CODEX Digest 05.28.26

AI may undermine reasoning—or strengthen it. In a longitudinal study of 372 Chinese medical students using an AI-assisted diagnosis tool during supervised rotations, greater use correlated with higher AI literacy and critical thinking. AI literacy partially mediated this link. Although this study lacked a non-AI comparator group, the results identify some potential drivers of how learners best build critical thinking with AI.

Autopsy decline has long concerned pathologists; in the AI era it also threatens the validity of clinical machine learning. This commentary from a forensic pathologist argues that to outperform—not scale—human diagnostic error, algorithms must be grounded in pathological truth, requiring data quality commensurate with AI ambitions.

Colorectal cancer screening remains low in the US, especially in low-income and minoritized groups. In this quality improvement trial, behavioral economics–informed automated texts resulted in better at-home colon cancer screening completion compared to nurse-led phone outreach. Automated messaging may be a scalable, low-cost way to boost preventive care and reduce staff burden in underserved populations.

Timely stroke diagnosis helps prevent harm. This Swiss retrospective study tracks 24-year trends and drivers of pediatric diagnostic delays in and out of hospital. Most diagnoses occurred after optimal treatment windows, with delays increasing over time, primarily due to pre-hospital cases. The authors highlight patient, system, and access contributors and call for greater awareness of stroke in younger children and more direct-to-stroke center pathways.

Rapid sepsis detection and treatment improve inpatient outcomes. This multi-methods study examines how digital alert systems affect outcomes and why. Across four NHS trusts, alert system effectiveness varied by patient age and care setting. Findings suggest using broader EHR data and tailoring alerts to specific patient groups.

Emergency radiology demands high accuracy under time pressure, and rising workloads increase error risk. This Chinese study compared five LLMs with 12 board-certified radiologists for detecting errors in Chinese emergency radiology reports. DeepSeek-R1 showed the most promise, with performance and speed supporting use as a real-world proofreading aid.

Monetary incentives can influence healthcare behaviors of both patients and clinicians. This randomized vignette experiment provided scenarios of using AI versus eye care professionals (ECPs) to patients living with Type 1 Diabetes, for required screening for diabetic retinopathy. Scenarios included AI or ECP, with or without a copay. Results showed cost removal boosted AI preference, but patients still requested human verification of the results even when normal. Incentives can promote AI adoption, yet integration strategies must address the persistent preference for human validation.

LLMs alone have illustrated the potential to outperform clinicians using the same LLM. This UK study confirmed prior US-based findings that physicians assisted by an LLM scored significantly worse than the LLM alone, although better than when using conventional non-LLM resources. This study provides a more in-depth look at the wide variation of how the physicians are engaging with LLMs which could explain their poorer performance.

Physicians must disclose the risks and benefits of recommended tests, but must they also disclose results they believe are erroneous or clinically irrelevant? This law review examines informed-consent ethics, diagnostic testing error, and how patient portals create a default expectation of sharing all results. It argues that even disputed results may warrant disclosure, with clinicians accountable for explaining their doubts and the implications of mandating such disclosure.

Missed mammograms in the primary care safety net can increase national breast cancer diagnostic disparities. This study applies continuous process improvement cycles to design and evaluate a digital reminder program’s impact on Screening Mammography Missed Care Opportunity (SM-MCO) rates at federally qualified health centers and mammography van using a combination of texts, standard reminders, and videos. Paradoxically, no-show rates were higher with text reminders and fell with digital engagement, highlighting a digital divide. The results illustrate that quality improvement frameworks can help monitor and refine digital strategies to expand radiology access.

Patient-reported measures can illuminate emergency diagnostic experiences. This research letter summarizes cross-site PRIME-ED (Patient-Report to IMprove Diagnostic Excellence in Emergency Department settings) results from four health systems. Implementation rationales varied, but results support the feasibility of capturing patient-reported diagnostic excellence in real-world emergency settings and highlight both its promise and complexity for learning and accountability.

As AI use in healthcare expands, this systematic review of qualitative studies synthesizes patients’ concerns about AI from a multi-level perspective. The themes identified include privacy, reliability, diagnostic accuracy, and “black box” decisions, effects on physician–patient relationships, trust and accountability, ethics and equity, and fears about diffusion and workforce displacement.