The present invention is provided with: a learning data input unit 10 that inputs m sentences included in an electronic health record of a patient; a similarity index value calculation unit 100 that extracts n words from the m sentences and calculates a similarity index value reflecting a relation between the m sentences and the n words; a classification model generation unit 14 that, on the basis of a sentence index value group comprising n similarity index values regarding one sentence, generates a classification model for classifying the m sentences into a plurality of phenomena; and a dangerous behavior prediction unit 21 that predicts a possibility of overturn occurrence from a to-be-predicted sentence by applying, to the classification model, the similarity index value calculated by the similarity index value calculation unit 100 from the sentence inputted by a prediction data input unit 20. The present invention generates a high-precision classification model by using the similarity index value indicating the extent to which a given word contributes to a given sentence.
展开▼