Using Natural Language Processing to identify symptoms and risk factors associated with lung cancer in primary care records (360G-Wellcome-228290_Z_23_Z)

Background: Lung cancer is the third most common cancer in England, but the leading cause of cancer death. Most patients with lung cancer are not diagnosed until the disease is advanced, which contributes to the poor patient outcomes. Patients with lung cancer usually visit their GP with symptoms of the disease before diagnosis, and symptom based electronic tools are used to help GPs make decisions about urgent cancer referrals. However, such tools often rely on codes entered by clinical staff, ignoring the vast amount of information entered as free text. My approach: I will use an approach called Natural Language Processing (NLP) to automatically pull-out key information on risk factors (e.g. smoking) and symptoms of lung cancer from the GP records of around 150,000 patients who were risk assessed for cancer. I will compare this with coded information to determine whether there are differences in the frequency or type of information recorded. I will compare how the accuracy of commonly used lung cancer risk assessment tools changes when using coded and NLP information vs coded information alone. Impact: This research will help determine whether NLP could improve the accuracy of lung cancer risk assessment and cancer detection in primary care