The research, conducted at Samsung Medical Center, focused on diagnosing pheochromocytomas and paragangliomas (PPGL). These tumors cause an overproduction of stress hormones, which standard plasma-free metanephrine tests detect. However, the test frequently produces false-positives because mild hormone elevations often occur in patients without tumors. Analyzing data from 20,516 patients, scientists sought to determine if machine learning could filter out these errors by integrating electronic health records.
AI diagnostic tools face hurdles in adrenal tumor testing
Machine learning shows promise in refining blood tests for rare adrenal tumors, but researchers warn that algorithms often rely on clinical shortcuts rather than pure biochemical data. A study presented at the ADLM 2026 meeting reveals that models can inadvertently mirror physician ordering habits, potentially skewing diagnostic accuracy.
Initial results appeared highly successful, with the software achieving excellent performance by incorporating biomarkers, medications, and comorbid conditions. Yet, subsequent audits revealed that the algorithm was performing "shortcut learning." Instead of identifying tumor-specific biochemical signals, the model was effectively tracking which follow-up tests were ordered by clinicians who already suspected a diagnosis. Se-eun Koo, a co-author of the study, emphasized that this finding highlights a critical need for rigorous auditing in laboratory medicine. To be truly effective, developers must ensure that algorithms learn intended clinical signals rather than merely reflecting existing patterns of medical workups.
Comments (0)
No comments yet. Be the first!