Semantic and phonetic markers in schizophrenia-spectrum disorders; a combinatory machine learning approach

IntroductionSpeech is a promising marker for schizophrenia-spectrum disorder diagnosis, as it closely reflects symptoms. Previous approaches have made use of different feature domains of speech in classification, including semantic and phonetic features. However, an examination of the relative contribution and accuracy per domain remains an area of active investigation. Here, we examine these domains (i.e. phonetic and semantic) separately and in combination. MethodsUsing a semi-structured interview with neutral topics, speech of 94 schizophrenia-spectrum subjects (SSD) and 73 healthy controls (HC) was recorded. Phonetic features were extracted using a standardized feature set, and transcribed interviews were used to assess word connectedness using a word2vec model. Separate cross-validated random forest classifiers were trained on each feature domain. A third, combinatory classifier was used to combine features from both domains. ResultsThe phonetic domain random forest achieved 81% accuracy in classifying SSD from HC. For the semantic domain, the classifier reached an accuracy of 80% with a sparse set of features with 10-fold cross-validation. Joining features from the domains, the combined classifier reached 85% accuracy, significantly improving on models trained on separate domains. Top features were fragmented speech for phonetic and variance of connectedness for semantic, with both being the top features for the combined classifier. DiscussionBoth semantic and phonetic domains achieved similar results compared with previous research. Combining these features shows the relative value of each domain, as well as the increased classification performance from implementing features from multiple domains. Explainability of models and their feature importance is a requirement for future clinical applications.