Files

Abstract

The Health On the Net’s Foundation (HON) Code of Conduct, HONcode, is the oldest and the most used ethical and trustworthy code for medical and health related information available on the Internet. Until recently, websites voluntarily applying for the HONcode seal were evaluated manually by an expert medical team according to 8 principles, referred to as criteria, and associated published guidelines. In the scope of the European project Kconnect, HON is developing an automated system to identify the 8 HONcode criteria within health webpages. When the research on the development of such a system evolved from simple algorithmic testing to a real full-content setting, it revealed a number of issues. The preceding study consisted in taking a set of 27 health-related websites and having them assessed for their compliance to each of the 8 HONcode criterion, first manually by senior HONcode experts, and then through supervised machine learning by the automated system. The results showed disc repancies mainly for two criteria: “submerged content” under the Complementarity criterion and “extremely low recall” under the Date Attribution criterion. In this article, the authors investigate different approaches to solve the problems related to each of these criteria, namely a customized Named Entity Recognition Model instead of a machine learning component for Date Attribution, and a sliding window instead of the whole document as a unit of detection for Complementarity. The results obtained show that the newly adapted automated system greatly improves accuracy: 74% vs. 41% for the Date Attribution criterion and 74% vs. 22% for the Complementarity criterion.

Details

Actions

PDF