New referrals
A substantial part of outpatient care in head and neck surgical oncology work involves assessment of patients referred in with symptoms suspicious of HNC. Apart from a few malignancies (basal cell carcinomas, low-grade salivary gland tumours and well-differentiated thyroid cancer in young patients), most HNCs progress within a matter of months, and a delay in management can have an adverse effect on the prognosis. In 2000, the Department of Health in the UK developed national guidelines for referral of suspected HNC from primary care, updated by the National Institute for Health and Care Excellence (NICE) in 2005 and 2015. Systematic reviews have shown that the pooled detection rate of cancer in this population is between 8.8% 8 and 11.1%9. Thus, the vast majority of referrals may be safely triaged to a deferred assessment.
In 2016, our group used information from a cohort of 4715 patients referred in for suspected HNC to generate a 13-symptom inventory which, when combined with age, could generate a personalised risk of cancer with a high diagnostic efficacy (area under the curve (AUC) 0.77)10. Sensitivity analysis identified the 8% risk probability cut off to offer optimal performance for the calculator. This risk calculator was subsequently validated in an external cohort of 1998 patients, where it performed well compared to the inception cohort, with an AUC of 0.81 11.
After further refinements using additional information to the symptom inventory in a new cohort (n= 3531), we published the Head and Neck Cancer Risk Calculator (HaNC-RC)-V.2, which demonstrated increased diagnostic efficacy with an area under the curve of 0.89, and a sensitivity of 85% 12. The model performed optimally at a probability cut off of 7.1%, where the negative predictive value (NPV) was 98.6%. Thus, when the symptom inventory is applied to a patient who has been referred for suspected cancer and the calculator indicates a <7% risk probability for cancer diagnosis, the chance of missing cancer if the patient is not seen by a face-to-face conventional consultation is 1.4%. In all these instances, a logistic regression model was used to define personalised probability risk for each patient. When a variety of machine learning algorithms were used on a similar cohort of 5082 patients referred for suspected HNC, logistic regression, the technique used to create the calculator, offered one of the highest true negative rates, an essential characteristic of a test to triage patients out during resource constrained times13.