Clinical Utility of Multicancer Detection in Symptomatic Patients: A Decision-Making Perspective.

purposeThere is growing interest in multicancer detection (MCD) blood tests for diagnosing patients with cancer-related symptoms. However, recent studies suggest that MCD testing may not be sensitive enough to rule out cancer in the symptomatic population without retraining the underlying classifiers. On the basis of clinical guidelines for suspected cancer referral, here we cast these data into a formal diagnostic decision-making perspective to assess clinical utility.

methodsData were extracted from the SYMPLIFY study (ISRCTN10226380), which evaluated the performance of the Galleri test (GRAIL, LLC). The decision threshold for suspected cancer referral was extracted from the National Institute for Health and Care Excellence Guideline 12. Clinical utility was estimated using Bayesian decision curve analysis.

resultsFor the guideline-derived decision threshold of 3%, the Galleri MCD test avoided 18,005 unnecessary suspected cancer referrals per 100,000 symptomatic patients, with a 99.4% posterior probability of clinical utility. High probabilities of clinical utility were observed for gynecologic, lower GI, and upper GI referral pathways, avoiding between 25,414 and 62,501 unnecessary referrals per 100,000 symptomatic patients. The rapid diagnostic center and lung referral pathways showed negligible probabilities of clinical utility. The minimum diagnostic performance required for clinical utility varied significantly across referral pathways. The gynecologic pathway showed the lowest sensitivity requirement (under 30% for a highly specific test) and the lung pathway the highest (over 90% for any specificity level).

conclusionClinical utility of MCD testing for symptomatic patients in the United Kingdom varies substantially across referral pathways but is favorable for gynecologic and GI cancers. Future pathway-specific optimization of MCD tests must consider clinical utility explicitly and does not require retraining the underlying machine learning classifiers.