Harnessing acoustic speech parameters to decipher amyloid status in individuals with mild cognitive impairment

Abstract

Alzheimer's disease (AD) is a neurodegenerative condition characterized by a gradual decline in cognitive functions. Currently, there are no effective treatments for AD, underscoring the importance of identifying individuals in the preclinical stages of mild cognitive impairment (MCI) to enable early interventions. Among the neuropathological events associated with the onset of the disease is the accumulation of amyloid protein in the brain, which correlates with decreased levels of Aβ42 peptide in the cerebrospinal fluid (CSF). Consequently, the development of non-invasive, low-cost, and easy-to-administer proxies for detecting Aβ42 positivity in CSF becomes particularly valuable. A promising approach to achieve this is spontaneous speech analysis, which combined with machine learning (ML) techniques, has proven highly useful in AD. In this study, we examined the relationship between amyloid status in CSF and acoustic features derived from the description of the Cookie Theft picture in MCI patients from a memory clinic. The cohort consisted of fifty-two patients with MCI (mean age 73 years, 65% female, and 57% positive amyloid status). Eighty-eight acoustic parameters were extracted from voice recordings using the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS), and several ML models were used to classify the amyloid status. Furthermore, interpretability techniques were employed to examine the influence of input variables on the determination of amyloid-positive status. The best model, based on acoustic variables, achieved an accuracy of 75% with an area under the curve (AUC) of 0.79 in the prediction of amyloid status evaluated by bootstrapping and Leave-One-Out Cross Validation (LOOCV), outperforming conventional neuropsychological tests (AUC = 0.66). Our results showed that the automated analysis of voice recordings derived from spontaneous speech tests offers valuable insights into AD biomarkers during the preclinical stages. These findings introduce novel possibilities for the use of digital biomarkers to identify subjects at high risk of developing AD.

Keywords: Alzheimer's disease; automated pattern recognition; biomarkers; cerebrospinal fluid; early diagnosis; machine learning; mild cognitive impairment; speech acoustics.

Figure 1

Pipeline used to evaluate the goodness-of-fit of all the models used to predict amyloid status. Performance metrics and confidence intervals were calculated from the metric distribution obtained after 5,000 iterations of a nested leave-one-out cross validation (LOOCV) where the training set used for adjusting the models was generated by bootstrapping.

Figure 2

Receiver operating characteristic (ROC) curve for predicting amyloid status based on demographic and neuropsychological, acoustic, and a combination of demographic and acoustic variables. The results correspond to the best models presented in Table 3. For each model, the mean AUC calculated by 5,000 bootstrap iterations as described in Figure 1 is shown. AUC: area under the curve.

Figure 3

Uniform manifold approximation and projection (UMAP) (Sainburg et al., 2020) dimensionality reduction of the most discriminative feature set obtained by the VLPSO feature selection algorithm. (A) Projection highlighting positive and negative amyloid status. (B) Projection of the K-nearest neighbor decision boundary. On the left is the density plot representing the higher presence (red) or absence (blue) of amyloid-positive cases in the data. On the right are shown the KNN predictions of amyloid-positivity, where red indicates that the model assigns a higher probability of amyloid-positivity and blue a lower probability. UMAP hyperparameters: number of neighbors = 8 and minimum distance = 0.1; the rest of the hyperparameters were left as default.

Figure 4

SHapley Additive exPlanations (SHAP) values Lundberg and Lee (2017) of the best model for predicting amyloid status. The SHAP values were calculated on the test set using LOOCV. The best model corresponds to the combination of VLPSO with KNN using acoustic variables (see Table 3). The feature color indicates how it relates to the probability of being amyloid positive or negative. The red color is associated with higher feature values, while blue is associated with lower values. Thus, lower values in the F3-bandwidth (voiced) CoV (blue area) are associated with an increased likelihood of being amyloid positive; and having a lower mean of the spectral Hammarberg index (voiced) (red area) is associated with a lower probability of being amyloid positive. CoV: coefficient of variation; AMean: arithmetic mean; Std: standard deviation.