This study explored the use of artificial intelligence (AI)-based support in clinical decision-making in dermatology. The authors propose a framework for future research on image-based diagnostics to improve AI use in clinical practice.

Machine learning may have the potential to improve clinical decision-making and diagnosis. In this study, machine-learning algorithms generally performed better than human experts in accurately diagnosing 7 types of pigmented skin lesions and the top 3 algorithms performed better than the 27 physicians.

Measuring and addressing diagnostic error remains challenging. A prior study showed that when providers had similar individual diagnostic accuracy rates, pooling their assessments led to improved decision accuracy. This computer simulation study analyzed 1710 diagnoses provided by 285 medical students for 6 simulated patients presenting to the emergency room. Investigators found that pooling independent assessments led to enhanced diagnostic accuracy as compared to the average independent assessment, further supporting the idea that collective intelligence may help prevent diagnostic error.

… Journal of clinical pathology … J Clin Pathol … Second opinions can shed light on challenging … The authors suggest that such characteristics could prompt a second opinion in order to enhance diagnostic accuracy. …

… of the United States of America … Proc Natl Acad Sci U S A … Diagnostic error remains a significant source of preventable patient harm. Because … purpose of improving medical diagnosis. A previous WebM&M commentary discussed a case of diagnostic error. …

This study found that eliciting second opinions in pathology improved the accuracy of breast histopathology specimens. This work provides further evidence that diagnostic accuracy can be enhanced with second opinions. The authors suggest that implementing multiple clinician review may augment the diagnostic process.

Collective intelligence encompasses several methods for summarizing input from multiple individuals, which can often be more accurate than any one expert. In this study, investigators applied several collective intelligence algorithms to mammography interpretation. They found that aggregating the interpretations of multiple radiologists resulted in higher accuracy—fewer false positive results and more true positive results—than even the most accurate single radiologist. This work builds on earlier studies of diagnostic accuracy in imaging studies. This study has profound implications for improving diagnosis through collaboration between clinicians in real time, perhaps facilitated through technology, as a complement to the long-standing diagnostic safety strategy of morbidity and mortality conferences, which provide group feedback once a case has concluded.

Microscopic review of biopsy tissue is considered the gold standard for diagnosis of cancer and other diseases, but prior research has shown a small yet consistent rate of errors in cancer diagnosis that is attributable to misinterpretation of biopsy specimens. This study sought to quantify error rates in breast cancer diagnosis by having a broad sample of pathologists review a standardized set of biopsies whose diagnoses had been established by expert clinicians. Although biopsies with cancer were diagnosed very accurately, specimens with atypia (abnormal tissue that may be pre-cancerous) had substantial variability, with pathologists tending to overdiagnose these specimens (i.e., ascribe a diagnosis of cancer or pre-cancerous lesions when the correct diagnosis was benign). The authors caution that the specimens used in this study were intentionally chosen to be relatively difficult to interpret, and this may have resulted in overestimating the error rate. A related editorial notes that while the overall rate of diagnostic error in this study was low, misdiagnosis of atypia does have important prognostic and treatment significance for women, and therefore pathologists should systematically consult with colleagues in difficult cases, and more advanced molecular diagnostic methods should be applied in order to reduce subjectivity in biopsy interpretation.

Clinicians' diagnoses of simulated clinical cases with nonspecific symptoms were fairly accurate, especially when items of the history were present that allowed physicians to use heuristics appropriately.

… discuss mammogram results directly with patients, but only a minority would disclose any information about an error in … error and how it occurred. Errors in cancer diagnosis are a frequent cause of malpractice lawsuits, but in this study, … Dr. Thomas Gallagher, was interviewed for AHRQ WebM&M in January 2009. …