CODEX Digest 10.22.26

AI is increasingly shaping diagnostic quality and patient safety in healthcare. This review examined diagnostic, predictive, and decision-support applications, along with value assessment, across healthcare quality and safety domains. It identifies five persistent barriers to aligning AI with clinical value: limited real-world testing, poor integration with clinical workflows, inadequate interoperability, equity concerns, and uncertainty about costs and oversight.

Diagnostic uncertainty is unavoidable, but how it is recognized and shared can influence downstream care. This task-force review combined a targeted literature review with perspectives from students, residents, and radiologists to identify causes of uncertainty and potential improvements. The results share recommendations that address communication, education, reporting, quality improvement, research, and AI, emphasizing the importance of intentional conversations as an improvement strategy.

How clinicians reflect on diagnostic cases may influence learning. This Brazilian study engaged 67 medical students to test whether reviewing a case during deliberate reflection changed later diagnostic performance. The study found that both groups had the same overall accuracy, but the timing of their access to case information changed how they learned. Closing off case details forced participants to rely on memory and better recognize diseases they had already studied. Keeping case details visible, however, made it easier to apply that knowledge to new, related situations.

Overdiagnosis remains an under-recognized diagnostic harm. This qualitative Belgian study examined the perspectives of 34 health professionals involved in breast, lung, or prostate screening. Few participants naturally identified overdiagnosis as a screening harm, and their responses varied according to their broader views of screening. The findings illustrate how professional beliefs and context can shape diagnostic decisions.

Patient-facing AI is expanding faster than evidence about its safety and effectiveness. This scoping review identified 65 studies of AI in primary care, most involving chatbots, apps, or symptom/triage tools. Despite the wide availability and use of such AI tools, the findings highlight limited evidence evaluating safety, equity and access, noting diagnostic excellence concerns in their use. The authors conclude that stronger pre- and post-deployment evaluation is needed to support safe widespread implementation of patient-facing AI.

Claims about LLM diagnostic reasoning depend partly on whether models have encountered similar cases during training. This study compared five frontier LLMs on 2,000 clinical case reports from before and after presumed training cutoffs. Diagnostic accuracy was nearly identical between cohorts, with little lexical or semantic overlap. However, the evaluation relied partly on an AI judge, limiting how conclusively the findings establish “reasoning.”

Patient involvement in diagnosis can increase awareness, but self-detection has limitations. This Australian study evaluated 191 surfer and outdoor swimmer recognition of skin conditions using clinician examination and histopathology as reference standards. Self-examination had only modest sensitivity for melanoma and basal cell carcinoma, with substantial false-negative rates and low positive predictive values. The findings, while limited due to the distinct population studied, support self-examination for awareness while illustrating the value of clinician assessment.

Small changes in how diagnostic screening appointments are offered may affect participation and diagnostic timeliness. This UK NHS service evaluation quasi-randomized 17,965 women to different combinations of open and timed invitations. Attendance was highest with timed appointments, with substantial increases among women in the most deprived areas. The findings demonstrate how health-system design can influence access to diagnostic prevention and potentially address participation disparities.

As diagnostic care becomes increasingly technology-dependent, the physical examination remains an important source of clinical information. This narrative review synthesizes evidence linking effective examination skills with diagnostic accuracy, error reduction, cost, and the patient–clinician relationship. The authors argue for renewed emphasis on bedside examination while integrating—not simply opposing—technology and other diagnostic tools.For more on the topic from author Brian Garibaldi, MD, check out his CODEX webinar "Medical AI Needs Our Help - Here's Why."

The care setting in which cancer is diagnosed can reflect diagnostic timeliness. This single-center Canadian medical-records review compared cancers first diagnosed through the emergency department (12.8% of new cases) with those diagnosed in the community. Emergency-diagnosed cancers were far more often advanced (65% versus 17% stage IV) and deadlier within one year (39% versus 7%). Diagnosis was faster (19 versus 84 days), largely due to quicker imaging and biopsy. Dedicated diagnostic pathways could speed community diagnosis and improve outcomes.

Introducing AI into diagnosis may change how responsibility for errors is perceived. Across five experiments involving participants in China and six English-speaking countries, physician–AI collaboration reduced perceived physician responsibility after a diagnostic error. The findings raise important questions about accountability and governance, although perceived responsibility in scenarios should not be equated with legal or clinical responsibility.

Diagnostic problems extend beyond the clinical encounter. In this Gallup-Pivotal survey of 5,141 adults, 51% of women experienced at least one recent healthcare-access problem, 34% felt their symptoms were dismissed, and 47% of diagnosed women waited three months or longer for their most recent diagnosis. The survey documents that patient work life, caregiving, and relationships were impacted, highlighting diagnostic timeliness as a broader equity issue.