Machine-learning-based Phenomapping of Patients with Keratinocyte Carcinoma: Data-driven Subgrouping by Disease Burden, Comorbidities and Socioeconomic Status

Keratinocyte carcinoma (KC) places a considerable and growing burden on healthcare systems. Given the KC population's heterogeneity, tailored clinical pathways are needed to accommodate diverse management needs. This study applied machine learning (ML)-based phenomapping to identify distinct real-world subgroups within a national KC population using demographic and medical history variables. The study included KC patients treated in publicly-funded, office-based dermatology practices and registered in the Danish Skin Cancer Registry (2014-2022). In 106,490 KC patients, 7 ML-derived clusters were identified. One cluster (22.2%; n=23,590) consisted primarily of young, well educated high-income medically noncomplex females presenting with low-risk basal cell carcinomas (BCCs). Four clusters (55.2%; n=58,736) comprised patients with a greater dermatological disease burden, characterized by facial BCCs, multiple BCCs, a greater proportion of squamous cell carcinomas (SCCs), a history of skin cancer or previous actinic keratosis-related treatment. The remaining 2 clusters (22.7%; n=24,164) comprised highly comorbid patients with a greater proportion of SCCs, a high number of KCs on the head and neck and high immunosuppressive drug exposure. ML-derived phenomapping of a national KC population identified distinct, clinically relevant KC subgroups, providing a basis for tailored clinical pathways with differentiated resource needs.