Machine learning significantly improved the diagnostic accuracy of plasma metanephrine testing for rare adrenal tumors in a large real-world study, but researchers also found that the apparent gains could be misleading unless artificial intelligence models are carefully audited before clinical use.
The research, presented at the 2026 Association for Diagnostics and Laboratory Medicine (ADLM) meeting, evaluated more than 20,500 adults who underwent plasma-free metanephrine testing at Samsung Medical Center between 2010 and 2024. Plasma metanephrines are the recommended first-line laboratory test for pheochromocytomas and paragangliomas (PPGL), rare neuroendocrine tumors that produce excess stress hormones. Although the test is highly sensitive, false-positive results are common, creating uncertainty for clinicians and often leading to additional investigations.
The investigators examined whether combining laboratory results with structured clinical information could improve diagnostic interpretation. The machine learning models incorporated urine biomarkers, kidney function, medication history, and diagnostic information alongside plasma metanephrine results. The best-performing model increased the area under the receiver operating characteristic curve from 0.81 to 0.91, with the greatest improvement coming after clinical and medication data were added rather than laboratory results alone.
For diagnostic laboratories, however, the study's most important finding may be what happened next. Researcher Se-eun Koo explained, "After abstract submission, while testing the prototype app with the developed machine learning model, we realized that some of the apparent improvement might reflect patterns of clinical workup rather than independent biochemical information. Additional robustness analyses showed that much of the improvement was explained by shortcut learning from informative missingness,"
Rather than identifying biological patterns associated with PPGL, the algorithm had learned to recognize patterns in routine clinical care, including which follow-up tests clinicians ordered when they already suspected the disease.
The findings highlight an increasingly important issue as laboratories evaluate artificial intelligence tools for routine diagnostics. Strong statistical performance alone may not demonstrate that an algorithm is using clinically meaningful information. Instead, models require external validation and targeted assessments to determine whether they are relying on intended biological signals rather than hidden features of clinical workflows.
The authors conclude that integrating laboratory and clinical information can support more context-aware interpretation of plasma metanephrine testing, potentially reducing false-positive results and unnecessary follow-up. At the same time, the study suggests that successful implementation of machine learning in laboratory medicine will depend as much on rigorous validation and transparency as on improvements in diagnostic accuracy.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
