Applicability of accreditation requirements to machine learning-based methods in molecular oncology testing
A College of American Pathologists working group made up of molecular pathologists, pathology trainees, a bioinformatician, and two CAP staff members conducted a review in which they evaluated the extent to which CAP accreditation requirements can be applied to the clinical adoption of machine learning (ML)–based methods used in molecular oncology testing. They also provided examples of current and emerging ML applications for this area. As machine learning increasingly supports such tasks as tumor classification, variant interpretation, microsatellite instability detection, and methylation-based diagnostics, concerns have emerged regarding validation, quality management, and regulatory oversight in the absence of ML-specific accreditation standards. The authors demonstrated that CAP accreditation frameworks are broadly applicable to ML-based laboratory methods when interpreted through a test-centric rather than an algorithm-centric lens. Analytical and clinical validation requirements outlined in CAP checklists can be applied directly to ML-based assays, provided that the ML component is clearly defined as part of the test system. Performance characteristics, such as accuracy, precision, sensitivity, specificity, and reproducibility, are measurable and enforceable for ML-driven outputs. Quality control and quality assurance principles already embedded in CAP standards are sufficient to address ML deployment, including requirements for ongoing performance monitoring, documentation of failures, and corrective action processes. The authors emphasized the importance of tracking performance drift, particularly for algorithms trained on evolving data sets. Change management and revalidation requirements were deemed particularly relevant to machine learning. The review highlighted that updates to algorithms, including any changes in model architecture, training data, or software version, constitute test modifications and therefore trigger revalidation under CAP checklist requirements. The authors noted that adaptive algorithms, in particular, require especially robust governance structures. They further reported that transparency and reproducibility are essential to maintaining accreditation compliance. The review also found that ML-based methods can be regulated appropriately as laboratory-developed tests or test components, reinforcing the central role of the laboratory director in oversight, interpretation, and accountability. This framing aligns ML governance with established laboratory medicine practices rather than treating machine learning as a separate regulatory category. The authors concluded that CAP accreditation requirements, when thoughtfully applied, provide a robust framework for the validation, implementation, and ongoing oversight of ML-based molecular oncology tests. The primary challenges are not deficiencies in accreditation standards, but rather the need for clear definition of intended use, rigorous documentation, and disciplined lifecycle management for ML systems. Overall, the review positions ML adoption in molecular pathology as an extension of established laboratory quality principles rather than a disruptive exception, underscoring that successful clinical integration of ML depends more on governance, validation, and accountability than on the development of new regulatory paradigms.
Furtado LV, Ikemura K, Benkli CY, et al. General applicability of existing College of American Pathologists accreditation requirements to clinical implementation of machine learning-based methods in molecular oncology testing. Arch Pathol Lab Med. 2025;149(4):319–327.
Correspondence: Dr. Larissa V. Furtado at [email protected]