Over 7.4 million U.K. adults’ anonymized electronic health records were used to create two new algorithms to predict a seemingly cancer-free patient’s chances of actually having the disease. The predictive algorithms use information about a person’s health conditions and results from simple blood tests. They were developed by researchers from Queen Mary University of London and the University of Oxford.
The new models, they say, are much more sensitive than existing models and could thus “revolutionize” how cancer is detected in primary care, making it easier for patients to get treatment at much earlier stages.
The study appears in Nature Communications. The lead author is Julia Hippisley-Cox, MD, Professor of Clinical Epidemiology and Predictive Medicine at Queen Mary University of London. CA Coupland, University of Oxford, is the co-author.
Hippisley-Cox said, “These algorithms are designed to be embedded into clinical systems and used during routine GP consultations. They offer a substantial improvement over current models, with higher accuracy in identifying cancers—especially at early, more treatable stages.”
She added, “They use existing blood test results which are already in the patients’ records making this an affordable and efficient approach to help the NHS meet its targets to improve its record on diagnosing cancer early by 2028.”
The researchers developed and externally validated two diagnostic prediction algorithms to estimate the probability of having cancer for 15 cancer types, including hard-to-diagnose liver and oral cancers. The first algorithm incorporates multiple predictors including age, sex, deprivation, smoking, alcohol, family history, medical diagnoses and symptoms (both general and cancer-specific symptoms). The second also includes commonly used blood tests, including full blood count and liver function tests.
The NHS currently uses prediction algorithms, such as the QCancerscores, to combine relevant information from patient data and identify individuals deemed at high risk of having a currently undiagnosed cancer. GPs and specialists can then do further testing on these patients.
Crucially, in addition to information about a patient’s age, family history, medical diagnoses, symptoms, and general health, the new algorithms incorporated the results of seven routine blood tests (which measure a person’s full blood count and test liver function) as biomarkers to improve early cancer diagnosis.
Compared with the QCancer algorithms, the new models identified four additional medical conditions associated with an increased risk of 15 different cancers including those affecting the liver, kidneys, and pancreas. Two additional associations were also found for family history with lung cancer and blood cancer, and seven new symptoms of concern (including itching, bruising, back pain, hoarseness, flatulence, abdominal mass, dark urine) were identified as being associated with multiple cancer types.
The new algorithms appear to offer much improved diagnostic capabilities. They are the only ones currently that can be used in primary care settings to estimate the likelihood of having an existing but undiagnosed liver cancer.
Carol Coupland, senior researcher at the Queen Mary University of London and Emeritus Professor of Medical Statistics in Primary Care at the University of Nottingham, and co-author, said, “These new algorithms for assessing individuals’ risks of having currently undiagnosed cancer show improved capability of identifying people most at risk of having one of 15 types of cancer based on their symptoms, blood test results, lifestyle factors and other information recorded in their medical records.”
She added, “They offer the potential for enabling earlier cancer diagnoses in people from the age of 18 onwards, including for some rare types of cancer type.”
