Vocal expression of emotions discriminates dementia with Lewy bodies from Alzheimer's disease

Abstract

Dementia with Lewy bodies (DLB) and Alzheimer's disease (AD), the two most common neurodegenerative dementias, both exhibit altered emotional processing. However, how vocal emotional expressions alter in and differ between DLB and AD remains uninvestigated. We collected voice data during story reading from 152 older adults comprising DLB, AD, and cognitively unimpaired (CU) groups and compared their emotional prosody in terms of valence and arousal dimensions. Compared with matched AD and CU participants, DLB patients showed reduced overall emotional expressiveness, as well as lower valence (more negative) and lower arousal (calmer), the extent of which was associated with cognitive impairment and insular atrophy. Classification models using vocal features discriminated DLB from AD and CU with an AUC of 0.83 and 0.78, respectively. Our findings may aid in discriminating DLB patients from AD and CU individuals, serving as a surrogate marker for clinical and neuropathological changes in DLB.

Highlights

• DLB showed distinctive reduction in vocal expression of emotions.

• Cognitive impairment was associated with reduced vocal emotional expression in DLB.

• Insular atrophy was associated with reduced vocal emotional expression in DLB.

• Emotional expression measures successfully differentiated DLB from AD or controls.

1. INTRODUCTION

Dementia with Lewy bodies (DLB) is the second most prevalent neurodegenerative dementia after Alzheimer's disease (AD), estimated to affect up to 22.8% of people with dementia. ¹ DLB is characterized by four core clinical features including fluctuating cognition and parkinsonism. ² Compared with AD, it progresses faster and has a wider range of symptoms, resulting in lower quality of life and shorter life expectancy for patients, as well as an increased burden on caregivers. ³ , ⁴ However, despite its importance, DLB is often misdiagnosed with other conditions such as AD. ⁵ A study reported that 78% of DLB patients received an initial diagnosis other than DLB, ⁴ leading to delays in receiving appropriate care that includes community services and pharmacologic therapy, with different needs from AD care. ⁴ One reason hindering proper diagnosis is the limited awareness and knowledge of DLB among primary care physicians, thus failing to make appropriate referrals to specialists. ⁴ , ⁵ , ⁶ Additionally, the limited availability of established biomarkers for DLB, ⁷ coupled with invasive and expensive standard biomarkers for AD, ⁸ further hinders proper identification. Therefore, a screening tool capable of suggesting the possibility of DLB, even in nonspecialist settings, would facilitate timely referrals to specialists and enable early differential diagnosis of DLB.

Alterations in emotional processing, particularly in recognition of emotions, have been reported for several types of dementia. ⁹ , ¹⁰ In particular, it is commonly found in patients with behavioral and semantic variants of frontotemporal dementia. ¹⁰ , ¹¹ For AD patients, although it is less prominent, studies have reported their difficulty in recognizing emotional expressions across different types of emotions, where it was most pronounced in recognizing “fear” and the least for “surprised.” ¹² For DLB, a few studies have suggested deficits in recognizing emotional expressions depending on the type of emotion, for example, diminished recognition of “surprised” faces ¹³ note that investigations into emotional processing alterations in DLB are still limited. ¹⁴ Moreover, studies on the neural correlates of impaired emotional recognition have also suggested disease‐specific patterns. ¹⁰ For example, in AD, deficits in emotion recognition were associated with brain regions such as the anterior cingulate, ¹⁵ , ¹⁶ while in DLB, they were associated with regions including the middle frontal gyrus. ¹³ In summary, previous studies suggest the existence of distinct profiles of emotional processing across different types of dementia, both in terms of perception and neural correlates. Therefore, although emotional processing alterations may not be a representative feature of DLB, they may serve as a surrogate marker for the differentiation of DLB from other types of dementia exemplified by AD.

Contrary to emotional recognition, emotional expression in dementia patients, which is also crucial for social interactions, ⁹ , ¹⁴ , ¹⁷ affecting individuals' mental health and quality of life, ⁹ remains largely uninvestigated. To the best of our knowledge, no study has explored alterations in emotional expressions in DLB or compared with those in AD. In AD, a few studies reported reduced expression of emotions, ¹⁸ , ¹⁹ and this reduction correlated with cognitive impairment across multiple domains, including memory and attention. ¹⁸ While no study exists for DLB, reports on emotional expression alterations in Parkinson's disease (PD), another form of the Lewy body spectrum disorder, are abundant. ⁹ Alterations in vocal and facial expressions of emotions in PD are likely attributed to parkinsonian motor impairments, specifically hypokinetic dysarthria and facial bradykinesia, ⁹ although several studies have also suggested associations with executive dysfunction and alterations in the affective process. ⁹ , ²⁰ Taken together, emotional expression may be affected by multiple cognitive and motor functions. DLB patients exhibit a combination of multiple cognitive and parkinsonian motor impairments. Therefore, DLB patients may experience more pronounced alterations in emotional expression than AD patients.

Research in Context

1. Systematic Review: The authors reviewed the literature using medical and academic databases (e.g., PubMed, Google Scholar) and cited work. Alterations in vocal expression of emotions have been investigated for various neurological conditions including Alzheimer's disease (AD). However, its alterations in dementia with Lewy bodies (DLB), or differences between the two dementia types, remain uninvestigated.

2. Interpretation: Our results provide initial evidence of (a) DLB‐specific reduction in vocal expression of emotions during story reading and its associations with cognitive impairment and insular atrophy, and (b) the feasibility of machine‐learning models based on the characteristics of vocal expression of emotions to discriminate DLB from AD or controls as an easy‐to‐perform screening tool.

3. Future Directions: Our findings need cross‐cultural and cross‐linguistic validation with larger multicenter samples. Future studies may also address the applicability of our findings to emotional expression in everyday conversational speech.

Among the modalities of emotional expression, vocal expression, or emotional prosody, is particularly promising for applications in clinical practice, given readily available continuous scales in terms of high‐level emotional dimensions, ²¹ facilitating the capture of continuous changes in patients. In emotion theories, the commonly utilized two‐dimensional valence‐arousal model represents the negative‐to‐positive and calm‐to‐excited aspects of emotional expressions, respectively. ²² This model is increasingly in the assessment of patients with AD, depression, and other conditions, ²³ , ²⁴ owing to emerging digital technologies that enable automated measurement of valence and arousal scales, even in nonspecialist settings. ²¹ Therefore, by using the valence‐arousal model to analyze voice data obtained from older adults comprising DLB and AD patients as well as cognitively unimpaired (CU) controls, the present study aimed to examine the following hypothesis: DLB patients exhibit more pronounced alterations in emotional prosody across high‐level emotional dimensions, thereby making objective measurement useful for discriminating DLB from other conditions.

2. METHODS

2.1. Study participants and assessments

We recruited community‐dwelling older adults in Ibaraki, Japan, including outpatients from the Department of Psychiatry at the University of Tsukuba Hospital. All participants underwent cognitive and clinical assessments, as well as structural MRI to assess structural brain changes (Table 1 and Supplementary Method 1 for details). The patients met the standard research diagnostic criteria for prodromal/clinical AD ²⁵ , ²⁶ or DLB. ² , ²⁷ (Supplementary Method 2 for details). CU participants did not meet any of these criteria and were matched to the patients for age, sex, and years of education. Regarding the disease stage, patients in the AD and DLB groups ranged from mild cognitive impairment (MCI; hereafter MCI‐AD and MCI‐LB, respectively) to moderate dementia. ²⁶ , ²⁷ , ²⁸ All participants were native Japanese speakers and had no serious hearing, vision, and speech impairments that would interfere with the voice data collection.

TABLE 1.

Participant demographics and cognitive/clinical measures

	CU (n = 49)		AD (n = 76)		DLB (n = 27)		p‐Value
Disease stage, MCI	n/a		42	(55.3%)	19	(70.4%)	0.253
Age, yr	72.4	(4.1)	74.2	(6.0)	75.1	(5.0)	0.075
Sex, female	28	(57.1%)	33	(43.4%)	12	(44.4%)	0.299
Education, yr	13.3	(2.1)	13.2	(2.7)	12.7	(2.8)	0.624
Antipsychotic medication d	0	(0.0%)	3	(3.9%)	6	(22.2%)	<0.001	a,b
Mini‐Mental State Examination (MMSE)	27.8	(1.9)	23.4	(4.5)	26.5	(3.6)	<0.001	a,c
Logical Memory–immediate recall	11.2	(3.1)	4.5	(3.4)	7.9	(3.9)	<0.001	a,b,c
Logical Memory–delayed recall	9.5	(3.0)	2.1	(2.6)	5.9	(4.1)	<0.001	a,b,c
Frontal Assessment Battery	13.5	(2.5)	11.0	(3.7)	11.1	(4.1)	<0.001	b,c
Trail Making Test part A e	35.8	(11.3)	58.4	(47.2)	68.5	(54.0)	0.001	b,c
Trail Making Test part B f	89.7	(41.3)	177.3	(100.8)	180.8	(87.9)	<0.001	b,c
Clock Drawing Test g	6.8	(0.8)	5.9	(1.9)	6.6	(1.1)	0.007	c
Clinical Dementia Rating g	0.0	(0.0)	0.7	(0.3)	0.6	(0.4)	<0.001	b,c
Clinical Dementia Rating–Sum of Boxes g	0.0	(0.1)	2.7	(2.4)	2.6	(2.7)	<0.001	b,c
Geriatric Depression Scale (GDS) e	3.3	(2.9)	3.3	(3.0)	4.1	(3.9)	0.495
5‐Item Unified Parkinson's Disease Rating Scale (UPDRS)	n/a		n/a		2.1	(2.1)

Notes: Values are displayed as mean (SD) and n (%) for continuous and binary measures, respectively. Bold values highlight statistically significant differences examined by one‐way analysis of variance for continuous measurements and by chi‐squared test for binary measurements.

Abbreviations: AD, Alzheimer's disease; CU, cognitively unimpaired; DLB, dementia with Lewy bodies; MCI, mild cognitive impairment.

Significant differences between individual diagnostic groups (Tukey–Kramer or chi‐squared test, p < 0.05) are marked with a, b, or c (a: significant for DLB vs. AD; b: significant for DLB vs. CU; c: significant for AD vs. CU).

^d Data missing for six participants.

^e Data missing for two participants.

^f Data missing for 10 participants.

^g Data missing for one participant.

The study was conducted with the approval of the Ethics Committee, University of Tsukuba Hospital (H29‐065), and it followed the ethical code for research with humans as stated in the Declaration of Helsinki. All participants provided written informed consent.

2.2. Voice data collection and analysis

The participants were instructed to read aloud part of the picture book “Hanasaka Jiisan,” a well‐known Japanese folktale. This story includes a multitude of positive and negative words expected to elicit emotional responses (Supplementary Method 3 for details). Participants sat in front of an iPad Air 2 tablet, and voice data were recorded using the tablet's internal microphone (44.1 kHz, 16 bits). The experiment was conducted in the same room for all participants.

To compare emotional prosody across the diagnostic groups, voice data were characterized by emotional dimensions of valence and arousal (Figure 1A). To predict valence/arousal values from audio data without the use of linguistic information, we used the transformer‐based speech emotion recognition model developed by Wagner et al. ²¹ (Supplementary Method 4 for details). The model was applied to voice segments with a 3‐second window and 1.5‐second stride, resulting in a series of valence and arousal values. We then used the mean values of valence and arousal, as well as emotional expressiveness (i.e., variability in the valence‐arousal space) as affective features. The variability in the valence‐arousal space was measured using Shannon entropy, where a higher entropy indicates a larger variation in vocal emotional states (Supplementary Method 5 for details). Additionally, to further investigate alternation in low‐level acoustic characteristics, we characterized the voice data using seven acoustic features shown to be a robust representation of vocal prosody across the most common languages and countries. These features comprise voice quality, loudness, pitch and formants, rhythm and tempo, shimmer, pitch variation, and mel‐frequency cepstral coefficient 3 ²⁹ (Supplementary Method 6 for details).

FIGURE 1.

Overview of affective feature extraction (A) and analysis results (B–D). (A) (i) Voice data were split into 3‐second segments with a 1.5‐second stride (1.5‐second overlap); (ii) valence and arousal values were obtained for each segment through a deep neural network model developed by Wagner et al.²¹; and (iii) a series of valence/arousal values were used to calculate vocal affective features. (B) Group differences in vocal affective features. Horizontal bars indicate significant differences (Tukey–Kramer test, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001). Boxes indicate the 25th (Q1) and 75th (Q3) percentiles; whiskers indicate the upper and lower adjacent values that are most extreme within Q3+1.5(Q3‐Q1) and Q1‐1.5(Q3‐Q1), respectively; the line and notch in each box represent the median and its 95% confidence interval, respectively; the diamond represents the mean; and dots denote outliers. (C and D) Receiver operating characteristics (ROC) curves for DLB versus AD and DLB versus CU. Shades represent 95% confidence intervals. (E and F) Emotional expressiveness associated with MMSE and insular thickness in the DLB group. Lines and shades represent the fixed effect of MMSE or insular thickness and its 95% confidence intervals; and each data point corresponds to the utterance of one sentence by one participant. DLB, dementia with Lewy bodies; AD, Alzheimer's disease; CU, cognitively unimpaired; MMSE, Mini‐Mental State Examination

Next, to evaluate the feasibility of using voice data to differentiate DLB from AD and CU, we used supervised machine‐learning models for binary classification (Supplementary Method 7 and Supplementary Table 1 for details of the model algorithms and hyperparameters, respectively). The input variables for the model using vocal data were the aforementioned three affective and seven acoustic features (total 10 variables). The model performance was evaluated by the area under the receiver operating characteristic curve (AUC) obtained from 10 iterations of a leave‐two‐subjects‐out cross‐validation procedure. ³⁰ We also compared performance against a baseline model using clinico‐demographic variables (i.e., Mini‐Mental State Examination [MMSE] score, age, sex, and years of education).

2.3. Statistical analysis

Group differences in affective and acoustic features were examined using one‐way analysis of covariance (ANCOVA) tests adjusted for age, sex, years of education, and the use of antipsychotic medication as covariates, with post hoc Tukey–Kramer pairwise comparisons. We also elaborated on the results in terms of the difference in the linguistic content of reading (e.g., emotional vs. neutral, conversational vs. ground). Then, we conducted exploratory analyses on clinical and neural correlates of alterations in these features observed in the DLB group, using linear mixed‐effect regression models with the same covariates (i.e., age, sex, years of education, and the use of antipsychotic medication). First, we examined associations between the affective features and clinical measures representing cognitive, motor, and mental aspects of the participants. For the clinical measures, we used MMSE ³¹ for the cognitive aspect, the five‐item Unified Parkinson's Disease Rating Scale (UPDRS) ³² for the motor aspect, and the Geriatric Depression Scale (GDS) ³³ for the mental aspect. Second, we examined associations between the affective features and structural brain changes. To exclude brain regions irrelevant to structural changes in DLB, we only included the cortical and subcortical regions of interest (ROIs) showing statistical atrophy in the DLB group relative to CU. We carried out the mixed‐model analyses that modeled sentences as well as participants as random intercepts to address differences in linguistic and affective content per sentence. In all statistical analyses, the alpha value was set to 0.05, and Benjamini–Hochberg correction was applied for multiple testing.

3. RESULTS

We first investigated how vocal affective features differ across the DLB, AD, and CU groups. Consequently, compared with both AD and CU participants, DLB participants showed lower emotional expressiveness, even after correcting for multiple testing (one‐way ANCOVA, corrected p = 0.008; Figure 1B). In addition, the utterances of DLB participants were characterized by lower valence and arousal values, indicating more negative and calmer states (valence: corrected p = 0.008, arousal: corrected p < 0.001). These tendencies persisted when the DLB and AD groups were subdivided in accordance with disease stage (i.e., MCI or dementia), with dementia stage‐wise reductions in each patient group (Supplementary Figure 1). To further investigate the sex differences, we performed two‐way ANOVAs (diagnostic group × sex), showing that these affective features had significant interaction effects between diagnostic group and sex (p = 0.004 to 0.039), and the effect sizes across the diagnostic groups were larger in males (male: η ² = 0.150 to 0.243, female: η ² = 0.001 to 0.008). Regarding the low‐level acoustic features, the voice quality was lower in the DLB group than in the AD group, but this difference did not remain significant after correcting for multiple testing (uncorrected p = 0.012, corrected p = 0.083; see Supplementary Figure 2 for the full results). We compared phoneme rate as a measure indicating reading performance and found no statistical difference between the diagnostic groups (one‐way ANOVA, p = 0.257), suggesting that altered vocal expression of emotions could not be explained by the difference in reading performance.

To deepen our understanding of altered vocal expression of emotions in the DLB group, we further investigated the association of group differences in vocal affective features with the nature of linguistic content. Specifically, we investigated whether the following two factors were important to elicit emotional responses with discernable differences between the DLB and other groups: (i) positive and negative sentences and (ii) conversational sentences. When we compared the degree of group differences in affective features between conditions with and without positive/negative sentences, the results showed that the discriminative power of emotional expressiveness for differentiating DLB from AD or CU was reduced for utterances without positive/negative sentences (Δη ²: −0.049, ΔAUC for vs. AD: −0.017, ΔAUC for vs. CU: −0.053). Similarly, the discriminative power was reduced in ground sentences compared with conversational sentences (Δη ²: −0.012, ΔAUC for vs. AD: −0.094, ΔAUC for vs. CU: −0.008). Taken together, these results suggest that the reduction in vocal expressivity in the DLB group could be evident, particularly in emotion‐eliciting conversations such as those containing positive or negative words.

Next, we examined whether vocal affective/acoustic features could discriminate participants in the DLB group from those in the AD and CU groups. Through the cross‐validation procedure, the baseline models using clinico‐demographic variables (i.e., MMSE score, age, sex, and years of education) classified DLB and AD with an AUC of 0.75 (95% confidence interval [CI]: 0.75–0.76) and classified DLB and CU with an AUC of 0.65 (95% CI: 0.65–0.66). On the other hand, the models using vocal features could improve the AUC by +0.08 for DLB versus AD (AUC of 0.83; 95% CI: 0.83–0.83; Figure 1C) and +0.13 for DLB versus CU (AUC of 0.78; 95% CI: 0.77–0.78; Figure 1D). Furthermore, the performance was mostly preserved even for discriminating patients in the MCI stage: an AUC of 0.86 (95% CI: 0.86–0.87) for MCI‐LB versus MCI‐AD and 0.75 (95% CI: 0.75–0.75) for MCI‐LB versus CU.

The exploratory analyses to examine whether the vocal affective features could capture clinical and neuropathological changes in the DLB group revealed the following associations. For the clinical measures, the mixed‐model analysis showed significant positive associations of the MMSE score with emotional expressiveness (corrected p = 0.001; Figure 1E). No significant associations were found for the UPDRS or GDS scores (Supplementary Table 2 for the full results). The inclusion of the MMSE score increased the marginal R ² by up to +0.16 from a model not including MMSE. For the neuropathological measures, the following ROIs exhibited significant atrophies in the DLB group compared with CU: fusiform, insula, and accumbens area (Supplementary Table 3 for details). Among them, only the insular thickness showed significant positive associations with emotional expressiveness (corrected p = 0.021; Figure 1F) and valence (corrected p = 0.013), while there were no significant associations for the other ROIs (Supplementary Table 4 for the full results). The inclusion of the insular thickness increased the marginal R ² by up to +0.10. Although these associations should be considered exploratory due to the limited sample size, these results may suggest that alterations in vocal affective features could reflect certain aspects of the disease progression of DLB.

4. DISCUSSION

We investigated characteristics of vocal expression of emotions in story‐reading voices collected from older adults comprising DLB and AD patients as well as CU controls. As a result, three main findings emerged. First, the DLB group had different characteristics of vocal expression in comparison to AD and CU; specifically, the degree of emotional expressiveness based on high‐level emotional dimensions was statistically lower in the DLB group than in the other two groups. Second, in discriminating DLB from both AD and CU, machine learning‐based classification models using quantitative measures of emotional prosody outperformed baseline models using clinico‐demographic measures. Furthermore, exploratory mixed‐model analysis showed that reduced emotional expressiveness in the DLB group was associated with a lower cognitive score, as measured by MMSE, and with reduced insular thickness. These results suggest the potential of quantitative measures of emotional prosody as a digital marker for dementia, as well as its usefulness as a screening tool for DLB, a frequently overlooked form of dementia with common symptoms with AD. ⁴ , ⁵

Our statistical analysis found a DLB‐specific reduction in vocal expression of emotions. This alteration in the DLB group aligns with the clinical findings that DLB patients are frequently disrupted in vocal/facial expressivity. ⁹ , ³⁴ Our results elaborated on the finding with quantitative measures by showing multifaceted changes in the characteristics of vocal emotional expression in terms of both valence (positive to negative) and arousal (excited to calm) scales and their overall variability. In addition, with respect to speech during neuropsychological tasks, DLB patients' utterances have been reported to show distinct alterations compared with AD in terms of both linguistic and acoustic aspects. ³⁵ Our results extended the findings by showing that DLB patients also exhibited specific alterations in emotional prosody with respect to speech during non‐neuropsychological tasks, which would involve voices in a context closer to everyday social interactions. On the other hand, the AD group exhibited preserved emotional expressiveness in our analysis. This result, contradictory to several previous studies reporting a reduction in vocal expressivity in AD, ¹⁸ , ¹⁹ might be explained by differences in the task for collecting voices. For example, emotional expressiveness in directed production or imitation of emotional prosody was impaired in AD, ¹⁹ , ³⁶ , ³⁷ while it was preserved in spontaneous speech. ¹⁹ , ³⁷ , ³⁸ Also, Han and colleagues found that emotional expressiveness in an interview speech for recent events was reduced in AD, while it was relatively preserved for events in their childhood. ¹⁸ Given these mixed results regarding AD, future research would further investigate whether the DLB‐specific alteration in emotional expressiveness found in this study occurs consistently across different types of speech, thereby addressing the possibility of DLB screening using everyday speech, as well as the impact of reduced emotional expressivity on everyday social interactions.

In the present study, a classification model based on the characteristics of emotional prosody during an approximately 2‐minute story reading could discriminate DLB from AD with reasonable performance (AUC of 0.83), outperforming the baseline model using clinico‐demographic measures in our dataset. This performance is competitive with another speech analysis model using five speech‐based neuropsychological tasks (AUC of 0.833) ³⁵ and even with fluid biomarkers (AUC of 0.723–0.843). ³⁹ Thus, our findings have the potential to aid DLB screening in clinical practice by providing a less time‐consuming, less invasive tool. In addition to its use in clinical practice, speech analysis is also of interest as a means of automated screening and longitudinal monitoring in large‐scale cohort studies and clinical trials, ⁴⁰ which would benefit from our findings. For example, the AD Neuroimaging Initiative study uses speech analysis on story recall tasks for the initial screening of likely AD, MCI, or CU individuals from a large participant pool (N∼20000). ⁴¹ , ⁴² Adding a short story‐reading task may help conduct similar large‐scale studies for DLB or enable earlier exclusion of likely DLB cases from AD studies. Furthermore, our results showed that emotional expressiveness from conversational sentences better performed in discriminating DLB than using ground sentences, suggesting the promising nature of our approach to be applied to everyday conversations beyond controlled task settings.

This study showed that the reduction of emotional expressiveness in DLB was associated with the measure of cognitive impairment as well as with insular atrophy. The insula is typically affected in DLB ⁴³ , ⁴⁴ , ⁴⁵ and has been associated with deficits in emotion recognition in DLB. ¹³ , ⁴⁴ Our results extend these prior findings, suggesting that the insula can be a key component of the affective process in DLB, related to both recognition and expression aspects of emotional processing. Furthermore, our results, which revealed a DLB‐specific alteration compared with AD, suggest that the reduction in emotional expressiveness in DLB may arise through a different mechanism than in AD. Also, this alteration in DLB was not associated with motor impairment, suggesting that the reduction in emotional expressiveness in DLB may be more attributable to the affective process than to parkinsonism, as indicated by the work of Möbes et al. on PD patients without dysarthria. ²⁰ Regarding the mechanism, the process of vocal expression of emotions has been explained in terms of affective and sensorimotor components, ⁴⁶ where the former refers to access to abstract representations of emotional prosody and the latter to motor planning and implementation. According to a functional neuroimaging study, the basal ganglia were activated in both affective and sensorimotor components, while the insula was only involved in the former, and the hippocampus in the latter. ⁴⁶ These cortical and subcortical regions involve those with typically observed neuropathological changes in DLB, AD, and PD, that is, insula in DLB, ⁴³ , ⁴⁴ , ⁴⁵ hippocampus in AD, ⁴⁷ and basal ganglia in PD. ⁴⁸ Quantitative measures of multifaceted characteristics of emotional expression, therefore, may provide promising surrogate markers for disease‐specific patterns of neuropathological changes.

This study has several limitations. First, the small sample size resulted in the patient groups comprising both MCI and dementia patients. Although the majority of findings were preserved in the analysis restricting patients to MCI, the validity of our findings, particularly in the prodromal stage, needs to be confirmed with a larger sample. Second, our findings were derived from a single local community and the Japanese language. Hence, as culture and language affect emotional prosody, ¹⁷ , ⁴⁹ , ⁵⁰ cross‐cultural and cross‐linguistic validation is warranted. Third, the dataset was imbalanced in terms of diagnostic groups, which might also affect the generalizability of our findings. Finally, the diagnoses in our dataset were primarily based on clinical features, and thus the mixed DLB/AD pathology could not be evaluated, requiring further research with validated neuropathological biomarkers.

In conclusion, this study is the first to show a DLB‐specific alteration in the characteristics of vocal expression of emotions and its clinical and neural correlates. Our analyses revealed that (i) emotional expressiveness during story reading was specifically reduced in DLB, while this reduction was not observed in AD; (ii) models using these vocal affective features could discriminate DLB from AD or controls, with comparable or superior performance to neuropsychological testing; and (iii) the reduction in emotional expressiveness in DLB was associated with cognitive impairment and insular atrophy. These results suggest that emotion‐eliciting utterances can be useful not only for screening for DLB but also as a surrogate marker for disease progress and neuropathological changes in DLB.

CONFLICT OF INTEREST STATEMENT

M.K., Y.Y., and K.S. are employed by IBM Corporation. M.N. received funding from the Japan Society for the Promotion of Science, KAKENHI [grant number 19H01084]. M.O. has nothing to disclose. K.N. received funding from the Japan Society for the Promotion of Science, KAKENHI [grant numbers 19H01084 and 21K12153]. K.N. reports payment for lectures from Eli Lilly and Otsuka. T.A. received funding from the Japan Society for the Promotion of Science, KAKENHI [grant number 19H01084]. T.A. reports payment for lectures from Eisai and Sumitomo Pharma. Author disclosures are available in the supporting information.

CONSENT STATEMENT

The study was conducted with the approval of the Ethics Committee, University of Tsukuba Hospital (H29‐065), and it followed the ethical code for research with humans as stated in the Declaration of Helsinki. All participants provided written informed consent.

Supporting information

Supporting Information

Supporting Information

Kobayashi M, Yamada Y, Shinkawa K, et al. Vocal expression of emotions discriminates dementia with Lewy bodies from Alzheimer's disease. Alzheimer's Dement. 2024;16:e12594. 10.1002/dad2.12594