Development and validation of risk prediction algorithms to estimate future risk of common cancers in men and women: prospective cohort study

• Published in: BMJ open Vol. 5; no. 3; p. e007825
• Main Authors: Hippisley-Cox, Julia, Coupland, Carol
• Format: Journal Article
• Language: English
• Published: England BMJ Publishing Group LTD 17.03.2015; BMJ Publishing Group
• Subjects: Adult; Age Factors; Aged; Aged, 80 and over; Alcohol Drinking - epidemiology; Alcohol use; Algorithms; Antipsychotic Agents - therapeutic use; Bipolar Disorder - epidemiology; Body Mass Index; Breast cancer; Chronic illnesses; Codes; Comorbidity; Contraceptives, Oral - therapeutic use; Diabetes; Diabetes Mellitus, Type 1 - epidemiology; Diabetes Mellitus, Type 2 - epidemiology; Disease; Electronic health records; England - epidemiology; Epidemiology; Family medical history; Female; Forecasting - methods; Gynecology; Hormone Replacement Therapy; Hospitals; Humans; Incidence; Male; Medical diagnosis; Medical research; Middle Aged; Mortality; Neoplasms - epidemiology; Neoplasms - ethnology; Neoplasms - genetics; Neoplasms, Second Primary - epidemiology; Ovarian cancer; Patient admissions; Population; Poverty; Primary care; Proportional Hazards Models; Prospective Studies; Risk Assessment - methods; Risk factors; ROC Curve; Schizophrenia - epidemiology; Smoking; Smoking - epidemiology; Ulcers; Womens health; England; United Kingdom > UK; cancer; PRIMARY CARE; QResearch
• ISSN: 2044-6055; 2044-6055
• DOI: 10.1136/bmjopen-2015-007825

Objective To derive and validate a set of clinical risk prediction algorithm to estimate the 10-year risk of 11 common cancers. Design Prospective open cohort study using routinely collected data from 753 QResearch general practices in England. We used 565 practices to develop the scores and 188 for validation. Subjects 4.96 million patients aged 25–84 years in the derivation cohort; 1.64 million in the validation cohort. Patients were free of the relevant cancer at baseline. Methods Cox proportional hazards models in the derivation cohort to derive 10-year risk algorithms. Risk factors considered included age, ethnicity, deprivation, body mass index, smoking, alcohol, previous cancer diagnoses, family history of cancer, relevant comorbidities and medication. Measures of calibration and discrimination in the validation cohort. Outcomes Incident cases of blood, breast, bowel, gastro-oesophageal, lung, oral, ovarian, pancreas, prostate, renal tract and uterine cancers. Cancers were recorded on any one of four linked data sources (general practitioner (GP), mortality, hospital or cancer records). Results We identified 228 241 incident cases during follow-up of the 11 types of cancer. Of these 25 444 were blood; 41 315 breast; 32 626 bowel, 12 808 gastro-oesophageal; 32 187 lung; 4811 oral; 6635 ovarian; 7119 pancreatic; 35 256 prostate; 23 091 renal tract; 6949 uterine cancers. The lung cancer algorithm had the best performance with an R2 of 64.2%; D statistic of 2.74; receiver operating characteristic curve statistic of 0.91 in women. The sensitivity for the top 10% of women at highest risk of lung cancer was 67%. Performance of the algorithms in men was very similar to that for women. Conclusions We have developed and validated a prediction models to quantify absolute risk of 11 common cancers. They can be used to identify patients at high risk of cancers for prevention or further assessment. The algorithms could be integrated into clinical computer systems and used to identify high-risk patients. Web calculator: There is a simple web calculator to implement the Qcancer 10 year risk algorithm together with the open source software for download (available at http://qcancer.org/10yr/).