CODEX Digest 04.02.26

LLM-based AI systems may improve patient diagnostic conversations, but there is limited real-world evidence. This pre-print feasibility study tests Google’s Articulate Medical Intelligence Explorer (AMIE) for history taking and diagnosis presentation. The results suggest AMIE can safely collect histories and generate diagnostic suggestions, but the authors also acknowledge limited generalizability due to the study taking place at a single center and the 100 participant sample size.

Early-onset depression may persist into adulthood without timely identification and treatment. This retrospective cohort study evaluates adolescent screening in primary care to support early depression diagnosis and reduce long-term mental health issues. Findings indicate that using a standard patient health questionnaire can help to identify at-risk adolescents, however only 66% of patients received at least one screening after being recommended.

AI can enhance breast cancer screening by improving quality and consistency while helping address radiologist shortages. In this British study across five NHS sites, AI matched or outperformed specialist radiologists. The Google-developed AI second-reader system resulted in earlier diagnosis and reduced diagnostic errors.

Diagnostic uncertainty remains a major challenge for individual physicians. This commentary discusses physician diagnostic doubt, explores contributing factors through case examples, and presents a six-part behavioral framework that promotes cognitive flexibility, adaptive reasoning, and empathic communication to improve diagnostic decision-making and safety.

False-negative (FN) mammograms can delay breast cancer diagnoses to impact clinical outcomes. This large retrospective study examines FN rates in screening and diagnostic mammograms from the National Mammography Database, identifying links between FN rates and patient or facility characteristics. The findings highlight the need to investigate these associations further to improve quality assurance and reduce delayed breast cancer diagnoses.

Human-in-the-loop AI combines human expertise with machine intelligence to improve decision making and support ethical and safe patient care. This narrative review examines its effects on care processes such as diagnostic imaging and clinical decision support, addresses challenges like workflow integration, regulatory issues, and sustainability, and suggests future research for enhancing collaboration in healthcare.

Diagnostic errors are a significant healthcare concern, but difficult to assess due to the need for time-intensive chart reviews. This Swiss study examines an automated tool for detecting emergency medicine diagnostic diagnostic errors by searching for diagnostic discrepancies using emergency department discharge letters, hospital discharge letters, and general practitioner notes. The tool reveals high accuracy that could help prioritize cases for further review.

Skin bacteria contamination in blood samples often lead to extended hospital stays, unnecessary treatments, diagnostic delays, environmental waste, and higher healthcare costs. This review assesses various blood culture diversion methods, finding that both device-based and open approaches significantly reduce blood sample contamination rates.

Tracking and reviewing changes in pathology reports helps identify diagnostic errors. Over five years examined in this study, report changes were rare, but major mistakes usually involved mislabeling tissue as benign or cancer, which affects diagnosis. Regular monitoring and double-checking key cases can help lab teams catch serious errors.

AI is increasingly used in cancer care, but patient views are not well understood. This cross-sectional study finds patients to be most comfortable with AI for cancer screening (80.2%) and least comfortable (70.4% ) with AI aiding diagnosis. About half of patients were concerned about the risks of AI implementation.

Assessing accuracy and applicability to different populations is vital for reliable diagnostic test reviews. This article presents QUADAS-3, a six-phase tool designed to assess the risk of bias and applicability of primary test accuracy studies for diagnosis, screening, or staging of disease. Notable updates include creation of an ideal test accuracy trial concept, an estimate-level assessment, a new "Analysis" domain, rationale sections for judgments, and formal overall risk and applicability evaluations.

The delineation of accountability for AI use in decision making is emerging. This study explores how AI use affects mock jurors' perceptions of radiologist responsibility in malpractice cases. When radiologists reviewed scans only once with AI, 75% of participants assigned blame; reviewing twice (alone and with AI) lowered blame to 53%. Double-checking scans may lessen juror blame for doctors who miss findings flagged by AI.