# Are Generative AI systems Capable of Supporting Information Needs of Patients?

**Authors:** Shreya Rajagopal, Subhashis Hazarika, Sookyung Kim, Yan-ming Chiou, Jae Ho Sohn, Hari Subramonyam, Shiwali Mohan  
**arXiv:** [2402.00234](https://arxiv.org/abs/2402.00234)  
**Date:** 2024-01-31  
**Summary page:** https://www.summarizepaper.com/en/paper/2402.00234v1/

## Key points

- Patients managing complex illnesses like cancer face a daunting information challenge
- Close interaction with healthcare experts improves patient learning and disease outcome
- Generative AI models are being explored to support patient information needs in radiology imaging data
- Thematic analysis identified common themes in discussions between participants and radiologists
- Evaluation of ChatGPT-4V and MedFlamingo showed varying levels of response quality
- Current generative AI systems may not adequately address patients' information needs for medical scans and reports
- High error rates were observed during evaluation, raising concerns about misinformation being conveyed to patients
- Models struggled with relevance and frequently produced irrelevant elaborations

## Summary

Patients managing complex illnesses like cancer often face a daunting information challenge. They not only need to understand their condition but also learn how to effectively manage it. Close interaction with healthcare experts has been shown to improve patient learning and ultimately their disease outcome. However, this approach is resource-intensive and can take expert time away from other critical tasks. With recent advancements in Generative AI models aimed at improving the healthcare system, researchers set out to investigate whether and how generative visual question answering systems could responsibly support patient information needs in the context of radiology imaging data. Through a formative need-finding study involving discussions of chest computed tomography (CT) scans and associated radiology reports with a cardiothoracic radiologist, participants highlighted common themes such as clarifying medical terminology, locating issues mentioned in reports within scanned images, understanding disease prognosis, discussing diagnostic steps, and comparing treatment options. Thematic analysis of these interactions led to the identification of 91 content-based codes grouped into 10 broad themes that captured the essence of participant-radiologist conversations. Evaluation of two state-of-the-art generative visual language models - ChatGPT-4V and MedFlamingo - against responses provided by the radiologist revealed varying levels of response quality across different themes. While MedFlamingo tended to provide concise answers in clinical language that may be challenging for patients and caregivers to understand, ChatGPT-4V generated lengthy responses with generic descriptions without focusing on specific case details or filtering relevant information. The study's findings suggest that current generative AI systems may not adequately address patients' information needs when it comes to understanding medical scans and reports. High error rates were observed in both models during evaluation on real interaction questions, raising concerns about misinformation being conveyed to patients who may not possess medical expertise to discern inaccuracies. The models also struggled with relevance and frequently produced irrelevant elaborations rather than directly addressing questions asked. Overall, the research highlights a critical gap in evaluating generative AI systems for practical healthcare applications and underscores the challenges of ensuring accountable deployment in real-world scenarios where accuracy and relevance are paramount for supporting patients' informational needs effectively.