Early Identification of Speech Changes Due to Amyotrophic Lateral Sclerosis Using Machine Classification

Abstract

We used a machine learning (ML) approach to detect bulbar amyotrophic lateral sclerosis (ALS) prior to the onset of overt speech symptoms. The dataset included speech samples from 123 participants who were stratified by sex and into three groups: healthy controls, ALS symptomatic, and ALS presymptomatic. We compared models trained on three group pairs (symptomatic-control, presymptomatic-control, and all ALS-control participants). Using acoustic features obtained with the OpenSMILE ComParE13 configuration, we tested several feature filtering techniques. ML classification was achieved using an SVM model and leave-one-out crossvalidation. The most successful model, which was trained on symptomatic-control data, yielded an AUC=0.99 for females and AUC=0.91 for males. Models trained on all ALS-control participants had high diagnostic accuracy for classifying symptomatic and presymptomatic ALS participants (females: AUC=0.85; males: AUC=0.91). Additionally, probabilities from these models correlated with speaking rate (females: Spearman coefficient=-0.60, p<0.001; males: Spearman coefficient=-0.43, p<0.001) and intelligible speaking rate (females: Spearman coefficient=-0.65, p<0.001; males: Spearman coefficient=-0.40, p<0.01), indicating their possible use as a severity index of bulbar motor involvement in ALS. These results highlight the importance of stratifying patients by speech severity when testing diagnostic models and demonstrate the potential of ML classification in early detection and progress monitoring of ALS.

Detection of Bulbar Involvement in Patients With Amyotrophic Lateral Sclerosis by Machine Learning Voice Analysis: Diagnostic Decision Support Development Study

“…An et al [ 23 ] used convolutional neuronal networks (CNNs) to compare the intelligible speech produced by patients with ALS to that of healthy individuals. Gutz et al [ 24 ] merged SVM and feature filtering techniques (SelectKBest). In addition, Vashkevich et al [ 25 ] used linear discriminant analysis (LDA) to verify the suitability of the sustain vowel phonation test for automatic detection of patients with ALS.…”

“…Once the features were obtained, we used various classification algorithms to perform predictions based on supervised classification. In addition to traditional SVMs [ 9 , 16 , 21 , 22 , 24 ], NNs [ 9 , 16 , 23 ], and LDA [ 25 ], we used logistic regression (LR), which is one of the most frequently used models for classification purposes [ 29 , 30 ]; random forest (RF) [ 31 ], which is an ensemble method in machine learning that involves the construction of multiple tree predictors that are classic predictive analytic algorithms [ 22 ]; and naïve Bayes (NaB), which is still a relevant topic [ 32 ] and is based on applying Bayes’ theorem.…”

A Novel Hierarchical Framework for Measuring the Complexity and Irregularity of Multimodal Speech Signals and Its Application in the Assessment of Speech Impairment in Amyotrophic Lateral Sclerosis

“…Although most acoustic measures except the irregularity of the low-frequency subband signal did not show a substantial contribution to the classification between individuals with ALS and healthy controls, it by no means suggests that acoustic information is not useful in detecting speech impairment in ALS. In fact, previous studies have demonstrated success in using acoustic signals (e.g., melfrequency cepstral coefficients; filterbank energies) to differentiate individuals with ALS from healthy speakers (An et al, 2018;Gutz et al, 2019). Yet a limitation of these studies is that the acoustic features used to train the classification models are not interpretable from a physiologic perspective.…”

Detecting Bulbar Involvement in Patients with Amyotrophic Lateral Sclerosis Based on Phonatory and Time-Frequency Features

“…An et al [ 25 ] employed Convolutional Neural Networks to classify the intelligible speech of ALS patients compared to healthy people. Finally, Gutz et al [ 26 ] combined SVM and feature filtering techniques.…”