Visualizing the Interpretation of a Criteria-Driven System That Automatically Evaluates the Quality of Health News:...

Machine learning techniques have been shown to be efficient in identifying health misinformation, but the results may not be trusted unless they can be justified in a way that is understandable.

This study aimed to provide a new criteria-based system to assess and justify health news quality. Using a subset of an existing set of criteria, this study compared the feasibility of 2 alternative methods for adding interpretability. Both methods used classification and highlighting to visualize sentence-level evidence.

A total of 3 out of 10 well-established criteria were chosen for experimentation, namely whether the health news discussed the costs of the intervention (the cost criterion), explained or quantified the harms of the intervention (the harm criterion), and identified the conflicts of interest (the conflict criterion). The first step of the experiment was to automate the evaluation of the 3 criteria by developing a sentence-level classifier. We tested Logistic Regression, Naive Bayes, Support Vector Machine, and Random Forest algorithms. Next, we compared the 2 visualization approaches. For the first approach, we calculated word feature weights, which explained how classification models distill keywords that contribute to the prediction; then, using the local interpretable model-agnostic explanation framework, we selected keywords associated with the classified criterion at the document level; and finally, the system selected and highlighted sentences with keywords. For the second approach, we extracted sentences that provided evidence to support the evaluation result from 100 health news articles; based on these results, we trained a typology classification model at the sentence level; and then, the system highlighted a positive sentence instance for the result justification. The number of sentences to highlight was determined by a preset threshold empirically determined using the average accuracy.

The automatic evaluation of health news on the cost, harm, and conflict criteria achieved average area under the curve scores of 0.88, 0.76, and 0.73, respectively, after 50 repetitions of 10-fold cross-validation. We found that both approaches could successfully visualize the interpretation of the system but that the performance of the 2 approaches varied by criterion and highlighting the accuracy decreased as the number of highlighted sentences increased. When the threshold accuracy was ≥75%, this resulted in a visualization with a variable length ranging from 1 to 6 sentences.

We provided 2 approaches to interpret criteria-based health news evaluation models tested on 3 criteria. This method incorporated rule-based and statistical machine learning approaches. The results suggested that one might visually interpret an automatic criterion-based health news quality evaluation successfully using either approach; however, larger differences may arise when multiple quality-related criteria are considered. This study can increase public trust in computerized health information evaluation.