Close this search box.

A multi-cancer early detection blood test using machine learning detects early-stage cancers lacking USPSTF-recommended screening – npj Precision Oncology

We report early real-world experience with a validated MCED test designed for cancer screening and available for clinical implementation. Although the MCED test detects cancer across all stages, we describe the early detection (Stage I) of 3 cancers that are not covered by USPSTF-recommended screening or other routine screening tests in individuals without known risk factors. All 3 individuals were asymptomatic and thus unlikely to be otherwise diagnosed at an early stage, and 2 of them had no risk factors other than age. In all 3 cases, the CSOs were proven correct by pathology and helped guide efficient diagnostic evaluation. Diagnostic resolution ranged from 28–68 days, consistent with that reported in the PATHFINDER study12. All three were eligible for and underwent curative-intent treatment with guideline-concordant care.

The experiences of these individuals highlight this technology’s potential to detect early-stage cancers in asymptomatic individuals and illustrate the ability of CSO prediction capability to achieve diagnoses efficiently. With respect to the specific cancers discussed here, ovarian cancer tends to present at a late-stage16, is challenging to diagnose due to non-specific or absent symptoms17, and screening is mainly considered in a subset of individuals at genetically high risk, with marked limitations in screening modality performance (i.e., transvaginal ultrasound)18,19,20. Similarly, early-stage RCCs are usually asymptomatic and are generally detected incidentally on imaging with a notable potential for overdiagnosis21. In the case described here, the individual had a higher grade histology, which is predictive of more aggressive behavior. There are no current recommendations for RCC screening in individuals at average risk. Finally, no routine screening programs or tests exist for oropharyngeal cancers beyond findings noted through routine oral exams in dental offices or self exams15. Although HPV is a risk factor for oropharyngeal cancers, there is no approved HPV screening test for the throat, in contrast to cervical cancer15. Additionally, the detection of less common cancers such as oropharyngeal cancer in the real-world is particularly notable, given that it may not be viable to have single-cancer screening tests for less common cancers.

In the absence of screening, an undetected early-stage cancer can progress to a more advanced stage before the presentation of clinical symptoms that would lead to a diagnosis, by which point the prognosis may have become less favorable. The stage dependency for survival outcomes for these cancer types suggests that these three cases are likely to have favorable long-term outcomes (survival outcomes for these cancers at localized, regional, and distant stages, respectively, are: ovarian, 93.1%, 74.2%, and 30.8%; RCC, 93.3%, 74.7%, and 15.7%; and oropharyngeal SCC, 83.1%, 77.8%, and 48.7%)22,23.

The technology underlying MCED tests relies on the detection of tumor-associated circulating cfDNA. As such, not all cancers are detectable with this technology as not all tumors and tumor types shed cfDNA in quantities above the clinical limit of detection (LOD). For example, in the CCGA study, whereas overall sensitivity for head and neck cancer was 85.7% and ovarian cancer 83.1%, it was only 18.2% for renal cancer, which is known to be among the lower shedding tumor types. For low-grade prostate cancers, which are associated with indolent behavior, detection rates are under 4%7. The three cases described here benefited from the fact that their tumors shed cfDNA at a level above the clinical LOD for this test; several lines of evidence suggest that tumors that do so are associated with the potential for aggressive behavior even in early stages24,25.

Some considerations should be noted when weighing the clinical insights supported by these cases. First, in these 3 cases, CSO calls corresponded with the tumor type diagnosed. There was no evidence of a second type of cancer during comprehensive clinical evaluation and follow-up for at least one year in all 3 cases, indicating that tumor shedding was from the cancers that were diagnosed and treated. In addition, though the individual cases presented here are by design meant to serve as case illustrations, it is still appropriate to acknowledge that they represent only a small subset of a larger set of individuals who have received this test.

MCED represents a new paradigm with the potential to address a significant unmet need in cancer screening. By combining next-generation genome sequencing and machine learning, MCED tests can detect multiple cancer types, including those that are insufficiently prevalent to allow for efficient single-cancer screening26,27. Because this test detects a shared cancer signal across multiple cancer types, individual cancer prevalence can be aggregated across multiple cancers to improve screening efficiency, resulting in a much higher PPV and overall cancer detection rate than currently endorsed screening tests2,28. In addition, the machine-learning algorithms continuously learn from new data of the kind presented here, so the test performance characteristics can continuously improve.

Machine learning is a subcategory of the broader field of artificial intelligence and uses algorithms to automatically learn insights and recognize patterns from data, applying that learning to make increasingly better decisions29. In this case, to learn which cfDNA fragments may have originated from cancerous cells, the classifier algorithm was initially trained on sequencing data from more than 15,000 individuals in the CCGA study that enrolled participants between 2016 and 20186,7. This study comprised 6670 individuals without cancer and 8584 individuals with cancer for whom the cancer type was also recorded along with any comorbidities. The first step of the classifier training phase was deciding the right way to encode DNA methylation status so that it is computer-readable (“representation”). Second, the algorithm compared the patterns of methylation from individuals without cancer in CCGA to the individuals known to have cancer and derived a shared cancer signal (“learning”). This cancer signature is almost never observed in people known not to have cancer. Finally, the algorithm assigned a score to each individual that estimated the likelihood that they had cancer, and then assigned each of these likelihoods to one of two bins: cancer signal detected, i.e., test positive, or not, i.e., test negative (“thresholding and scoring”). Once the classifier was trained in this way and passed the representation, learning, and scoring stages, it was tested and validated on additional data that it had not seen yet. If the classifier returns a test positive, a second algorithm is triggered, to learn which cells the cancerous cfDNA fragments came from, resulting in the prediction of a CSO. The training stage runs on 1600 computer processors and takes four hours, while the day-to-day predictions run on 48 processors and take one minute. This approach was selected as it enables a continuous learning environment, where we can train the classifier on more diverse data driving improved performance over time.

Unlike current single-cancer tests, which are calibrated to maximize sensitivity and thus have higher false positive rates, MCED tests are designed for high specificity and very low false positive rates (<1%) with promise to minimize potential harms. Importantly, the MCED test used in these cases provides a prediction of the cancer signal origin, which can facilitate streamlined diagnostic evaluations. These 3 cases are not meant to stand on their own as evidence for clinical use but provide examples of the power and potential of the test for early-stage diagnosis and how new AI-based technologies can be directly applied to real-world clinical settings to optimize patient care. The cases should be reviewed in the context of robust clinical trial data and ongoing real-world evidence accrual, which support clinical use as an LDT. When used at a population level, MCED tests have the potential to reduce cancer mortality by intercepting cancers at earlier stages28.