Day 2: Panel Discussion: From Foundation to Bedside: Will AI Finally Deliver Health Impact

Value of the Information & Strength of the Argument

The panel provides valuable insights into the practical challenges of deploying AI in healthcare, moving beyond theoretical discussions. The arguments are well-founded, drawing on real-world experiences and specific examples. For instance, Thomas’s account of building an ECG foundation model with Google and the Ministry of Health illustrates the importance of stakeholder engagement and data harmonization. Hernan’s distinction between pattern recognition and causal inference is a crucial conceptual contribution, clarifying why many AI models fail to answer ‘what if’ questions. The panelists agree on the need for clinician involvement and alignment with clinical workflows, but they also highlight systemic issues like data silos and regulatory hurdles. The discussion is balanced, acknowledging both the potential and the limitations of current AI technologies.

Scientific Rigor, Source Quality, Title Accuracy

The panelists are recognized experts in their fields, lending credibility to the discussion. However, the conversation is largely anecdotal, with no formal citations or references to specific studies. The title accurately reflects the content, focusing on the gap between AI development and clinical impact. The discussion is scientifically rigorous in its conceptual framing, particularly Hernan’s explanation of causal inference. However, the lack of concrete data or references to published work limits the ability to verify claims. The panel does not address potential biases or limitations of the discussed approaches in depth, but overall, the scientific quality is high.

The title accurately reflects the panel's focus on the translation of AI from research to clinical practice.

7/10

Panel of experts with strong academic credentials, discussing real-world deployment challenges. No formal citations, but practical insights and references to ongoing projects. Some claims are anecdotal, but overall reliable.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction of panelists and framing of the discussion on AI's path to clinical impact.
• 05:00Farah Shamout discusses three main challenges: lack of evaluation ecosystem, data integration issues, and clinical utility.
• 08:00Huazhu Fu highlights the importance of involving clinicians in AI design and the need for clinically relevant metrics.
• 10:00Rajat Mani Thomas shares the ECG foundation model example and the importance of stakeholder management.
• 12:00Miguel Hernan clarifies that AI is data analysis and distinguishes pattern recognition from causal inference.
• 15:00Discussion on trust and validation, with Shamout emphasizing early clinician involvement and alignment with incentives.
• 18:00Thomas describes building an ECG foundation model at scale and focusing on a narrow use case to demonstrate value.
• 21:00Fu discusses the potential of AI in low-resource regions and the need for deployment where clinician shortages exist.
• 24:00Hernan explains the role of randomized trials as benchmarks for causal questions and the potential of digital twins.
• 27:00Panelists discuss how lagging countries can leapfrog by leveraging accessible data and building local solutions.

• ADIA Lab Symposium 2025 — The panel discussion is part of this symposium, and the description mentions the event.

• AI in Healthcare: Challenges and Opportunities — This article discusses similar challenges in AI deployment, aligning with the panel's points.

• Deep Learning for ECG Analysis: A Review — While the panel suggests that ECG models may not translate across populations, this review highlights the generalizability of deep learning models, presenting a contrasting view.