Striatal network dysfunction underlying speech impairment in early Parkinson’s disease

Demographic and clinical characteristics

A total of 148 right-handed participants, including 104 PD patients (68M, 36F) and 44 age-matched HC (23M, 21F), were analyzed (Table 1). No significant differences in age, years of education, MoCA score, or delays between voice recordings, DaT-SPECT, or MRI examinations were observed between patients and HCs. PD patients had significantly lower Mattis scores and higher MDS-UPDRS Part III scores in the OFF state. Within the HC group, males and females showed no differences in the clinical scores. However, in the patient group, males had a higher average MDS-UPDRS Part III score than females, as well as lower Mattis and MoCA scores. No significant difference in the disease duration, defined as time since formal diagnosis, was found between males and females.

Speech features

Group differences in speech features are presented in Table 2. PD patients overall had significantly more monotonous speech (lower SD log F0), longer pause durations, difficulty maintaining a steady rhythm (higher RSD), and a worse (higher) articulation score than controls (6.3 < F < 24.0; 10⁻⁵ < p_corr < 0.02), with medium to large effect sizes (0.04 < η² < 0.14). No significant difference between PD patients and controls was found after false discovery rate (FDR) adjustment for SD intensity, absolute rhythm, and phonation. The main effect of sex was significant for SD log F0 (higher in women) and median pause duration (shorter in women), with 10.0 < F < 17.8 and 0.0003 < p_corr < 0.007. Moreover, there was a significant interaction effect between disease and sex for the articulation score (greater PD effect in males than females; F = 17.57; p_corr = 0.0003).

Neuroimaging data

Group differences in DaT-SBR are presented in Table 3. There was a main effect of the disease with PD patients showing reduced SBR in both the left and right associative and sensorimotor territories of the putamen and the caudate nucleus (65.3 < F; 10⁻³⁹ < p_corr < 10⁻¹¹; 0.47 < η² < 0.91) and a main effect of sex in all regions (4.51 < F < 21.3; 10⁻⁴ < p_corr < 0.04), with higher SBR in females. There was no significant Disease-Sex interaction effect.

Group differences in functional connectivity are presented in Table 4. The ANOVA tests revealed a significant disease main effect (Fig. 1A) with decreased functional connectivity in PD observed between both the left and the right associative caudate nuclei and the right superior temporal gyrus (Fig. 1B). There was no significant effect of sex alone, but significant Disease-Sex interaction effects were observed (Fig. 2A). In two networks involving the caudate nucleus, we observed a PD effect in males (lower connectivity in PD than HC), which was reversed in females (Fig. 2B). The first network was composed of the left sensorimotor caudate, functionally connected to the left precentral gyrus, left supplementary motor area, and right putamen (red blobs in Fig. 2A). The second network was composed of the right associative caudate functionally connected to the left middle temporal gyrus and the pars triangularis of the right inferior frontal gyrus (yellow blobs in Fig. 2A). No significant group differences were observed in the functional connectivity with the putamen.

It should be noted that the term “PD effect” or “disease effect” refers here to group differences between PD patients and HCs and does not imply a direct causal effect of the disease, as these differences may also reflect compensatory mechanisms or even medication effects.

Correlations between speech features and neuroimaging

DaT-SBR only correlated with prosody, and only in males (Table 5). SD log F0 positively correlated with DaT-SBR in the associative and sensorimotor territories of the putamen and caudate nucleus in the right hemisphere and the associative territory of the caudate nucleus in the left hemisphere (0.52 < r < 0.63; 0.01 < p_corr < 0.04). No significant correlation was found in females. No significant correlations were found between DaT-SBR and pause duration, RSD rhythm or articulation score.

All the significant correlation results between fMRI striatal connectivity and speech features are listed in Table 6 and described below.

Prosody (SD log F0) was positively correlated, in females, with connectivity between (1) the left sensorimotor putamen and the right middle frontal gyrus and the left ventrolateral thalamus; and (2) the left associative putamen and the left superior frontal gyrus (medial part) and the opercular part of the right inferior frontal gyrus. In addition, negative correlations were found with connectivity between (1) the left associative caudate and the left superior and middle temporal gyri; (2) the left sensorimotor caudate and the right middle frontal gyrus; (3) the right sensorimotor caudate and the right middle frontal gyrus; and (4) the right associative caudate and the triangular part of the right inferior frontal gyrus, and the left middle temporal gyrus. In males, a negative correlation was observed with connectivity between the left associative putamen and the right middle temporal gyrus.

Median pause duration was negatively correlated, in females, with connectivity between the left associative putamen and the triangular part of the inferior frontal gyrus, and the middle temporal gyrus in the left hemisphere, as well as between the right associative putamen and the left superior temporal gyrus. In males, a negative correlation was found with connectivity between the left associative caudate and the left superior frontal gyrus, and a positive one with connectivity between the left associative caudate and the left putamen.

RSD rhythm measures were positively correlated, in females, with connectivity between the right associative putamen and the opercular part of the right inferior frontal gyrus. No significant correlations were observed in males.

Articulation values were positively correlated, in females, with connectivity between the left associative putamen and the left middle temporal gyrus, as well as between the right associative caudate and the right middle temporal gyrus. In males, a negative correlation was observed with connectivity between the left associative caudate and the left middle temporal gyrus.

All correlation analyses presented here were controlled for age and education level when appropriate, thereby reducing additional sources of unexplained variability and spurious correlations. We also repeated the analyses with and without adjustment for the time delay between voice recording and imaging acquisition, and the main findings remained materially unchanged. Given its ability to account for an additional potential source of variability, the adjusted analysis was retained in the present article.

Participants

Subjects were prospectively enrolled as part of the ICEBERG longitudinal study (ClinicalTrials.gov Identifier: NCT02305147). This cohort includes HCs and patients with PD. Inclusion criteria were a clinical diagnosis of idiopathic PD according to the United Kingdom Parkinson’s Disease Society Brain Bank⁵⁹ with less than 4 years of disease duration and minimal or no cognitive disturbances (i.e., a Mini-Mental State Examination score ≥26/30)⁶⁰. HCs had to be free of any neurological disorders. Left-handed subjects and participants with a gap of over 1 year between voice recordings and neuroimaging were excluded from the analyses to minimize variability in hemispheric lateralization and ensure temporal alignment between behavioral and neurofunctional data. Participants with PD were pharmacologically treated and performed the audio recordings and neuroimaging acquisitions during the 6 h following their last dopaminergic medication intake, so under the active effect (at least partial) of their treatment, to avoid the potential distress or motor complications associated with an interruption of their medication and because of logistic constraints. No dyskinesia was observed (during MDS-UPDRS III motor examination) at the time of the visit. Local ethics committee approved this study (IRB Paris VI, RCB: 2014-A00725-42) and all participants gave written informed consent.

A total of 148 participants (one session per subject) were kept for the analyses, comprising 104 PD (68M, 36F) and 44 HC (23M, 21F). Among them, 146 underwent rs-fMRI (103 PD patients [67M, 36F] and 43 HC [22M, 21F]), and 78 completed a DaT-SPECT (48 PD [28M, 20F] and 30 HC [13M, 17F]).

Clinical examination

Clinical examination included the Hoehn and Yahr scale⁶¹, the Movement Disorder Society Unified Parkinson’s Disease Rating Scale (MDS-UPDRS; parts I, II, III in the OFF and ON states, and IV)⁶², the Mattis dementia rating scale (DRS)⁶³, and the Montreal Cognitive Assessment score (MoCA)⁶⁴. All subjects’ clinical scores were measured on the same day as the voice recordings.

Speech data

All participants were recorded in consultation rooms in the clinical investigation center at the Paris Brain Institute (ICM), or in the Pitié-Salpêtrière Hospital sleep disorder unit (Paris, France). The microphone was a professional head-mounted omnidirectional condenser microphone (Beyerdynamic Opus 55 Mk II) located approximately 10 cm from the subject’s mouth and connected to a professional sound card (Scarlett 2i2, Focusrite). Speech data were sampled at a rate of 96,000 Hz with 24-bit resolution, covering a frequency bandwidth of 50–20,000 Hz. Speech recordings were obtained during a single 20-min session per subject, with the assistance of a medical experimenter. The speech tasks were presented in a randomized order via a graphical user interface and included vocalizing sustained vowels, spontaneous speech, rapid and slow syllable repetitions at a specified pace, short sentence repetitions, and reading tasks⁵.

Different quantitative acoustic features were automatically extracted from the speech recordings using PRAAT⁶⁵, MATLAB (The MathWorks, Inc., Natick, MA,USA), and Kaldi⁶⁶ software. These features were related to voice, speech, and language production, including prosody, verbal fluency, rhythm-following ability, phonation, and articulation. Pre-processing steps involving spectral subtraction⁶⁷ were conducted before feature extraction to eliminate additive and stationary background noise. To quantify prosodic impairments, pitch variations (melody) and intensity variations (stress on words or syllables) were analyzed. Pitch variations were measured by the standard deviation of the base-10 logarithm of fundamental frequency (SD log F0), calculated across reading, sentence repetition, and free speech tasks. Intensity fluctuations were evaluated by the standard deviation of the sound intensity level (SD Int) during dialog reading. Mean intensity was excluded from the analysis because of methodological constraints (microphone placement variability, probable drift in microphone sensitivity over time) and environmental factors (hospital ambient noise), which prevented a reliable and interpretable measure. To examine verbal fluency disruptions, the median pause duration (Pause duration) during the free speech task was calculated. To evaluate rhythmic abilities, participants were instructed to repeat specific syllables (/pa/ and /ku/) for 30 s at a steady pace of one syllable per second, following a 5-s example provided beforehand. The ability to maintain a steady rhythm was analyzed as the relative standard deviation of pace (RSD rhythm), and the ability to globally keep an imposed pace was derived from the difference between the average pace and the given pace (Absolute rhythm). To assess phonation impairments, three types of voice measures were extracted using the sustained vowel /a/ tasks: jitter, shimmer, and Noise-to-Harmonics Ratio. These features characterized pitch instability, intensity instability, and voice harmony, respectively. Due to the inherent correlation among jitter, shimmer, and the Noise-to-Harmonics Ratio, a composite phonation feature (Phonation score) was constructed by averaging the standardized values of these features. Further details about the speech features and the signal processing algorithms used for their calculation are provided in a previous article⁵. Finally, to quantify articulatory impairments, we used the scores resulting from classifiers based on the MFCCs, which characterize the spectral envelope (Articulation score). We chose appropriate classification techniques depending on the sex, namely an MFCC-Gaussian Mixture Model for males⁶⁸ and a classifier using x-vectors for females⁷. The speech features used in this study are summarized and described in Table 7.

Neuroimaging

Participants underwent MRI on a Siemens Prisma 3 T system (Siemens Healthcare, Erlangen, Germany) using a 64-channel receive-only head coil. The following images were acquired: whole-brain T1-weighted three-dimensional (3D) magnetization-prepared two rapid gradient echo (MP2RAGE; 1-mm isotropic voxel size), and resting state blood oxygen level-dependent (BOLD) echo planar images (EPI) (44 oblique axial slices, repetition time/echo time/flip angle = 2.19 s/25 ms/90°, voxel size = 2.8 × 2.8 × 2.8 mm³, 220 volumes). During rs-fMRI, the participants kept their eyes closed and had to relax while remaining awake. A field map was acquired to correct for the EPI geometric distortion. More details about the parameter settings can be found in Table S1.

DaT-SPECT imaging (hybrid gamma camera Discovery 670 Pro system; GE Healthcare, Chicago, IL) was performed using the ¹²³I-FP-CIT tracer (DaTScan^TM; 185-200 MBq) (circular orbit, clockwise rotation, zoom factor = 1.5/1.5, matrix size = 128 × 128 voxels, number of slices = 128, voxel size = 2.9 × 2.9 mm², slice thickness = 2.9 mm, rotations = 1, number of frames in rotation = 120, actual frame duration = 30 s, energy window = 143.1–174.9 keV). Participants received thyroid gland blockers via one dose of Lugol regimen and an intravenous injection of ¹²³I-FP-CIT tracer, followed by a SPECT scan performed between 3 and 4 h after the tracer injection.

When not otherwise stated, all pre-processing and analyses were carried out using MATLAB (R2017b, The MathWorks, Natick, MA, USA).

DaT-SPECT images were reconstructed on the GE Healthcare Xeleris Workstation using an iterative algorithm with spatial filtering (fourth-order low-pass filter with a cut-off frequency equal to 3.5 mm⁻¹) and Chang attenuation correction (μ = 0.12 cm⁻¹)⁶⁹. The DaT-SPECT images were rigidly aligned to the corresponding T1-weighted anatomical MRI via the corresponding computed tomography image. All image coregistrations were performed using NiftyReg^70,71.

Resting state fMRI data were preprocessed using Statistical Parametric Mapping (SPM12; http://www.fil.ion.ucl.ac.uk/spm/) with spatial realignment for head motion correction, slice timing, spatial normalization, and smoothing. Nuisance regressors were individually defined using the PhysIO toolbox⁷² in the Translational Algorithms for Psychiatry-Advancing Science (TAPAS) software package⁷³. This toolbox is based on Framewise Displacement (FD) that evaluates abrupt volume-to-volume displacements >0.5 mm, which was inputted into the first-level analyses. The nuisance regressors included the white matter signal, cerebrospinal fluid signal (CSF), motion parameters with their derivatives (24 regressors), and stick functions to exclude volumes with FD > 0.5 mm. Individual anatomical and realigned functional images were normalized to the Montreal Neurological Institute (MNI) space using DARTEL and smoothed with an 8-mm Gaussian filter. Temporal bandpass (0.01–0.1 Hz) filtering was then used to remove low-frequency drifts and physiological high-frequency noise.

Our analyses were focused on the striatum and its relationship to cortical language networks using a region of interest (ROI)-based approach. For striatal regions, the sensorimotor and associative territories of the left and right putamen and caudate nucleus were automatically segmented by aligning a digital template of the human basal ganglia (YeB atlas)⁷⁴ to the anatomical T1-weighted MRI (Supplementary Fig. S1). For cortical regions, two masks were defined: the first included bilateral brain areas belonging to the sensorimotor language network, and the second included bilateral brain areas belonging to the associative language network. Both these networks were defined based on the language-specialized SENtence Supramodal Areas AtlaS (SENSAAS atlas)⁷⁵, the Atlas of Intrinsic Connectivity of Homotopic Areas (AICHA atlas)⁷⁶, and Automated Anatomical Labelling Atlas 3 (AAL3) for the thalamic relays. The associative mask was composed of the following bilateral regions: the pre-SMA, the inferior frontal sulcus, the superior and inferior frontal gyri (including the pars orbitalis and pars triangularis), the precentral gyrus, the supramarginal gyrus, the angular gyrus, the anterior insular gyrus, the superior and middle temporal gyri, and the ventral anterior thalamic nucleus (Fig. S2). The sensorimotor mask was composed of the following bilateral regions: the precentral gyrus, the SMA, the posterior putamen, and the ventrolateral thalamic nucleus (Fig. S3).

For each subject, the DaT-SPECT image was processed according to a previously detailed pipeline^77,78. Briefly, after identification of the striatal ROIs, the following steps were performed: correction for partial volume artifacts, quantitative DaT referencing relative to a region in the occipital cortex with non-specific tracer uptake, and regional calculation of the average striatal binding ratio (SBR).

A seed-based connectivity analysis was performed using SPM12 (first-level analysis) between the eight striatal seeds (two territories: associative, sensorimotor; two striatal regions: caudate, putamen; two hemispheres: left, right) and the corresponding (associative or sensorimotor) language masks. For each participant, general linear modeling of the BOLD signal of each voxel within the language masks was performed using the time course of the corresponding striatal seed (averaged over the voxels in the seed mask) as a regressor, alongside nuisance covariates. For each subject and each striatal seed, a functional connectivity map was obtained. These individual functional connectivity maps, hereafter referred to as “striatal maps,” were transformed into individual z-score maps and included as inputs in the group analysis.

Statistical analysis

Prior to conducting the statistical analyses, the normality of the data distribution was assessed (Shapiro-Wilk test). Based on these results, parametric or non-parametric tests were applied accordingly.

Group analyses were conducted using univariate two-way analyses of variance (ANOVAs), with disease status and sex as factors, and including the interaction between both in the model. As age and education level did not show any significant difference between PD and HC, they were not included as nuisance parameters for the group analyses. For speech analyses, ANOVAs were performed for each speech feature. Regarding DaT-SPECT, ANOVAs were computed from the SBR of the eight previously defined anatomical and functional regions: associative and sensorimotor putamen (left and right) and associative and sensorimotor caudate nucleus (left and right). P values were adjusted with the FDR method to account for multiple comparisons. Differences were considered significant at a corrected p < 0.05. For rs-fMRI statistical analysis, ANOVAs were defined within the SPM12 framework, incorporating the individual striatal maps as input in the model (second-level analysis). Functional connectivity group differences were considered significant at the threshold of p < 0.001 uncorrected at the voxel-based level and p < 0.05 with family-wise error (FWE) correction at the cluster level. Effect sizes for each significant effect were estimated using partial eta-squared (η²), derived from the corresponding F-statistics (from DaT-SPECT regional analyses or at the peak voxel of significant fMRI clusters) and their associated degrees of freedom. For each η², a 95% confidence interval was computed using the noncentral F distribution.

Correlation analyses were performed in PD patients, between the discriminant speech features (according to the two-way ANOVA) and neuroimaging data. Males and females were analyzed separately to examine the neural correlates, potentially sex-dependent, of PD-related speech impairments.

For DaT-SBR, partial Pearson correlations were calculated considering the individual DaT-SPECT SBR values of the four striatal regions (two territories: associative, sensorimotor; two striatal regions: caudate, putamen). The correlations were controlled for age, as it may affect both speech⁷⁹ and SBR values⁸⁰. Control was also applied for the time delay between the voice recording and DaT-SPECT to prevent an additional source of unexplained variability. P values were FDR adjusted to compensate for the multiple comparisons. Differences were considered significant at corrected p < 0.05.

Regarding functional connectivity, multiple regression analyses were conducted for each striatal seed region, with the discriminant speech features as predictors and the individual functional connectivity maps as dependent variables. Both age and education level were included as covariates of no interest to control for their effects, as they may influence both speech⁸¹ and the functional connectivity⁸² in the corresponding neural networks. For the same reasons as with the DaT-SPECT, the correlations were also controlled for the time delay between the voice recording and the fMRI scan. Correlations were considered significant at the threshold of p < 0.001 uncorrected at the voxel level and p < 0.05 with FWE correction at the cluster level.

Effect sizes for each significant effect were estimated using the partial correlation coefficient (r), derived from the corresponding t-statistics (from DaT-SPECT regional analyses or at the peak voxel of significant fMRI clusters) and their associated degrees of freedom. For each r, a 95% confidence interval was computed using Fisher’s z transformation.

We did not include the Levodopa Equivalent Daily Dose (LEDD) as a covariate in our main analyses because of its collinearity with disease severity. By controlling for LEDD, we would risk removing a significant portion of the variance directly related to the disease progression and its underlying pathology.