CODEX Digest 09.24.26

Accuracy statistics alone don’t show how often radiology AI will be wrong in real practice or support sound purchasing decisions. This commentary explains how to calculate local error rates using false discovery and false omission rates, and offers the risk-based evaluation matrix (RADDE) as a framework for weighing the medical, legal, and other impacts to inform AI implementation decisions.

Diagnostic error remains a persistent patient safety challenge that linear improvement strategies alone haven't solved. This study uses ecological psychology and Safety II principles to examine diagnostic "good catch" events, interviewing four clinicians across three sessions. Facilitators of effective diagnosis included "trust but verify" admission practices, point-of-care ultrasound-enhanced exams, and the use of open spaces to foster team collaboration. The authors present the results as a framework that combines Safety II principles and ecological psychology to develop robust and pragmatic interventions to mitigate diagnostic errors.

Faster diagnostic work does not necessarily mean better diagnosis. This preprint reports on a German, multicenter randomized controlled trial, in which 82 physicians used conventional resources with or without Prof. Valmed, the first LLM-based decision-support system to receive European Conformity certification as a medical device. The results show that AI substantially reduced case-processing time on three rheumatology cases but did not improve diagnostic accuracy. The authors highlight that overconfidence and AI over-reliance remain concerns in the use of such systems.

Childbirth-related post-traumatic stress disorder (PTSD) may be clinically important yet difficult to recognize in primary care. This commentary highlights the gap of actionable primary-care evidence and practical guidance for general practitioners (GPs) to diagnose childbirth-related PTSD. Although GPs are positioned to identify persistent psychological symptoms after birth, uncertainty about recognition and management may contribute to missed diagnosis and inadequate support.

Diagnostic tests can create harm as well as information. This conceptual analysis examines how testing can trigger false positives, incidental findings, overdiagnosis, and patient anxiety. The authors argue that diagnostic excellence requires clinicians to consider whether a test’s expected benefits outweigh its potential harms—and, when appropriate, deliberately choose not to investigate.

Delayed childhood cancer diagnosis is a global diagnostic excellence challenge, but educational solutions remain poorly evaluated. This scoping review identified 29 studies from 20 countries examining educational interventions for timely pediatric cancer diagnosis. Most used in-person training, including workshops, courses and lectures, but approaches and outcome measures varied substantially, limiting conclusions about effectiveness.

Emergency diagnosis may be an important marker of diagnostic vulnerability in a wide range of conditions. This population-based English cohort study uses linked electronic health records from 1.7 million patients with 13 non-neoplastic conditions to examine occurrence and outcomes of such conditions. Emergency diagnoses were common across the explored conditions and associated with poorer subsequent outcomes compared to patients whose conditions were diagnosed in non-emergency encounters. The results highlight opportunities to address diagnostic delay, primary-care recognition, and health-system barriers as factors affecting diagnostic excellence.

More cancer detection does not automatically mean better outcomes. This qualitative study interviewed 15 board-certified dermatologists about melanoma screening and overdiagnosis. Participants described screening’s perceived benefits while grappling with limited evidence that population-based screening reduces mortality. The findings highlight challenges in communicating overdiagnosis and balancing early detection against potential overtreatment.

The quality of clinical information reaching radiologists can influence diagnostic interpretation. This multireader study compared LLM-generated radiology-relevant histories with clinician-provided indications for 241 examinations. Claude 3.5 Sonnet and Qwen 2.5-7B Instruct produced more comprehensive and factual summaries. The study however only assesses information quality and usefulness—not whether LLMs improve diagnostic accuracy or patient outcomes.

Diagnostic reliability requires consistency—not just a plausible answer. This study tested 10 large language models across 120 clinical vignettes and 9,600 responses. The study finds that if a prompt was shared to ask a model “Are you sure?”, correct answers were often changed to incorrect ones, with correct-to-incorrect transitions substantially exceeding the reverse. GPT-5 performed best overall, but instability remained a broader concern.

AI may need to support clinicians differently depending on how they are reasoning. This commentary proposes adaptive AI support across three diagnostic reasoning modes used in emergency care, rather than a one-size-fits-all “copilot” to address the limitations of cognition in a high stress, uncertain environment. The framework is innovative, but it remains conceptual: no clinical dataset or intervention was tested, so claims about improved diagnostic performance remain unproven.

Reducing unnecessary screening requires understanding what patients actually value. This national discrete-choice experiment surveyed 673 women aged 70 years and older about strategies for de-implementing mammography over screening. Participants preferred discussions with clinicians and balanced information about benefits and harms, suggesting that effective de-implementation may depend more on shared decision-making than calculators or checklists.