Association of acoustic perturbation measures with dementia status: A Framingham Heart Study

Abstract Background Diagnosis of dementia due to Alzheimer’s disease (AD) is a resource‐intensive process, often requiring clinical assessments and neuroimaging which may fall beyond the reach of patients in remote settings. Given the challenges of traditional dementia work‐up, identification of disease‐relevant signatures from low‐cost modalities has attracted interest as a means to improve disease detection. Here, we describe a machine learning strategy to associate acoustic perturbations in voice with dementia status using digital recordings of neuropsychological exams administered in the Framingham Heart Study (FHS). Method We used 118 voice recordings from 79 participants to extract jitter, shimmer, and haronics‐to‐nosie ratio (HNR) from the ComParE 2016 dataset. Each of these acoustic perturbation measures are commonly used in the clinical context to characterize pathological voice. For each voice transcript, these features were computed and and fed as input to a one‐dimensional convolutional neural network to predict dementia status. The data was randomly split in the ratio of 1:1 (at the participant‐level), where 50% of the cases were used for training and the remaining for testing. This process was repeated 10 times and the average model performance was reported. For each model run, the neural network was trained for 500 epochs without early stopping and the model obtained after the final epoch was used for testing. Result The mean (±std) age of the participants included in the study was 82.6±8.3 years, and 21 of them were ApoE positive. The mean (±std) area under the receiver operating characteristic curve and the mean (±std) area under the precision‐recall curve of the model classifying participants with dementia from those with normal cognition was 0.673±0.043 and 0.791±0.052, respectively ( Fig. 1 ). Also, the mean accuracy (±std) was 0.615±0.025, the mean precision (±std) was 0.714±0.121, and the mean sensitivity/recall (±std) was 0.721±0.179. Conclusion Our study presents a proof‐of‐principle that deep learning on acoustic perturbation measures can predict dementia status. Further studies with additional participants and an expanded set of acoustic, speech and linguistic features are needed to evaluate the full potential of machine learning to derive disease‐relevant signatures from audio data, and potentially expand the reach of early detection and intervention efforts.