024.

AAIC 2024Dementia CareLLM Fine-tuning
Abstract
🍸

MoodShaker

An applied generative-AI system that turns mood, flavor preferences, soberness level, and allergy constraints into personalized cocktail recipes with music pairing. 600+ recipes served in its first week at Penn.

2nd · Wharton Hack-AI-ThonApplied LLM System
Demo Code

📄 Research topics: benchmarking conversational mental-health assistance (KDD ’25) · compassionate mental-health support with fine-tuned LLMs (AMIA 2024) · fine-tuned LLMs for dementia talk therapy using GPT self-generated data (AAIC 2024)

02

Medical Imaging and Computer Vision

Deep learning and ensemble methods for disease detection and image analysis across heterogeneous clinical imaging sources.

👁️

Cross-Source Diabetic Retinopathy Screening

Attention-enhanced ensemble learning for robust detection of referable diabetic retinopathy across heterogeneous fundus image sources, presented at the 55th ARVO Annual Meeting.

ARVO 2026Medical Imaging
🧩

Early Alzheimer’s Prediction

MRI-based classifiers that predict dementia development across cross-sectional and longitudinal cohorts.

Machine LearningNeuroimaging
🧲

BrainSuite Demonstration

Automated MRI processing of the human brain with the open-source BrainSuite toolkit: surface extraction, registration, and analysis.

MRI / fMRIDeep Learning

📄 Research topics: robust referable diabetic retinopathy detection via attention-enhanced ensemble learning (ARVO 2026, The 55th ARVO Annual Meeting)

03

Fairness and Trustworthy Machine Learning

Methodological work on fairness-aware representation learning for equitable analysis of biomedical data.

⚖️

Fair Canonical Correlation Analysis

A fairness-aware formulation of canonical correlation analysis that embeds fairness constraints directly into the CCA objective while preserving cross-view correlation structure, with theoretical guarantees for equitable multiview representation learning. Published at NeurIPS 2023.

NeurIPS 2023FairnessMethod
Paper Code
04

Computational Aging and Population Health

Quantifying biological aging and neurodegeneration with causal inference, mediation analysis, and machine learning on population cohorts, and building informatics tools for healthy longevity.

🔍

Penn AITech Portal

Text mining to build an Alzheimer’s research portal: a knowledge base for identifying technologies that support healthy aging.

NLPKnowledge Base
🏡

ElderlyCo

A dual-market growth strategy for China’s ¥7T senior-care industry: AI-enabled financial planning and personalized care targeting 3-year breakeven with 90%+ occupancy.

1st · Bain Project StarStrategy × AI

📄 Research topics: mediation analysis and structural equation modeling of phenotypic age (IJMI 2025) · biological markers and sociodemographic determinants of the phenotypic vs chronological age gap (IHSC 2024) · causal effects of environmental exposures on aging (IEEE BIBM 2023) · determinants of age-gap prediction in the NHANES cohort (AMIA 2023) · clustering Alzheimer’s disease subtypes via similarity learning and graph diffusion (ICIBM 2023)

05

Spatial and Social Data Science

Geospatial and computational social-science methods applied to health equity, environmental exposure, and public discourse.

🌳

Uncovering Inequities in Green Space

Geospatial analysis of urban green-space access and its community-health implications. 1st place in the Green Space Data Challenge, organized by Georgetown University’s Massive Data Institute; invited to present the winning project at the APDU Annual Conference in Arlington, VA (Jul 2023).

1st · Data ChallengeArcGIS
Code Penn SP2 News Challenge Site
📻

Venezuelan Political Discourse NLP

A GCP pipeline for multi-source political data: fine-tuning Whisper for Spanish political speech and analyzing 100k+ news articles to quantify sentiment shifts in the Chávez era. With UPenn, Columbia and UC Berkeley.

Speech-to-TextNLP at Scale
💨

Air Quality Index Prediction, India

Big-data analysis of air pollution across India, winner of the Best Visualization award in Penn’s CIS 545 Big Data Analytics.

Best VisualizationBig Data