The present paper collected the asthma related 4,762 Web pages from 1,759 sites using 6 queries. Each site is manually categorized by the standard topics of description and information dissemination, diary and idle talk and Q&A. By careful analysis, it turned out that the pages can be classified in non-topical categories such as 'reading level', 'objectivity/subjectivity' and 'reliability'. The manually assigned labels of non-topical categories are then used as learning data to apply SVM (support machine vector). The prediction performance (F-measure) were below 50% with the naive application of SVM. However, the prediction performance was improved over 50% by feature selection except for reading level.