CODEX Digest 07.16.26

EMR-based sepsis alerts can speed care for children but risk overdiagnosis and alert fatigue. This review examines pediatric outcomes after alert implementation. Findings show alerts do not raise hospital admissions or antibiotic use and may improve time to fluids and antibiotics, without increasing admissions nor antibiotic use. Their effect on mortality remains unclear. For multidisciplinary teams, balancing timeliness with safety is key to responsible use.

Policies shaping AI in radiology directly affect diagnostic safety. This review proposes a practical, lifecycle-based framework aligning AI governance with clinical practice. It clarifies what AI must deliver to clinicians and patients, and what institutions need for safe implementation. Covering data stewardship, development, validation, deployment, and monitoring, it defines shared responsibilities among vendors, institutions, and clinicians—grounded in trustworthy AI principles to protect patient safety.

Many patients find it hard to understand lab and test results. This cross-sectional study looks at how patients felt about using an AI assistant built into their patient portal to help explain results in plain language. Although still a new tool, findings show patients are willing to use it—especially those managing multiple health conditions—suggesting its potential value for improving understanding of medical results.

The National Academies’ Forum on Advancing Diagnostic Excellence hosted a hybrid workshop on strengthening patient, caregiver, and family engagement in diagnosis to reduce diagnostic errors. Discussions emphasized active participation, stronger clinician communication, partnership with families, and intentional inclusion of patients and caregivers in health system design. Key approaches included communication tools, caregiver involvement, transparent discussion of diagnostic uncertainty, responsible use of digital tools and AI, and systems that ensure patients are heard, respected, supported, and included throughout diagnosis.

There is racially inequitable implementation of lung cancer screening in the United States, with lower screening for Black/African-American people particularly. This mixed-methods study identifies factors shaping screening participation of Black veterans and highlights the need for implementation strategies that address both individual barriers and structural inequities to improve equitable access to this distinct patient population that may be applied to screening improvement efforts in other communities.

Safety II focuses on improving care by learning from success rather than failure. This commentary applies to a structured Safety II approach to a complex metastatic breast cancer case undetectable by imaging. The authors discuss three contributors to diagnostic success in this example: assessing disease likelihood, engaging the patient, and collaborating across specialties. It shows how studying successful diagnoses can generate practical strategies for difficult cases.

As LLMs are embedded throughout healthcare, physicians may overrely on AI advice—a risk to diagnostic safety. This randomized trial preprint of 72 physicians reviewing clinical vignettes shows that a behavioral nudge may reduce overreliance. The findings suggest that embedding cues into clinical interfaces could help preserve human judgment, offering a potential way to minimize automation bias in clinical decision making.

Clear communication is central to diagnostic excellence, especially for patients at risk of not receiving a timely, accurate explanation of their condition. This prospective Dutch study examines diagnostic agreement among various adult emergency department (ED) patients and associated factors. It found high patient-physician diagnostic agreement, showing that clear diagnostic explanations are possible even under ED time pressure and without an established patient-physician relationship.

Patients using LLMs for diagnostic consultations may unintentionally bias the conversation based on how they phrase their questions. This study tests six systems to find that even small input biases can reduce diagnostic accuracy. Standard prompt rewording ("prompt-engineering") didn't improve response correctness, but pairing a conversational AI with a reasoning-focused one restored most lost accuracy. This two-part design offers a testable approach to make medical AI more reliable and sensitive to the biases humans can bring.

Patient portals can provide timely test result communication, but direct online access of serious diagnostic findings can be psychologically harmful without a clinician interface. This survey of more than 2,400 patients at a comprehensive cancer center finds that, although most learned their cancer diagnosis from a clinician, a growing share especially those with recurrent cancer, first learned it through the portal. Despite valuing rapid electronic access to results, three-quarters preferred to receive new or recurrent cancer diagnoses directly from their clinical team.

Delayed diagnosis is common, but its full impact varies across conditions and is difficult to measure. Through clinician expert interviews, this qualitative study identified four broad categories of costs: health, healthcare, nonhealthcare, and psychosocial. The findings provide a practical framework to help researchers better estimate the economic and human burden of delayed diagnosis and support future improvements in diagnostic evaluation.

Improving diagnosis in children requires pediatric-specific priorities, yet consensus has been lacking. This study reports on the work of a multidisciplinary panel (of family partners, researchers and safety experts) to identify key research and operational priorities, including high-risk pediatric conditions and care scenarios, stronger feedback mechanisms, and reporting of missed diagnostic opportunities.