Automated Detection of Health Websites' HONcode Conformity: Can N-gram Tokenization Replace Stemming?

Area of Science:

• Computational linguistics
• Health informatics
• Web science

Background:

• Evaluating healthcare website compliance with standards like HONcode is crucial for user trust and information accuracy.
• Manual review processes are time-consuming and resource-intensive.
• Automated classification methods are being explored to streamline this evaluation.

Purpose of the Study:

• To assess the efficacy of supervised automatic classification algorithms for determining healthcare web page compliance with HONcode criteria.
• To compare the performance of character n-gram vectors against traditional word-based and stemmed word-based classification methods.
• To investigate the potential of language-independent approaches for health website classification.

Main Methods:

• Utilized supervised machine learning algorithms, specifically a Naive Bayes classifier.
• Employed varying length character n-gram vectors to represent healthcare web page documents.
• Compared n-gram tokenization with document words and Porter-stemmed document words, using Document Frequency (DF) for dimensionality reduction.

Main Results:

• Character n-gram tokenization demonstrated comparable performance to document word and stemmed document word classification.
• Using 5-grams as features, n-gram tokenization proved to be a potentially viable alternative to document word stemming.
• Compared dimensionality reduction techniques (DF, Chi-square, Z-score), indicating n-gram's robustness.

Conclusions:

• Character n-gram tokenization presents a promising, language-independent method for automated healthcare web page compliance assessment.
• This approach may offer a safe and efficient alternative to traditional word-based classification methods.
• Further research could explore n-gram models for broader applications in health information quality assessment.