Accuracy of an AI prediction platform in predicting tumor origin: A real-world study.

Abstract

74 Background: Only four cancers are covered by screening programs. The remaining cancers are diagnosed after a point-of-care presentation to a primavry care physician (PCP). Patients with these cancers typically present with a combination of vague and non-specific symptoms, lacking specificity to a particular tumor type. The difficulties in interpreting these symptoms by PCPs leads to increased diagnostic intervals and the use of invasive diagnostic modalities with unnecessary radiation exposure. Localizing the cancer origin is critical to reduce time to diagnosis in order to improve earlier diagnosis and survival in these cancer types. C the Signs is a clinical platform integrated with electronic health records (EHR) that uses artificial intelligence to help identify patients at risk of cancer. Used by PCPs during a point-of-care assessment, C the Signs analyses the symptoms the patient presents with, alongside clinical data in the EHR, to identify current cancer risk, predict the tumor site of origin, and recommend the most effective diagnostic or referral pathway. The aim of this study is to evaluate the accuracy of C the Signs in predicting cancer origin for patients in a real-world setting. Methods: An observational study was conducted between 1st January 2021 and 31st October 2022 in the National Health Service in England utilizing all patients who were risk-assessed with the C the Signs EHR integrated platform. There was no pre-selection criteria for the patient population assessed. All patients were followed up with for 6 months post-risk assessment to determine if they had a cancer diagnosis. Analysis was performed on the patients who were identified at risk of cancer by the C the Signs platform and subsequently diagnosed with cancer, to establish the cancer origin accuracy. Results: 111,421 patients were risk-assessed by C the Signs, of which 7,360 were diagnosed with cancer. Of the patients diagnosed with cancer, 7,257 patients were identified at risk of cancer by C the Signs (98.67%). The platform predicted cancer origin accurately in 93.2% of patients. The platform was most accurate in predicting cancer origin for brain and CNS, breast, lung, gynecological, head & neck, bowel, skin, and prostate cancers. Conclusions: This study demonstrates how a cancer prediction platform such as C the Signs can be used to accurately detect tumor origin in patients with vague symptoms and facilitate targeted investigations appropriate to the tumor type. Further research is required to identify which cancers can be most accurately detected taking a computational approach.