This webinar introduces the newly formed SCOP Artificial Intelligence (AI) working group, dedicated to integrating AI safely and equitably into pediatric cancer care and research globally. The discussion covers the widespread adoption of Large Language Models (LLMs) in clinical practice, highlighting their benefits for tasks like research and email drafting, while also addressing significant concerns regarding data privacy, "hallucinations" (false information), and the need for human validation. A central feature is the live demonstration of "Capricorn," a specialized AI-powered search tool developed in collaboration with Google. Capricorn aims to assist pediatric oncologists by rapidly analyzing patient data against a vast medical literature database (PubMed), providing evidence-based treatment suggestions, with a strong emphasis on transparency, data security, and physician oversight. The session concludes by emphasizing AI's role as a supportive tool to enhance, rather than replace, human expertise, and encourages participation in the working group to foster responsible AI development and education.
Summarizing the AI in Pediatric Oncology Webinar: SCOP Working Group & Capricorn Tool
0:00:54 Introduction & Logistics: Speakers confirm audio, discuss slide sharing and pre-webinar polls, and note current heatwave conditions.
0:02:47 Webinar Attendance: 85 people registered, with typically half attending live.
0:04:48 Future AI Initiatives: Plans for a second webinar and an in-person event during SCOP AI in Amsterdam, hosted at Google offices, focusing on new AI developments in health.
0:05:45 Event Registration Forms: Discussion on using tools like App Connect or Survey Monkey/Woofu for on-site event registrations, differentiating from Zoom webinars.
0:15:56 SCOP AI Working Group Introduction: Pablo Elasco (pediatrician, Balbron Hospital, Barcelona) introduces the newly created SCOP Artificial Intelligence Working Group, co-coordinated with Friso Kalcoen and Uri Elan from Princess Maxima Center.
0:16:51 AI's Impact & Challenges: AI has revolutionized science and daily life, with thousands of publications showing advances in pediatric cancer (diagnosis, omics, imaging, synthetic data). However, it's hard to stay updated and safely implement AI in clinical practice.
0:17:35 Group's Purpose & Objectives: To create a multi-stakeholder platform connecting developers, experts, clinicians, and tech companies. Aims include improving AI education/training (webinars, meetings), promoting collaboration and research, engaging industry for sustainable projects, and ensuring tools are accessible to all countries, regardless of resources.
0:19:42 Focus on Large Language Models (LLMs): Differentiates complex AI projects from more accessible LLMs (ChatGPT, Perplexity, Google Gemini, Grok) that are increasingly used for daily tasks like writing and searching.
Grok's suggestion: More detailed, mentioning specific chemotherapy (FLAGA), targeted therapies, CAR T-cell therapy, supportive care, and clinical trials.
Correctness & Trust: Participants questioned the veracity of AI's output without references and the risk of patient data input.
Anonymization Concerns: Even general patient details can potentially de-anonymize in rare cases (e.g., 7-month-old male with ALL relapse post-transplant in the Netherlands).
Human Oversight is Crucial: Consensus that AI is a tool, not a replacement; human validation, tumor boards, and checking references are essential. AI cannot account for individual patient clinical situations or tolerance.
0:35:34 Role of Champions: The attendees, as "champions of AI," are tasked with spreading knowledge on proper and safe AI usage, especially among younger doctors who may already be using LLMs with patient data.
Uri Elan introduces Capricorn, a tool developed with Google to address the limitations of traditional search engines (like PubMed) for complex patient data.
Goal: To allow unstructured patient data input and have AI read papers, analyze, and summarize evidence-based treatment options.
Funding Model: Developed pro-bono by Uri (medical input) and Google engineers (technical support); aims to be free for pediatric oncologists worldwide, sustained by Princess Maxima and future grants.
Security & Privacy: Enterprise version, data is kept safe, Google Gemini not trained on it, no one can access user data, user can review previous cases. Includes a "redaction" feature to remove identifying information.
Functionality: Processes patient history and lab results, searches 37 million PubMed papers in real-time, extracts "actionable events" (e.g., KMT2A, CD22, FLT3 ITD), and ranks papers based on an algorithm (journal impact, year, disease match, clinical trial vs. case report, drug testing).
Transparency: Users can see the scoring for each paper and directly access the original paper for validation.
Output: Provides a summary of the case, actionable events, and suggested treatments with evidence, including potential toxicities and response signs (e.g., VOD for inotuzumab). Can export as PDF for tumor board discussions.
Improved Accuracy: More data input leads to better, more holistic results, targeting disease pathways in multiple ways.
Availability: Currently a test version, link to be shared in the AI working group page for pediatric oncologists to try and provide feedback.
AI's Tendency to Answer: Acknowledged that LLMs "try to please," suggesting prompts like "if you don't know, say I don't know" for better results.
Capricorn's Role (not a recommender): Designed as a tool for pediatric tumor boards to explore possibilities, not to provide direct treatment suggestions that bypass physician judgment.
Toxicity Information: Capricorn does highlight expected toxicities in its output.
Integration with PubMed: In talks with NIH to implement Capricorn's search methodology directly into PubMed.
Performance Comparison: Early data shows AI found 10% more targets and 20% more drugs than human analysis in specific cases.
User Access: Currently, self-identification; future plans for Maxima/SCOP environment requiring login and approval for pediatric oncologists.
Open Sourcing: Discussion on open-sourcing Capricorn or running it freely on the Maxima environment to prevent dependency on Google and ensure long-term availability.
"Hallucination" Prevention: Capricorn is designed to be "grounded" (only provide data from papers) and not hallucinate references.
0:59:07 Parental Use of AI: Debate on whether parents should use such tools, acknowledging their deep engagement but emphasizing that physicians have the expertise to evaluate information more critically. Suggestion of separate patient/doctor versions for AI tools.
Future aims include wider global implementation via Google Health and integrating with clinical guidelines.
Upcoming SCOP talk on using AI for tumor boards (humans watching, AI discussing patients).
AI as an Enhancer: AI is not here to replace physicians but to help them be better. While some studies suggest AI can outperform human-AI combinations (due to human bias), AI offers opportunities for education and decision support.
"You will not be replaced by AI, but you will be replaced by someone using AI." Encouragement to actively engage with and learn to use AI.