Clairity’s FDA Evidence: What Remains Missing for Payers
The evidence supporting Clairity’s May 2025 FDA authorization establishes that its mammogram-based artificial intelligence can stratify breast cancer risk. From a payer perspective, however, it leaves a central question unresolved: how much useful information does the commercial product add beyond an established clinical risk assessment, and does that information improve care enough to justify its cost?
The pivotal study was substantial: a retrospective analysis of 77,511 examinations from 44,112 patients at five validation sites independent of the algorithm’s development sites. Its principal weakness for coverage assessment is the comparator. Clairity’s Allix5 achieved a five-year AUC of 0.70, exceeding a prespecified performance benchmark of 0.64 derived from published traditional risk models. But the study did not report the performance of Gail, Tyrer–Cuzick, or another clinical model in these same patients. Nor did it compare clinical assessment alone with clinical assessment plus Allix5. Consequently, the apparent difference between 0.64 and 0.70 cannot be interpreted as the demonstrated incremental benefit of purchasing Clairity.
Allix5 uses mammographic images without separately entered age, family history, or genetic results. That simplifies data collection, but it does not resolve the question of added value. Images contain breast density and other characteristics associated with conventional risk factors. An image-only score may therefore reproduce some information already available to clinicians. A payer needs to know whether it improves prediction and decisions after that existing information has been taken into account.
The headline AUC also includes cancers identified during the current screening episode. When the investigators excluded “current cancers”—defined as cancers diagnosed within 90 days after a positive index screening assessment—the AUC fell to 0.66. This analysis is particularly relevant to a service intended to guide future screening after the present examination has been resolved. It still exceeded the FDA benchmark, but showed more modest discrimination. Neither AUC represents the percentage of women whose outcomes were correctly predicted.
Calibration presents another concern. Among patients without current cancer, the model significantly overestimated risk in the highest predicted-risk decile. Acceptable calibration was also not demonstrated in several subgroups, including Asian patients and women younger than 50 or aged 65 and older. These findings do not prove failure in every subgroup, but they limit confidence in using the reported percentages to trigger additional care. FDA addressed these uncertainties through labeling and other safeguards; payers must also consider unnecessary imaging, biopsies, preventive treatment, and anxiety.
Finally, retrospective prediction does not establish clinical utility or economic value. The FDA summary does not demonstrate that an Allix5-guided care pathway reduces advanced cancers, improves survival, or produces a favorable balance of benefits, harms, and costs compared with existing assessment. Earlier research on related software can support plausibility, but cannot establish those outcomes for commercial Allix5. The coverage gap is therefore specific: the product needs evidence of incremental value over a credible clinical comparator, followed by evidence that acting on its predictions improves care.
