Digital Speech Analysis in Progressive Supranuclear Palsy and Corticobasal Syndromes

Abstract

Background:

Progressive supranuclear palsy syndrome (PSPS) and corticobasal syndrome (CBS) as well as non-fluent/agrammatic primary progressive aphasia (naPPA) are often associated with misfolded 4-repeat tau pathology, but the diversity of the associated speech features is poorly understood.

Objective:

Investigate the full range of acoustic and lexical properties of speech to test the hypothesis that PSPS-CBS show a subset of speech impairments found in naPPA.

Methods:

Acoustic and lexical measures, extracted from natural, digitized semi-structured speech samples using novel, automated methods, were compared in PSPS-CBS (n = 87), naPPA (n = 25), and healthy controls (HC, n = 41). We related these measures to grammatical performance and speech fluency, core features of naPPA, to neuropsychological measures of naming, executive, memory and visuoconstructional functioning, and to cerebrospinal fluid (CSF) phosphorylated tau (pTau) levels in patients with available biofluid analytes.

Results:

Both naPPA and PSPS-CBS speech produced shorter speech segments, longer pauses, higher pause rates, reduced fundamental frequency (f0) pitch ranges, and slower speech rate compared to HC. naPPA speech was distinct from PSPS-CBS with shorter speech segments, more frequent pauses, slower speech rate, reduced verb production, and higher partial word production. In both groups, acoustic duration measures generally correlated with speech fluency, measured as words per minute, and grammatical performance. Speech measures did not correlate with standard neuropsychological measures. CSF pTau levels correlated with f0 range in PSPS-CBS and naPPA.

Conclusion:

Lexical and acoustic speech features of PSPS-CBS overlaps those of naPPA and are related to CSF pTau levels.

INTRODUCTION

Progressive supranuclear palsy syndrome (PSPS) and corticobasal syndrome (CBS) are both associated with frontotemporal lobar degeneration (FTLD) with underlying 4-repeat tau pathology. Both disorders are noted for prominent extrapyramidal symptoms. PSPS is characterized by postural instability and oculomotor dysfunction with axial rigidity; CBS is a lateralized motor disorder with dystonia, limb rigidity, tremor, and/or myoclonus. While PSPS and CBS are distinguishable, they frequently co-occur. Cerebral atrophy and hypometabolism are typically detected in subcortical as well as cortical structures, including frontal, temporal, and parietal cortices [1–4].

While researchers often focus on motor symptoms in PSPS and CBS, cognitive symptoms are increasingly recognized. These symptoms may include deficits in executive, visuo-spatial, memory, social, and language functions. Speech and language abnormalities often focus on apraxia of speech (AoS) that is distinguishable from dysarthria and aphasia involving motor-speech errors, sound distortions, and changes in speech timing, but can also include difficulty with naming, word finding, and phonological expression as well as dysfluent, slow, slurred, and hypophonic speech [5–9]. These cases correspond to PSP with predominant speech-and-language disorder (PSP-SL), identified recently as one presentation of PSPS [3]. Some patients with PSPS or CBS present with concomitant non-fluent/agrammatic variant of primary progressive aphasia (naPPA) [1–3], another FTLD spectrum disorder associated with underlying tau pathology [10, 11]. naPPA is characterized by slowed, effortful speech with agrammatism, and speech-sound errors with AoS [12–20]. Primary progressive AoS (PPAoS), a progressive motor speech disorder, also has been observed to occur independently of the linguistic changes found in naPPA [21].

There is an ongoing debate regarding the source of speech and language impairments in PSPS, CBS, and naPPA, attributing them to either cognitive or motor mechanisms or both. Case studies and small series report cases presenting with either PSPS or CBS who eventually develop naPPA, or the converse [22–33]. Such studies are rare and often rely on neuropsychological tests, and manual and subjective speech assessments [27, 30, 32–34]. Detailed, objective speech and language characterizations of patients with PSPS-CBS without naPPA are rare and have been directly compared with those observed in naPPA very infrequently. In this study, we used a novel, automated algorithm to analyze the acoustic and lexical characteristics of patients with PSPS-CBS and compared these directly to acoustic and lexical features produced by patients with naPPA. Finally, we aimed to test our novel automated techniques of speech analyses from a one-minute picture description task. These methods are objective and highly reproducible, enabling the quantification of speech markers, including acoustic measures such as pitch, speech and pause durations, and pause rate, and lexical measures such as verb and noun production and partial words. We also related speech markers to neuropsychological measures. We hypothesized that lexical and acoustic speech features in PSPS-CBS would reveal a spectrum of speech-language impairments that partially overlaps features seen in naPPA, and that these speech features could not be easily attributed to other neuropsychological deficits.

In order to obtain a clearer understanding of the underlying neurobiological associations of these speech features, we also examined cerebrospinal fluid (CSF) levels of phosphorylated tau (pTau). PSPS, CBS, and naPPA are each associated with tau pathology, and we examined CSF levels of pTau because this appears to reflect the cerebral burden of tau pathology [35]. Previous work has related specific acoustic features of speech to CSF pTau in naPPA [18]. We hypothesized that CSF pTau levels would be associated with speech features identified in PSPS-CBS and naPPA.

METHODS

Participants

We examined 153 native English speakers with PSPS (n = 41), CBS (n = 46), and naPPA (n = 25), and healthy controls (HC, n = 41). Patients were examined at the Hospital of the University of Pennsylvania by experienced neurologists (MG, DJI) and the clinical phenotype was reviewed in a multidisciplinary consensus conference according to published clinical research criteria [1, 3, 15, 36]. All patients were assessed between January 2000 and June 2019. Thirty-one of these cases had a neuropathological diagnosis of a tauopathy (PSP, CBD, or Pick’s disease). Some of the naPPA patients have been reported previously, but not in comparison to PSPS-CBS [12–14, 17, 18, 37]. We excluded cases with other neurological (e.g., vascular disease, hydrocephalus, head trauma), medical (e.g., infectious, inflammatory, metabolic), or primary psychiatric (e.g., psychosis, major depression, bipolar) conditions that could confound our observations. We excluded 4 patients with PSPS-CBS diagnosis because it was not clear whether the primary diagnosis was PSPS or CBS. Demographic and clinical characteristics of the patient groups are summarized in Table 1. We combined patients with PSPS and CBS who did not have concomitant naPPA (PSPS-CBS, n = 87) since these groups were matched for sex, age, education, and disease duration, and their speech characteristics were similar (Supplementary Table 1). Mini-Mental State Exam (MMSE) was used to measure overall disease severity and was obtained within 6 months of the speech sample. However, since MMSE scores differed in PSPS and CBS patients (t = 2.71, p = 0.009, 95% confidence interval (CI) 0.74 to 4.88), we also compared PSPS and CBS groups separately for speech and language impairments (see Supplementary Table 1). The PSPS-CBS and the naPPA groups were matched for education, sex, age, disease duration, and MMSE scores. The HC group matched both patient groups in age, sex, and education.

Table 1.

Demographic and clinical characteristics of patient groups and healthy controls

Mean (Standard Deviation)	Healthy controls	PSPS-CBS	naPPA	p
n	41	87	25
Age (y)	68.94 (7.57)	68.16 (7.53)	70.91 (8.93)	0.298
Education (y)	15.80 (2.42)	14.75 (2.69)	15.32 (3.00)	0.108
Disease duration (y)	NA	4.11 (2.41)	3.32 (1.91)	0.132
Sex: M (%)	18 (43.9%)	40 (46%)	11 (44%)	0.969
n	35	74	24
MMSE (max=30)	29.11 (0.99)	24.92 (4.75)*	23.04 (5.89)*	<0.001

PSPS, progressive supranuclear palsy syndrome; CBS, corticobasal syndrome; naPPA, non-fluent/agrammatic primary progressive aphasia; y, year; MMSE, Mini-Mental State Examination; M, males.

^*Differs from HC (p < 0.001).

In the course of preparing for this study, we discovered a small group of patients with PSPS-CBS and concomitant naPPA (PSPS-CBS + naPPA, n = 8). We did not include these patients in the PSPS-CBS and naPPA patient groups studied below, but we analyzed them separately. We excluded one patient with CBS, naPPA, and concomitant behavioral variant frontotemporal degeneration to constrain this secondary analysis to PSPS-CBS with a concomitant well-defined naPPA diagnosis. This subgroup of patients with both naPPA and PSPS-CBS is summarized in Supplementary Table 2 and Supplementary Figures 1 and 2.

To determine the frequency of clinical AoS features in our cohort, we reviewed all available clinical charts (n = 82) between 2009 and 2019 (earlier charts were not available). We found a total of five patients with features clinically judged to be consistent with AoS. This included two patients from the PSPS-CBS group, one patient from the naPPA group, and one from the PSPS-CBS + naPPA group. One patient from the PSPS-CBS + naPPA group presented with PPAoS. Patients with PPAoS have features of AoS, but these features are isolated from other disorders, such as aphasia [21]. Due to these small sample sizes, we were unable to pursue a statistical analysis of AoS in this study.

This study complies with guidelines on human experimentation. All participants were enrolled in study protocols and participated in an informed consent procedure in accordance with the Declaration of Helsinki approved by the Institutional Review Board of the University of Pennsylvania.

Experimental speech task

We used the same methodology to acquire a semi-structured speech sample as employed in our previous studies [13, 14, 18, 38–40]. Briefly, an interviewer instructed participants to describe the Cookie Theft scene from the Boston Diagnostic Aphasia Examination [41] in as much detail as they could and using full sentences. The interviewer prompted the participants to continue speaking if they were silent for more than a few seconds, while aiming to minimize interruptions and speech overlap with participant’s speech. Total recording time averaged 70.9 s (range 15.1–156.6 s). Only the earliest recordings of participants were analyzed. Recordings were obtained digitally in a quiet room with minimal background noise and stored as.wav files.

Acoustic analysis

We used the same methodology as published previously to analyze the digitized speech samples [18, 38–40]. Briefly, a speech activity detector (SAD), developed at the University of Pennsylvania Linguistic Data Consortium (LDC), automatically segmented the acoustic signal into speech and nonspeech (silent pause) segments [42, 43]. The output and sound files were manually reviewed in Praat [44] to validate sound quality and accuracy of the automated segmentation and pitch tracking. Manual corrections were sparse. We manually labelled out interviewer’s speech as well as cough, laughter, and background noise that interrupted speech in order to minimize pitch tracking errors.

Using an R script, we calculated acoustic measures, including mean speech segment duration, mean pause segment duration, and pause rate (number of pauses per minute over total speaking time) as these have been informative in previous studies of PPA [12, 14, 18]. We excluded silent pauses at the beginning and end of recordings and those following an interviewer’s prompt, as those are either artifactual or represent response time. We tracked the fundamental frequency (f0) of the participant using Praat’s pitch tracker. Pitch tracking provided f0 estimates for each 10 ms in Hertz (Hz) during continuous speech segments only; any artifactual pitch estimates during pause segments were not included. We transformed the raw f0 estimates from Hz to a semitone (ST) scale, a relative logarithmic scale selecting each participant’s 10th percentile f0 estimate as the reference point. Thus, the 90th percentile represented the speaker’s f0 range as described in our previous studies [18, 38–40], We reviewed all participants with an f0 range 1.5 SD above or below the group mean and found 23 participants with substantial “creaky voice” or substantial background noise. These vocal characteristics carry a high probability for pitch-tracking errors; thus we excluded these f0 outliers from further analyses of f0 range (HC, n = 5; PSPS, n = 5; CBS, n = 8; naPPA, n = 5; PSPS-CBS + naPPA, n = 1).

Lexical-semantic and grammatical analyses

The speech samples were transcribed by an expert linguist (SA). We used a previously reported methodology to automatically tag part of speech (POS) [37, 40] in the transcribed picture descriptions [45, 46]. Briefly, spaCy [46] uses the Penn Treebank [47] to map words onto tags from the Google Universal POS tag set [48]. In previous studies using the same automated algorithms, we found specific lexical features characterized PPA patients’ speech [18, 37]. Speech rate (words per minute, wpm), a measure of speech fluency, for which deficits are also characteristic of naPPA [15–17] was calculated automatically by adding up word and partial word counts for each recording from the spaCy output, then dividing by total recording time, excluding interviewer speech. The POS tags of spaCy are detailed in the library’s documentation [46]. Tokens annotated as “X,” “VB”, and “NOUN” were manually inspected for validation. The “X” tag consisted mostly of partial words and thus is referred to as “partial words” in our report. In our analyses, instances of partial words, verbs, and nouns per 100 words were calculated to control for total word production.

We previously performed manual coding by an expert linguist (SA) for measures of “grammatical performance,” reflecting the grammatical complexity of participants’ utterances [13]. We selected two grammatical measures that were previously linked to naPPA speech [12–14, 18]: dependent clauses per 100 utterances (DC) and percentage of utterances that are well-formed sentences (WFS). We calculated the average of these two measures for each speaker, in order to minimize ceiling and floor effects. This combined grammatical measure was used for clinical correlations (described below) to validate our novel automated speech markers and relate them to underlying neurolinguistic mechanisms.

Neuropsychological assessment correlations

In order to determine the potential role of cognitive deficits in the speech impairments in our cohort, we investigated several domain-specific cognitive measures that were available in most patients within 6 months of their speech sample (Table 2), including the Boston Naming Test (BNT) [49], which measures confrontation naming; delayed recall from the Philadelphia (repeatable) Verbal Learning Test (PrVLT) [50] to assess episodic verbal memory; modified Rey-Osterrieth Complex Figure test for visuoconstructional functioning [51]; and letter-guided category naming fluency using the target letter F for executive functioning [52]. F letter scores were collected within 12 months of speech sample collection to increase sample size.

Table 2.

Acoustic, lexical, and cognitive characteristics of patient groups and healthy controls

Mean (Standard Deviation)	Healthy controls	PSPS-CBS	naPPA	p 1
n	41	87	25
Speech segment duration (s)	2.00 (0.57)	1.42 (0.48)*,P	1.17 (0.50)*	<0.001
Pause segment duration (s)	0.93 (0.44)	1.82 (1.01)*	1.65 (1.10)*	<0.001
Pause rate (ppm)	31.61 (9.51)	45.10 (13.43)*,P	57.81 (19.95)*	<0.001
n (f0 range)	36	74	20
f0 range (ST)	5.87 (1.94)	4.97 (2.03)*	4.30 (1.59)*	0.010
Speech rate (wpm)	144.95 (32.12)	89.03 (36.70)*,P	65.15 (25.55)*	<0.001
Partial words	0.52 (0.96)	1.33 (2.93)P	3.67 (5.25)*	0.013
Verb production	23.07 (3.45)	23.19 (4.56)P	19.99 (4.46)*	0.013
Noun production	20.06 (4.53)	22.20 (4.98)*	21.40 (8.33)	0.055
Grammatical performance score2	0.00 (0.78)	−0.84 (1.13)*	−1.48 (1.71)*	<0.001
n	26	67	18
Boston Naming Test score	28.12 (2.45)	24.46 (6.19)*	24.17 (4.69)*	0.001
n	26	62	18
Delayed Verbal Recall score	7.04 (1.82)	5.55 (2.56)*	5.67 (2.93)	0.034
n	26	57	19
Rey-Osterrieth Complex Figure	34.54 (2.11)	22.21 (10.90)*	25.58 (9.83)*	<0.001
n	4	50	15
F-letter Fluency score	17.75 (8.10)	6.72 (3.57)*	5.93 (3.39)*	0.038

PSPS, progressive supranuclear palsy syndrome; CBS, corticobasal syndrome; naPPA, non-fluent/agrammatic primary progressive aphasia; s, seconds; ppm, pauses per minute of speech time; ST, semitones; wpm, words per minute; f0, fundamental frequency.

¹p-values for all contrasts in main groups: HC, PSPS-CBS, and naPPA:

^*differs from HC

^Pdiffers from naPPA.

²Values are z-scores of averaged dependent clauses per 100 utterances and percentage of utterances that are well-formed sentences, using control mean and standard deviation.

Cerebrospinal fluid phosphorylated tau analysis

We analyzed CSF samples that were available in a subset of patients, including PSPS-CBS (n = 50; PSPS = 26 and CBS = 24), naPPA (n = 11), and PSPS-CBS+naPPA (n = 5). As previously reported [35], CSF was analyzed using two platforms, Luminex xMAP or Innotest ELISA (which is then transformed to the Luminex scale). We previously associated CSF levels of pTau directly with cerebral burden of tau in our autopsy cohort [35]. We excluded patients who had a CSF profile suggestive of underlying AD pathology (pTau/Aβ<0.09 [35]). Thus, this CSF analysis included 47 patients (PSPS-CBS, n = 34, including PSPS = 21 cases and CBS = 13; naPPA, n = 9; PSPS-CBS + naPPA, n = 4) whose CSF profiles suggested underlying FTLD pathology. This subset of patients did not differ demographically from the rest of the patients.

Statistical analysis

We evaluated normal distribution of speech variables using Kernel density and Q-Q plots. Levene’s test revealed violations of homogeneity of variance for pause segment duration, pause rate, partial words, and nouns, so we used non-parametric tests, including the Kruskal-Wallis test to compare groups on acoustic, lexical, and grammatical performance. Pairwise comparisons between groups were calculated with the Mann-Whitney test with a 95% CI. For comparisons between groups, unless otherwise stated, all differences are significant at least at the two-tailed level p < 0.05 and the p-value was adjusted using a False Discovery Rate (FDR) correction for multiple comparisons.

To investigate the relationship between acoustic-lexical markers and specific language measures [15], we performed univariate linear regression analyses where each acoustic or lexical marker was the outcome (dependent) measure, and grammatical performance or speech rate (each was used in a separate model) were the main predictors (independent variables) in a cohort that included both main patient groups (PSPS-CBS and naPPA). We first examined group interactions with grammatical performance or speech rate (e.g. [speech marker] ~ [grammatical score] + [group] + [grammatical performance X group]). We eliminated the interaction term when no interaction was found. We used natural logarithmic transformations to normalize skewed data when examining speech and pause segment durations. Regression models were validated by viewing the residual plots. We used Spearman correlations with Bonferroni correction for multiple tests to correlate speech markers with BNT total scores, PrVLT scores, Rey visuoconstructional total scores, and F-letter fluency scores in PSPS-CBS and naPPA patients.

To relate speech markers with probable underlying pathology, we used linear regression models where CSF pTau level (transformed to the natural logarithm) was the dependent variable. We ran a separate model for each speech marker as the main predictor (independent variable). We controlled for potential confounders including age and the time interval between the speech sample collection and CSF collection (range −62.7–106.7 weeks). We checked residual plots to validate the models.

In Supplementary Table 2 and Supplementary Figures 1 and 2, we compared speech markers in PSPS-CBS + naPPA to those of the main patient groups, using the Kruskal-Wallis test. Due to the small sample size, we report the results with uncorrected p-values.

All statistical analyses were conducted using R v.3.6.3, RStudio v. 1.2.5033 [53], and Rstudio with additional packages [54–60].

Data availability statement

Access to all data anonymized for purposes of replicating procedures and results is available upon request. We share anonymized data with qualified investigators who have appropriate regulatory approval and transfer agreements.

RESULTS

Acoustic results

Table 2 summarizes the acoustic, lexical, and grammatical characteristics of the participants. Pairwise comparisons of the acoustic markers revealed that, compared to HC, PSPS-CBS had shorter speech segments (W = 722, p < 0.001, 95% CI −0.77 to −0.4, Fig. 1A), longer pause segments (W = 2914, p < 0.001, 95% CI 0.46 to 0.98, Fig. 1B), higher pause rate (W = 2814, p < 0.001, 95% CI 9.29 to 18.15, Fig. 1C), and reduced f0 range (W = 975, p = 0.03, 95% CI −1.84 to −0.18, Fig. 1D).

Fig. 1.

Speech and pause durations and f0 range by clinical phenotype. A) Mean speech segment duration measured in seconds. B) Mean pause segment duration measured in seconds. C) Pause rate measured in pauses per minute of speech time (ppm). D) fundamental frequency (f0) range in semitones (ST): the 90th percentile represents the trimmed f0 range. HC, healthy controls; naPPA, non-fluent/agrammatic primary progressive aphasia; PSPS-CBS, progressive supranuclear palsy and corticobasal syndrome spectrum disorders; s, seconds.

Compared to HC, naPPA also had shorter speech segments (W = 110, p < 0.001, 95% CI −1.1 to −0.59, Fig. 1A), longer pause segments (W = 761, p = 0.001, 95% CI 0.18 to 0.74, Fig. 1B), higher pause rate (W = 899, p < 0.001, 95% CI 17.93 to 33.38, Fig. 1C), and reduced f0 range (W = 186, p = 0.007, 95% CI −2.62 to −0.62; Fig. 1D).

Compared to naPPA, PSPS-CBS patients produced longer speech segments (W = 671, p = 0.004, 95% CI 0.09 to 0.41) and had reduced pause rate (W = 1526, p = 0.002, 95% CI −20.42 to −4.37).

Lexical results

Compared to HC, PSPS-CBS had a slower speech rate (W = 435, p < 0.001, 95% CI −72.38 to −47.80, Fig. 2A) and produced more nouns (W = 2271.5, p = 0.038, 95% CI 0.53 to 4.11, Fig. 2B). They produced a similar number of partial words and verbs (Fig. 2B).

Fig. 2.

Lexical measures by clinical phenotype. A) Speech rate in words per minute (wpm). B) Partial words, verb and noun production counts per 100 words. HC, healthy controls; naPPA, non-fluent/agrammatic primary progressive aphasia; PSPS-CBS, progressive supranuclear palsy and corticobasal syndrome spectrum disorders.

Compared to HC, naPPA also had a slower speech rate (W = 30, p < 0.001, 95% CI −95.79 to −66.35, Fig. 2A). They produced fewer verbs than HC (W = 316.5, p = 0.015, 95% CI −4.87 to −0.75, Fig. 2B) and more partial words (W = 719.5, p = 0.011, 95% CI 0 to 1.64, Fig. 2B). They did not differ from HC (p = 0.54) or PSPS-CBS (p = 0.54) in their noun production (Fig. 2B).

Compared to naPPA, PSPS-CBS had a faster speech rate (W = 669, p < 0.003, 95% CI −34 to −6.88), produced more verbs (W = 689, p = 0.015, 95% CI −5.11 to −0.91), and produced fewer partial words (W = 1371.5, p = 0.054, 95% CI 0 to 1.32).

Clinical correlations

We used linear regression models to test the hypothesis that our digitized, automated, acoustic and lexical markers were associated with underlying grammatical impairment and slowed speech rate as the most distinctive linguistic manifestations of naPPA [15].

Greater grammatical impairment was associated with longer pause segment duration (beta = −1.35, p < 0.001, Fig. 3B), higher pause rate (beta = −21.73, p = 0.013, Fig. 3C), and reduced verb production (beta = 7.86, p = 0.003, Fig. 3D) for both PSPS-CBS and naPPA. Greater grammatical impairment was associated with briefer speech segment duration only for PSPS-CBS (beta = 0.69, p = 0.001, Fig. 3A). Grammatical impairment correlated with speech rate (beta =1.19, p < 0.001), independent of phenotype. Grammatical impairment was related to noun use (beta = −11.48, p < 0.001), but did not correlate with partial words or f0 range.

Fig. 3.

Linear regression models relating grammatical performance and speech rate to acoustic and lexical features of speech. Grammatical performance, transformed by z-score relative to healthy controls, related to (A) natural log of mean speech segment duration measured in seconds, (B) natural log of mean pause segment duration measured in seconds, (C) pause rate measured in pauses per minute of speech time, (D) verb count per 100 words. Speech rate in words per minute (wpm) related to (E) natural log of mean speech segment duration, (F) natural log of mean pause segment duration, and (G) pause rate. naPPA, non-fluent/agrammatic primary progressive aphasia; PSPS-CBS, progressive supranuclear palsy and corticobasal syndrome spectrum disorders; s, seconds; wpm, words per minute; ppm, pauses per minute.

Higher speech rate was associated with longer speech segment duration (beta = 0.003, p < 0.001, Fig. 3E) independent of phenotype, and a lower pause rate (beta = −0.15, p < 0.001. Fig. 3G) only in PSPS-CBS. Pause segment duration was related to speech rate in both groups, and this was more pronounced in naPPA (beta = −0.018, p < 0.001, Fig. 3F) than PSPS-CBS (beta = −0.011, p < 0.001). Speech rate was not associated with f0 range, partial words, verb or noun production in any group.

In relating speech measures to the neuropsychological assessment scores, we found only that lower F-letter score was related to longer pause segment duration (rho = −0.387, p = 0.047) and lower speech rate (rho = 0.479, p = 0.002) in PSPS-CBS and naPPA.

CSF analyses

We found that narrow f0 range (beta = −0.069, p = 0.002, Fig. 4) correlated with higher CSF levels of pTau in all patients.

Fig. 4.

CSF pTau level related to f0 range in PSPS-CBS and naPPA. CSF pTau levels is inversely correlated with f0 range, controlling for age and time interval between speech sample recording and CSF collection. CSF pTau, cerebrospinal fluid phosphorylated tau; naPPA, non-fluent/agrammatic primary progressive aphasia; PSPS-CBS, progressive supranuclear palsy and corticobasal syndrome spectrum disorders; PSPS-CBS + naPPA, PSPS-CBS with concomitant naPPA; f0, fundamental frequency; ST, semitones.

DISCUSSION

We used a novel, automated speech analysis of natural speech to characterize acoustic and lexical markers in PSPS-CBS and naPPA. We found that speakers with PSPS-CBS produce shorter speech segments, longer and more frequent pauses, reduced f0 range, reduced speech rate, and greater noun use in comparison to normal speakers. naPPA speakers showed a similar pattern of acoustic markers, but their speech segments were even shorter and pause rates higher. More frequent production of partial words and reduced verb production were found only in naPPA speakers. Longer and more frequent pauses were related to impaired grammaticality in both PSPS-CBS and naPPA, and shorter speech segments were related to impaired grammaticality in PSPS-CBS. Narrow f0 range was related to higher CSF pTau levels in patients. We discuss these findings in detail below.

We found that PSPS-CBS and naPPA show similar patterns of impairment in acoustic properties of speech. These findings suggest that PSPS-CBS speakers lie on a spectrum of speech impairments with naPPA, and that the slow, non-fluent speech and other speech impairments we find in naPPA may be properties of the PSPS-CBS phenotype as well. This has been recognized in the speech-and-language presentation of PSPS [3]. A previous study reported similar impressions of dysprosodic speech in PSPS and naPPA [61]. PSPS speakers were also found to have reduced speech production and impaired pitch range, with higher pause rate compared to Parkinson’s disease [62]. These impressions of acoustic impairment in PSPS are in line with our report. The findings in the present study are cross-sectional. Others have described the longitudinal progression to naPPA in patients presenting with PSPS-CBS, and conversely in patients presenting with naPPA who eventually demonstrate features of PSPS-CBS [22, 23, 28–30, 32, 63, 64]. The cross-sectional nature of our study limits our ability to report on the longitudinal evolution of speech in PSPS-CBS in such matters, such as determining whether these individuals eventually exhibit the criteria necessary for the diagnosis of naPPA, as in the 8 cases reported in Supplementary Table 2 and Supplementary Figures 1 and 2.

While PSPS-CBS and naPPA exhibited many similar speech characteristics, it is important to highlight that they do not have identical speech patterns. naPPA speakers are more impaired in their acoustic markers and have distinct lexical impairments. This is in keeping with previous publications indicating that naPPA is characterized by effortful, agrammatic speech, with reduced pitch range and impaired verb production, higher pause rate and slower speech rate [12–18, 37]. We found that speech segment duration, pause rate, speech rate, and production of partial words and verbs are not as compromised in PSPS-CBS as in naPPA and thus can be used as markers to distinguish between PSPS-CBS and naPPA.

While our automated assessment of acoustic and lexical properties of digitized speech is objective and reproducible, we are unable to perform an automated assessment of speech errors, due in part to the difficulty detecting and quantifying speech sounds that are not from the corpus of speech sounds in English, the native language of our speakers. It is important to be able to quantify speech errors since this represents an important characteristic of AoS [20]. However, partial words may in part be a surrogate for detecting AoS because this characteristic may be related in part to articulatory disruptions such as mid-word pauses that are another characteristic of AoS. Our rate of clinical detection of AoS in this cohort is lower than expected, possibly related to incomplete assessments of speech in a busy movement disorders clinic. In another study, we estimated a 21% frequency of AoS in naPPA [65].

Our cohort of 8 participants with PSPS-CBS + naPPA is a small but informative group. This sub-group, summarized in the Supplementary Table 2 and Supplementary Figures 1 and 2 because it was too small to analyze statistically, has been described previously [23–33]. We found that this group of PSPS-CBS speakers is similar to naPPA in their acoustic markers, with shorter speech segments, higher pause rate, and reduced f0 range compared with PSPS-CBS without naPPA. These results align with earlier case studies on patients with concomitant PSPS-CBS and naPPA, where researchers subjectively observed that speech was more severely impaired than might be expected in PSPS-CBS alone [22–31]. Future longitudinal studies of these cases would be informative to learn about the evolution of phenotypic heterogeneity in the tauopathies.

Reduced grammatical complexity and dysfluency are said to be characteristic of naPPA [12–18]. The results of our regression analyses support the hypothesis that longer pauses, shorter speech segments and higher pause rate relate to impaired grammaticality and dysfluency in patients with naPPA as well as patients with PSPS-CBS. Reduced speech rate and verb production were also related to grammatical impairment in both PSPS-CBS and naPPA. Reduced noun production was associated with a grammatical impairment in PSPS-CBS as well, and the basis for this unexpected finding requires additional investigation. These observations reinforce the assertion that speech in naPPA and PSPS-CBS can be considered to be on the same spectrum of impairment.

Pause segment duration and speech rate were related to F-letter scores in PSPS-CBS and naPPA, but no other speech characteristics were related to neuropsychological measures. While this may suggest that these speech features are closely linked to executive dysfunction in our patients, pause segment duration and speech rate are timed measures like F-letter fluency, and no other acoustic or lexical measures of speech were related to F-letter fluency. It is thus unlikely that speech measures can be fully explained by an executive deficit [17]. In fact, no other neuropsychological measure correlated with our speech markers, suggesting that these speech measures provide unique information not otherwise provided by traditional neuropsychological measures. We assessed only a subset of possible neuropsychological measures, and further studies relating neuropsychological to speech measures are needed to determine more precisely the unique data provided by quantified speech measures.

Finally, we found an association between increased CSF pTau levels and limited f0 range. This resembles our previous finding where we reported that elevated CSF pTau is related to limited f0 range in naPPA patients [18]. Here, we extend this finding to PSPS-CBS, which is also frequently a tauopathy. Previous work has demonstrated a direct relationship between CSF pTau levels and the burden of tau pathology in the cerebrum [35]. These findings therefore suggest a link between our novel automated speech markers and underlying tau pathology. The presence of a constrained f0 range, combined with elevated CSF pTau level, thus would provide additional confidence in estimating the likelihood of an underlying tauopathy, and f0 range may be an inexpensive, non-invasive marker that can screen for underlying tau pathology in PSPS-CBS and naPPA. Since speech is easily collected and its analysis is reliable and reproducible, this adds a potential role for these speech markers in patient screening and monitoring during clinical trials for novel interventions targeting tau pathology.

Strengths and limitations

The speech analyses described in this study are reliable, objective, and highly reproducible since they were based on automated algorithms assessing digitized speech samples. No subjective ratings were used which could confound or obscure the data. The SAD is accurate in extracting pause and speech durations from the natural speech of patients. The POS tagger has a high accuracy in tagging parts of speech in the transcripts, and we confirmed this in the present stuidy with a manual assessment of the lexical measures we analyzed. Speech measures extracted this way have an advantage over more commonly used clinical measures, such as neuropsychological tests, because they are obtained without the need for trained personnel, judgments are objective, and they provide unique information on speech and language impairment. Acquisition is efficient and requires only about one minute of patient time in performing a familiar picture description task, highly reproducible, and the data can be obtained remotely.

Despite these strengths, there are also some limitations that should be acknowledged when considering our results. Pitch-tracking may be inaccurate, due in part to voice characteristics and uncontrolled environmental noise that can make it difficult to estimate the true pitch. While PSPS, CBS, and naPPA are relatively rare conditions, we included a reasonable number of patients to power our statistical analyses, although the very small number of PSPS-CBS + naPPA participants limited our ability to analyze these patients statistically. Investigation of a larger group is needed to better understand the manifestations of these individual and combined phenotypes. Another limitation is that we did not have a sufficient number of motor scores on our participants to investigate possible motor effects on speech in PSPS-CBS and naPPA. We examined only a limited number of neuropsychological measures because of limited availability, and only a subset of patients with digitized speech had neuropsychological measures since data was sparse for earlier recordings of our patients. Available data on AoS was also limited. Future work is needed to focus on the development of an objective marker of speech-sound errors. Longitudinal data would significantly enhance our observations of speech deficits in PSPS-CBS and naPPA.

With these limitations in mind, our findings provide a quantitative comparison of speech and language impairments in PSPS-CBS and naPPA. Both groups produce shorter speech segments, longer and more frequent pauses, as well as reduced f0 range and speech rate, compared to normal speakers. Patients with naPPA produced speech segments that were even shorter, pause rate even higher, and speech rate even slower, and additionally show increased production of partial words and reduced verb production. These impaired speech markers were related to impaired grammaticality in both patients with naPPA and patients with PSPS-CBS. These observations are consistent with the idea that these patient groups have speech characteristics that are on a similar spectrum of impairment. This study also demonstrates the use of automated speech analyses to characterize speech patterns in impaired speakers suffering from a neurodegenerative disease.