D Tighe, D Howell, A George, R Buttery
This study was conducted to predict future prevalence patterns of head and neck cancer and skin cancer within the Kent and Medway region using data derived from the Kent Integrated Dataset. Machine learning methodologies were employed to model the prevalence of disease, and geospatial analysis techniques were applied to visually represent the predicted patterns of that prevalence across the region. The Kent integrated dataset (KID) was searched for instances of head and neck cancer and skin cancer diagnoses. Within a machine learning environment, we built prevalence prediction models using associated demographic, comorbidity, lifestyle, and exposure data of patients with/without a diagnosis of head and neck cancer, or skin cancer between 2014-2019 recorded on the KiD. Over two million care episodes were analysed. Of these 1 249 had a diagnosis of head and neck cancer and 11 152 had a diagnosis of skin cancer. We demonstrated that demographic and comorbidity data were features worthy of inclusion in prevalence prediction models. We did not identify socio-economic, environmental or coastal vs non-coastal geography as independent predictors for either diagnosis. Logistic regression models were chosen as the most suitable predictive models. Model performance is reported at different population screening thresholds (1%, 5%, 10% and 25%). Prevalence prediction based on input data alone was used to produce geo-spatial heatmaps in Kent & Medway. Model build is challenging for rare diagnoses but we were able to demonstrate a data-driven process that highlights the highest risk geographic regions of head and neck cancer and skin cancer in Kent and Medway.