CODEX Digest - 6.18.26

1) Clarifying the muddle: towards a comprehensive taxonomy of cognitive biases in medicine.

Amoretti MC, Lalumera E. Med Health Care Philos. Epub 2026 Apr 9.

Cognitive biases are a major contributor to diagnostic error, but their effects depend not just on how they work cognitively, but also on the medical situations in which they occur. This commentary offers a framework that distinguishes constructive from hindering biases and calls for a more nuanced, context-sensitive approach to improving diagnostic accuracy.

2) Evaluation of symptom checker formats to support health literacy and trust in AI: results from an online randomised-controlled trial. (This is a preprint that has not yet gone through the peer review process.) 

Ayre J, Gallagher K, Smith J, et al. medRxiv. Epub 2026 Mar 12.

Digital symptom checkers are increasingly used by patients for self-diagnosis and disease management, yet the impact of explicitly disclosing generative AI involvement remains underexplored. This Australian preprint randomized trial of adults given a case of hypothetical symptoms finds that AI assisted advice improved self-management knowledge and lowered stated intention to go to the doctor for a low-acuity illness. Importantly, disclosing AI use did not reduce user trust or acceptability, supporting safe integration of AI-powered tools without undermining patient confidence.

3) Implementation of the NHS England Lung Cancer Screening Programme over 5 years.

Lee RW, Nair A, Balata H, et al. Nat Med. 2026;32(5):1817-1826.

Low-dose CT screening improves timely diagnosis of lung cancer and all-cause mortality. This commentary reviews the first five years of England’s national program, which invited more than two million people and diagnosed 7,193 lung cancers with an increase in early-stage diagnosis and improvements among high-risk and underserved groups. The report highlights large-scale implementation policies necessary to achieve such a a program.

4) An ethics-informed computable audit framework for monitoring misdiagnosis risk in AI-assisted diagnosis.

Li Y, Yuan M, Yang Y, et al. Sci Rep. Epub 2026 Apr 6.

Diagnostic AI can fail when real-world patients diverge from training data. This study tested a real-time monitoring framework tracking data drift, fairness, calibration, and clinician-AI interactions via a composite health score and alerts. Simulations showed reliable detection of emerging issues, stability under normal conditions, and actionable oversight—offering a practical tool to monitor and audit diagnostic AI.

5) Using generative artificial intelligence to identify central line-associated bloodstream infections.

Morgan DJ, AlShanqeeti S, Coffey KC, et al. Clin Infect Dis. 2026;82(4):e773-e780.

CLABSI is a preventable hospital-acquired infection whose surveillance is required and publicly reported in U.S. hospitals but usually depends on manual chart review. In this retrospective cohort study, AI detected CLABSI at least as accurately as diagnostic codes and chart review, while being faster, more consistent, and preferred by experts. AI-assisted review could strengthen infection surveillance and reporting.

6) Strategies for teaching nursing diagnosis in undergraduate nursing education: a scoping review.

Napoleão AA, Santos Diogo RCD, Eduardo AHA, et al. Nurse Educ Pract. 2026;94:104854.

Accurate nursing diagnosis is a primary element of safe nursing care, yet the best ways to teach this skill remain unclear. This scoping review found that active, student-centered approaches may strengthen critical thinking, reflection, and diagnostic competence. It also highlights the need for standardized evaluation tools and longitudinal research to assess whether these methods are sustainable and scalable across settings.

7) Women's screening preferences shift after being informed about the 2024 USPSTF breast cancer screening guideline.

Parmet T, Yoder G, McCaffery K, et al. Patient Educ Couns. 2026;148:109581.

After the 2024 US guideline shift to biennial breast cancer screening, a study of 300 screening-eligible women were shown more likely to prefer screening starting in their 40s. Once introduced to the guidelines changes, preferences were shaped by understanding of benefits, personal risk, and family history. Women also wanted clearer information about screening benefits and harms before deciding when and how often to screen.

8) Standardizing ICU-to-ward transitions: introducing and refining ICU-PAUSE in the process of medical intensive care unit downgrade.

Perez E, Tasevac B, Kuchera T, et al. Am J Med Qual. Epub 2026 May 15.

Standardized handoff tools for diagnostic pauses enhance communication between teams when patients transfer in-hospital. This single-center study found that an ICU-written handoff tool (ICU-PAUSE) did not change overall provider satisfaction. However, most clinicians preferred ICU authorship and, after targeted training, reported greater confidence in care transitions.

9) Ableism, ageism, and other biases in healthcare: the impact on young adult patients.

Rosewater J-T, Dave S, Didier M, et al. Health Care Transitions. 2026;4:100136.

Young adults, especially those with chronic or rare diseases, face bias and discrimination that impede timely diagnosis. In a roundtable discussion of five patients and six healthcare professionals, participants envisioned a more just healthcare system, highlighting diagnostic delays driven by ageism, ableism, and symptom invisibility. These intersecting barriers demand systemic attention to improve diagnostic equity for vulnerable patients across age groups.

10) CMS Potentially Overpaid Medicare Advantage Organizations $462 Million Based on Certain Unsupported Acute Stroke Diagnosis Codes. 

US Office of Inspector General; 2026.

Under Medicare Advantage (MA), CMS links payments to enrollee risk factors, including diagnoses. This audit examines acute stroke diagnosis codes submitted by MA organizations without corresponding hospital records in 2020. The results estimate potential mismatches between physician and hospital coding resulting in $462 million in overpayments to MA providers in 2021.

11) Systematic evaluation of the DeepSeek large language model for clinical diagnostic reasoning.

Wang Y, He Y, Qin X, et al. PLoS ONE. 2026;21(5):e0346078.

AI demonstrates potential to improve disease detection, diagnosis, and treatment. This preliminary study evaluates the widely used DeepSeek model in acute care scenarios, assessing diagnostic reasoning, consistency over time, and adherence to evidence-based critical care guidelines from Merck Manual cases. Results include 83% accuracy in final diagnosis and weaker accuracy in generating differential diagnoses, highlighting ongoing challenges in probabilistic reasoning and integration of multimodal clinical data.

12) Using large language models to determine reasons for missed colon cancer screening follow-up.

Williams CYK, Sarkar U, Adler-Milstein J, et al. Jt Comm J Qual Patient Saf. Epub 2026 April 13.

Identifying reasons for missed follow-up is key to improving diagnostic quality and patient safety. This cross-sectional proof of concept study of patients with a documented positive colon cancer stool screening test suggests LLMs can detect whether reasons for missing colonoscopy after abnormal noninvasive screening are documented in clinical notes and identified reasons for not obtaining colonoscopy in 20% of cases; this suggests AI can support guide diagnostic result follow-up improvement efforts.

*UCSF CODEX’s co-chair, Julia Adler-Milstein, PhD, is an author on this publication.