Emotional prosody recognition is impaired in Alzheimer’s disease

One hundred and two participants from the database of the Czech Brain Aging Study, a longitudinal, memory clinic-based study on aging and cognitive impairment [32], were investigated [33]. Of these, 43 participants were aMCI with high (30%) and intermediate (70%) biomarker probability of underlying AD pathology, 36 were diagnosed with dementia due to AD with high (25%) and intermediate (75%) biomarker probability of AD etiology, and 23 were cognitively normal participants. Biomarkers used included cerebrospinal fluid levels of amyloid beta, total tau, and phosphorylated tau proteins. In participants with aMCI, memory impairment was established when the participant scored more than 1.5 standard deviations below the mean of age and education-adjusted norms on any memory test, and activities of daily living were preserved to meet the Petersen et al. 2004 criteria [32, 34]. The aMCI group included both aMCI single-domain and aMCI multiple-domain phenotypes. All participants involved in this study signed written informed consent approved by the Motol University Hospital ethics committee.

Participants were excluded from the study if they reported a history of major neurological or psychiatric disorders, hearing difficulties, depression (≥ 6 points on the 15-item Geriatric Depression Scale) [35], or had significant vascular impairment on brain MRI (Fazekas scale more than 2) [36].

The Emotional Prosody Recognition Test was designed according to the methodology of Ariatti et al. published in 2008 [37]. The experimental battery was prepared in collaboration with four professional actors who produced 200 recordings in total using two Czech sentences with neutral meaning (“The table has four legs” or “Dogs that bark do not bite”). These sentences were spoken by two male and two female performers, native speakers, who were instructed to produce a specific emotional tone of voice (3-s duration) representing five emotions: sadness, fear, happiness, disgust, and anger.

From this large dataset the most appropriate emotions were chosen by 88 healthy volunteers recruited from the clinical staff of Motol University Hospital (mean age 31.6, M to F 1:1) to build the experimental battery. Volunteers invited to validate this test were relatively young because originally this test was prepared for examination of patients with epilepsy who are younger than patients with AD.

The final experimental test included 25 short recordings (spoken by one male and one female performer) each with a 3-s duration. Although the sentences held neutral semantic meaning, these recordings were presented with emotionally charged voices representing happiness, sadness, fear, disgust, and anger; thus, each emotion was represented 5 times. Recordings were presented to subjects on a computer using headphones. The 25 recordings were presented in the same order to each participant. Participants had to select the appropriate emotion from the list of emotions after each of the 25 recordings. There was no time limit to reduce stress during testing. When the participants hesitated, the examiner repeated the instruction to choose one answer from the list and waited until the participant made a choice. The test was scored as correct or incorrect after each recording and the maximum score was 25 points.

Participants’ brain MR scans were performed on a 1.5T system (Siemens, Erlangen, Germany). A T1 weighted, 3-dimensional high resolution magnetization-prepared rapid acquisition with gradient echo (MPRAGE) was acquired with TR/TE/TI = 2000/3.08/1100 ms, flip angle 15, 192 continuous partitions, slice thickness 1.0 mm, and in-plane resolution 1 mm [38]. Participants’ scans were visually inspected to determine sufficient technical quality and to exclude participants with radiologic findings that could interfere with cognitive functioning (i.e., cortical infarctions, tumors, subdural hematomas, hydrocephalus or more extensive white matter hyperintensities equal to Fazekas scale above 2). To measure right- and left-sided amygdala volume and thickness of the temporal pole, superior temporal sulcus, and rostral and caudal parts of the AC (RAC and CAC), we used an automated algorithm from FreeSurfer, version 5.3. (http://​surfer.​nmr.​mgh.​harvard.​edu), described in detail elsewhere [39, 40].

Amygdala volumes were normalized for the differences in head size by regressing the estimated total intracranial volume (eTIV) among participants, as previously described [41, 42]. Temporal pole, superior temporal sulcus, and RAC and CAC thickness were not eTIV adjusted.

A one-way analysis of variance (ANOVA) with post hoc Sidak’s test was used to evaluate differences between the groups in continuous demographic variables. The χ² test was used to evaluate differences in gender proportions. An analysis of covariance (ANCOVA) with post hoc Sidak’s test was used to evaluate differences between the groups in emotional prosody scores. The analysis was controlled for age (mean-centered) and years of education (mean-centered). The receiver operating characteristic (ROC) analysis was used to assess the ability of the Emotional Prosody Recognition Test to differentiate the control group from the aMCI and dementia groups. Sizes of the areas under the ROC curves (AUCs) and optimal sensitivity and specificity based on the Youden’s index were calculated. The Pearson’s correlation coefficients were calculated to explore the bivariate relationships between emotional prosody scores, structural brain measures, and Mini-Mental State Examination (MMSE) scores (a measure of disease severity). Holm-Bonferroni correction for multiple comparisons was used. Next, significant associations were tested using the hierarchical linear regression models adjusted for demographic characteristics, age (mean-centered), and years of education (mean-centered). A two-tailed p value < 0.05 was considered statistically significant. Analyses were performed using R statistical language environment [43] and IBM SPSS 25.0 software.

Group demographic and clinical characteristics are reported in Table 1. There was no difference in sex distribution among the groups. There was a significant group effect for age indicating that the control group was younger than the aMCI and dementia groups (both p < 0.001). There were no differences between the aMCI and dementia groups (p = 1.000).

The dementia group had significantly fewer years of education compared to the control group (p < 0.001) and the aMCI group (p = 0.012). The difference in education between the control group and the aMCI group was not significant (p = 0.362). MMSE score was significantly lower in the aMCI and dementia groups compared to the control group (both p < 0.001). The dementia group had lower MMSE scores than the aMCI group (p < 0.001).

The total score for EPR is shown in Fig. 1. There was a significant group effect in total EPR score (p = 0.002). The post hoc analysis controlling for age and education confirmed that EPR scores were lower in the dementia group compared to the control group (p = 0.002). The dementia group also had lower total EPR score compared to the aMCI group (p = 0.033) (Fig. 1).

In the ROC analyses, the total score for EPR differentiated the control group from the patients’ groups (pooled aMCI and dementia groups) with AUC values of 0.85 (p < 0.001), sensitivity of 87.0% and specificity of 72.2%, the control group from the aMCI group with AUC values of 0.80 (p < 0.001), sensitivity of 73.9% and specificity of 74.4%, the control group from the dementia group with AUC values of 0.91 (p < 0.001), sensitivity of 87.0% and specificity of 83.3%, and the aMCI group from the dementia group with AUC values of 0.67 (p = 0.009), sensitivity of 62.8% and specificity of 61.1%.

EPR total score significantly correlated with disease severity according to MMSE score in the entire sample (r = 0.471, p < 0.001). The association remained significant in the regression analysis adjusted for age and education, where lower EPR total scores were associated with lower MMSE scores (ß = 0.35, p < .001, 95% CI [0.07, 0.27]).

EPR total score significantly correlated with left and right TP thickness (both r ≥ 0.35, p < .001), left and right STS thickness (both r ≥ 0.52, p < .001), and left and right RAC thickness (both r ≥ 0.35, p < .001). The correlation between ERP total score and left and right amygdala volumes (both r ≥ 0.23, p ≤ .026) did not survive the correction for multiple comparisons.

The associations with right-sided structural brain measures of TP and STS and bilateral structural brain measures of RAC remained significant in the regression analysis adjusted for age and education, where lower EPR total scores were associated with smaller right TP (ß = 0.33, p = .002, 95% CI [0.41, 1.71]), right STS (ß = 0.37, p < .001, 95% CI [1.32, 4.29]) and right (ß = 0.31, p < .001, 95% CI [1.85, 5.70]) and left (ß = 0.29, p = .001, 95% CI [1.39, 5.09]) RAC thickness. The associations between EPR total scores and left-sided structural brain measures of TP and STS were not significant in the regression analysis (all ß ≤ 0.14, p ≥ .199, 95% CI [− 0.39, 0.94] and [− 0.58, 2.76]).

One possible limitation of our study is the significant difference in age and education between controls and individuals with aMCI and dementia. By adjusting our analyses for age and education we tried to address this issue. Additionally, we focused only on specific brain regions in volumetry analysis. A notable strength of our data set is the homogeneity of the study groups, provided by strict CBAS criteria including AD biomarkers. In order to quantify the predictive power of EPR for the identification of individuals at risk of AD, larger group studies would need to be performed and machine learning approaches utilized. Other limitations could include the sample size and lack of controls for other covariates, which may have impacted the EPR and volume associations (e.g., depression).