Fusing voice and query data for non-invasive detection of laryngeal disorders

Topic of this study is exploration and fusion of non-invasive measurements for an accurate detection of pathological larynx. Measurements for human subject encompass answers to items of a specific survey and information extracted by the openSMILE toolkit from several audio recordings of sustained phonation (vowel /a/). Clinical diagnosis, assigned by medical specialist, is a target attribute distinguishing subject as healthy or pathological. Random forest (RF) is used here as a base-learner and also as a meta-learner for decision-level fusion. 5 RF classifiers, built separately on 3 variants of audio recording data (raw and after two types of voice activity detection) and 2 variants of questionnaire (with 9 and 26 questions) data, are fused selectively by finding out the best combination of all possible. Before fusion, due to presence of missing values in query modalities, several imputation techniques were evaluated besides the complete-case analysis by listwise deletion. Out-of-bag equal error rate (EER) was found to be higher for audio data and lower for query, but each variant was outperformed by the decision-level fusion. Fusion after listwise deletion provided EER of 4.84%, meanwhile imputation was found to improve detection slightly and helped to achieve EER of 4.55%. Variable importance, as permutation-based mean decrease in RF accuracy, was reported for query and audio data. Finally, regarding the noteworthy performance of the query data, 22 association rules (11 healthy + 11 pathological) using 17 questions were obtained for comprehensible insights.