CODEX Digest - 8.13.26

1) The 5R’s of large language model-assisted diagnosis: a practical framework for hospitalists.

Barish P, Auerbach AD, Ranji SR. J Hosp Med. Epub 2026 Jul

LLMs are available to support diagnostic reasoning, but safe implementation is still evolving. This framework proposes five principles—using AI for the right patient, right question, right data, right workflow, and right oversight—to guide responsible clinical use. Rather than promoting indiscriminate use of AI, the authors emphasize maintaining clinician judgment, recognizing AI limitations, protecting patient privacy, and monitoring performance to reduce diagnostic risk.

**UCSF CODEX’s director Sumant Ranji, MD, is an author for this publication. 

2) A framework to characterize the costs of delayed diagnoses.

Berdahl CT, Hero JM, McCleskey SG, et al. Jt Comm J Qual Patient Saf. Epub 2026 Jun 17.

Diagnostic delays result in substantial but often overlooked costs extending beyond direct medical care. This conceptual scaffold classifies consequences for patients, families, clinicians, healthcare organizations, and society, including financial burdens, psychological harm, productivity losses, and wasted resources. By providing a common language for measuring these impacts, the framework supports more thorough evaluation of diagnostic improvement initiatives and strengthens the case for investing in diagnostic excellence.

3) What the tests missed: a journey through misdiagnosis.

Cogna W. NPJ Dement. 2026;2(1):61.

Patient and family experiences of the diagnostic process provide valuable insights for improvement. This perspective follows one patient’s long and confusing path to a diagnosis, showing that even advanced brain scans and fluid tests may not give clear answers on their own. The final diagnosis was confirmed only after autopsy, highlighting the need to combine test results with symptom history, caregiver observations, and lived experience.

4) From uncertainty to understanding: The Focus on Follow-Up initiative.

Corines MJ, Taya M, Sinha V, et al. Clin Imaging. 2026;136:110843.

Receiving feedback on diagnostic calibration in radiology residency can be challenging. The Focus on Follow-Up quality improvement initiative is a prospective educational approach to improve feedback to residents on imaging assessment and diagnostic uncertainty. Radiology residents at a single academic center prospectively tracked weekly cases involving diagnostic uncertainty, followed outcomes, and discussed selected cases in peer-learning conferences. Feasibility, participation, and resident perceptions were assessed, demonstrating how structured feedback can be integrated into workflow to enhance training.

5) Socioeconomic disparities in lung cancer screening participation.

Delilovic S, Yacaman Mendez D, Markholm Nordgren N, et al. JAMA Netw Open. 2026;9(7):e2623504.

Early cancer detection depends on equitable access to screening, yet participation in national programs remains uneven. This population-based Swedish cohort study finds that lung cancer screening attendance among eligible women to be equitable, but participation in the initial risk-selection stage varied by socioeconomic factors, highlighting the importance of equitable screening design.

6) Prospective patient-reported reasons for delayed diagnosis of spontaneous subarachnoid haemorrhage.

Hall S, Suresh VA, Bandyopadhyay S, et al. Emerg Med J. 2026;43(7):390-396.

Subarachnoid hemorrhage diagnosis timing can be suboptimal because early symptoms can resemble more benign conditions. In this prospective UK interview study, patients identified factors contributing to delayed diagnosis after seeking care, including symptom misinterpretation and diagnostic testing errors. The interviews surfaced data missed during chart review, underscoring the value of seeking patient experiences data to identify opportunities to enhance timely recognition of stroke.

7) Diagnostic reasoning with and without AI: automation bias in pre-clerkship medical students.

Hayden R, Blanco M, Ramesh S, et al. BMC Med Educ. Epub 2026 Jul 21.

AI tools may improve diagnostic reasoning but can also encourage overreliance on incorrect recommendations. In this educational study, pre-clerkship medical students demonstrated evidence of automation bias (overreliance on AI output) when GPT-4o chatbot generated suggestions conflicted with clinical reasoning. The findings reinforce the need to teach critical appraisal and independent reasoning alongside AI literacy rather than assuming AI will improve diagnostic performance.

8) Radiologic error as an emergent property of complex adaptive systems: implications for diagnostic safety and governance.

Karamoskos P. AJR Am J Roentgenol. Epub 2026 Jul 22

Diagnostic errors in radiology often reflect interactions among people, technology, workflows, and organizational factors rather than isolated individual mistakes. Applying complex adaptive systems theory, this review argues for shifting from person-centered blame toward systems-based governance, resilience, and continuous learning. The concepts extend beyond radiology and offer a safety science approach for improving diagnosis.

9) The impact of a student-led initiative to improve cancer screenings in primary care.

Klein KV, Srivastava PV, Sauer JM, et al. Am J Prev Med. 2026;70(5):108262.

Primary care screening rates can improve through innovative workforce approaches. This study discusses a student-led initiative that increased completion of recommended cancer screening by assisting with outreach, education, and care coordination. The results suggest that engaging trainees in population health activities may strengthen preventive care while expanding opportunities for experiential learning.

10) Diversion and deviation: how distracting contextual factors amplify cognitive biases in clinical reasoning.

Loncharich MF, Durning SJ, Merkebu J. J Eval Clin Pract. 2026;32(4):e70503.

Clinical reasoning occurs in busy environments where interruptions, workload, and contextual pressures can distort thinking. This simulation study examines how distractions amplify cognitive biases and increase diagnostic vulnerability. Physicians watched video encounters of common disorders with or without added distracting influences, then completed a post-session form and a think-aloud exercise to render their thinking process transparent to allow for analysis. The authors argue that improving diagnosis requires addressing both individual reasoning skills and the workplace conditions in which decisions are made.

11) Can AI assist in reducing diagnostic error? A narrative review.

Scott IA. Diagnosis (Berl). Epub 2026 Jul 22.

AI has the potential to assist diagnosis but also introduces new challenges, including bias, automation errors, poor calibration, and workflow disruption. This narrative review provides a solid introduction to the current state of AI use in diagnostic processes to improve safety. The author provides a comprehensive reference list and summarizes their findings across the diagnostic process—from history taking to organizational applications of AI technology to improve safety. The review suggests that AI should complement—not replace—clinical expertise, emphasizing the importance of rigorous evaluation, human oversight, and careful implementation before widespread adoption.

12) Concordance between an artificial intelligence self-triage programme and physical triage.

Wempe M, Holleman F, Schinkel M, et al. Emerg Med J. Epub 2026 Jun 23.

AI-based symptom-checkers are increasingly used to guide patients before seeking care. This Dutch observational study compares recommendations from Symptomate, a proprietary AI self-triage chatbot, with in-person clinical emergency department triage finding moderate agreement between the two approaches. While the results suggest AI may support initial symptom assessment, further research is needed to determine whether self-triage improves patient outcomes, safety, or healthcare utilization.