Unlocking the Potential of Patient Metadata for Skin Cancer Detection: An AI Framework

Early detection of suspicious skin lesions can significantly increase the five-year survival rates of the patients. Advancements in computer vision techniques facilitate the use of artificial intelligence (AI) models along with image data for skin cancer detection. However, there is limited work done on skin cancer detection solely based on patient metadata. The 7-point checklist (7PCL) and Williams methods use a limited number of meta-features to calculate skin lesion risk scores and to find a patient at risk of developing skin cancer, respectively. This study attempts to fill the gap and proposes an AI-based framework for classifying skin lesion metadata into binary classes: Suspicious vs Non-suspicious. The developed framework has been evaluated using real-world skin lesion metadata sourced from a network of private skin diagnostic clinics across the UK. We have collected and analyzed 54,000 skin lesions metadata, from 25,214 patients undergoing teledermatology assessment after clinical examination and imaging, comprising 25 features including patient age, gender, and lesion location. The metadata has been pre-processed through encoding, followed by feature selection using wrapper, Shapley, and Pearson correlation methods. Finally, five different predictive models were utilized and optimized to classify skin lesion metadata into Suspicious vs Non-suspicious classes. Our proposed approach achieved 83.53(±0.03) % sensitivity in detecting suspicious lesions using only metadata and outperformed the 7PCL and Williams methods. We believe this AI-based framework is unique in classifying skin lesions based solely on metadata and has significant potential to improve the performance of current AI models that are based on image assessment alone.