Quantifying disease-interactions through co-occurrence matrices to predict early onset colorectal cancer.

Colorectal cancer (CRC) is the third most common cancer in terms of the number of cases and deaths in men and women in the USA. According to the Centers for Disease Control and Prevention, the CRC screening compliance rate remains low in the United States. It is even more concerning that the number of cases and deaths due to CRC is increasing in the younger population, for which there is no guideline to get colonoscopy screening. In this paper, we develop a novel network-based model to identify patients under 50 years of age having a high risk of CRC, particularly those who do not have a family history. Our model can help predict which patients are at risk of developing colorectal cancer to aid the practicing primary care physician and gastroenterologist to triage patients into meaningful risk groups to provide accelerated diagnostic steps and care to these patients. The model uses our proposed variables created through comorbidity network matrices obtained from CRC and non-CRC patients as inputs. We used the electronic medical records of thousands of CRC and non-CRC patients to develop and validate the models. Our model for younger patients correctly predicted 73.2% of the patients with future diagnoses with an area under the ROC curve of 0.81. At the 50% sensitivity, the false positive rate was 11.5%. The performance of our model is the highest in the current state-of-the-art. Our proposed variables quantifying the interactions between multiple diseases can also be adapted in future predictions of other diseases in a patient.