Characterizing Subtypes and Neural Correlates of Receptive Aprosodia in Acute Right Hemisphere Stroke

Abstract

Introduction:

Speakers naturally produce prosodic variations depending on their emotional state. Receptive prosody has several processing stages. We aimed to conduct lesion-symptom mapping to determine whether damage (core infarct or hypoperfusion) to specific brain areas was associated with receptive aprosodia or with impairment at different processing stages in individuals with acute right hemisphere stroke. We also aimed to determine whether different subtypes of receptive aprosodia exist that are characterized by distinctive behavioral performance patterns.

Methods:

Twenty patients with receptive aprosodia following right hemisphere ischemic stroke were enrolled within five days of stroke; clinical imaging was acquired. Participants completed tests of receptive emotional prosody, and tests of each stage of prosodic processing (Stage 1: acoustic analysis; Stage 2: analyzing abstract representations of acoustic characteristics that convey emotion; Stage 3: semantic processing). Emotional facial recognition was also assessed. LASSO regression was used to identify predictors of performance on each behavioral task. Predictors entered into each model included 14 right hemisphere regions, hypoperfusion in four vascular territories as measured using FLAIR hyperintense vessel ratings, lesion volume, age, and education. A k-medoid cluster analysis was used to identify different subtypes of receptive aprosodia based on performance on the behavioral tasks.

Results:

Impaired receptive emotional prosody and impaired emotional facial expression recognition were both predicted by greater percent damage to the caudate. The k-medoid cluster analysis identified three different subtypes of aprosodia. One group was primarily impaired on Stage 1 processing and primarily had frontotemporal lesions. The second group had a domain-general emotion recognition impairment and maximal lesion overlap in subcortical areas. Finally, the third group was characterized by a Stage 2 processing deficit and had lesion overlap in posterior regions.

Conclusions:

Subcortical structures, particularly the caudate, play an important role in emotional prosody comprehension. Receptive aprosodia can result from impairments at different processing stages.

1.0. Introduction

Emotional prosody refers to the variations in pitch, volume, rate, and rhythm of speech used to convey emotions. Speakers naturally produce prosodic variations depending on their emotional state. For communication to be effective, it is essential for listeners to parse emotional prosody because the meaning of an utterance can vary greatly depending on the speaker’s sentiment. For example, the sentence “I didn’t know you would be here!” spoken with pleasant surprise has a very different meaning than if it is spoken in anger. Impaired emotional prosody is a fairly common consequence of right hemisphere stroke (Ross & Monnot, 2008; Sheppard et al., 2020; Tippett & Ross, 2015) and is even more common than visuospatial neglect (Dara et al., 2014), which is frequently associated with right hemisphere stroke. Even though emotional prosody deficits are more common following right hemisphere stroke, they have received far less attention in the research literature compared to studies of visuospatial deficits. Impaired emotional prosody, termed aprosodia, can have many negative consequences, including reduced relationship satisfaction (Blonder et al., 2012). Moreover, impaired recognition of emotions of others has been rated by caregivers of patients with right hemisphere stroke as the most important residual consequence of stroke (Hillis & Tippett, 2014).

1.1. The Neural Architecture of Receptive Emotional Prosody

1.1.1. Brain Areas Subserving Receptive Emotional Prosody - Functional Neuroimaging Evidence in Unimpaired Populations

Many functional neuroimaging studies investigating receptive emotional prosody (e.g., listening to sentences spoken with emotional prosody and selecting the emotion that matches the tone of voice) in neurologically unimpaired populations show bilateral activation with greater activation in right vs. left temporoparietal regions (Beaucousin et al., 2007; Buchanan et al., 2000; Ethofer et al., 2006, 2012; Grandjean et al., 2005; Mitchell et al., 2003; Seydell-Greenwald et al., 2020; Wiethoff et al., 2008; Wildgruber et al., 2005). In contrast, Kotz and colleagues (2003) found nearly equivalent activation in both hemispheres and found evidence that the basal ganglia played an important role in emotional prosody discrimination. A recent meta-analysis of functional neuroimaging studies in neurotypical populations noted that right posterior superior temporal gyrus (pSTG) was important for both emotional and linguistic prosody, suggesting it does not exclusively process emotional voices (Belyk & Brown, 2014). Right supplementary motor area (SMA) and bilateral inferior frontal gyrus (IFG) pars orbitalis were also found to be critical areas for emotional prosody processing, and the authors propose that IFG pars orbitalis may act as an interface between limbic and sensorimotor networks. A certain amount of bilateral processing is unsurprising given that several studies have demonstrated that each hemisphere preferentially processes different acoustic features of speech, such that the right hemisphere preferentially analyzes pitch (Meyer et al., 2002; Tzourio et al., 1997; VanLancker & Sidtis, 1992; Zatorre & Belin, 2001), and the left preferentially analyzes and integrates temporal information (Meyer et al., 2005; Schirmer, 2004; VanLancker & Sidtis, 1992).

1.1.2. Brain Areas Subserving Receptive Emotional Prosody - Evidence from Lesions in Individuals with Right Hemisphere Stroke

Even though it does appear that bilateral regions are important for receptive emotional prosody, studies of patients with right hemisphere stroke suggest the right hemisphere plays a critical role as emotional prosody recognition is a common consequence of right hemisphere stroke, but is typically spared following a left hemisphere stroke (Schmitt et al., 1997). Numerous subsequent studies of patients with impaired emotional prosody comprehension following right hemisphere stroke at the subacute or chronic stages of recovery implicate right temporoparietal regions (Baum & Pell, 1997; Darby, 1993; Gorelick & Ross, 1987; Ross & Monnot, 2008). In contrast, Cancelliere and Kertesz (1990) found that the basal ganglia were implicated in impaired receptive emotional prosody. Specifically, they investigated comprehension of emotional prosody and recognition of emotional situations and facial expressions in three groups: patients with right hemisphere stroke, patients with left hemisphere stroke, and controls. Impaired comprehension of emotional prosody was associated most frequently with basal ganglia damage in either hemisphere, and to a lesser degree, anterior temporal lobe and insula were also implicated in both hemispheres. Furthermore, many patients with basal ganglia damage also experienced impaired comprehension of emotional facial expressions and situations. However, this study did not control for lesion volume, and patients with larger strokes (and likely more severe emotional prosody and emotion recognition deficits compared to other patients) may have had lesions that involved the basal ganglia, insula, and anterior temporal lobe.

Most aprosodia studies have investigated patients at the subacute or chronic stage of recovery (Baum & Pell, 1997; Darby, 1993; Hughes et al., 1983; Ross, 1981; Ross & Monnot, 2008) while very few studies have investigated aprosodia at the acute stage (approximately < 10 days after a stroke) (Sheppard et al., 2020; Starkstein et al., 1994; Wright et al., 2018). For lesion-symptom mapping purposes, it is important to consider studies at the acute stage of recovery before functional reorganization of behavioral processes has occurred during spontaneous recovery (Jarso et al., 2013; Marsh & Hillis, 2006). Lesion symptom-mapping and case studies at the acute stage of recovery have also primarily implicated right hemisphere temporoparietal regions as well as right hemisphere basal ganglia (Sheppard et al., 2020; Starkstein et al., 1994; Wright et al., 2018). For example, Starkstein (1994) investigated processing of emotional prosody and facial expressions in a group of patients with right- and left-hemisphere stroke within ten days of stroke. They found that in addition to right temporoparietal and basal ganglia lesions, subcortical atrophy was also predictive of impaired receptive emotional prosody. Furthermore, emotional facial expression comprehension was frequently impaired in those with impaired comprehension of emotional prosody. Sheppard and colleagues (2020) investigated patients with right hemisphere stroke within five days of stroke and found that right posterior superior temporal gyrus was associated with impaired comprehension of overall emotional prosody (collapsed across six different emotions). The amygdala was specifically implicated in the recognition of fear but not any other emotion.

When investigating deficits associated with acute stroke, it is important whenever possible to take into account both lesioned areas (core infarct) as well as surrounding hypoperfusion. Hypoperfused areas can extend well beyond the core infarct and have significant impact on functioning and in some cases is the primary cause of impairments rather than damage in the infarct core (Hillis et al., 2001, 2002b, 2002a). Acute deficits associated with hypoperfusion often resolve with reperfusion of those areas (Hillis et al., 2000, 2006; Hillis & Heidler, 2002).

1.1.3. Neurobiological Models of Emotional Prosody Comprehension

The combination of evidence from functional imaging and lesion studies has led to the development of neurobiological models of receptive emotional prosody, which include stages with bilateral processing as well as some right-lateralized functions (Grandjean, 2020; Schirmer & Kotz, 2006). For example, Schirmer and Kotz (2006) described a three-stage mode: 1) Stage 1 (sensory processing) bilateral auditory processing of speech; 2) Stage 2 (integration) encoding emotionally meaningful acoustic cues along the auditory “what” stream originating in superior temporal gyrus and projecting to anterior superior temporal sulcus; 3) Stage 3 (cognition) evaluative judgments and higher-order cognitive processing of emotionally meaningful information in right IFG and orbitofrontal cortex, and integration of emotional prosody with language in left IFG. Additionally, they propose that subcortical structures support receptive emotional prosody but do not discuss their role in detail. Similarly, Brück and colleagues (2011) proposed a neurobiological model of multimodal emotion information processing. This model proposes that emotional prosody is first processed in the brainstem and thalamus. The auditory signals containing prosodic information are relayed through the thalamus to structures that are involved in either explicit or implicit processing. The explicit pathway (middle STG, posterior STG, bilateral dorsolateral prefrontal cortex and orbitofrontal cortex), involves conscious evaluation of emotional prosody. In contrast, an emotional response is induced via the automatic unconscious implicit pathway, which projects from the thalamus to the amygdala, insula, nucleus accumbens, and medial frontal cortex. The basal ganglia are proposed to support basic executive functions that subserve emotional prosody processing.

Recently, Grandjean (2020) described five networks essential for emotional prosody processing. The first is a subcortical network (involving auditory thalamic projections, primary auditory cortex, and the amygdala) for the coarse encoding of emotional auditory stimuli. The second involves temporal cortical areas, including temporal voice areas, which are specific areas in temporal cortex that are particularly sensitive to processing human speech. These temporal areas mediate processing of information fed forward along the ascending auditory pathway, and an integrated concept is formed based on the received auditory information. The third system comprises inferior frontal regions, which categorize auditory emotion information and are involved in feedback information flow. Similarly, the fourth system encompasses orbitofrontal cortex with connections to subcortical structures and temporal voice areas. The fourth system integrates and evaluates contextual information derived from specific sound objects and is important for feedback information flow. Finally, the fifth system includes subcortical nuclei, the basal ganglia, and the cerebellum, which aid in the organization of sound processing. According to this model, the lateralization of each system is proposed to depend on the time scale necessary to represent the given information.

It should also be noted that recently, converging evidence has also supported a right hemisphere dual stream organization for emotional prosody processing analogous to the left hemisphere dual streams for language processing (Sammler et al., 2015; Sheppard et al., 2020; Wright et al., 2018). A great deal of evidence exists supporting the left hemisphere dual stream model of language processing, with a dorsal stream for sound-to-articulation mapping and a ventral stream for sound-to-meaning processing (Fridriksson et al., 2016; Hickok & Poeppel, 2004, 2007; Saur et al., 2008; Scott et al., 2000). The concept of right hemisphere streams for emotional prosody processing developed from past research primarily implicating right hemisphere temporoparietal regions for emotional prosody recognition and right hemisphere inferior frontal regions for emotional prosody generation (Hughes et al., 1983; Ross, 1981; Ross & Monnot, 2008; Schirmer & Kotz, 2006; Sheppard et al., 2020; Starkstein et al., 1994; Wright et al., 2016).

For example, in a study of 23 patients following acute right hemisphere stroke, Sheppard and colleagues (2020) found evidence using multivariable regression that a model including percent damage to ventral stream regions predicted impaired receptive emotional prosody, but a model including dorsal stream regions did not. In a study using both functional imaging and tractography in unimpaired participants, Sammler and colleagues (2015) also found evidence for strong right-lateralization (with a substantially smaller level of left-hemisphere activation) of a dorsal stream for expressive emotional prosody generation and ventral stream for comprehension. In contrast, Seydell-Greenwald (2020) evaluated receptive emotional prosody using functional imaging in neurologically unimpaired individuals and found strong right lateralized activation in frontotemporal areas, with bilateral activation in IFG pars orbitalis, anterior insula, and amygdala. They propose that processing is initially right-lateralized in temporal areas, but once emotions are identified, bilateral integrative and evaluative processes are subsequently activated.

Neurobiological processing models have helped illuminate the specific neural regions essential for processing of emotional prosody. However, our recent work (Wright et al., 2018) suggests that a comprehensive model of receptive emotional prosody should present a more detailed cognitive architecture that includes more explicit information about how specific acoustic features in different emotions are processed and mapped onto emotion representations.

1.2. Proposed Cognitive Architecture of Receptive Emotional Prosody

Ross (1981) proposed that there are eight distinct aprosodic syndromes, similar to the neoclassic aphasia syndromes (e.g., Wernicke’s aphasia, Broca’s aphasia, etc.), that result from right hemisphere stroke. These aprosodias were based on performance of spontaneous prosody and gesturing, repetition, prosodic comprehension, and comprehension of emotional gesturing tasks. The patient with sensory aprosodia had particular difficulty with emotional prosody comprehension and repetition and had a lesion affecting posterior superior temporal and posterior inferior parietal lobe. Similarly, Hughes and colleagues (1983) reported three Mandarin-speaking patients with sensory aprosodia, all of whom had temporoparietal lesions. Ross’s syndromes provide aprosodia profiles but are not informative for understanding why comprehension or production of emotional prosody is impaired following a right hemisphere stroke.

In order to remedy this, we proposed a cognitive architecture of the representations and processes underlying both recognition and generation of emotional prosody (Wright et al., 2018). Figure 1 shows our proposed updated three-stage model of receptive emotional prosody. Stage 1: Acoustic Analysis involves post-perceptual recognition of frequency, intensity, rate, and rhythm information contained within speech signals. During Stage 2: Abstract Representations of Acoustic Characteristics that Convey Emotion (ARACCE), abstract representations of the different acoustic features (e.g. happy: <high pitch> <fast rate>) of each emotion are accessed. Stage 3: Semantic Representation Access entails access and retrieval of the semantic representation of the emotion. Multimodal domain-general emotion knowledge and processing interacts with Stage 2 and Stage 3 processing to subserve emotional prosody recognition. Additionally, feedback information flow occurs between Stages 2 and 3.

Figure 1.

Three-stage model of receptive emotional prosody. A) The three stages of the cognitive architecture model depicting Stage 1 (Acoustic Analysis), Stage 2 (Analysis of ARACCE), and Stage 3 (Semantic Representation Access). Domain-general emotion knowledge & processing interacts with Stages 2 and 3 of receptive emotional prosody. B) Example process of recognizing happy prosody.

In the Wright et al. (2018) study, we reported a case series that included three patients with receptive aprosodia. Aprosodia was assessed in each patient using a battery of tasks equivalent to the battery used in the current study. Two of these patients had acute right hemisphere stroke and were both selectively impaired on Stage 1 processing where they were asked to listen to semantically neutral sentences spoken with emotional prosody and identify prosodic features (e.g., does this sentence have a low pitch or a high pitch?). Lesions in each of these patients affected temporal and insular regions. A third patient with frontotemporal dementia (FTD) also presented with receptive aprosodia and was able to successfully identify prosodic features in sentences. However, he was impaired on Stage 2 processing as evidenced by his performance on a task during which he was asked to match acoustic features associated with a particular emotion (e.g., Which of these features (high pitch, low pitch, etc.) are associated with a sad tone of voice?). This third patient had diffuse atrophy that was greater in the right than left temporal lobe. Wright et al. (2018) demonstrated that patients can be selectively impaired at different stages, yet we still do not know whether specific deficit profiles exist based on patterns of performance at each stage. It is also unclear how likely it is for patients with emotional prosody impairments to have a general emotion recognition deficit that also extends to other modalities, like recognizing emotions in faces.

1.3. Current Study

In the current study we investigated the proposed three-stage model of receptive emotional prosody by investigating processing at each stage within a group of individuals with receptive aprosodia following acute right hemisphere stroke. We administered a receptive emotional prosody task as well as tasks to assess processing at each stage of the proposed three-stage cognitive architecture. We also administered a task investigating emotional facial expression recognition to determine whether patients, or a subset of patients, had a multimodal domain-general emotional comprehension deficit that encompassed both emotional prosody and facial expression recognition. Recall that we have previously demonstrated, in a small group of patients, that right hemisphere damage can lead to selective deficits associated with receptive emotional prosody (Wright et al., 2018). In the current study, our first aim was to conduct lesion-symptom mapping to determine whether damage (core infarct or surrounding hypoperfusion) to specific brain areas was associated with receptive aprosodia overall or with impairment at different stages of the three-stage model. We hypothesized that impairment at each processing stage would be predicted by damage to different brain regions along the right hemisphere ventral stream. Second, we aimed to determine whether different subtypes of receptive aprosodia exist. We hypothesized that receptive aprosodia could result from breakdown of any processing stage in the proposed three-stage model. Therefore, we predicted we would find different subtypes of aprosodia that were characterized by distinctive behavioral performance patterns. We used k-medoids clustering to identify patient subgroups that exhibit different subtypes of receptive aprosodia. We also hypothesized each subtype of receptive aprosodia would be associated with distinctive lesion patterns.

2.0. Methods

We report how we determined our sample size, all data exclusions, all inclusion/exclusion criteria, whether inclusion/ exclusion criteria were established prior to data analysis, all manipulations, and all measures for this study. No part of the study procedures or analysis was pre-registered prior to the research being conducted.

2.1. Participants

Every participant provided informed consent or consent was obtained from their legally authorized representative. This study was approved by the Johns Hopkins Institutional Review Board.

2.1.1. Participants with Right Hemisphere Stroke

Twenty patients (10 women, 10 men) with receptive aprosodia following ischemic right hemisphere stroke were enrolled in this study at Johns Hopkins Hospital. Note that a larger sample of 35 participants (18 women, 17 men) with acute ischemic right hemisphere stroke completed the testing battery, but we only enrolled the 20 participants with receptive aprosodia. Receptive aprosodia was defined as patients who had a mean score on our emotion recognition task that was two or more standard deviations below the mean performance of age-matched healthy controls. Patients had a mean age of 64.9 years (SD = 14.6) (range: 28–85 years) and a mean education level of 13.4 years (SD = 2.9). All patients were right-handed and native speakers of English. Each participant had an acute ischemic stroke affecting their right hemisphere that was visualized on diffusion-weighted imaging (DWI). The patients had no history of previous symptomatic stroke or other neurological disease affecting the brain and had normal or corrected-to-normal hearing and vision. Testing was completed within five days of hospital admission.

2.1.2. Healthy Age-matched Controls

A group of seventeen healthy age-matched controls (six women, 11 men) with no history of stroke or other neurological disease were enrolled. They had a mean age of 57.2 years (SD = 16.0) (range: 28–82 years) and a mean education level of 16.4 years (SD = 2.9). Age matching was established statistically with a two-tailed independent samples t-test (t (35) = 1.32, p = 0.20).

2.2. Prosody Testing

Participants were given six behavioral assessments, which were all presented on a laptop. Impaired performance was defined as more than two standard deviations below the mean accuracy of the control group for each task. For assessments where participants could choose from a set of answers, the choices were presented in the center of the computer screen, and the order of choices remained the same for each trial. In cases where the patient had signs of left visuospatial neglect (8/20 participants), the laptop was oriented on the patient’s right side. An example of each of the six assessments is depicted in Figure 2. Assessment 1 tested receptive emotional prosody, and assessments 2–6 evaluated each component of three-stage model of receptive emotional prosody (Figures 1 and 2). Materials are available to access at this link: https://osf.io/95z83/?view_only=3aed7c92348f4bfaa3e8f28332eb7be6.

Figure 2.

Depiction of six assessments included in the receptive prosody battery of tests. The processing stage that each assessment is measuring is also depicted.

Assessment 1- Receptive Emotional Prosody:

First, receptive emotional prosody was assessed by presenting 24 semantically-neutral sentences (e.g., “They went to the house next door”) spoken with emotional prosody. Participants were asked to choose which emotion the speaker was feeling based on their tone of voice from a set of choices (happy, surprised, angry, sad, afraid, or bored). Impairment on this task could indicate impairment at any proposed stage of emotional prosody processing. As we aimed to only include participants with receptive aprosodia in the current study, this was the task used to screen participants for inclusion.

Assessment 2 - Acoustic Discrimination in Tones (Stage 1):

Stage 1 (Acoustic Analysis) was tested using two assessments (Assessments 2 and 3). First, the Acoustic Discrimination in Tones task asked participants to identify specific differences between nine pairs of tones, wherein each pair differed by only one acoustic feature. After listening to each pair of tones, participants indicated how the second tone differed from the first from a list of provided choices: higher pitch, lower pitch, higher volume, lower volume, faster rate, slower rate, shorter duration, or longer duration.

Assessment 3 - Acoustic Discrimination in Speech (Stage 1):

The second assessment of the Acoustic Analysis stage was an Acoustic Discrimination in Speech task. Here patients listened to 24 semantically-neutral sentences, each of which conveyed one of six emotions. After listening to each sentence, they were asked to select a specific number (2 or 3 depending on the emotion) of acoustic features from a set of choices displayed on the screen. The acoustic feature choices varied by emotion. Emotions associated with two acoustic features were presented with two pairs of acoustic features from which to choose, and emotions associated with three acoustic features were presented with three pairs of features from which to choose. For example, if the sentence was spoken with happy prosody, participants would be asked to select the correct answer from each of these two pairs: fast rate or slow rate?; high pitch or low pitch?. If a sentence was spoken with sad prosody they would be asked to select the correct answer from each of these three pairs: fast rate or slow rate?; high pitch or low pitch?; Loud or quiet?. See Table 1 for the specific features associated with each emotion. Note that the sentences were not labeled with the corresponding emotion.

Table 1.

Acoustic Features of Each Emotion

Emotion	Features
Happy	1. Fast rate
	2. High pitch
Sad	1. Slow rate
	2. Low pitch
	3. Quiet
Angry	1. Fast rate
	2. Low pitch
	3. Loud
Surprised	1. Fast rate
	2. Rising Pitch
Afraid	1. Fast rate
	2. Quiet
Bored	1. Slow rate
	2. Flat pitch

Assessment 4 - Matching Features to Emotions (Stage 2):

Stage 2 Access to ARACCE was assessed using a Matching Features to Emotions task during which patients were presented with each of the six emotions, one at a time. For each trial, they were asked to select two or three phrases (depending on the emotion) that best describe how voices sound when speakers express a particular emotion. After being given the target emotion, they were instructed to select features from the following list: high pitch, low pitch, rising pitch, flat pitch, fast rate, slow rate, loud, quiet. Each emotion was associated with a different set of features as listed in Table 1.

Assessment 5 – Emotion Synonyms (Stage 3):

In order to assess Stage 3 (Access to the Semantic Representation of the Emotion), participants were given an Emotion Synonyms task. During this task, participants were presented with an emotion word at the top of the screen (e.g., Surprised) and were asked to choose which of the two words presented below it was most closely related in meaning (e.g., Irritated or Astonished).

Assessment 6 – Emotional Facial Expression Recognition (Visual Nonverbal Emotion Recognition):

In order to assess whether patients had a multimodal emotion recognition deficit that extended beyond receptive aprosodia, we also assessed their ability to recognize emotional facial expressions. Participants were presented with pictures in which people portrayed one of five different emotions (i.e., happy, sad, surprised, disgusted, and angry). The picture was presented along with the five written choices, and participants were instructed to select the emotion word that best matched how the person in the picture was feeling.

2.3. Imaging

Several MRI sequences were acquired for each patient: DWI to visualize acute lesions; fluid attenuated inversion recovery (FLAIR) imaging to visualize and rule out old lesions and to estimate regions and volume of hypoperfusion by evaluating hyperintense vessels; susceptibility weighted imaging (SWI) to rule out hemorrhage; and T2-weighted imaging to evaluate any additional structural abnormalities. Scans were acquired clinically within 24 hours of admission for stroke on several 3.0T Siemens Trio scanners. Scans were acquired using multiple clinical protocols, and the number of slices on DWIs varied from 25–48 slices. The scan resolution ranged from (min-max) (0.599 − 1.799) × (0.599 − 1.799) × (3.675 − 6.500) mm. Technicians blinded to the prosody evaluation results who were extensively trained and supervised by a neurologist identified the presence or absence of tissue dysfunction on DWI images and manually traced stroke lesions slice-by-slice on the DWI trace images using MRIcron software (Rorden et al., 2007). Once tracings were completed, the images were normalized to standard space using SPM12 (Statistical Parametric Mapping; www.fil.ion.ucl.ac.uk/spm/software/spm12/). The normalization transforms were computed for the DWI b0 image to a template that was based on a group of age-matched controls (Rorden et al., 2012), and then these normalization parameters were applied to the lesion tracings. Figure 3 shows the overlay of all patient lesions onto a template image. Specific regions of interest (ROIs) included right hemisphere ventral stream and dorsal stream structures as well as right hemisphere subcortical and limbic regions that have been previously implicated in emotional prosody studies (Cancelliere & Kertesz, 1990; Gorelick & Ross, 1987; Hughes et al., 1983; Ross, 1981; Ross & Monnot, 2008; Sheppard et al., 2020; Starkstein et al., 1994; Wright et al., 2016). The proposed ventral stream projects from STG to anterior temporal lobe and subsequently to frontal regions where evaluative judgments are made, with subcortical structures subserving receptive prosody. These regions included caudate, putamen, globus pallidus, thalamus, amygdala, anterior cingulate cortex, superior temporal gyrus (STG), middle temporal gyrus (MTG), temporal pole, inferior frontal gyrus (IFG; including pars opercularis, triangularis, and orbitalis combined), middle frontal gyrus (MFG), insula, arcuate fasciculus, sagittal stratum, orbitofrontal cortex, supramarginal gyrus, and angular gyrus.

Figure 3.

Lesion overlay map of 20 participants with receptive aprosodia. Map shows the number of participants with damage to each right hemisphere region.

The percent of damage was calculated for each participant in every gray matter anatomical parcel of the Automated Anatomical Label (AAL) (Tzourio-Mazoyer et al., 2002) and for major white matter pathways defined by a tractography-derived white matter tract atlas (Catani & De Schotten, 2008). Additionally, because the white matter atlas does not specifically define sagittal stratum, and sagittal stratum is theorized to be an important region for emotional prosody processing (Davis et al., 2016), percent damage to the sagittal stratum was calculated using the Johns Hopkins University (JHU) atlas (Faria et al., 2012; Mori et al., 2008).

2.4. FLAIR Hyperintense Vessels Ratings

Hypoperfusion is often evaluated clinically using dynamic contrast perfusion weighted imaging (PWI), or CT perfusion (CTP), which require a contrast agent. Yet, in many cases, PWI and CTP cannot be acquired due to renal failure, contrast allergy, or lack of IV access. An alternative to PWI/CTP was recently developed by Reyes and colleagues (2017), where ratings signifying the location and severity of hyperintense vessels visualized on FLAIR (FHV) reflect hypoperfused areas. Separate FHV ratings are assigned to different areas including four areas within the middle cerebral artery (MCA) territory (MCA-Frontal, MCA-Temporal, MCA-Parietal, and MCA-Insular), as well as anterior cerebral artery (ACA) and posterior cerebral artery (PCA) territories. The resulting NIH-FLAIR Hyperintense Vessel (NIH-FHV) score is highly correlated with hypoperfusion volume as measured on PWI (Reyes et al., 2017). The territory is given a score of 0 if no FHV are present, a score of 1 if there are 1–2 FHV in a slice and 1–2 slices have FHV, and a score of 2 if there are 3 or more FHV on 1 slice or 3 or more slices contain FHV. A technician trained by a neurologist (AEH) to score FHV evaluated the FLAIR scans from each participant in each vascular territory. In order to calculate interrater reliability, Cohen’s kappa (0.74) was calculated from 50% of scans in the study sample between the technician trained by the neurologist, and a second technician who was trained by the first technician. For our analyses we used ratings from the technician who was directly trained by the neurologist. Territories that overlapped with ROIs in this study and where at least two patients had FHV were included in LASSO regression models.

2.5. Statistical Analyses

2.5.1. LASSO Regression Analyses

Mean accuracy on each behavioral task was computed for the group of healthy controls and the patient group. To address aim 1 and build predictive models for each behavioral assessment using damage to specific brain regions and demographic variables, we used Least Absolute Shrinkage and Selection Operator (LASSO) regression (Tibshirani, 1996). LASSO allowed us to evaluate whether damage to specific ROIs (lesioned tissue or presence of hypoperfusion) was related to performance on each of the six behavioral assessments. LASSO regression is ideal for use in the current study because it is particularly effective when there is a large number of potential predictors combined with a small sample size (Meinshausen & Yu, 2009). It is also useful for situations where there is high multicollinearity, which is a concern in lesion-symptom mapping studies as it is common for people with damage in one brain region to have damage to neighboring brain regions. LASSO regression performs regularization that shrinks coefficients toward zero. The resulting models are simple sparse models with few coefficients that have maximal prediction capacity. The glmnet package (https://cran.r-project.org/web/packages/glmnet/index.html) (Friedman et al., 2010) as implemented in R software (R Core Team, 2020) was used to conduct leave-one-out cross validation LASSO using 5000 permutations. We included percent damage to ROIs and FHV scores for regions where at least two patients had damage. This resulted in the following 14 right hemisphere ROIs being included in analyses: Caudate, putamen, thalamus, globus pallidus, ACC, STG, MTG, IFG, MFG, insula, arcuate fasciculus, sagittal stratum, supramarginal gyrus, and angular gyrus. FHV scores were included for the right ACA, MCA-temporal, MCA-parietal, and MCA-insula territories. Additional variables entered into the LASSO models included age, education, sex, and overall lesion volume.

2.5.2. k-medoid cluster analyses

Recall that our second aim was to determine whether unique subgroups of patients with different subtypes of receptive aprosodia exist based on impairments within each stage of the proposed cognitive architecture. In order to accomplish this aim, we performed k-medoid clustering analyses using the ‘cluster’ package (Maechler et al., 2019) in R (R Core Team, 2020). The Partitioning Around Medoids (PAM) algorithm was used to determine how patients clustered based on their performance on the emotional prosody recognition assessment as well as the four behavioral assessments associated with different stages of prosody processing. Additionally, we included performance on the facial expression recognition task so cluster analyses would account for patients with a specific prosody emotion recognition deficit versus a multimodal deficit that extended beyond prosody recognition.

The k-medoids algorithm partitions a dataset into subgroups by minimizing the distance between points within a particular cluster and a point at the center of that cluster (i.e., medoid). K-medoids clustering is more robust than k-means clustering as it is less sensitive to noise and outliers (Kaufman & Rousseeuw, 1990). We ran the k-medoids analysis using the Euclidean distance to specify membership within identified clusters. In order to determine the optimal number of clusters, we used the NbClust package (Charrad et al., 2014), which provides 30 indices to determine the best clustering scheme. In cases in which the data contain more than two variables, as was the case in the current study, Principal Component Analysis (PCA) is used to reduce the dimensionality of the data. In order to interpret the dimensions (aka principal components), we conducted an additional PCA retaining only principal components with eigenvalues over 1. The loadings of each of the six behavioral tests onto the resulting dimensions revealed deficit patterns that determined clustering patterns. This allowed for the evaluation of clustering patterns in light of the three-stage model of receptive prosody.

Next, because the resulting three patient clusters each had a relatively small n (≤ 8 patients per cluster), we lacked sufficient power to conduct regression analyses that would allow us to investigate lesion patterns in each cluster. However, we did create and visualize lesion subtraction plots, where we created a lesion overlay map for each group, and from that we subtracted the lesion overlay maps from the other two groups. For example, to calculate the lesion subtraction map for patient cluster 1 we took the lesion overlay from group 1 and subtracted the lesion overlays from groups 2 and 3. This procedure left us with lesioned voxels that were uniquely associated with each patient cluster. When visual inspection of subtraction maps implicated ROIs included in our LASSO regression analyses in one group, independent two-sample t-tests were used to assess whether there was a statistically significant difference in percent damage to that ROI, or overall lesion volume, between that group and the other two groups. Bonferonni correction was used to account for multiple comparisons (corrected alpha = 0.025).

2.5.3. Statement of Data Availability

The conditions of our ethics approval do not permit public archiving of any anonymized individual data or analysis code related to patient information. Readers who would like to access the data should contact the corresponding author, Shannon M. Sheppard (ssheppard@chapman.edu). Access will be granted to named individuals in accordance with ethical procedures governing the reuse of sensitive data. Specifically, requestors must complete a formal data sharing agreement.

3.0. Results

3.1. Behavioral Assessment Results

3.1.1. Healthy age-matched controls

On the emotional prosody comprehension task, controls attained a mean accuracy (standard deviation (SD)) of 78.4% (SD 6.3%). For the assessments investigating Stage 1 (Acoustic Analysis), controls had a mean accuracy of 89.5% (SD 17.3%) on the Acoustic Discrimination in Tones task, and a mean accuracy of 87.7% (SD 12.4%) on the Acoustic Discrimination in Speech task. For the Matching Features to Emotions assessment that mapped onto Stage 2 (Access to ARACCE) processing, controls had a mean accuracy of 72.5% (SD 10.9%). For the Stage 3 (Access to Semantic Representation of Emotions) assessment, the Emotion Synonym task, controls achieved a mean accuracy of 97.8% (SD 3.3%). Finally, on the emotional facial expression recognition task the control group had a mean accuracy of 92.4% (SD 4.4%).

3.1.2. Group of Individuals with Receptive Aprosodia

Individuals with receptive aprosodia by definition all had impaired emotional prosody comprehension with a mean accuracy of 44.5% (SD 11.8%) (chance performance = 16.7%). For Stage 1 (Acoustic Analysis) assessments, the patient group achieved 51.1% (SD 28.7%) (chance performance = 50%) accuracy on the Acoustic Discrimination in Tones task, and 73.9% (SD 15.2%) (chance performance = 50%) on the Acoustic Discrimination in Speech task. For the Stage 2 (Access to ARACCE) Matching Features to Emotions task, individuals with receptive aprosodia had a mean accuracy of 60.8% (SD 11.8%). On the Emotion Synonym task (indexing Stage 3 Access to ARACCE abilities), the patient group achieved 82.9% (SD 13.2%) (chance performance = 50%) accuracy. On the emotional facial expression recognition task, the mean accuracy was 67.6% (SD 13.4%) (chance performance = 20%).

3.2. LASSO Lesion Symptom Mapping Results

We aimed to determine how the proportion of damage to 14 right hemisphere ROIs, hypoperfusion in four vascular territories, age, education, sex, and overall lesion volume predicted performance on six behavioral assessments related to emotional prosody perception. LASSO regression was used to determine the set of variables that best predicted performance on each assessment. LASSO regression results are summarized in Table 2. The final model for receptive emotional prosody included percent damage to caudate and female sex; within this model, percent damage to caudate was the only independent predictor of performance. The final model for Acoustic Discrimination in Tones included age, female sex, percent damage to thalamus, ACC, MFG, and SMG, as well as MCA-parietal hypoperfusion. Older age, greater percent damage to MFG, and greater hypoperfusion in the MCA-parietal territory were independent predictors of difficulty discriminating acoustic differences between tones. The models for the Acoustic Discrimination in Speech, Matching Features to Emotions and Emotion Synonyms task were null. Finally, the model for performance on the Emotional Facial Expression Recognition task included both age and percent damage to the caudate as significant predictors. In other words, older age and greater damage to the caudate predicted poorer performance on recognizing emotional expressions in faces.

Table 2.

LASSO Regression Results in Individuals with Receptive Aprosodia

	Receptive Emotional Prosody		Acoustic Discrimination in Tones		Acoustic Discrimination in Speech		Matching Features to Emotions		Emotion Synonyms		Emotional Facial Expression Recognition
	Adjusted Coefficient	p value	Adjusted Coefficient	p value	Adjusted Coefficient	p value	Adjusted Coefficient	p value	Adjusted Coefficient	p value	Adjusted Coefficient	p value
Model Intercept	2.58 × 10−16		−1.47 × 10−17		-	-	-	-	-	-	−4.17 × 10−17
Age			−5.97 × 10−1	<0.001 *							−0.140	0.017 *
Education
Sex - Female	−9.38 × 10−3	0.054	−0.069	0.074
Sex - Male
Overall Lesion Volume
% Damage to Right Hemisphere:
Caudate	−0.371	0.001 *									−0.130	0.022 *
Putamen
Thalamus			−0.126	0.065
Globus Pallidus
ACC			−0.090	0.09
STG
MTG
IFG
MFG			−0.222	0.032 *
Insula
Arcuate Fasciculus
Sagittal Stratum
SMG			−0.023	0.073
Angular Gyrus
Hypoperfusion for Acute time point (FHV scores):
MCA-temporal
MCA-parietal			−0.171	0.046 *
MCA - insula
ACA

Note.

^*denotes significance at p ≤ 0.05.

3.3. K-medoid Cluster Analysis Results

We hypothesized that receptive aprosodia could result from breakdowns at any stage of the cognitive architecture for emotional prosody comprehension. Therefore, we believe there are different subtypes of receptive aprosodia that vary depending on ability to successfully perform each stage of the proposed cognitive architecture. We conducted a k-medoid cluster analysis to determine whether we could identify subgroups of patients who exhibit different subtypes of receptive aprosodia based on performance on the six behavioral assessments, each of which correspond to a different aspect of emotional prosody recognition (or emotion recognition in general). After visualizing the plot generated by the NbClust package, three patient clusters were found to be most representative of the data. Cluster 1 included eight patients, cluster 2 had six patients, and cluster 3 also had six patients (Figure 4A). Recall that the six assessments load onto dimensions (factors). As depicted in Figure 4A, the first dimension explained 43.9% of the variance and the second dimension explained 21.1% of the variance. Additionally, the third dimension (not depicted in Figure 4A as the figure is 2-dimensional) explained 17.2% of the variance. Factor loadings were extracted, which gives information about how each individual behavioral assessment loaded onto each dimension (Figure 4B). Factor loadings ≤ −0.50 or ≥ 0.50 were considered to reflect high loadings of a given subtest onto a given principal component. Factor loadings indicated that the first dimension was primarily influenced by performance on the receptive emotional prosody task and the emotional facial expression recognition task. The second dimension primarily reflected performance on ARACCE (Stage 2) processing (Matching Features to Emotions), and the Stage 3 Emotion Synonyms task. Finally, the third dimension was primarily influenced by performance on the two Stage 1 tasks - Acoustic Discrimination in Tones and Acoustic Discrimination in Speech.

Figure 4.

K-medoid cluster analysis results. Red loading indicates positive loading and blue indicates negative loading. A) Results of the k-medoids cluster analysis using performance on all six assessments. As a whole, patients in cluster 1 had a milder form of receptive aprosodia that was associated with impaired access to ARACCE (Stage 2 processing). The patient cluster 3 group tended to have a moderate receptive aprosodia with more difficulty with Acoustic Analysis (Stage 1 processing). Overall, patients in cluster 2 had the most severe receptive aprosodia in the current sample and an emotion recognition deficit that extended to facial expressions. Dim = Dimension. B) Factor loadings for Principal Components (PC) 1–3, which correspond to clustering dimensions 1–3 in the k-medoid analysis, for each of the six behavioral subtests.

3.3.1. Characteristics of Patient Clusters

Here we discuss the characteristics that differentiated each of the three patient clusters. By definition, all patients in the current study had impaired ability to recognize emotional prosody, yet we still saw differences in terms of aprosodia severity. As demonstrated in Table 3, patients in cluster 1 were overall less impaired on receptive emotional prosody and all other assessments, with the exception of Matching Features to Emotions. Their relatively poor performance on the Matching Features to Emotions subtest suggests they have particular difficulty with access to and/or processing of ARACCE (Stage 2). Their emotional prosody recognition deficit did not appear to extend to other modalities as their mean accuracy on the facial expression task (78.8%) was relatively high.

Table 3.

Individual patient performance and demographic variables by cluster compared to control group

	Age (Years)	Education (Years)	Sex	Overall Lesion Volume (Voxels)	Receptive Emotional Prosody (% Accuracy)	Acoustic Discrimination in Tones (Stage 1) (% Accuracy)	Acoustic Discrimination in Speech (Stage 1) (% Accuracy)	Matching Features to Emotions (Stage 2) (% Accuracy)	Emotion Synonyms (Stage 3) (% Accuracy)	Emotional Facial Expression Recognition (% Accuracy)
Controls
Mean	57.2	16.4	11 M 6 F		78.4%	89.5%	87.7%	72.5%	97.8%	92.4%
SD	16.0	2.9	-		6.3%	17.3%	12.4%	10.9%	3.3%	4.4%
Patient Cluster 1:
P1	71	10	F	10,587	29.2%	55.6%	93.8%	41.7%	95.8%	65.0%
P3	37	12	M	1,000	41.7%	66.7%	85.4%	66.7%	79.2%	75.0%
P14	42	9	F	5,792	50.0%	88.9%	64.6%	58.3%	75.0%	80.0%
P6	74	10	M	64,335	54.2%	66.7%	81.3%	58.3%	95.8%	82.5%
P9	64	16	M	2,959	58.3%	88.9%	91.7%	58.3%	91.7%	85.0%
P13	28	12	F	104,669	58.3%	55.6%	81.3%	66.7%	100.0%	77.5%
P16	61	13	M	92	58.3%	88.9%	91.7%	58.3%	91.7%	82.5%
P15	59	16	M	558	62.5%	77.8%	62.5%	58.3%	87.5%	82.5%
Mean	54.5	12.3	-	23,437	51.6%	73.6%	81.5%	58.3%	89.6%	78.8%
SD	16.8	2.7	-	39,167	11.1%	14.5%	12.1%	7.7%	8.6%	6.4%
Patient Cluster 2:
P19	69	14	M	6,964	25.0%	55.6%	66.7%	33.3%	50.0%	42.5%
P18	66	16	F	3,158	33.3%	88.9%	66.7%	58.3%	87.5%	50.0%
P8	75	12	F	16,681	37.5%	22.2%	75.0%	66.7%	70.8%	46.7%
P11	56	12	M	17,070	37.5%	55.6%	81.3%	75.0%	79.2%	60.0%
P2	78	16	F	14,255	41.7%	0.0%	37.5%	41.7%	58.3%	60.0%
P10	73	9	M	36,337	50.0%	22.2%	85.4%	58.3%	66.7%	45.0%
Mean	69.5	13.2	-	15,744	37.5%	40.7%	68.8%	55.6%	68.8%	50.7%
SD	7.9	2.7	-	11,528	8.3%	31.9%	17.1%	15.5%	13.6%	7.6%
Patient Cluster 3:
P4	64	12	F	2,820	20.8%	44.4%	89.6%	66.7%	87.5%	67.5%
P5	85	20	F	9,710	41.7%	0.0%	50.0%	75.0%	87.5%	67.5%
P12	76	16	F	14,544	41.7%	22.2%	62.5%	66.7%	91.7%	67.5%
P20	61	13	M	41,222	41.7%	55.6%	56.3%	58.3%	79.2%	80.0%
P7	80	16	F	60,208	54.2%	22.2%	79.2%	66.7%	91.7%	70.0%
P17	60	13	M	3,412	54.2%	44.4%	75.0%	83.3%	91.7%	65.0%
Mean	71.0	15.0	-	21,986	42.4%	31.5%	68.8%	69.4%	88.2%	69.6%
SD	10.7	3.0	-	23,449	12.2%	20.4%	15.0%	8.6%	4.9%	5.3%

SD = standard deviation; M = male, F = female

The mean accuracy on emotional prosody recognition (37.5%) for patients in cluster 2 was the lowest of all the clusters, and they performed poorly on all other subtests as well, attaining the lowest mean scores on every subtest except for Acoustic Discrimination in Tone. Patients in cluster 2 had a more severe form of receptive aprosodia as well as an impaired ability to recognize emotional facial expressions, indicative of a domain-general, multimodal receptive prosody impairment.

Patients in cluster 3 had a mean accuracy on the emotional prosody recognition task that was between patients in clusters 1 and 2. They had a relatively low accuracy on the Acoustic Discrimination in Tones (Stage 1) task, accompanied by the highest accuracy on the Matching Features to Emotions task (Stage 2). This was the opposite of the pattern seen in patient cluster 1, which included patients who demonstrated good performance on the Acoustic Discrimination in Tones task but poor performance on the Matching Features to Emotions task. Receptive aprosodia in patient cluster 3 appears to be explained primarily by a deficit in Stage 1 (acoustic analysis) processing, with relatively spared ARACCE (Stage 2) processing.

3.3.2. Visual Inspection of Lesion Subtraction Maps Associated with Each Cluster

Visual inspection of the lesion subtraction map for patients in cluster 1 (Figure 5A) reveals that lesions were primarily located in frontotemporal areas including MFG and STG. In contrast, damage in cluster 2 was primarily characterized by subcortical lesions impacting putamen, internal capsule and superior corona radiata (Figure 5B). Finally, patients in cluster 3 showed predominantly posterior lesions affecting middle occipital gyrus, and fusiform gyrus (Figure 5C). Several patients in cluster 3 also had some subcortical damage, particularly the thalamus and putamen. No statistically significant between-group differences were found in percent damage to these ROIs or overall lesion volume using independent two-sample t-tests with Bonferonni correction, likely due to the small number of patients within each cluster (n = 6 or 8).

Figure 5.

Lesion Subtraction Maps for each Patient Cluster. A) Lesion subtraction map for 8 patients in Cluster 1 with maximal overlap frontotemporal regions. B) Lesion subtraction map for 6 patients in patient cluster 2 with maximal overlap in subcortical regions. C) Lesion subtraction map for 6 patients in patient cluster 3 with maximal overlap in posterior regions as well as in the thalamus and putamen.

4.0. Discussion

We investigated our proposed cognitive architecture of emotional prosody comprehension by investigating specific impairments corresponding to each proposed processing stage. Our three-stage model of receptive emotional prosody (Figure 1) includes the following stages: Stage 1 - Acoustic analysis; Stage 2 – Access to ARACCE; and Stage 3 – Semantic representation access. Stages 2 and 3 also interact with and contribute to domain-general emotion processing abilities. Our first aim was to investigate whether damage to specific regions of interest (lesioned or hypoperfused tissue) was associated with impairment for overall emotional prosody comprehension, impairment at any specific processing stage, or emotional recognition deficits outside the auditory modality (i.e., emotional facial expressions). We hypothesized that behavioral performance on each of these tasks would be predicted by different patterns of brain damage. Our second aim was to determine whether we could identify different patient subgroups with different subtypes of receptive aprosodia using k-medoid cluster analysis based on performance at each stage of processing. We hypothesized we would find patient clusters with selective impairments, and that each patient cluster would be associated with specific lesion patterns.

For our first aim, we conducted lesion symptom mapping analyses using LASSO regression to investigate which combination of factors were most predictive of performance on each behavioral assessment. A larger proportion of damage to the caudate was the only independent predictor of impairment on the emotional prosody recognition task. Next, lesion symptom mapping analyses indicated that older age, greater damage to MFG, and severity of hypoperfusion in parietal regions each independently predicted poorer performance on the Acoustic Discrimination in Tones task (Stage 1). Models were null for the second Stage 1 task, Acoustic Discrimination in Speech, as well as the Stage 2 Matching Features to Emotions and Stage 3 Emotion Synonyms task. Finally, older age and greater caudate damage were independent predictors of impaired Emotional Facial Expression Recognition.

For aim 2, k-medoid cluster analyses identified three subgroups of patients with receptive aprosodia. Patient cluster 1 had the mildest receptive aprosodic deficits of the three groups. They had particular difficulty with Stage 2 processing (Matching Features to Emotions), relative to the other behavioral tasks and presented with primarily frontotemporal lesions. Patient cluster 3 had more severe receptive aprosodic deficits than patients cluster 1, but they achieved higher accuracy on the Matching Features to Emotion subtest than patient cluster 1. Instead, they had the most difficulty with Stage 1 processing (Acoustic Discrimination in Tones). Visual inspection of lesion subtraction maps confirmed patients in cluster 3 had primarily posterior and thalamic lesions. Finally, patient cluster 2 demonstrated the most severe receptive aprosodic deficit, and they had primarily subcortical lesions. However, t-tests did not provide statistical evidence that lesions in these particular ROIs were more or less likely to occur in each patient cluster group. This was not surprising given the small sample in each patient cluster group. Additionally, participants in cluster 2 had difficulty with every behavioral task, including recognizing emotional facial expressions. This pattern indicates they likely have a multimodal emotion recognition deficit reflecting difficulty with domain-general emotional processing and not just emotional prosody.

In the current study, we used a cluster analysis and found groups of patients with acute right hemisphere stroke who had selective deficits at Stage 1 (patient cluster 3), and Stage 2 (patient cluster 1). These findings align with and expand upon our previous work by Wright et al. (2018), who found two patients with a selective Stage 1 processing deficit and one patient with a selective Stage 2 processing deficit. Thus, our findings also demonstrate that there are different levels of emotional prosody processing, and that there are distinct subtypes of receptive aprosodia corresponding to deficits at each stage. Interestingly, even though patients in cluster 2 had the smallest mean overall lesion volume, they had the most impaired prosody. Additionally, patients within cluster 1, who had the least severe emotional prosody comprehension deficit, also had the largest mean lesion volume of the three patient clusters. Overall, patients in cluster 1 had a mean age that was approximately 15 years younger than the patients in clusters 2 and 3. Studies of visuospatial neglect following right hemisphere stroke have found older individuals are much more likely to experience neglect, which has been attributed to the presence of pre-stroke age-related cerebral atrophy (Gottesman et al., 2008; Levine et al., 1986). Perhaps receptive emotional prosody deficits are similar to neglect, wherein older patients with premorbid cerebral atrophy are more likely to experience impairment, regardless of lesion size.

The lesion-symptom mapping analyses in the current study implicated caudate for overall emotional prosody processing, which was likely driven by patients with severe emotional prosody processing deficits in cluster 2 who had significant subcortical damage. Patient cluster 2 exhibited poor performance on every behavioral task, including emotional facial expression recognition. The caudate was also an independent predictor of impaired facial expression recognition. These results are in line with the Cancelliere and Ketesz (1990) findings that many patients with basal ganglia damage experience both impaired emotional prosody and emotional facial expression recognition. Emotional facial expression deficits following subcortical stroke does not appear to be simply attributed to cognitive deficits (Cancelliere & Kertesz, 1990; Cheung et al., 2006; Starkstein et al., 1994; Yip et al., 2004). Our results, however, diverge from Pell and Leonard (2005), who investigated patients with early stage Parkinson’s disease (and thus basal ganglia dysfunction) on emotional prosody versus emotional facial expression recognition. The patients had significant prosody recognition deficits but only appeared to struggle with recognizing facial expressions depicting specific emotions, particularly disgust. Correspondingly, emotion recognition has also been studied in patients with Huntington’s disease, which is associated with atrophy of the caudate and putamen. Patients with Huntington’s disease have been found to have multimodal emotion recognition deficits, particularly for negative emotions such as anger, fear, and disgust, in both facial expressions (Henley et al., 2008; Montagne et al., 2006; Sprengelmeyer et al., 1996, 2006) and prosody (Speedie et al., 1990; Sprengelmeyer et al., 1996). Our finding that the caudate was implicated in receptive emotional prosody and emotional facial expression recognition deficits, combined with prior research in Parkinson’s disease and Huntington’s disease, strongly suggests the caudate is important for general emotion recognition across multiple modalities and not to emotional prosody alone.

Most neurobiological models of emotional prosody recognition acknowledge that subcortical structures, including the caudate, play an important role likely with domain-general processing for executive functions (Brück et al., 2011; Grandjean, 2020; Schirmer & Kotz, 2006) but often do not provide explicit information about their role. Evidence suggests that the basal ganglia play an important role in sequencing and decoding spectrotemporal information, which facilitates higher-order processing in frontal regions (Pell & Leonard, 2003; Stirnimann et al., 2018). Garrido-Vásquez and colleagues (2013) investigated the specific role of basal ganglia in emotional prosody processing by distinguishing between implicit early stage processing and explicit evaluative judgments. They found evidence that patients with right hemisphere striatum damage had an implicit processing deficit but not an explicit processing deficit. They proposed that the right striatum (caudate + putamen + globus pallidus) is involved in early emotional salience detection. Furthermore, several studies demonstrate that the dorsal and ventral basal ganglia are involved in different aspects of emotional prosody processing. The dorsal basal ganglia is proposed to decode temporal variations of acoustic features and anticipate decoding of temporal patterns along with the auditory cortex (Ethofer et al., 2012; Frühholz & Grandjean, 2012; Hass & Herrmann, 2012; Kotz et al., 2009). In contrast, the ventral basal ganglia may be involved in recognition of anger/aggression (Calder et al., 2004), and has been shown to play a role in inducing feelings of pleasure while listening to music (Chanda & Levitin, 2013). The ventral basal ganglia may also be associated with pleasurable feelings while listening to voices (Frühholz et al., 2016).

Many past studies have also implicated temporoparietal regions for receptive emotional prosody (Hughes et al., 1983; Ross, 1981; Sheppard et al., 2020; Starkstein et al., 1994; Wright et al., 2018). Our lesion-symptom mapping analyses did not reveal any evidence that temporal damage was an independent predictor of impairment, which was surprising given our previous findings that damaged to right posterior STG was an independent predictor of receptive aprosodia (Sheppard et al., 2020). The current study included participants with acute receptive aprosodia, whereas Sheppard and colleagues (2020) investigated lesion contributions among all participants with acute right hemisphere lesions who may or may not have demonstrated emotional prosody recognition impairments. It may be the case that individuals with acute right hemisphere damage with and without receptive aprosodia demonstrated posterior STG damage, and as this region has been implicated in the recognition of prosody in general and not just emotional prosody (Belyk & Brown, 2014), selective damage to this location may not be the best predictor of receptive emotional aprosodia presence, or severity.

While we did not find that temporal regions served as independent predictors for performance on any subtest, we did find that FHV in MCA-parietal cortex (as well as greater proportion of damage to MFG and older age) were independent predictors of impairment on the Stage 1 Acoustic Discrimination in Tones task. Unfortunately, with the FHV approach, we were not able to localize parietal hypoperfusion to the temporal-parietal interface, which would be theorized to be part of the ventral stream for emotional prosody comprehension, or to more dorsal structures theorized to be important for emotional prosody generation (e.g., angular gyrus, supramarginal gyrus). We may also have seen right MFG (and possibly parietal regions) implicated here because they belong to a fronto-parietal network that subserves several cognitive functions including top-down attentional and working memory processing (Fassbender et al., 2006; Schneiders et al., 2012). Furthermore, past studies have also found that auditory perception, and particularly pitch identification, declines with age even in neurologically unimpaired populations (Clinard et al., 2010), which could account for the poorer performance by older participants on the Acoustic Discrimination in Tones task.

It should also be noted that while we identified patient clusters with Stage 1 and Stage 2 deficits, we did not find a patient cluster who had difficulty with Stage 3 processing (Semantic Representation Access) in the three-stage model of receptive emotional prosody. Wright and colleagues (2018) found similar results with the same subtests that were used in the current study. It is possible that the Emotion Synonyms task we used to probe semantic representation access was not sensitive to Stage 3 deficits. Alternatively, deficits at this stage may be rare or non-existent after unilateral stroke but instead require bilateral damage because the conceptual representations of emotions are bilaterally represented. The only case of Stage 3 impairment to date was reported in a patient with frontotemporal dementia (Gorno-Tempini et al., 2004).

Regardless, the identification of different subtypes of receptive aprosodia has important clinical implications. Forthcoming clinical research can focus on developing formal diagnostic assessments that evaluate each stage of the cognitive architecture for emotional prosody comprehension. Additionally, treatments can be developed that will target specific underlying deficits, which will help patients regain their emotional prosody comprehension skills. In the future, clinicians may be able to select individualized treatment programs based on the receptive aprosodia subtype that directly addresses a patient’s specific deficit locus or loci.

There are several limitations of the current study. First, we have a relatively small number of patients in the study, and, therefore, we were not able to conduct lesion-symptom mapping analyses within each patient cluster. It is necessary for future studies to enroll a larger number of patients to validate the current findings. Second, we did not assess cognitive functioning, which is likely to impact performance on the various assessments. Future work will include cognitive measures of attention, working memory, and other executive functions in analyses to determine how cognitive functioning affects prosodic processing. Third, we only reported acute deficits, but we do not know if patients in each patient cluster will be more or less likely to recover emotional prosody comprehension over time. Future studies will use a longitudinal design to investigate recovery patterns across the different subtypes to determine whether subtype serves as a predictor of emotional prosody comprehension recovery. We also did not have direct imaging to more specifically localize hypoperfusion. While the FHV ratings have been shown to correlate with volume of hypoperfusion within vascular territories, these ratings do not provide the same level of granularity that the damaged ROI metrics do. Additionally, we did not include participants with right hemisphere damage that did not result in receptive aprosodia. It is important that future lesion-symptom mapping studies include participants with right hemisphere damage without aprosodia as it is also important to evaluate whether damage to specific areas is not associated with a behavioral aprosodic deficit. Furthermore, it is possible that other processes are involved in the completion of each subtest that are not wholly captured by our three-stage model. Future research is required to further evaluate and validate our model. Finally, we did not capture measures regarding the impact of emotional prosody deficits on patient’s quality of life and relationships or the impact on their families. It is important for future work to assess how these deficits affect patients’ lives outside of the research laboratory.

5.0. Conclusions

Our results confirm that emotional prosody comprehension is a complex process relying on many different brain regions. Our results also establish the existence of different subtypes of receptive aprosodia. Following a right hemisphere stroke, some patients may be selectively impaired on acoustic analysis (Stage 1) processes, or on the ability to access ARACCE (Stage 2). Alternatively, patients may experience a domain-general emotion recognition deficit that extends beyond recognizing emotional prosody to recognizing emotions in faces as well. Future studies can use a longitudinal design to determine if each subtype of receptive aprosodia is characterized by unique recovery patterns and outcomes. Additionally, the results demonstrate that subcortical structures, and particularly the caudate, play an important role in emotional prosody comprehension. Patients with the most severe receptive aprosodia were more likely to have subcortical lesions encompassing the caudate and were more likely to experience domain-general emotion processing impairments. It is important for clinicians to assess emotional prosody comprehension in patients following a right hemisphere stroke, particularly if they are older and have subcortical damage.