CODEX Digest 06.25.26

Growing demand for chest X-rays has contributed to reporting backlogs in many health systems. In this year-long prospective multicenter study of more than 63,000 chest radiographs across five UK hospitals, an AI tool was tested for identifying likely normal examinations. The system showed high sensitivity and a low rate of clinically significant missed abnormalities, although most studies were still flagged as abnormal. The findings suggest AI could help prioritize imaging workflows, but further evaluation is needed before widespread clinical use.

Patient history-taking is central to diagnosis. In 104 first-year medical student interviewing sessions, an ambient AI scribe helped educators produce better written feedback on the students' interviewing skills without increasing effort. In this small single-site study, occasional inaccuracies still required review, but the approach may improve feedback documentation.

How diagnostic safety is translated into policy and practice remains unclear. This study compares the policy trajectories of the US and Australia over 25 years to show how diagnostic safety can be embedded as a core patient safety priority. The authors share recommendations for improvement.

Accurate stroke identification by emergency dispatch is critical, but stroke mimics can drive overtriage. This Norwegian study found that its central dispatch system stroke criteria had low positive predictive value, indicating substantial early overtriage, while universal criteria better distinguished confirmed strokes. The findings underscore the operational burden of stroke mimics and the need to refine dispatch strategies.

Overdiagnosis is a key harm of prostate cancer screening. Using long-term trial data and mortality estimates, this study finds that overdiagnosis was relatively low overall after a one-time PSA screening but rose sharply with age due to competing causes of death. Policies encouraging opportunistic screening, especially in older men, should be reconsidered.

Classic symptom patterns can aid diagnosis, but atypical presentations raise the risk of diagnostic error. This study applies a classification system for atypical presentations that can be used across specific diseases or more broadly. The authors argue that a detailed, nonbinary framework is needed to study how atypical presentations contribute to diagnostic errors and how to reduce them.

As AI becomes more involved in diagnosis, questions remain about who is responsible when diagnostic harm occurs. In a study of 652 adults in the US and UK, participants reviewed a scenario involving a delayed diagnosis. People assigned less responsibility when AI acted alone than when physicians were involved. Public expectations consistently placed responsibility on clinicians, even when AI contributed to decision-making. Providing explanations of how AI reached its conclusions did not substantially change views about physician accountability after adverse events.

Diagnostic timeouts (DT) may help teams revisit uncertain cases. In this pilot across four children’s hospitals, DTs were implemented during unplanned PICU transfers. After DTs, clinicians were surveyed and rated team understanding improvement as an important impact of DT, but timeouts less often changed the differential diagnosis. Use during unplanned PICU transfers varied by site but was feasible in under five minutes. Perceived value was highest when patients, families, or bedside nurses took part.

Early cancer detection depends on screening programs that are both effective and accessible. In this study, 22 adults from low-income communities helped identify practical ways to reduce disparities in mammography screening. Participants prioritized clear, patient-centered communication and lower out-of-pocket costs as high impact. The project also illustrates that quality-improvement methods can help communities and healthcare organizations identify and prioritize strategies to improve screening equity.

LLMs are being explored as tools to support psychiatric diagnosis, but concerns remain about fairness and reliability. This study evaluated six LLMs using identical psychiatric cases that varied only by gender, race, or ethnicity. The models often produced different diagnoses and reasoning despite unchanged symptoms, reflecting and sometimes amplifying existing biases. Findings highlight risks of overdiagnosis and unequal assessments across groups, underscoring the need for careful oversight before using LLMs in mental health decision-making.

AI skin cancer tools do well in controlled settings but often fall short in routine care. In this multi-institutional study, three AI algorithms were compared with 652 physician evaluators using realistic, including rare and atypical cases. The general-purpose model beat physician readers with less than three years of experience, matching those with three to 10 years, and lagged behind experts with more than 10 years.

Colorectal cancer is the second deadliest cancer in the US and hits underresourced communities hardest. This trial compared two mail-based outreach approaches to boost screening in community health centers, including among English- and Spanish-speaking patients. At 90 days, overall screening participation was low, but the more advanced test including phone, text, and email reminders had greater participation than the test with only text reminders. Follow-up colonoscopy rates were poor. The findings demonstrate that test choice and outreach strategies must fit local context to support successful screening.