JMIR Publications

ABSTRACT

Background:

An increasing number of users are turning to online sources which is a significant source of patient information for healthcare guidance. Thus, trustworthy sources of information should be automatically identifiable using objective criteria.

Objective:

The purpose of this study was to automate the assessment of the HONcode criteria, enhancing our ability to pinpoint trustworthy health information sources.

Methods:

A dataset of 538 webpages displaying health content was collected from 43 health-related websites. HONcode criteria have been split into two levels ((1) webpage and (2) website levels). For the website-level criteria (confidentiality, transparency, financial disclosure, and advertising policy), a bag of keywords has been identified to assess the criteria, and for webpage-level criteria (authority, complementarity, justifiability, and attribution) a machine learning (ML) approach has been used and 200 webpages were manually annotated. For website-level criteria, a list of words for each criterion was generated and validated by consensus. Three ML models -Random Forest (RF), Support Vector Machines (SVM), and BERT- were applied on the initial annotated data and evaluated. A second step of training was implemented for the complementarity criterion using the BERT model for multiclass classification of the complementarity sentences obtained by annotation and data augmentation (Positive, Negative, and non-committal complementarity sentences). Finally, the remaining web pages were classified using the selected model and 100 sentences were randomly selected for manual review.

Results:

The evaluation of the website's criteria was effectively conducted using a bag-of-words approach for confidentiality and transparency. However, this method was less effective for the other website-level criteria. For webpage level criteria, the RF model showed a good performance for the attribution criterion while displaying subpar performance in the others. BERT and SVM had a stable performance across the different criteria. BERT had a better area under the curve (AUC) of 0.96, 0.98, and 1.00 for neutral sentences, justifiability, and attribution respectively. SVM had the overall better performance for the classification of complementarity with the AUC equal to 0.98. Finally, SVM and BERT had an equal AUC of 0.98 for the authority criterion. Also, for the identification of website criteria the model was able to retrieve webpages with an accuracy of 97% for confidentiality, 82% for transparency, 51% for financial disclosure, and 51% for advertising policy. The final evaluation of the sentences determined 0.88 of precision and the agreement level of reviewers was computed 0.82.

Conclusions:

Our results showed the potential power of using a BERT model in automating the classification of HONcode criteria in web content. It achieved higher performance compared to traditional approaches using RF and SVM. This approach could be utilized with different types of pre-trained models to accelerate the text annotation, and classification and to improve the performance in low-resource cases.