Natural speech markers of Alzheimer’s disease co-pathology in Lewy body dementias

Abstract

Introduction:

An estimated 50% of patients with Lewy body dementias (LBD), including Parkinson’s disease dementia (PDD) and Dementia with Lewy bodies (DLB), have co-occurring Alzheimer’s disease (AD) that is associated with worse prognosis. This study tests an automated analysis of natural speech as an inexpensive, non-invasive screening tool for AD co-pathology in biologically-confirmed cohorts of LBD patients with AD co-pathology (SYN + AD) and without (SYN-AD).

Methods:

We analyzed lexical-semantic and acoustic features of picture descriptions using automated methods in 22 SYN + AD and 38 SYN-AD patients stratified using AD CSF biomarkers or autopsy diagnosis. Speech markers of AD co-pathology were identified using best subset regression, and their diagnostic discrimination was tested using receiver operating characteristic. ANCOVAs compared measures between groups covarying for demographic differences and cognitive disease severity. We tested relations with CSF tau levels, and compared speech measures between PDD and DLB clinical disorders in the same cohort.

Results:

Age of acquisition of nouns (p = 0.034, |d| = 0.77) and lexical density (p = 0.0064, |d| = 0.72) were reduced in SYN + AD, and together showed excellent discrimination for SYN + AD vs. SYN-AD (95% sensitivity, 66% specificity; AUC = 0.82). Lower lexical density was related to higher CSF t-Tau levels (R = −0.41, p = 0.0021). Clinically-diagnosed PDD vs. DLB did not differ on any speech features.

Conclusion:

AD co-pathology may result in a deviant natural speech profile in LBD characterized by specific lexical-semantic impairments, not detectable by clinical disorder diagnosis. Our study demonstrates the potential of automated digital speech analytics as a screening tool for underlying AD co-pathology in LBD.

1. Introduction

Lewy body dementias (LBD), which encompasses Parkinson’s disease (PD) with dementia (PDD) [1], and Dementia with Lewy bodies (DLB) [2], are a clinically and pathologically heterogeneous group of disorders characterized by intracellular alpha-synuclein (SYN) pathology [3]. Co-existing Alzheimer disease (AD) pathology is common in LBD where up to 50% of cases have sufficient amounts of amyloid and tau pathology at autopsy to receive a secondary diagnosis of AD [4]. AD co-pathology is strongly associated with worse prognosis in LBD, characterized by an earlier onset and more severe dementia, and a shorter survival time [5–7]. AD co-pathology also explains much of the clinical heterogeneity in dementia profiles in LBD [6]. Thus, in-vivo detection of AD co-pathology is necessary for predicting prognosis and in clinical care decision-making. Patient stratification based on pathology is also critical for protein-targeting drug trials, and ultimately for the development of more targeted and effective therapies [8].

Some studies suggest a link between concomitant AD in LBD and the DLB clinical disorder [9], distinguished from PDD based on an arbitrary “1 year rule” between the relative onset of cognitive and motor disorders [2]. However, large-scale harmonized autopsy studies studying the full LBD spectrum report shared underlying biology and overlapping symptomology in PDD and DLB [10,11]; the clinically-diagnosed distinction between DLB and PDD has not proven useful for identifying patients with AD co-pathology [11]. Currently, AD co-pathology can be detected in vivo using costly or invasive tests, such as molecular PET imaging and CSF AD biomarkers. Blood p-Tau is also used as a biomarker of AD, but performance in LBD leaves room for improvement [12]. There remains a need for non-invasive and inexpensive clinical markers that can serve as a screener for concomitant AD in LBD.

In this study, we evaluate the performance of an automatic natural speech analysis tool during picture descriptions in its ability to screen for AD co-pathology in biologically-confirmed cohorts of SYN + AD and SYN-AD, stratified by CSF AD biomarkers or neuropathological diagnosis. A comprehensive study of natural speech is warranted considering recent clinical studies that report impaired language function, specifically impaired confrontation naming and semantic fluency in SYN + AD compared to SYN-AD [13–15]. Based on these studies, we hypothesized differences in lexical-semantic properties during picture descriptions, particularly of nouns which may reflect deficits in naming, and acoustic measures such as longer pauses, consistent with word retrieval deficits. We further hypothesized that speech differences we find for SYN + AD vs. SYN-AD would not be explained by the clinical syndromic distinction of DLB and PDD. We validated speech markers of AD co-pathology with patients’ CSF tau levels.

2. Methods

2.1. Participants

We examined digital audio samples of picture descriptions collected from 60 patients with LBD and clinical evidence of dementia diagnosed by experienced neurologists (M.G., D.J.I, M.S., A.D.). Patients were followed and evaluated at the University of Pennsylvania’s Frontotemporal Degeneration Center, Parkinson’s Disease and Movement Disorders Clinic, or Alzheimer’s Disease Research Center. Patients were retrospectively selected from the University of Pennsylvania Integrated Neurodegenerative Disease Database (INDD) as of December 31st, 2021. Inclusion criteria were English as a first language; availability of at least one picture description speech sample; disorder-meeting clinical criteria for an LBD with cognitive impairment (i.e., PDD [1] or DLB [2]); and at least one lifespan CSF sample or a post-mortem autopsy report, which was necessary to identify likely AD co-pathology. Participants were analyzed cross-sectionally. When multiple speech samples were available, the most recent one was chosen to be closest to autopsy.

We focused on patients with dementia and thus excluded PD patients without any cognitive difficulties, i.e., PD patients who maintained MoCA score >26 over lifespan testing. For the included patients, we confirmed that dementia onset preceded date of speech sampling. We excluded patients with autopsy evidence of a non-synucleinopathy (e.g., PSP) or a non-AD tauopathy (e.g., argyrophilic grain disease), or evidence of small vessel ischemic disease on MRI (Fazekas score >1).

Clinical tests included Mini Mental State Examination (MMSE), which served as a measure of global cognitive severity, Boston Naming Test (BNT), F letter phonemic fluency, animal naming semantic fluency, and word memory recall (MoCA delayed recall score). Motor symptoms were assessed using the MDS Unified PD Disease Rating Scale (UPDRS) Part III motor assessment. Clinical data were collected within 3 months of speech sampling (mean 2.4 ± 1.7 months).

2.1.1. Biological grouping

Patients were stratified into biologic groups of SYN + AD (n = 22) and SYN-AD (n = 38) using an autopsy-confirmed CSF AD biomarker [16] for majority of patients (n = 46), and by neuropathological diagnostic criteria when available (n = 14), described further below. For patients with both types of data available (n = 7), the autopsy designation was preferred as ground truth evidence for AD co-pathology: four of 7 had CSF and autopsy designations in accordance with each other; the remaining 3 had AD co-pathology at autopsy but did not meet AD CSF biomarker cut-off at time of speech sampling.

We reviewed patients’ charts to gather a list of active medications that could modify patients’ behaviour at time of testing, specifically the use of dopaminergic therapies (including Levodopa (25–200 mg), Ropinirole (2–12 mg), Amantadine (100 mg), and Azilect (1 mg)), cholinesterase inhibitors (including Donepezil (5–10 mg), Galantamine (8 mg), and Memantine (10 mg)), and antidepressants/neuroleptics (including bupropion (300 mg), clonazepam (0.5 mg), QUEtiapine (25 mg), and sertraline (50–100 mg)).

Demographics, clinical characteristics, and medications of SYN + AD and SYN-AD are summarized in Table 1. SYN + AD group had fewer years of education (|d| = 0.59), lower BNT scores (|d| = 0.61), and lower semantic fluency scores (|d| = 0.61) than SYN-AD group, which we adjusted for in our models.

TABLE 1.

Demographics, clinical characteristics, and CSF data for the study cohort, stratified by biological criteria (SYN+AD vs. SYN−AD).

Variable		SYN+AD group (n=22)	SYN−AD group (n=38)	p value
Clinical Disorder, n (%)	DLB	11 (50)	12 (32)	0.25 a
	PDD	11 (50)	26 (68)
Sex, n Male (%)		17 (77)	27 (71)	0.82 a
Education, y (SD)		14.7 (2.4)	16.1 (2.3)	0.03 b
Age, y (SD)		70.8 (7.8)	68.2 (6.5)	0.25 b
Disease duration, y (SD)		6.8 (4.9)	9.4 (6.6)	0.17 b
Cognitive symptom duration, y (SD)		2.2 (3.5)	2.9 (2.3)	0.17 b
Neuropsychological & Neurological Examination Scores
n		21	38	p value
MMSE, max=30 (SD)		24.7 (5.9)	26.5 (2.9)	0.47 c
F letter fluency, words/min (SD)		8.9 (5.1)	13.2 (5.5)	0.11 c
Animal semantic fluency, words/min (SD)		12.1 (5.4)	16 (7.3)	0.04 c
BNT, max=30 (SD)		24.3 (4.5)	26.9 (4.0)	0.03 b
MoCA Delayed Word Recall, max=5 (SD)		1.2 (1.9)	2.0 (1.8)	0.35 c
MDS-UPDRS Part III total, max=132 (SD)		32.5 (11.1)	34.9 (9.8)	0.48 c
MDS-UPDRS Part III rigidity, max=4 (SD)		1.1 (1.4)	1.1 (1.1)	0.96 c
MDS-UPDRS Part III tremor at rest, max=4 (SD)		0.1 (0.4)	0.3 (0.6)	0.43 c
MDS-UPDRS Part III hand action/postural tremor left + right, max=8 (SD)		1.2 (1.4)	1.8 (1.3)	0.37 c
CSF AD biomarker levels
n		18	35	p value
Time between CSF and speech (months)		31.0 (26.7)	18.2 (14.6)	0.06 c
CSF p-Tau, pg/mL		20.1 (11.7)	19.5 (9.0)	0.86 c
CSF t-Tau, pg/mL		47.3 (21.2)	55.4 (32.6)	0.28 c
CSF Aβ, pg/mL		158.5 (47.8)	296.9 (69.1)	-
Active medications
n		22	38	p value
Dopaminergic therapy = yes (%)		11 (50)	25 (66)	0.35
Cholinesterase inhibitor = yes (%)		7 (32)	5 (13)	0.16
Anti-depressants/ neuroleptics = yes (%)		4 (18)	7 (18)	1.00
Clinical Syndrome, n (%)	DLB	11 (50)	12 (32)	0.25 a
	PDD	11 (50)	26 (68)
Sex, n Male (%)		17 (77)	27 (71)	0.82 a
Education, y (SD)		14.7 (2.4)	16.1 (2.3)	0.03 b
Age, y (SD)		70.8 (7.8)	68.2 (6.5)	0.25 b
Disease duration, y (SD)		6.8 (4.9)	9.4 (6.6)	0.17 b
Cognitive symptom duration, y (SD)		2.2 (3.5)	2.9 (2.3)	0.17 b
Neuropsychological & Neurological Examination Scores
n		21	38	p value
MMSE max=30 (SD)		24.7 (5.9)	26.5 (2.9)	0.47 b
F letter fluency words/min (SD)		8.9 (5.1)	13.2 (5.5)	0.11 c
BNT max=30 (SD)		24.3 (4.5)	26.9 (4.0)	0.03 b
MoCA Delayed Word Recall max=5 (SD)		1.2 (1.9)	2.0 (1.8)	0.35 c
MDS-UPDRS Part III max=132 (SD)		32.5 (11.1)	34.9 (9.8)	0.48 c
CSF AD biomarker levels
n		19	33	p value
Time between CSF and speech (months)		31.0 (26.7)	18.2 (14.6)	0.060 c
CSF p-Tau, pg/mL		20.1 (11.7)	19.5 (9.0)	0.86 c
CSF t-Tau, pg/mL		47.3 (21.2)	55.4 (32.6)	0.28 c
CSF Aβ, pg/mL		158.5 (47.8)	296.9 (69.1)	-

^a.p value reflects χ² test estimate.

^b.p value reflects H test estimate.

^c.p value reflects t test estimate.

Abbreviations: SYN+AD= LBD patients with autopsy/CSF evidence of significant AD co-pathology; SYN-AD= LBD patients without AD co-pathology; DLB=dementia with Lewy Bodies; PDD=Parkinson’s disease dementia; disease duration=interval from disease onset to speech collection date; Cognitive symptom duration=interval from dementia onset to speech sampling date; MMSE = Mini Mental State Examination; BNT = Boston Naming Test; MoCA = Montreal Cognitive Assessment; MDS-UPDRS Part III= The MDS-sponsored Revision of the Unified Parkinson’s Disease Rating Scale motor examination.

2.1.2. Clinical grouping

The same 60 patients were also independently stratified by their clinical presentation into 37 PDD (25 males, age: 71 ± 7 years, education: 16 ± 2.2 years, MMSE: 26.9 ± 3.8) and 23 DLB (19 males, age: 66 ± 6 years, education: 15 ± 2.7 years, MMSE: 22.2 ± 5.4) to test a syndromic effect on our speech features. This clinical categorization was performed blinded to likely pathology. PDD group was significantly older than DLB group (|d| = 0.84, p = 0.003). DLB group had shorter disease durations (|d| = 0.76, p = 0.048), and had significantly lower MMSE (|d| = 0.92, p = 0.001) and BNT scores than PDD (|d| = 0.58, p = 0.040). Clinical groups did not differ on all other demographic and clinical variables. Clinical groups did not differ on cholinesterase inhibitor medications or anti-depressant/neuroleptics; PDD group had significantly more patients on dopaminergic therapy than DLB (84% vs 22%, p < 0.001). We reviewed the DLB patients’ medical records to ensure they met current clinical criteria using the one-year interval between parkinsonism motor symptoms and dementia onset to differentiate DLB from PDD [10].

2.2. CSF collection and analysis

CSF data were available in 53 of 60 patients. CSF was collected under standard operating procedures and analyzed with a Luminex xMAP immunoassay platform (Luminex, Austin, TX) to measure CSF t-Tau, p-Tau (phosphorylated at threonine-181), and Aβ1–42 (Aβ). Aβ ≤192 pg/mL, an autopsy-confirmed AD biomarker cut-off [16], was used to identify patients likely harboring AD co-pathology (n = 13 SYN + AD; n = 33 SYN-AD). Tau levels were used to relate AD pathology with speech features. Table 1 summarizes CSF levels by group.

2.3. Neuropathologic examination

A subset of patients had autopsy data available (n = 14). Expert neuropathologists (Authors EBL, JQT) applied currently validated diagnostic criteria to determine AD neuropathologic change (ADNC), and SYN Lewy body stages, as well as the final diagnosis for each case. SYN cases with concomitant intermediate or high ADNC were categorized SYN + AD (n = 9), and no or low ADNC were categorized SYN-AD (n = 5).

2.4. Digital speech data collection

All participants provided a natural speech sample by describing the “Cookie Theft” picture scene from the Boston Diagnostic Aphasia Examination [17]. Participants were asked to describe the picture in as much detail as they could. The speech samples were recorded on a single channel without compression at a sampling rate of 16 KHz and bit depth of 16 in a quiet room with minimal background noise. The average duration of the speech samples was 80 s (±24.86). Recordings were orthographically transcribed. The most recent recording of each participant was analyzed with our automated lexical and acoustic pipelines, to derive 28 automated digital speech measures, described further below.

2.5. Speech analysis - lexical pipeline

Detailed description of the lexical pipeline and validation of the part-of-speech (POS) tagging accuracy have been previously published [18]. Briefly, we automatically tagged the POS category of all tokens and calculated patients’ POS usage, normalized to total number of words (e. g., number of verbs per 100 words). Dysfluency markers were counted, including repetitions and partial words. Mean Length of Utterance (MLU) was estimated by dividing the number of inflected-verbs (calculated by summing the number of modal auxiliaries, past-tense, and present-tense verbs) by the total number of words.

Each noun was rated automatically for concreteness, semantic ambiguity (number of a given word’s meanings in a context), frequency (defined as word frequency per million words on a log₁₀ scale), age of acquisition (AoA, the typical age a word is learnt), familiarity (z-standardized measure of the number of people who know a given word), and word length by the number of phonemes based on published norms. We calculated the mean scores of these measures per participant.

We included lexical measures previously shown to be implicated in amnestic AD during picture descriptions - lexical diversity and lexical density [19,20]. Lexical diversity refers to how varied or complex the vocabulary is, whereas lexical density refers to how “rich in information” the vocabulary is. We measured lexical diversity using the moving-average type-token ratio (MATTR), calculated as number of unique words over total words averaged across a sliding window, yielding the most unbiased estimate of lexical diversity; the window length was set to 15 words, the shortest recording was 42 words long. We calculated lexical density as the proportion of content words (i.e., POS with lexical information – nouns, verbs, adjectives, and adverbs) to total words.

2.6. Speech analysis - acoustic pipeline

Detailed description of our acoustic pipeline has been published previously [21]. Briefly, we used an in-house automated speech activity detector to segment the audio into speech and silent pause segments. We extracted speech and pause durations and derived the following acoustic features: mean speech segment duration; mean pause segment duration; percent of speech time (sum of patient’s total speech and pause time) and pause rate (number of pauses per minute of speech). We visually validated the segmentation output. Pause segments at the beginning and end of each recording and interviewer’s prompts were excluded from the analysis. We used a Praat script to extract percentiles of the fundamental frequency (f0) for each speech segment, which most closely approximates the pitch of speech. We analyzed f0 range using semitones to minimize confounds associated with speaker sex.

We calculated “speaking rate” in words per minute (WPM) = total words divided by total time; and “articulatory rate” in syllables per second (SPS) = total syllables divided by total speech in seconds. The latter measure excludes pause time, and thus is more indicative of actual articulation rate, while WPM is sensitive to prolonged pausing. The automated speech analysis software is available to be shared for scientific research upon request.

2.7. Statistical considerations

All statistical tests were two-sided with a significance threshold of α = 0.05. Statistical analyses were performed with R version 4.1.0 and RStudio version April 1, 1717. The primary goal of the study was to assess the performance of the automated natural speech protocol as a screening instrument for AD co-pathology in LBD. We determined the critical speech measures associated with AD co-pathology using best subset regression modelling [22] which compared logistic regression models classifying SYN + AD vs. SYN-AD across all possible combinations of predictors. The full model with all possible predictors included: all 28 automated speech measures; age, sex, and years of education; MMSE total scores; BNT scores; animal fluency scores; and clinical diagnosis (i.e., PDD vs. DLB). The best-fit model was chosen based on the smallest Mallow’s Cp which estimated the model with the largest R² with the fewest number of variables. Because the two biological groups differed in years of education (see Table 1), we forced-in education as a predictor. Some speech measures were correlated with each other (see Fig. S1), however multicollinearity has minimal impact on best subset regression modelling [23]. Our feature selection approach was validated using k-fold cross-validation, summarized in Supplementary Materials A.

Using the best-fit model, we performed a multiple logistic regression distinguishing SYN + AD vs. SYN-AD for formal comparison of the relative discriminatory strengths of the critical speech features, and conducted receiver operating characteristic curve (ROC) analyses on the predicted values to calculate area under the curve (AUC), sensitivity, and specificity. We performed two post-hoc supplementary analyses: (1) we compared the best-fit speech model with univariate logistic regressions of individual critical speech measures to test whether the variables might be useful as predictors of SYN + AD by themselves; and (2) we tested if adding BNT or animal semantic fluency as predictors improved discriminatory performance, by comparing them to the best-fit speech-only model using DeLong’s test for two uncorrelated (unpaired) ROC curves.

We compared speech measures between SYN + AD and SYN-AD using analyses of covariance (ANCOVAs), covarying for years of education, MMSE total scores, BNT, and animal fluency to adjust for their effects on speech measures. We confirmed normality per measure per group: partial words, repetitions, interjections, and MLU were moderately left-skewed which we reflected, and square-root transformed; MATTR was moderately right-skewed which we natural log transformed. Group comparisons of critical speech measures identified by the first analysis are summarized in the main manuscript. We report Cohen’s effect sizes for group differences in critical speech measures using Cohen’s d. Group means and comparison results for all 28 measures can be found in Supplementary Materials B.

Medication on/off and fluctuation status information was available for half of the cohort, of which n = 21 (70%) were non-fluctuators. As a supplementary analysis, we performed a sub-analysis in the non-fluctuator subset comparing critical measures between n = 13 SYN + AD and n = 8 SYN-AD.

In our CSF analysis, we tested associations between critical speech measures and patients’ CSF t-Tau and p-Tau levels using Pearson correlation tests. CSF Tau levels were log-transformed to normalize the data. We tested if patients’ clinical diagnosis and time interval between speech sampling and CSF sample collection were significant factors using linear regression models. Since the two factors were not significant, we report the results of simple correlations to simplify the models.

We also examined group differences between PDD and DLB using ANCOVAs, covarying for group differences in age, MMSE total and BNT scores. Clinical group means and comparisons across the 28 speech measures are summarized in Supplementary Materials B. Although disease duration was significantly different between PDD and DLB, we did not covary for it in our models because it is inherently tied to the “1-year” clinical criteria distinguishing these clinical disorders, as patients were recruited and enrolled upon onset of cognitive symptoms and presentation to the Cognitive Neurology Clinic.

2.8. Standard protocol approvals, registrations, and patient consents

The Institutional Review Board of the Hospital of the University of Pennsylvania approved the study of human subjects, and all participants agreed to participate in the study by informed written consent.

3. Results

3.1. Feature selection using best subset regression and final logistic regression model

Table 2 summarizes the best-fit logistic regression model with the lowest Mallow’s C_p (C_p = 4.01) distinguishing SYN + AD vs. SYN-AD. Two independent lexical-semantic speech features were identified as the optimal subset of measures: a lower lexical density (strongest predictor) and a younger AoA of nouns (2nd strongest). Clinical grouping of PDD vs. DLB was not chosen as a predictor of SYN + AD by best subset regression analysis. K-fold cross-validation of the best subset regression analysis confirmed an optimal model size of 3 (including education) for distinguishing SYN + AD vs. SYN-AD, and consistently chose AoA of nouns and lexical density as the two most important speech predictors across 4/5 folds (Supplementary Materials A).

TABLE 2.

The multiple logistic regression model with the best subset of measures classifying SYN+AD vs. SYN−-AD (standardized betas).

Variable	Standardized β coefficient	Std. error	z- value	p-value
SYN+AD vs. SYN−AD (reference group)
(Intercept)	−0.75	0.33	−2.28	0.0011
AoA of nouns	−0.72	0.38	−1.87	0.042
Lexical density	−0.90	0.37	−2.46	0.014
Education (forced in)	−0.55	0.36	−1.55	0.12

Abbreviations: SYN+AD= LBD patients with autopsy/CSF evidence of significant AD co-pathology; SYN-AD= LBD patients without AD co-pathology; AoA = Age of acquisition; Lexical density = number of content words/total words.

3.2. ROC analyses

ROC analysis of the best-fit speech model showed sensitivity of 95% and specificity of 66% in classifying SYN + AD vs. SYN-AD (AUC = 0.82).

ROCs for AoA of noun and lexical density separately (covarying for education) both had AUCs<0.71, indicating that the combined best-fit model performed better than individual speech measures. There was no significant difference between the speech-only best-fit model and the model with speech and BNT (p = 0.79) nor the model with speech and animal semantic fluency (p = 0.66), indicating that the speech features perform just as well alone as with the inclusion of cognitive testing.

3.3. Group differences by ANCOVAs

Patients with SYN + AD had a significantly lower AoA of nouns (p = 0.0025, |d| = 0.77) and lexical density (p = 0.0022, |d| = 0.72) compared to patients with SYN-AD, covarying for years of education, MMSE total, BNT, and animal semantic fluency scores (Fig. 1). Education was a significant predictor of both measures (β = 0.0447, p = 0.028; β = 0.0092, p = 0.0051, respectively). BNT, animal semantic fluency, and MMSE were not significant predictors of either speech measures.

Fig. 1.

Age of acquisition (of nouns) and lexical density are significantly reduced in SYN + AD compared to SYN-AD. Clinically-defined PDD and DLB syndromes did not significantly differ on speech measures.

There was no significant group difference in the total number of words. Acoustic measures related to speech durations, pausing, and pitch, did not significantly differ between SYN + AD vs. SYN-AD. Group means and p-values for all 28 speech measures can be found in Supplementary Materials B.

We compared AoA of nouns and lexical density in the non-fluctuators subset between n = 13 SYN + AD vs. n = 8 SYN-AD; numerically, SYN + AD had lower age of acquisition of nouns (|d| = 0.66) and lexical density (|d| = 0.20) than SYN-AD but group differences were not statistically significant in this small sample.

3.4. Associations with CSF t-tau levels

CSF t-Tau levels were inversely associated with lexical density, where CSF t-Tau levels were higher for patients with reduced lexical density (Fig. 2). AoA of nouns did not correlate with CSF tau levels. CSF p-Tau levels did not correlate with speech measures.

Fig. 2.

Lower lexical density scores are associated with increased levels of CSF total Tau (t-Tau) in LBD.

3.5. Speech similarities between PDD and DLB

Clinically-defined PDD and DLB groups did not significantly differ on any speech measures after covarying for differences in age, MMSE total, and BNT scores (Supplementary Materials B).

4. Discussion

There is considerable evidence that AD co-pathology, present in 50% of patients with LBD [4], has a significant impact on disease presentation and survival [5–7]. Valid, quantitative, behavioural-based clinical markers that can serve as screening tools for AD co-pathology prior to the use of more expensive and invasive biomarkers are needed. This study employed automated natural language and speech analysis tools and identified two distinct lexical-semantic measures with good discrimination of patients with SYN + AD from SYN-AD. Age of acquisition (AoA) of nouns and lexical density during picture descriptions were both reduced in SYN + AD than SYN-AD, beyond differences in education and despite a similar level of overall cognitive impairment. By contrast, none of the speech features that we examined showed significant group differences when comparing clinical disorders of PDD vs. DLB. The findings support our hypothesis that AD co-pathology results in a deviant natural language profile in LBD. These findings have important implications for the utility of automated natural speech analysis as a screening tool for AD co-pathology in LBD.

Picture descriptions of patients with SYN + AD had lower lexical density, which approximates how conceptually informative a description is by the proportion of content words (i.e., nouns, verbs, adjectives, and adverbs which contain lexical-semantic information, as opposed to function words which contain less lexical-semantic information) to the total number of words [19]. Reduced lexical density during picture descriptions have also been reported in amnestic AD, and this was related to fewer information content units in the description (i.e., describing fewer elements or themes in the picture) [24,25]. Our finding suggests that AD co-pathology results in less detailed, emptier speech in LBD. Traditional methods for measuring content density are labour-intensive, requiring manual identification of individual propositions or ideas within a text [19]; our study uses automated part-of-speech tagging to approximate lexical density which produces rapid, reproducible measurements.

Reduced lexical density in early life has been linked to greater tau pathology at autopsy in AD, as described in the Nun Study analyzing written letters [19]. In LBD, we report a moderately strong association between lexical density of oral picture descriptions and total Tau (t-Tau) CSF levels, which is thought to reflect non-specific neurodegeneration, and not CSF p-Tau, which is thought to be more specific to AD-related tau pathology. Our finding is not surprising considering the suggested distinct characteristics of CSF tau in LBD compared to AD: our previous work found a direct linear relationship between CSF t-Tau levels and cerebral tau burden at autopsy in LBD, and no such associations for CSF p-Tau [14]. Future studies should confirm associations between speech features and tau burden at autopsy in LBD.

The other significant speech predictor of SYN + AD was a younger age of acquisition (AoA), which refers to the average age a word is learned during language development. In general, earlier-learned words are easier and faster to retrieve from the lexicon than later-learned words, as they refer to items which are easier to imagine and that are more typical exemplars of their category [26]. Our finding is in line with confrontation naming and picture description studies in patients with amnestic AD, which also report better preservation and retrieval of words acquired earlier in life than later-acquired words, with a number of suggested explanations for the phenomenon in AD including the gradual loss of episodic and semantic memory [26,27]. Sources of AoA impairment may be similar in SYN + AD as deficits in both episodic and semantic memory have been previously reported in this population [13–15], and linked to tau pathology in temporal lobe [14]. An investigation into the sources of lexical impairment was beyond the scope of this study and needs to be studied in future investigations by associations with regional atrophy via MRI or post-mortem regional tau analysis. AoA of nouns did not relate to patients’ CSF tau levels, possibly due to the narrow range of variability in AoA of nouns during picture descriptions. By contrast, more elaborate pictures or narrative discourse may probe for larger AoA effects on noun retrieval [28].

Our two automated lexical measures distinguished SYN + AD cases from SYN-AD with high sensitivity of 95%, suggesting that natural speech analysis could be useful as a screening tool prior to the use of more expensive and invasive markers for detecting AD co-pathology in LBD. Adding naming and semantic fluency scores did not improve discriminatory performance, suggesting that the speech features perform just as well alone as with the inclusion of cognitive testing, which further emphasizes the potential clinical utility of our automated speech analysis tool. Automated speech assessment has experimentally-meaningful and practical advantages to assessing language function than traditional clinical tests, such as BNT. Natural speech production has strong face validity with “real-life” communication, unlike cognitive tests that assess individual language sub-domains separately and in more restrictive “unnatural” contexts, and as such may reveal unique impairments that only surface during tasks requiring concerted effort across multiple cognitive domains, like picture descriptions. We also expect no learning effects with natural speech, in contrast to highly structured neuropsychiatric tests which cannot be repeated frequently. Furthermore, the automated nature of the analysis, and the ease and reproducibility of this approach, emphasizes the utility of digital speech markers in remote settings. While we used picture descriptions in this study, the natural speech analysis tool can be easily adapted to other digitized speech tasks, such as story retelling and conversations which may reveal additional impairments in these patients, such as previously reported deficits in semantic and episodic memory, lexical retrieval, and narrative organization in SYN + AD [13–15,29].

Acoustic duration and pitch measures did not differ between SYN + AD and SYN-AD, which in picture descriptions reflect prosodic changes to speech timing, rate, and intonation. In contrast, our recent work in amnestic AD reported distinct acoustic changes during natural speech, including shorter speech segments, frequent pausing, and slow speech rate in AD (Cho, in press). We believe our finding of overlapping acoustic features in SYN + AD and SYN-AD may be due in part to the two groups’ shared motor speech disorder, specifically hypokinetic dysarthria of Parkinsonian etiology which results in short “rushed” speech utterances, frequent pausing, and monotonous speech [30]. This suggests that acoustic measures that reflect prosodic properties of natural speech can be affected by both cognitive and motor symptoms for different reasons.

Our findings show that clinical stratification of LBD patients into DLB vs. PDD based on the arbitrary “1-year rule” is insensitive to meaningful speech features associated with AD co-pathology. Moreover, PDD and DLB appear to be largely similar across all aspects of natural speech production after controlling for confounding group differences in MMSE and age. Our findings are consistent with previous work indicating that a clinical distinction between DLB and PDD may not be very useful in addressing pathological heterogeneity in LBD [10,11].

A major strength of our analysis is the implementation of an automated approach to analyze digitized speech signal in the study of neurodegenerative speech, making it objective and highly reproducible. Our method offers an opportunity to biologically-stratify LBD patients participating in treatment trials. Nevertheless, our study has some limitations. Our sample sizes were relatively small, and it would be important to study larger samples with known or likely pathology. A majority of our patients were categorized based solely on CSF analytes which are valid and reliable, but we cannot rule out the presence of other co-pathology that can be identified only at autopsy and was available for only a subset of the patients. SYN + AD patients stratified by ADNC at autopsy (~50% of group) may not have yet developed AD at time of speech sampling; we analyzed the most recent speech sample to increase the likelihood of patients harbouring AD pathology. It is important to confirm the utility of the speech markers in an independent sample, and determine relationships with burden of AD pathology. It is important to consider the effect of medication state and fluctuation status on motor and non-motor symptoms, which was only available for a small subset of this cohort. It would be useful to elicit a semi-structured speech sample using other tasks such as story recall, which better probes for impaired episodic memory, narrative organization, and learning in AD.

The present study provides evidence that the subset of patients with LBD who have significant AD co-pathology may be characterized by a pattern of lexical-semantic impairments during picture descriptions that diverge from patients with dementia due to relatively pure synuclein pathology. Biology-based stratification has important clinical and research implications, and our work demonstrates that automated analyses of digitized speech show promising utility for this purpose.