A panel of seven protein tumour markers for effective and affordable multi-cancer early detection by artificial intelligence: a large-scale and multicentre case-control study

Abstract

Background: Early detection of cancer aims to reduce cancer deaths. Unfortunately, many established cancer screening technologies are not suitable for use in low- and middle-income countries (LMICs) due to cost, complexity, and dependency on extensive medical infrastructure. We aimed to assess the performance and robustness of a protein assay (OncoSeek) for multi-cancer early detection (MCED) that is likely to be more practical in LMICs.

Methods: This observational study comprises a retrospective analysis on the data generated from the routine clinical testings at SeekIn and Sun Yat-sen Memorial Hospital. 7565 participants (954 with cancer and 6611 without) from the two sites were divided into training and independent validation cohort. The second validation cohort (1005 with cancer and 812 without) was from Johns Hopkins University School of Medicine. Patients with cancer prior to therapy were eligible for inclusion in the study. Individuals with no history of cancer were enrolled from the participating sites as the non-cancer group. One tube of peripheral blood was collected from each participant and quantified a panel of seven selected protein tumour markers (PTMs) by a common clinical electrochemiluminescence immunoassay analyser. An algorithm named OncoSeek was established using artificial intelligence (AI) to distinguish patients with cancer from those without cancer by calculating the probability of cancer (POC) index based on the quantification results of the seven PTMs and clinical information including sex and age of the individuals and to predict the possible affected tissue of origin (TOO) for those who have been detected with cancer signals in blood.

Findings: Between November 2012 and May 2022, 7565 participants were enrolled at SeekIn and Sun Yat-sen Memorial Hospital. The conventional clinical method, which relies only on a single threshold for each PTM, would suffer from a high false positive rate that accumulates as the number of markers increased. OncoSeek was empowered by AI technology to significantly reduce the false positive rate, increasing the specificity from 56.9% (95% confidence interval [CI]: 55.8-58.0) to 92.9% (92.3-93.5). In all cancer types, the overall sensitivity of OncoSeek was 51.7% (49.4-53.9), resulting in 84.3% (83.5-85.0) accuracy. The performance was generally consistent in the training and the two validation cohorts. The sensitivities ranged from 37.1% to 77.6% for the detection of the nine common cancer types (breast, colorectum, liver, lung, lymphoma, oesophagus, ovary, pancreas, and stomach), which account for ∼59.2% of global cancer deaths annually. Furthermore, it has shown excellent sensitivity in several high-mortality cancer types for which routine screening tests are lacking in the clinic, such as the sensitivity of pancreatic cancer which was 77.6% (69.3-84.6). The overall accuracy of TOO prediction in the true positives was 66.8%, which could assist the clinical diagnostic workup.

Interpretation: OncoSeek significantly outperforms the conventional clinical method, representing a novel blood-based test for MCED which is non-invasive, easy, efficient, and robust. Moreover, the accuracy of TOO facilitates the follow-up diagnostic workup.

Funding: The National Key Research and Development Programme of China.

Keywords: Artificial intelligence (AI); Liquid biopsy; Low- and middle-income countries (LMICs); Multi-cancer early detection (MCED); Protein tumour markers (PTMs).

Fig. 1

Schematic representation of clinical implementation workflow of OncoSeek test. 8 ml peripheral blood sample was collected from the individual in a cell-free DNA blood collection tube and mailed to the central lab. This is a special tube that proteins are stabilised stored at room temperature for seven days which makes it remote accessible as long as there's a local nurse who can draw blood. After plasma separation by centrifugation in the lab, PTM levels were measured by an electrochemiluminescence immunoassay analyser. OncoSeek was established using AI to distinguish cancer from non-cancer individuals by calculating the probability of cancer (POC) index based on the plasma levels of seven PTMs and clinical information including sex and age of the individuals. Then using another model to predict the possible affected TOO who has been detected with a cancer signal. PTMs, protein tumour markers. ECLI, electrochemiluminescence immunoassay. TOO, tissue of origin.

Fig. 2

Quantification of PTMs in different cancer types. Quantification value of each PTM (y-axis) based on healthy or individual cancer types (x-axis). The black horizontal lines are cut-off values that are recommended by the manufacturer. ∗∗ indicates P-value <0.01, ∗∗∗ indicates P-value <0.001.

Fig. 3

The performance of OncoSeek assay. (A) The receiver operating characteristic (ROC) curve evaluated the performance of OncoSeek in the training and independent validation cohorts. The area under the curve (AUC) of the three cohorts was depicted in the figure. The dotted vertical line in the ROC figures represents a 90.0% specificity. (B) The sensitivity of OncoSeek in individual tumour types. Sensitivity (y-axis) by cancer class based on individual cancer classes (x-axis), including multiple cancer types. Cancer classes are ordered based on sensitivity reducing, bars indicate 95% CI. The numbers in parentheses indicate the samples for each cancer class. (C) The sensitivity of OncoSeek in each clinical stage. Sensitivity (y-axis) based on individual cancer stage (x-axis), bars indicate 95% CI. The numbers in parentheses indicate the samples for each clinical stage.

Fig. 4

Receiver operating characteristic (ROC) curves of these seven PTMs and OncoSeek in individual cancer types. The area under the curve (AUC) of these seven PTMs and OncoSeek were depicted in the figure. The dotted vertical line in the ROC figures represents a 90.0% specificity.

Fig. 5

TOO accuracy by individual cancer type. Confusion matrices representing the accuracy of TOO localisation. Agreement between the actual (x-axis) and predicted (y-axis) TOO per sample using the OncoSeek model was depicted. Colour corresponds to the proportion of predicted TOO calls. Included 976 participants were those with cancer predicted as having cancer at 92.9% specificity.