Leveraging e-commerce user behavior data for common chronic diseases prediction

Background and Purpose Over half of people worldwide frequently shop on e-commerce platforms. By analyzing behavioral patterns on these platforms, it may be possible to predict chronic diseases in users. Methods We collected data from one of China’s largest e-commerce and internet healthcare platforms, JD.com. The number of purchases, cart additions, and browses for each user across 13,354 product categories over the past two years was collected, and a score was calculated as 10×purchase + 5×cart + browse for each category. Age, gender, marital status, education level, and delivery address were also included. The eight chronic diseases studied were overweight/obesity, hypertension, diabetes, lipid metabolism disorders, hyperuricemia/gout, anemia, chronic liver disease, and chronic kidney disease, diagnosed through real-name purchases of relevant prescription medications or online consultations. We selected users with behavior counts >200 within two years as the inclusion threshold, randomly including 500,000 patients and 500,000 undiagnosed users for each disease. Mann-Whitney U tests were used to compare behavioral differences between the two groups for each disease, with p-values adjusted using Bonferroni correction. Models were built using features with significant differences, excluding disease-specific medications and devices when modeling each disease. Since it is a positive-unlabeled learning problem and we couldn’t confirm whether undiagnosed users had the disease, patients were considered as initial positive samples and undiagnosed users as initial negative samples. We used XGBoost and a two-step method to iteratively update positive and negative samples. After iteration, all samples were divided in a 7:3 ratio for training and internal validation sets to build the final disease prediction models. Additionally, external validation was conducted using a dataset of over 196,778 individuals, with diagnoses determined by self-reported medical history or abnormal physical examination results. Results The top 10 most important features for each disease prediction model are shown in Figure 1. On average, 896 categories showed significant differences between patients and undiagnosed users. The average AUC for diseases in internal validation was 0.83, while in external validation it was 0.68 (see Figure 2). The performance drop might be related to differing diagnostic standards between the training and external validation sets. In the external validation set, the risks of the eight diseases for predicted positive users compared to predicted negative users were 1.78, 2.11, 2.84, 1.98, 3.39, 3.07, 2.51, and 3.01 times higher, respectively. Conclusions User behavior patterns on e-commerce platforms are associated with chronic diseases and may be used for personal health risk assessment and public health management.The top 10 most important features ROC curve for validation sets