Automatic Voice Analysis for Dysphagia Screening in Neurological Patients

Background:

Oropharyngeal dysphagia, defined as any alterations in swallowing abilities during the transit of food or liquids from the oral cavity to the esophagus, is an insidious complication of many neurological diseases. This condition can seriously lead to severe complications such as malnutrition, dehydration, and pneumonia, which overall has a huge impact on management costs and the number of hospitalization days. In this context, it is essential to immediately recognize the risk factors and the first signs of dysphagia to take prompt adequate actions and request further clinical and instrumental evaluations. Rapid, quantitative, and effective dysphagia screening methods are not currently available to support healthcare personnel. To date, only clinical rating scales or expensive invasive tests that require specialized personnel are adopted in clinical scenarios, whereas no objective tools are still available in extra-hospital contexts to alert patients of risk situations.

Current Gaps in Knowledge and Aim:

Since oropharyngeal dysphagia is caused by an impaired coordination control of the swallowing muscles and these muscles play also an important role in the phonation process, investigating voice alterations could be a screening option to recognize dysphagia in patients with neurological diseases. In the current literature, automatic voice analysis and the use of machine learning algorithms have given relevant findings in the discrimination between neurological diseases and healthy subjects, and there are also interesting preliminary data on dysphagia. The goal of this study is to the development a machine learning classification algorithm for dysphagia screening in neurological patients using automatic voice analysis.

Study Involvement:

The study involves patients with neurological diseases (Parkinson's disease, stroke, amyotrophic lateral Sclerosis) with or without dysphagia and healthy individuals. The participants are asked to perform some vocal tasks (sustained vocal phonation, diadochokinetic tasks, production of standardized sentences, free speech) in a single experimental session at the enrolment. Voice recordings will be automatically proceeded to derive acoustic voice features, used as input for the machine learning classification algorithm. The evaluation of the participants to characterize the studied sample is carried out with the collection of anamnestic and clinical data.