Precision Medicine Series: Diagnostics, Artificial Intelligence, and the Future of Clinical Practice

This summary was automatically produced by experimental artificial intelligence (AI)-powered technology. It may contain inaccuracies or omissions; see the full presentation before relying on this information for medical decision-making. If you see a problem, please report it to us here.

Introduction

This presentation by Peter McCaffrey, M.D., a pathologist and AI leader at UTMB and the UT System outlines a bullish perspective on the value and transformative potential of AI in healthcare, particularly focusing on Large Language Models (LLMs) and their implications for clinical practice.

The speaker acknowledges the anxieties surrounding impending regulation and the ethical considerations but emphasizes that the healthcare field has always dealt with powerful tools that require professional discretion.

The Evolving Landscape of AI, Especially LLMs

• Accelerated Technological Improvement: Technology, especially AI, is improving at an ever-quickening rate, driven by increased data availability, advancements in GPUs, and closed feedback loops from deployment.
• Model Capacitance and Emergent Intelligence: Modern deep learning models, particularly LLMs, have billions or even trillions of parameters, allowing them to discover complex functions and connect disparate concepts, leading to "emergent intelligence".
• Hallucinations as an Inherent Property: LLMs are fundamentally "hallucinatory modules" that "riff" on text predicates. Their non-deterministic nature means it is impossible to guarantee perfect performance, making rare errors (e.g., 1 in 1,000 or 1 in a million) significant in a healthcare context.
• The Importance of Context Augmentation (RAG): The true value and future of AI in healthcare, beyond basic linguistic improvement, lie in "grounding" or "context augmentation" (Retrieval Augmented Generation - RAG). This involves providing LLMs with additional, relevant data to guide their responses and constrain discussions, which is critical for utility in healthcare settings.
• The "Dumbing Down" Risk: If AI models are increasingly trained on content generated by other AI models, rather than human-generated "ground truth," it could lead to diminishing returns or even a "dumbing down" of intelligence over time.

Transformative AI Use Cases in Diagnostics and Screening

The core idea is that AI can provide "inference" – the valuable, thought-based work of healthcare professionals – which is currently a limited and costly resource. Even partial or sub-optimal AI inference can be transformative by addressing situations where no human inference currently occurs.

The speaker presents a matrix of AI use cases based on existing/new diagnostic methods and classical/new modalities:

• Existing Methods on Classical Modalities: AI can automate and scale existing practices, like estimating gestational age from sonography or detecting coronary artery calcification (CAC) on CT scans.
  • Incidental Coronary Artery Calcification (ICAC): AI can act as a "safety net" by identifying high CAC scores on routine, non-cardiac CTs. This flags patients who may be at high risk but have never seen a cardiologist, potentially increasing statin adherence and facilitating earlier interventions.
  • Mammography: Studies show that a human plus AI double-read is non-inferior to a human plus human double-read for mammogram interpretation, potentially alleviating the burden of human double-reads.
• New Methods on Classical Modalities: Using existing data for new purposes, such as screening for Type 2 Diabetes on a chest X-ray.
  • Subclinical Anemia (LGI flag): AI can analyse subtle trends in common, inexpensive CBC (complete blood count) tests to predict slow trends towards anemia, allowing for earlier triage for conditions like colon cancer.
  • Radiomics: AI can detect subtle, imperceptible patterns in imaging (e.g., high-resolution EKG data or unreleased blood sample values from lab tests) that humans cannot visually discern, leading to early detection of conditions like low ejection fraction.
• New Methods on New Modalities: Emerging use cases involving novel data sources, such as assessing early Parkinson's disease risk from keystroke patterns or depression recognition from voice patterns.
• Liquid Biopsy (MCED): While promising for early cancer detection from peripheral blood samples, these tests are expensive and produce a significant number of false positives, creating logistical and financial burdens related to follow-up care.

The speaker suggests that the future trend will be towards AI-driven screening and multi-AI workflows, with humans progressively less involved in the initial screening steps, acting more as confirmatory adjudicators.

Challenges and Implications for Healthcare Practice

• The Information Hazard: If AI tools can discover latent conditions, even if healthcare institutions choose not to run them, patients (or legal firms) could obtain their medical data, run AI tools themselves, and identify "missed" diagnoses. This creates a significant legal and ethical obligation for healthcare providers to address discovered conditions.
• Deluge of Adjudication Tasks and Bandwidth Limitations: The widespread use of AI for screening will inevitably generate a massive volume of "flags" or potential diagnoses, many of which will be false positives. Current healthcare systems lack the bandwidth to manage this deluge of adjudication tasks, potentially leading to the emergence of new roles focused solely on clinical adjudication.
• Impact on Physician Workflow and Reimbursement (RVUs):
  • AI can significantly automate the "preparative steps" of diagnosis (e.g., information retrieval and synthesis from EHRs), which can consume up to half of a physician's time.
  • Since the RVU (Relative Value Unit) reimbursement model often ties compensation to the time taken for a task, AI's efficiency gains could reduce the perceived time component of physician work, potentially lowering per-case reimbursement.
  • However, this efficiency could also allow physicians to process higher volumes of cases or shift their focus to new, complex diagnostic management tasks, broadening their scope.
• The "Jagged Technological Frontier": While AI can standardize work products, potentially diminishing the "uniqueness" of individual physician's reporting style, standardization is often preferred for quality and consistency in clinical settings.
• Reliability and Trust: Given the non-deterministic nature of generative AI, ensuring reliability and building trust is paramount. The speaker suggests that for chat-based AI in healthcare, the focus should be on user interfaces that facilitate efficient citation and ground truth comparison, allowing physicians to quickly verify AI-generated information.
• Validation Challenges: Scaling validation for AI models, particularly generative ones, is difficult. Solutions may include human ground truth annotation (potentially compensated) and adversarial AI pipelines (where other AIs critique the initial AI's output).
• Anchoring Bias: The order in which AI and humans interact matters. If AI provides its opinion first, it can anchor human judgment, potentially leading to missed alternative findings. Rigorous studies are needed to understand and mitigate this bias.

UTMB's Governance and the Regulatory Outlook

UTMB has established a structured approach to AI deployment:

• Centralized Intake and Championing: All AI tools undergo a centralized intake process to define requirements and identify clinical champions. Without strong champions, adoption does not proceed.
• Validating Pilots and Strategic Oversight: Successful proposals proceed to validating pilot programs before capital investment. An AI General Council oversees large capital outlays, liability issues, and high-level strategic prioritization.
• Impending Regulation: Texas legislation (e.g., TAIGA, HB1709) is in session, aiming to define "high-risk" AI use cases (including healthcare) and establish requirements for "developers" and "deployers." These regulations will likely mirror CLIA standards, requiring auditability of AI use, data, accuracy, and bias detection.
• Need for Dedicated Validation Environments: To meet future regulatory demands and ensure safe, effective deployment, healthcare systems will need dedicated, real-data-simulating environments (e.g., within Epic) for continuous pre- and post-deployment validation of AI tools.

An Optimistic Future for Healthcare Professionals

The speaker concludes on an optimistic note, viewing AI as a "cognitive utility" or "power company" that infuses intelligent capabilities into healthcare workflows. This means:

• Greater Access to Inference: AI will enable more people to receive more medical inference in more situations, bridging the current gap between desired and deliverable healthcare.
• Broadened Scope for Professionals: Instead of diminishing their role, AI will supercharge healthcare professionals, broaden their scope and allowing them to become "navigators and managers" of diagnostic journeys and the tapestry of intelligence flowing through the community. This shifts the role from artful, point-in-time consultation to comprehensive diagnostic management.
  

Artificial intelligence (AI) was used to transcribe the presentation’s contents and create a summary of the information contained in the presentation. This is an experimental process, and while we strive for accuracy, AI-generated content may not always be perfect and could contain errors. This summary has not been reviewed by the presenter to guarantee completeness or correctness of the content, so it should not be used for medical decision-making without reviewing the original presentation. If you have feedback, questions, or concerns, please contact us here.