The case for caution: current benchmarks often reflect pattern recognition rather than true reasoning. Concerns include context matching instead of reasoning, systematic base rate neglect, and reasoning traces that do not reflect actual model computation. #LLM #AI #TMC #MSBMI #Clinicians
Benchmark and clinical studies show strong performance in some settings, but results vary. Evidence highlights both potential strengths and ongoing limitations in diagnostic accuracy and reasoning. #LLM #AI #TMC #MSBMI #Clinicians
LLMs are being used across a wide range of tasks, including triage, differential diagnosis, documentation, test ordering, referral decisions, coding, and treatment planning. #LLM #AI #TMC #MSBMI #Clinicians
Research activity has also increased significantly. Publications evaluating LLMs in clinical medicine grew from 1 paper in 2019 to over 500 per year by 2024. #LLM #AI #TMC #MSBMI #Clinicians
Adoption is especially visible in clinical documentation. A 2024 survey found all responding health systems reported active use of AI tools such as ambient listening for documentation. #LLM #AI #TMC #MSBMI #Clinicians
LLMs are being adopted across clinical workflows at a rapid pace. In 2024, 31.5% of U.S. hospitals reported using generative AI, with another 24.7% planning near-term adoption. #LLM #AI #TMC #MSBMI #Clinicians
Diagnostic reasoning involves identifying hidden states from observable evidence. This process is central to clinical diagnosis, fault detection, and troubleshooting across multiple domains. #LLM #AI #TMC #MSBMI #Clinicians
Reporting from the MSBMI research seminar!
We’re joined by Todd Johnson, PhD, who is presenting: "Black Box as a Diagnostic Reasoning Benchmark for LLMs" #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC #LLM
The project proposes a structured set of domains and subdomains to capture a wide range of override rationales. This approach supports more consistent analysis across different settings. #CDS #MSBMI #UTHealth #AI #Clinicians
Current override analyses are difficult to compare across institutions because there is no consistent way to categorize user-reported reasons. This limits the ability to draw broader conclusions from existing data. #CDS #Clinicians #Healthcare #UTHealth #MSBMI
Why should we collect and analyze override rates? Some reasons include: Understanding why providers or patients are not following recommendations, measuring and understanding the burdens of CDS on providers/end users, and capturing patient and provider issues with recommendations. #CDS #MSBMI #AI
Override rates vary but have been shown to be as high as 90% for some CDS. High rates of incorrect, irrelevant, uninformative, or non-specific CDS alerts and recommendations contribute to alert fatigue and clinician burnout. #CDS #AI #Clinicians #Healthcare #MSBMI #UTHealth #TMC
This work focuses on developing a standardized taxonomy to describe why users override clinical decision support alerts. Understanding these reasons is important for improving safety, relevance, and adoption of CDS tools. #CDS #AI #Clinicians #Healthcare #MSBMI #UTHealth #TMC
Reporting from the MSBMI research seminar!
We’re joined by Rachel Richesson, PhD, MPH, FACMI, FAMIA, who is presenting: "Approaches to Standardizing Override Reasons for Clinical Decision Support" #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC #CDS
Effective communication depends on accurately assessing shared understanding with an audience. Framing provides context that guides how a message is interpreted. Clear and intentional language is especially important when communicating with non-technical audiences. #AI #UTHealth #MSBMI #Informatics
Just as with audience design, prompt optimization involves adjusting an existing prompt to improve output. Prompt engineering formalizes something humans already do intuitively. #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC #Promptengineering #Informatics #ArtificialIntelligence
The seminar explains how informatics is presented across different contexts, including funding proposals, academic collaboration, student recruitment, and public communication. The focus currently is on adapting language and framing to suit specific audiences. #MSBMI #AI #Communication
Reporting from the MSBMI research seminar!
We’re joined by Amy Franklin, PhD, who is presenting: "Prompt Engineering for People: Designing How We Talk About Informatics" #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC #Promptengineering #Informatics #ArtificialIntelligence #Seminar
An important component of the study is the use of standardized representations for contextual factors. This helps support interoperability across organizations and makes independent analyses more comparable while preserving local control of data. #TMC #MSBMI #UTHealth #Healthcare #Informatics #AI
This seminar explores how to use information already captured in EHRs to identify unwarranted variation at the site level. By focusing on local contextual factors, the approach aims to detect absolute variation rather than relative differences alone. #TMC #MSBMI #UTHealth #Informatics
Centralized analyses also raise practical concerns. Clinical variation data can be sensitive, and organizations may be hesitant to share operational details. When access is restricted, reporting and comparison become difficult. #TMC #MSBMI #UTHealth #Healthcare #Informatics #Research #Seminar #AI
A challenge with identifying unwarranted variation is that current approaches tend to focus on regional or site-level patterns. These methods can show where variation exists, but often cannot explain what is driving it. #TMC #MSBMI #UTHealth #Healthcare #Informatics #Research #Seminar #AI #health
Unwarranted clinical variation refers to care that does not align with a patient’s needs or clinical characteristics. It often stems from local contextual factors and can lead to higher costs, unnecessary interventions, and departures from evidence-based care. #UTHealth #MSBMI #Informatics #EHR
Reporting from the MSBMI research seminar!
We’re joined by Apollo McOwiti, who is presenting: "Predictive Machine Learning Algorithm for Identifying Unwarranted Clinical Variation from Contextual Factors Derived from EHR Encounter Data" #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC
Building on that point, Dr. Roberts discussed an alternative approach: using the same large language model multiple times and in different ways. This enables researchers to more effectively surface and work with the information already embedded in the model. #LLMK #MSBMI #AI #NLP #UTHealth #TMC
LLMs can make everyone an NLP expert- or can it?
Dr. Roberts believes accessibility is NOT the same as expertise. While people can do something, does that mean they really know how to do it correctly? #NLP #LLM #AI #BiomedicalInformatics #Informatics #UTHealth #MSBMI #TMC #Largelanguagemodel
Reporting from the MSBMI research seminar!
We’re joined by Kirk Roberts, PhD, MS, who is presenting: "The Rise of LLMs from an NLP Methodologist’s Perspective" #MSBMI #AI #BiomedicalInformatics #Healthcare #UTHealth #TMC
Closing out with a better understanding of how the medical humanities can help us think more carefully. We covered ethical concerns such as privacy risks, misinformation, underrepresentation in datasets, as well as the importance of bringing unique perspectives into the development process. #MSBMI
Several ethical concerns can emerge when new technologies are used in healthcare, including privacy risks, misinformation, underrepresentation, uneven treatment recommendations, and the potential to worsen existing disparities. These issues underscore the need for careful design and overview #MSBMI
This seminar touches on how the medical humanities can help us look more closely at the assumptions that are changing technology in healthcare. These approaches encourage teams to pause, ask different questions, and consider how tools will be experienced by people. #MSBMI #AI #UTHealth #Healthcare
