VoiceLens: A multi-view multi-class disease classification model through daily-life speech data

Abstract

Biomarkers in the human voice can offer insight into neurological disorders because voice signals are influenced by the underlying cognitive and neuromuscular functions. In the past decade, there is an increasing research attention on voice-based neural disorder detection using machine learning techniques. However, existing works only attempt to detect a single neurological disorder (e.g., Parkinson’s or Huntington’s). In this work, we present the first computational model, namely VoiceLens , that detects multiple neurological disorders at the same time. The proposed VoiceLens framework combines the effectiveness of the powerful Mel-Frequency-Cepstral-Coefficients (MFCC) within a two-phase multi-class classification module to build an accurate voice-based disease prediction model. The first phase captures the fine-grained details of these disorders and their sequential variation patterns within a stacked Long-Short-Term-Memory (LSTM) network to make the baseline disease detection, i.e., healthy v.s. pathology. In the second phase, the detected pathology samples are further analyzed by a deep multi-layer learned descriptor to identify the disease types. The VoiceLens method is developed and evaluated using a large-scale Saarbruecken-Voice-Database comprising of samples from 2000 individuals with multiple disease patterns, including Laryngeal Cancers, Dish-Syndrome, and Parkinson’s disease. Experimental results show the remarkable performance of VoiceLens by reporting Accuracy up to 97.5% in the disease detection, where the model also obtains 98.00% and 97.13% for F1-Score and Recall . Also, compared with existing machine learning models, the proposed VoiceLens system demonstrates around 15% (and 12%) average gain in the Accuracy (and F1-score ) in a multi-disease identification test including six (6) different pathology classes and one (1) healthy class.