FDA-authorized oncology artificial intelligence and machine learning devices and their clinical evidence: A cross-sectional analysis

BACKGROUND: Artificial intelligence (AI) and machine learning (ML) tools are increasingly embedded in cancer care, yet the scope of U.S. Food and Drug Administration (FDA) oncology-specific authorizations and the clinical evidence described in publicly available decision documentation remain unclear. METHODS: We conducted a cross-sectional analysis of FDA-authorized AI/ML-enabled devices with oncology-specific indications from January 12, 2021, to September 12, 2025. Devices were identified from the FDA AI-Enabled Medical Device List and linked to FDA decision documents. Using a prespecified coding manual, we abstracted cancer type, clinical domain, whether indications were screening-, diagnosis-, and/or treatment-related, how devices fit into the FDA's computer-aided detection, diagnosis, and triage (CAD) taxonomy, and evidence features: (1) clinical testing using patient-derived data, (2) clinician-in-the-loop testing (studies assessing clinician performance with device output), and (3) prospective testing. RESULTS: Of 1008 FDA-authorized AI/ML devices, 149 (15%) had oncology-specific indications. Indications clustered in radiology (69/149, 46%) and radiation oncology (57/149, 38%). Of the 149 devices, 113 (76%) reported clinical testing, 31 (21%) clinician-in-the-loop testing, and 7 (5%) prospective testing. Higher-tier evidence (clinician-in-the-loop and/or prospective testing) was significantly more common among CAD devices (20/43, 47%) than non-CAD devices (11/106, 10%; p < 0.001). CONCLUSIONS: Oncology AI/ML device authorizations are concentrated in imaging and radiation oncology domains, with publicly described clinician-in-the-loop and prospective evaluations remaining uncommon overall, though higher-tier evidence was more frequent among CAD devices designed to directly aid clinician interpretation. Evidentiary expectations should be calibrated to device function and clinical risk, with stronger evaluation requirements for devices that directly shape decision-making.