Validating Behavioral Proxies for Disease Risk Monitoring via Large-Scale E-commerce Data

TL;DR

This study addresses the challenge of validating digital traces from e-commerce data for disease risk monitoring by introducing a behavioral proxy: switching from regular to therapeutic diets in large-scale cat-food purchases. It validates the proxy against two independent insurance-derived ground-truth datasets for feline FLUTD, reporting a strong ingredient-level correlation of $r = 0.74$ and a seasonal-component correlation of $r = 0.82$, and demonstrates a dose-response relationship with wet-food consumption. The results show that, once validated, EC data can provide reliable, scalable signals that complement traditional surveillance, particularly for lifestyle-related and chronic conditions. The framework offers a general approach to bridging web-scale behavioral signals and clinical outcomes, with potential extensions to human health surveillance.

Abstract

Digital traces of everyday behavior, such as e-commerce (EC) purchase logs, provide scalable signals for population-level monitoring, yet their epidemiological validity remains unclear due to weak links to clinical outcomes. We propose a behavioral proxy for disease onset based on transitions from regular to therapeutic diets observed in EC purchase histories, and evaluate its validity through large-scale cross-domain analysis. Using EC purchase data (N = 55,645 users) and independent insurance-derived clinical records, we compare ingredient-level risk patterns and seasonal disease dynamics in feline lower urinary tract disease (FLUTD) as a case study. The proxy-based estimates show strong agreement with clinical data, with correlations of r = 0.74 for ingredient-level risk patterns and r = 0.82 for seasonal variation. Both data sources consistently capture elevated disease risk during winter months. Moreover, analysis using EC data alone reproduces established domain knowledge, including the association between higher wet food consumption and lower disease risk. Our results demonstrate that behavioral signals derived from large-scale EC data can serve as validated, cost-effective complements to traditional surveillance systems, and suggest broader applicability to monitoring lifestyle-related and chronic conditions.

Figures (5)

• Figure 1: Correspondence between clinical records and EC purchase data. The two data sources are not directly linked at the individual level.
• Figure 2: Definition of Case Group and Control Group extracted from time-series insurance records.
• Figure 3: Correlation between Claim Rate (Ground Truth) and Switch Rate (Proxy) across ingredient categories ($r = 0.74$).
• Figure 4: Comparison of seasonal components of FLUTD insurance claims and first-time therapeutic diet purchasers on EC platforms ($r = 0.82$).
• Figure 5: Relationship between wet food rate and disease risk (Switch Rate).