JMIR Publications

ABSTRACT

Background:

Digital biomarkers are increasingly being used to provide clinical decision support for various health conditions. Speech signals as one of these biomarkers can potentially offer insights into underlying biological processes due to the complexity of speech production. Speech production involves the physical speech mechanisms, the respiratory system, brain-cantered cognition, and motor systems for the preparation and execution of speech movements. Therefore, deficits in any of these systems can cause changes in speech signal patterns. Advances in speech processing techniques, artificial intelligence, and data analytics potentially can lead to the invention of a new generation of speech-based clinical decision support systems.

Objective:

This systematic scoping review aimed to investigate the technological revolution and trends in recent digital clinical speech signal analysis to understand the key concepts and research processes from both clinical and technical perspectives.

Methods:

A systematic search was undertaken in six databases, guided by a set of research questions. Articles were identified that focused on speech signal analysis for clinical decision-making. The included studies were analyzed quantitatively. A narrower scope of studies investigating neurological diseases was analyzed using qualitative content analysis

Results:

A total of 389 articles met the initial eligibility criteria of which 77 articles focused on neurological diseases were included in the qualitative analysis. Within the included studies, Parkinson's disease, Alzheimer’s disease, and cognitive disorders were the most frequently investigated. The literature investigated the potential for using speech signal analysis to support diagnosis, differential diagnosis, severity assessment, and treatment monitoring of neurological conditions. The most common speech tasks used to extract speech features were sustained phonations, diadochokinetic tasks, reading tasks, activity-based tasks, picture descriptions, and prompted speech tasks. From these tasks, classical speech features, advanced digital signal processing-based speech features, and audio images in the form of spectrograms were analyzed. Traditional machine learning and deep learning approaches were applied to make clinical predictions from those features while statistical analysis techniques were used to test statistical relationships and reliability of speech features. Model evaluations had mostly been confined to analytical validations. The need for a guided research process to lead research studies toward potential technological intervention in clinical settings was identified as a major research gap. A research framework was proposed adapting a research methodology in design science to guide research studies systematically.

Conclusions:

The findings reflect the contribution of the data science approach to exploiting the potential for speech signal analysis to aid clinical decision-making. Connecting interdisciplinary knowledge through a guided research process could make research efforts more clinically relevant in the future.