St. Jude Researchers Leverage Complex AI Prompting to Detect High-Risk Symptoms in Childhood Cancer Survivors

St. Jude scientists use complex AI prompting to analyze patient conversations, helping doctors identify long-term health risks in childhood cancer survivors.

By: AXL Media

Published: Mar 31, 2026, 8:54 AM EDT

Source: Information for this report was sourced from St. Jude Children's Research Hospital.

Unlocking Conversational Data to Improve Survivorship Care

The transition from active cancer treatment to long-term survivorship is often marked by a "ripple effect" of health challenges that can persist for decades. While 40% to 60% of a clinical encounter consists of patients discussing their symptoms, much of this nuanced conversational data remains underutilized because it is trapped in lengthy transcripts or open-ended survey responses. Researchers at St. Jude Children’s Research Hospital have developed a proof of concept showing that artificial intelligence (AI) can bridge this gap. By analyzing interviews with survivors and their caregivers, large language models (LLMs) can now help physicians detect symptom severity and functional disruptions that might otherwise go unnoticed in a standard review.

The Necessity of Sophisticated AI Prompting Strategies

A central finding of the research is that the effectiveness of AI in a clinical setting is dictated by the "art of the prompt." The study compared four distinct styles of instruction given to models like ChatGPT and Llama: zero-shot, few-shot, chain-of-thought, and generated knowledge. According to I-Chan Huang, PhD, simple prompts that provided minimal context were found to be unstable and inaccurate. In contrast, complex strategies performed significantly better, showing a high level of concurrence with human experts. This suggests that future clinical AI integrations must move beyond basic instructions toward more logical, step-by-step frameworks to be considered reliable.

Methodology and Gold Standard Human Comparison

To test the models, scientists interviewed 30 survivors between the ages of 8 and 17, along with their primary caregivers. Two human experts performed a "gold-standard" analysis of the resulting transcripts, identifying over 800 pieces of information related to excessive pain and fatigue. These experts categorized symptoms based on their physical, cognitive, and social impacts. When the same data was processed by the AI models using sophisticated prompting, the machines demonstrated a comparable ability to categorize physical and cognitive disruptions. However, the models showed only moderate success in detecting social impacts, highlighting an area where human intuition still maintains a clear advantage.