Connected speech markers of amyloid burden in primary progressive aphasia

Abstract

Introduction:

Positron emission tomography (PET) amyloid imaging has become an important part of the diagnostic workup for patients with primary progressive aphasia (PPA) and uncertain underlying pathology. Here, we employ a semi-automated analysis of connected speech (CS) with a twofold objective. First, to determine if quantitative CS features can help select primary progressive aphasia (PPA) patients with a higher probability of a positive PET amyloid imaging result. Second, to examine the relevant group differences from a clinical perspective.

Methods:

117 CS samples from a well-characterised cohort of PPA patients who underwent PET amyloid imaging were collected. Expert consensus established PET amyloid status for each patient, and 40% of the sample was amyloid positive.

Results:

Leave-one-out cross-validation yields 77% classification accuracy (sensitivity: 74%, specificity: 79%).

Discussion:

Our results confirm the potential of CS analysis as a screening tool. Discriminant CS features from lexical, syntactic, pragmatic, and semantic domains are discussed.

1. Introduction

Primary progressive aphasias (PPA) are a group of neurodegenerative clinical syndromes characterized by the insidious and gradual onset of predominantly language-related impairments, of which three clinical variants are recognized: the non-fluent (nfvPPA), semantic (svPPA), and logopenic (lvPPA) variants (Gorno-Tempini et al., 2011). Among all patients who meet the criteria for PPA, up to 55% present with amyloidosis (Villarejo-Galende et al., 2017). AD pathology is observed in the majority of lvPPA cases (Bergeron et al., 2018; Montembeault, Brambati, Gorno-Tempini, & Migliaccio, 2018), a clinical syndrome characterized by profound word retrieval difficulties in spontaneous speech and sentence repetition deficits (Gorno-Tempini et al., 2011; McKhann et al., 2011). The prevalence of AD pathology in lvPPA cases is variable across different expert centers, ranging from 57% to 100% (Bergeron et al., 2018). The detection of amyloidosis in PPA samples represents an important and current issue for treatment and enrollment in clinical trials (Rabinovici et al., 2008; Rogalski et al., 2016).

Amyloid imaging positivity is established in vivo with positron emission tomography (PET) by the standardized uptake value ratios (SUVR) of a radioligand (e.g., Pittsburgh Compound B, Florbetapir, Florbetaben, or Flutemetamol). Appropriate use criteria state that amyloid imaging is indicated for atypical presentations of AD such as lvPPA (Dubois et al., 2010; Johnson et al., 2013) and in vivo amyloid PET measures have shown high correlation with post-mortem histopathological reports (Fink et al., 2020; Santose–Santos et al., 2018; Tan et al., 2017). However, amyloid PET scanning is 1) invasive, 2) remains prohibitively expensive for most care settings and is 3) incompatible with telemedicine. There is thus a pressing need for cost-effective methods to select patients for PET amyloid imaging and novel disease-modifying treatments.

Current research focuses on the non-invasive detection of biomarkers in individuals most likely to benefit from a full clinical and biomarker assessment (Villemagne, Doré, Burnham, Masters, & Rowe, 2018). The potential of automated approaches to connected speech (CS) analysis in the search for early and reliable markers of neurodegeneration has been underlined (Battista, Salvatore, Berlingeri, Cerasa, & Castiglioni, 2020; Beltrami et al., 2018; Clarke, Foltz, & Garrard, 2020; Laske et al., 2015; Myszczynska et al., 2020).

Previous studies on CS in PPA have shown that the three clinical variants of PPA present specific CS profiles expressing their characteristic cognitive and speech impairments (Ash et al., 2013). In patients with lvPPA, CS is marked by false starts, hesitations, filled pauses and repairs, as well as a different rate of verbs, nouns and pronouns, syntactic errors and embeddings compared to svPPA or to nfvPPA groups (Ash et al., 2013; Boschi et al., 2017; Wilson et al., 2010). The fact that lvPPA patients produce different CS features from other PPA variants and that they are disproportionately more likely to be amyloid-positive raises the possibility that CS could be used to screen for amyloid positivity. However, it is still unknown whether such a classification is possible and, if so, which CS feature combination enables accurate screening of PPA patients for amyloidosis without first relying on a syndromic variant classification, a complex diagnostic procedure that requires access to clinical expertise, batteries of cognitive/language tests and complementary instrumental measures that may not be available in all clinical and research settings.

The general hypothesis of the study is that CS features can discriminate PPA with or without amyloidosis and that the most discriminant CS features would be related to the core problem affecting spontaneous speech in these patients, namely word retrieval difficulties, i.e., (Gorno-Tempini et al., 2011; McKhann et al., 2011). To address this issue, we investigated 117 transcripts of picture descriptions by a cohort of PPA patients who also underwent carbon 11-labeled Pittsburgh Compound-B PET (PiB). From the transcripts, we extracted a large set of features covering lexical, syntactic, semantic, and pragmatic aspects of CS. Using the currently growing approach in the field (Clarke et al., 2020), we combined CS features available in advanced natural language processing tools with other selected CS features based on recent clinical evidence on patients with PPA and/or AD pathology (Boschi et al., 2017; Fraser et al., 2014; Kamath, Sutherland, & Chaney, 2020; Slegers, Filiou, Montembeault, & Brambati, 2018). First, we based a classification experiment on this set of features to identify which patients have PET-confirmed amyloid deposits and plaques. Second, we identified the set of discriminant features that differentiate PiB+ from PiB− transcripts at the group level.

2. Methods

2.1. PPA participant database

Data used in the preparation of this article were obtained from the UCSF Memory and Aging Center database. We retrospectively identified all participants who received a diagnosis of PPA, according to current diagnostic criteria (Gorno-Tempini et al., 2011). To be included in the study, participants must have a CS sample from the picture description task from the Western Aphasia Battery (Kertesz, 1982) and carbon 11-labeled Pittsburgh Compound-B PET imaging (PiB). The general exclusion criteria for all participants were as follows: native language other than English, developmental learning disabilities, history of a psychiatric disorder, history of traumatic brain injury, and uncorrected hearing or vision problems.

All participants provided written informed consent, which was approved by the Internal Review Board of the University of California, San Francisco.

2.2. Amyloid imaging: positron emission tomography with Pittsburgh Compound-B (PiB)

Carbon 11-labeled Pittsburgh Compound-B PET (PiB) was performed at Lawrence Berkeley National Laboratory as previously described (Villeneuve et al., 2015). Native space SUVRs were created by normalizing mean images (at 50–70 min post-injection) by mean activity in cerebellum gray matter. Visual reads of native space SUVR images were performed by experienced investigators blinded to clinical data using published criteria (Rabinovici et al., 2011). Visual inspection based on these criteria has been validated previously as reproducible and reliable estimate of increased tracer uptake when compared with quantitative analysis. Every patient was thus assigned an amyloid positive (PiB+) or amyloid negative (PiB−) status.

2.3. Cognitive and sociodemographic data

As part of their assessment, all patients underwent a multi-disciplinary evaluation including a neurological examination, neuropsychological assessment, caregiver interview, screening laboratory tests, and high-resolution structural T1-weighted magnetic resonance imaging (MRI). Neuropsychological assessment included a previously described battery of standardized tests assessing memory, visuospatial abilities, and executive functions, global cognition and functional status (Gorno-Tempini et al., 2004).

2.4. Connected speech sample

A total of 117 CS transcripts were included in the study, including 47 PiB+ and 70 PiB− PPA patients. Table 1 shows the characterization of the sample by amyloid status (PiB).

Table 1 –

Average sociodemographic information and severity of impairment by group. Welch’s t-test with significance level, uncorrected. MMSE: Mini-Mental State Evaluation; CDR: Clinical Dementia Rating scale.

Sociodemographic information and severity	PiB negative N = 70	PiB positive N = 47	Welch’s t-test	p value
Age	65.83	63.71	−1.38	.17
Handedness (% right-handed)	77	77	.07	.97
Gender (% female)	53	60	.69	.49
Education (years)	15.72	16.91	2.05	.04
MMSE	24.03	22.28	−1.38	.17
CDR	.59	0.6	.14	.89
CDR Box Score	2.98	3.33	.73	.46

CS samples were obtained from a total of 113 PPA patients. We included two transcripts for four patients who produced picture descriptions on two separate visits. The present study aims at predicting the amyloid status (PiB). We purposefully bypass clinical diagnosis in our classification experiment. However, for sample characterization, we report that the PPA group included 20 svPPA with predominantly left anterior lobe atrophy (svPPA-l; n = 20), 12 svPPA with predominantly right anterior lobe atrophy (svPPA-r; n = 12), 38 nfvPPA, 43 lvPPA.

CS was elicited through the “picnic scene” picture description task from the Western Aphasia Battery (Kertesz, 1982). Patients were instructed the following: “take a look at this picture, tell me what you see, and try to talk in sentences”. The audio was imported into Audacity (http://audacity.sourceforge.net) and manually transcribed in English orthography. Unintelligible words were removed in preprocessing. Previous reports have shown highly reliable transcriptions using similar guidelines (Wilson et al., 2010).

2.5. CS feature extraction

We contribute to the growing trend in CS research (Clarke et al., 2020) with a hybrid approach similar to Fraser et al. (2014); Garrard, Rentoumi, Gesierich, Miller, and Gorno-Tempini (2014). We first extract a large array of features focusing exclusively on variables that can be computed automatically with natural language processing tools and add features from experimental (Ash et al., 2013; Fraser et al., 2014; Wilson et al., 2010), review (Auclair-Ouellet, 2015; Boschi et al., 2017; Slegers et al., 2018) and meta-analytic (Kamath et al., 2020) studies comparing groups of PPA subtypes to one another and/or amnestic AD to controls. We then perform a priori feature selection. This framework is common in studies deploying AI on language samples of AD patients and ensures that the model includes features describing clinically relevant CS characteristics (cf. de la Fuente Garcia, Ritchie, and Luz (2020) for a systematic review).

Text pre-processing was performed using spaCy (https://spacy.io/) version 2.3.2 and the “en_core_web_lg” model with standard pre-processing steps of tokenization, Named Entity Recognition (NER), Part of Speech tagging (POS) and dependency parsing (DEP). Variables computed with spaCy covered many lexical and syntactic features typical for the analysis of CS in neurodegenerative diseases. We supplemented psycholinguistic, semantic, and pragmatic features that were found to be crucial for the study of patients with possible amyloid pathology (or more generally PPA) and that are not already implemented in standard NLP libraries. In total, 261 features were computed: lexical = 53; syntactic = 157; semantic = 37; pragmatic = 14. Table A1 provides the list of extracted features with their definition and operationalisation. Analysis code is available (https://osf.io/uf3zb/).

2.5.1. Lexical features

Part-of-speech tags (POS) were extracted. We examined accuracy of the tags within a predefined (random seed set at 42) random 10% subset of transcripts. Upon close inspection of the spacy POS-tags, we identified two anomalies that required correction. First, 70% of gerunds (i.e., fishing, sailing, drinking, swimming, flying, reading) were wrongfully tagged as “nouns” and were manually changed to “verbs”. Second, 2.4% of the hesitation markers “uhm” were wrongfully tagged and was manually assigned an “interjection” tag.

Several different measures of lexical diversity (i.e., the ratio of different words relative to the total number of words used) have been reported in the study of CS in neurodegenerative diseases. Here, we extracted three popular metrics, namely Type-token ratio (TTR), Honoré (R) and Brunet (W) as in the study of Bucks, Singh, Cuerden, and Wilcock (2000). Brunet’s W is an inverted scale: a higher Brunet’s W indicates lower lexical diversity.

The absolute number of inflected verbs and their proportion relative to sample length were computed. Multiple psycholinguistic databases were queried for ratings of imageability and familiarity (Scott, Keitel, Becirspahic, Yao, & Sereno, 2019), concreteness (Brysbaert, Warriner, & Kuperman, 2014), age of acquisition (Kuperman, Stadthagen-Gonzalez, & Brysbaert, 2012), frequency (Brysbaert & New, 2009), valence, arousal, and dominance (Warriner, Kuperman, & Brysbaert, 2013). These psycholinguistic properties concern the conceptual and emotional content of individual words as they were subjectively rated by a group of participants in the published norms. Averages of these psycholinguistic variables were computed for 1) all words and for 2) nouns, 3) verbs, 4) adverbs, and 5) adjectives separately.

2.5.2. Syntactic features

Universal dependency¹ relationships (DEP) were extracted and proportions were established relative to both the number of words and the number of sentence roots in the transcript. For example, the word family in the utterance “the family is having a picnic” is assigned by spaCy the part-of-speech tag “NOUN” and the universal dependency “nsubj” (nominal subject), reflecting its syntactic function in the utterance. The number of children, i.e., the direct dependents of a word, which are connected to it by a single arc to the left or to the right in the dependency tree, was averaged. Mean sentence length and “noun chunks” were extracted using spaCy’s sentence and noun_chunk boundaries, respectively.

2.5.3. Semantic features

Information content units (ICU) were computed based on the list proposed by Jensen, Chenery, and Copland (2006). The ICU list includes 36 elements, which denote subjects, objects, places, and actions in the picture. ICUs are extracted as 36 separate Boolean features i.e., as they were mentioned (True) or not (False) in the patient’s description. The total number of ICUs mentioned was also tallied.

2.5.4. Pragmatic features

Features for pragmatic elements of speech were inspired by the findings of Garrard et al. (2014); Rentoumi, Raoufian, Ahmed, de Jager, and Garrard (2014) and were counted as follows, where words in brackets denote one instance of the pragmatic event: filled pauses (“uhm”), word-finding difficulties (“know”, “remember”, “unable”), deictic words (“this”, “that”, “here”), and uncertainty (“think”, “look”, “like”, “kind”, “see”, “maybe”, “can”, “something”). We acknowledge that these groupings and the labels “word-finding difficulty” and “uncertainty” are simplifications of diverse clinical phenomena that we postulate to be reflected in transcriptions of CS. These simplifications - or groupings - were a compromise to include relevant words highlighted in previous literature while not inflating the number of individual features by inputting single words as features. We implemented a measure of semantic idea density by computing the average cosine distance between all pairwise combinations of words within a sliding window (Ivensky, 2019). The word embeddings to average cosine distances were extracted from the same spacy “en_core_web_lg” model that supported grammatical dependency and POS tagging. Sliding windows of 3, 10, 20, and 50 words with a step size of half the window length were implemented, outputting four idea density scores per participant.

2.6. Classification and significant group differences

2.6.1. Classification with leave-one-out cross-validation

In total, 261 features were extracted (Table A1, appendix). To reduce dimensionality of the classification model, feature selection is performed within each of the 117 folds of the cross-validation scheme. For our first objective of classification, feature selection operates as follows: 1) perform Welch’s t-tests across groups; 2) correct the associated p values for multiple comparisons by the Benjamini-Hochberg method for False Discovery Rate (FDR) at .05; 3) subset only significant features at p < .05 after correction for FDR; 4) starting with the largest Welch’s coefficient, enter the features in the model in a stepwise manner at the condition that the added feature is not correlated >.75 to a previously entered feature with a higher t statistic. The classifier model is built with the selected features and then used to predict the unseen, “left-out” sample. This procedure is repeated within each fold of the leave-one-out cross-validation. As no dedicated test set is established for final model evaluation, we report performance metrics (accuracy, sensitivity, specificity) as averages of the leave-one-out cross-validation on the whole dataset (117 folds).

Data analysis was performed using Python with scikit-learn (version .21.2) implementations of two classifiers: Logistic Regression with “liblinear” solver and Support Vector Machine (SVM) with linear kernel. The classifiers and kernels were chosen for three reasons. First, using two classifiers demonstrates robustness and stability of classification accuracy (Fraser, Lundholm Fors, Eckerström, Öhman, & Kokkinakis, 2019b), i.e., ensuring that results are not specific to a single classifier. Second, linear kernels yield readily interpretable results (Fraser, Linz, Li, Fors, Rudzicz, & Königet, 2019a). Third, linear kernels tend to perform well when the number of features is relatively high compared to the number of data points (Hsu, Chang, & Lin, 2003), which is arguably the case in this experiment.

2.6.2. Significant and discriminant group differences

For our second objective of discussing group-level differences, we considered the whole sample (117 transcripts) instead of selecting features within each fold (116 transcripts in each of the 117 folds) as in Section 2.6.1. We identify discriminant group differences with a similar procedure: a series of Welch’s tests corrected for multiple comparisons with the Benjamini-Hochberg method for False Discovery Rate (FDR) at .05 and stepwise removal of multicollinearity, i.e., removal of features when correlated >.75 with any feature with a higher t statistic.

3. Results

3.1. Classification of amyloid burden (PiB) from CS features

We sought to benchmark our approach against other biomarkers (cf. Discussion), which show relatively balanced sensitivity and specificity figures (sensitivity: 78%; specificity: 80% in Giannini et al., 2017; sensitivity: 78%; specificity: 77% in Ashton et al., 2019). Initially, our unweighted (100%) classification scheme yielded an unbalanced ratio (sens: 70%; spec: 77%). We thus a posteriori weighted the PiB + class in two increments of 50% (150%, 200%) to skew the sensitivity to specificity ratio toward benchmark figures. We report the results from the 150% weight for the PiB + because it yielded satisfying results (acc: 77%; sens: 74%; spec: 79%) and was stable on both classifiers. SVM returned an average classification accuracy of 77% with a sensitivity of 74% and specificity 79%. As Table 2 shows, Logistic Regression returned similar, but slightly weaker results.

Table 2 –

Averaged classification outcomes for leave-oneout cross-validation.

Classifier	Accuracy	Sensitivity	Specificity
Logistic Regression	.74	.72	.74
SVM	.77	.74	.79

3.2. Group differences: discriminant CS features for PiB + PPA patients

Fig. 1 shows distributions of the discriminant features that differentiate between PiB+ and PiB− groups on the whole dataset. In line with our objectives and methods, we only report and discuss discriminant features that survived the filter for multicollinearity.² The twelve discriminant features form an array that spans multiple linguistic domains: eight lexical features, one syntactic feature, one semantic feature, and two pragmatic features. In the lexical domain, four features were lower in the PiB + group: average imageability of all words, average imageability of nouns, number of nouns per sentence root, and average concreteness of verbs. Conversely, adjective dominance and frequency of verbs and adjectives, as well as Brunet’s W index were higher in the PiB + group. In the syntactic domain, PiB + patients had a lower ratio of the average number of children to the total number of sentence roots. In the semantic domain, PiB + patients were less likely to precisely name the drink being poured by the woman in the picture. In the pragmatic domain, PiB + patients produced more words denoting uncertainty (e.g., “know”, “remember”, “looks like”) relative to their total output in number of words (proportion of uncertainties). The PiB + group also exhibited lower semantic idea density. To ensure that the difference in years of education between the groups does not skew results, post-hoc regression analyses were carried out and showed that all variables retained a significant association to PiB status beyond the effect of education. Fig. 1 and Table 3 summarize these findings.

Fig. 1 –

Significant group differences for discriminant features. All group differences are significant at p < .05 after Benjamini-Hochberg correction for False discovery rate of .05. Most Y-axis labels were omitted to lighten the plot. Imageability, Dominance, Frequency, Concreteness are based on Likert scales from the psycholinguistic databases listed in Section 2.5.1. Proportion of Uncertainties is the ratio of number of uncertainties in the transcript divided by the total number of words. Nouns/number of roots is the ratio of number of nouns divided by the number of sentence roots in the transcript. Number of left children/number of roots is the ratio of number of left children divided by number of sentence roots in the transcripts. Brunet’s W index results from the equation in Table A1 and yields an inverted scale: a higher Brunet’s W index corresponds to lower lexical diversity. Mention of the drink is plotted as the percentage of transcripts in which the drink was mentioned.

Table 3 –

Significant group differences for discriminant features. Corrected p values with Benjamini-Hochberg correction and False discovery rate of .05.

Connected speech feature	PiB− mean	PiB+ mean	Welch’s t-test	p value	corr. p value
Imageability (all word classes)	5.62	5.19	−5.05	.000002	.0004
Imageability (nouns)	6.32	6.04	−4.59	.00002	.001
ICU: drink	.60	.26	−3.95	.0001	.005
Dominance (adjectives)	3.91	5.40	3.66	.0004	.01
Uncertainties (Proportion of)	.02	.04	3.52	.0007	.01
Frequency (verbs)	6.13	6.65	3.43	.0009	.01
Nouns/Number of roots	2.61	1.73	−3.43	.0009	.01
Semantic idea density (window length = 10)	.64	.61	−3.41	.0009	.01
Frequency (adjectives)	4.13	5.34	3.03	.003	.03
Number of left children/number of roots	.06	.04	−3.02	.003	.03
Brunet’s W index	6.30	6.55	2.91	.004	.04
Concreteness (verbs)	3.20	2.99	−2.81	.006	.045

4. Discussion

The present study confirms our hypotheses by providing two main results. First, we showed that a short, non-invasive, and inexpensive sample of connected speech (CS) can identify PiB + patients with 77% accuracy, sensitivity of 74% and specificity of 79%, confirming the potential of CS in detecting the presence of amyloid in the brain of patients with PPA. Although methodological and sampling differences preclude a direct comparison of results with previous research, our analysis compares favorably to other approaches to identification of underlying pathophysiology such as clinical diagnosis and cognitive profiling (sensitivity: 78%; specificity: 80%) (Giannini et al., 2017) and plasma biomarkers (sensitivity: 78%; specificity: 77%) (Ashton et al., 2019).

Second, we demonstrated that a relatively small array of discriminant lexical, semantic, and pragmatic CS features, mostly attributable to word retrieval difficulties, play a central role in determining the CS profile characteristic of PIB + patients. The set of discriminant features indicates that PiB + patients present lower idea density, struggle to produce concrete nouns (imageability, frequency and concreteness), and that the nouns they manage to access are more abstract and are interspersed among many other words (nouns proportion), which tend to be less varied on average (Brunet W index). In addition, PIB + PPA patients present more uncertainties (e.g., think, maybe, looks like, etc.), which likely reflect CS attempts at circumventing word retrieval problems. Thus, this aspect of the laborious speech production of PiB + PPA patients can be captured by analyzing lexical features of CS transcripts as in Garrard et al. (2014); Rentoumi et al. (2014). The groups did not differ in the number of word-finding pauses, filled pauses, or other disruptions to fluency typically associated with word retrieval difficulties (Ash et al., 2013; Wilson et al., 2010). This latter set of features may not be sensitive markers since all PPA patients experience - albeit for different pathophysiological and neuropsychological reasons – a significant degree of pragmatic difficulties and disruptions to fluency. Beyond word retrieval difficulties, these features could reflect semantic processing struggles, as in svPPA, or hesitation in grammatical construction, as in nfPPA. The pragmatic features included in the analysis were crafted from observations garnered in related CS studies of svPPA (Garrard et al., 2014) and multifactorial AD (Rentoumi et al., 2014). These features are more subjective than the remainder of the feature set.

Future studies should include more automatic and fewer task-dependent features describing the characteristics (filled and unfilled) and durations of the pauses. In addition, the inclusion of features related to distortions and phonological paraphasia, which are known to occur in lvPPA, should also be considered to better capture word retrieval difficulties in PIB + PPA (Wilson et al., 2010). The inclusion of more sensitive variables could help improve classification accuracy in future studies, considering that the present experiment achieves 77%, from 60% chance accuracy.

So far, the set of discussed features can be traced to word retrieval difficulties previously described in PIB + PPA patients but can also be observable in patients with semantic deterioration (PiB− svPPA patients). Interestingly, the analysis also revealed the presence of a discriminant syntactic feature that is not shared by svPPA patients, namely a higher ratio of the number of average left “children” to the total number of sentence roots. Short sentences as a marker of PiB + speech could appear surprising considering they do not usually present with agrammatism. However, consistent with the absence of agrammatism, PIB + patients do not produce fewer verbs. In this framework, the lower ratio of the average number of children to the total number of sentence roots suggests that PiB + patients start many sentences with root verbs but that these have few dependents. This simplification of syntax in the absence of agrammatism is likely to reflect verbal working memory deficits that are well-documented in PIB + PPA but also described in AD patients. The pattern of CS features is generally compatible with the neuroanatomical correlates of lexical retrieval and verbal working memory deficits, namely atrophy in the posterior superior and middle temporal gyri and inferior parietal lobule, and damage to the posterior inferior longitudinal fasciculus, anterior inferior longitudinal fasciculus, uncinate fasciculus, and superior longitudinal fasciculus (Montembeault et al., 2018).

In addition, PIB + PPA patients also displayed higher average dominance of adjectives that is driven by the presence of emotional words such as “happy”, “good”, “strong”, “active”, and “nice”. PiB + patients tended to include more vivid adjectives related to characters’ feelings and general atmosphere of the picnic scene. This finding is compatible with the better processing of emotional concepts by left anterior insular regions in lvPPA compared to nfPPA patients (Piguet, Leyton, Gleeson, Hoon, & Hodges, 2015). Finally, PiB + patients were significantly less likely to identify the drink being poured by the woman in the picture. Indeed, 60% of the PiB− group specified what the drink could be, whereas only 26% of the PiB + group did. Patients of the latter group often substituted the information with uncertainty, e.g., “not sure what she is drinking” (not specifying the drink) instead of “she may be pouring coffee or tea” (specifying the drink). A drawback of this marker is its specificity to the picnic picture description task, and in this context, we cannot exclude that its significance is artefactual.

5. Conclusion

For patients presenting with language symptoms suggestive of PPA, a full clinical examination is indicated before scanning per current recommendations on clinical amyloid imaging (Laforce et al., 2016; Witte et al., 2015). When extensive testing is not feasible or delayed, an advantageous approach could be the collection of a short sample of CS. We showed that CS analysis has good sensitivity and specificity to AD pathology.