Measures of Voice Quality as Indicators of Mild Cognitive Impairment

Abstract

AbstractBackgroundNoticeable decline in communication skills of patients with Alzheimer’s disease are often observed in the early stages of the disease, years before the dementia emerges. Prior literature has identified language abnormalities such as decline of linguistic ability or syntactic deficits influenced by the cognitive impairment. While language‐related measures are able to characterize spoken words of speech samples, they are inherently incapable of quantifying the voice‐related properties such as emotion or voice quality. Focusing on automatically derived voice‐based biomarkers that reflect the voicing quality of speakers, our objective is to explore the usefulness of these measures at detecting mild cognitive impairment (MCI) as early indicators of MCI in machine learning models. We hypothesize that cognitive impairment influences the individual’s voice quality resulting in an abnormal periodicity in speech often known as vocal dysperiodicity.MethodWe analyzed acoustic properties of audio‐recordings of Craft Story test administered to 107 subjects (50 MCI and 57 healthy controls, mean age 81 yrs) from a recently completed randomized control trial aimed to enhance cognitive reserve by increasing social interactions using video‐chats (www.I-conect.org). In the Craft Story test, the examiner reads a short story to the participant who is asked to retell the story first immediately after hearing the story and next retell it after a 20‐minute delay. Using cepstral analysis method, we extracted cepstrum from both immediate recall (IR) and delayed recall (DR) tests, and utilized statistical measures to represent their acoustic characteristics. We also used recursive feature elimination to only select informative measures and utilized the support vector machine for building classification models.ResultCepstrum‐based Measures derived from the IR test achieved higher predictive ability than those derived from the DR test. Combining the measures derived from both IR and DR tests increased the classification results and achieved ROC AUC of 98.83% and 97.09% on gender‐balanced models trained on males and females, respectively.ConclusionOur results indicate that voice‐based measures are capable of detecting MCI from audio samples of the Craft Story capitalizing their potential use in automatic screening and telemonitoring of the disorder in research, clinical trials and ultimately clinical practice.