Longitudinal changes in qualitative aspects of semantic fluency in presymptomatic and prodromal genetic frontotemporal dementia

We included longitudinal data of 118 participants from the FTD Risk Cohort of the Erasmus MC University Medical Center (Rotterdam, the Netherlands). This is an ongoing study in which first-degree family members of FTD patients due to a pathogenic mutation are followed on a 1- or 2-year basis [16]. Participants were recruited between February 2010 and October 2019. DNA genotyping at study entry assigned participants to the mutation carrier (n = 63; MAPT n = 20, GRN n = 43) or non-carrier group (controls; n = 55). Upon study entry, all mutation carriers were presymptomatic according to clinical diagnostic criteria for FTD [2, 17, 18], and had global CDR^®-plus-NACC-FTLD [19] scores of 0. Ten mutation carriers (MAPT n = 6, GRN n = 4) developed symptoms during follow-up (phenoconverters). Diagnoses were made in multidisciplinary consensus meetings, using information from the standardized clinical assessment (see below). Phenoconverters met the following criteria: [i] progressive deterioration of behaviour, language and/or motor functioning; [ii] functional decline, evidenced by multiple study visits with global CDR^®-plus-NACC-FTLD [19] ≥ 0.5 without reversing back to 0; and [iii] cognitive decline, evidenced by ≥ 1.5 SD below age-, sex- and education-specific means in ≥ 1 domain on neuropsychological assessment. Eight phenoconverters had clinical features of bvFTD (MAPT n = 6, GRN n = 2), and two had non-fluent variant PPA (GRN n = 2). The presymptomatic mutation carriers that did not develop FTD symptoms are referred to as non-converters (n = 53).

Every 1–2 years, all participants underwent a standardized clinical assessment, consisting of a structured interview with the participant and a knowledgeable informant (incorporating the CDR^®-plus-NACC-FTLD [19]), medical history taking, neurological examination, neuropsychological assessment, and brain MRI. The neuropsychological assessment consisted of cognitive screening tests (Mini-Mental State Examination, MMSE, and Frontal Assessment Battery, FAB), and tests within the major cognitive domains. Neuropsychiatric symptoms were assessed with the Beck’s Depression Inventory (BDI) and the brief questionnaire form of the Neuropsychiatric Inventory (NPI-Q) (see [7, 8] for full test battery).

The semantic fluency task was part of the neuropsychological assessment at each study visit. In this task, participants were asked to verbally generate as many different animals as possible in 60 s [20]. The total score is the number of animals produced minus errors. Additionally, we calculated five qualitative fluency measures from the output: LF, AoA, number of clusters, cluster size and number of switches (Box 1, Appendix 1 for scoring guidelines).

Phenoconverters, non-converters and controls were compared at five time points: baseline, and follow-up after 2, 4, 5 and 6 years, for the purpose of this study restructured as (Fig. 1):

On each study visit, we performed volumetric T1-weighted MRI scanning on a Philips 3 T Achieva MRI scanner (Philips Medical Systems, Best, the Netherlands) using either an 8- or 32-channel SENSE head coil and the following scan parameters: inversion/repetition time = 933/2200 ms, flip angle = 8°, voxel size = 1.1 × 1.1 × 1.1 mm, matrix size = 256 × 256 × 208, total scan time = 4.43 min. All scans underwent visual quality control. The DICOM images were subsequently corrected for gradient nonlinearity distortions and converted to NifTI format. They were then pre-processed in the Voxel-Based Morphometry (VBM) pipeline in Statistical Parametric Mapping 12 (SPM12; Functional Imaging Laboratory, University College London, London, UK; www.​fil.​ion.​ucl.​ac.​uk/​spm) implemented in Matlab R2018a (Mathworks, USA). First, the T1-weighted images were normalized to a template space and segmented into GM, white matter (WM) and cerebrospinal fluid (CSF), after which they were rigidly aligned. We calculated total intracranial volume (TIV) by adding GM, WM and CSF. Second, the segmentations were spatially normalized to a DARTEL template by applying the flow fields of all the individual scans. Images were smoothed using a 6 mm full width at half maximum (FWHM) isotropic Gaussian kernel. After every preprocessing step, images were visually inspected.

We performed statistical analyses using SPSS Statistics 25.0 (IBM Corp., Armonk, NY). Alpha was set at 0.05 across all comparisons (two-tailed). We compared continuous demographic data between groups using one-way ANOVAs for normally distributed data (with Bonferroni post-hoc tests), or Kruskal–Wallis tests for non-normally distributed data (with Mann–Whitney U post-hoc tests). We analysed between-group differences in sex and gene with Pearson Χ² tests. Interrater reliability was explored using intraclass correlation analysis. We assessed the co-correlation of the five qualitative measures at each time point by means of principal component analysis using Varimax rotation. Only factors accounting for 3% or more of variance and Eigenvalues > 1 were retained. Factor loadings were only considered meaningful when r > 0.450, and any item that did not load sufficiently onto a factor was removed [23]. For ease of interpretation, we first converted raw fluency and relevant neuropsychological measures (Boston Naming Test (BNT), Semantic Association Test (SAT), Trailmaking test B (TMT-B), Stroop colour-word test card III) into z-scores by subtracting the mean of controls from each individual’s raw score at that time point, divided by the SDs of controls at that time point. We then used multilevel linear regression modeling to investigate longitudinal decline in total and qualitative fluency measures. We performed two separate analyses to assess longitudinal change in qualitative fluency measures per [i] clinical status (phenoconverters, non-converters, controls) and [ii] gene (MAPT, GRN). We entered [i] or [ii], time, and first-order interactions, with age, sex, education and total number of words generated as covariates. Assumptions were checked (non-linearity, dependence of errors, outliers, heteroscedasticity). Significant interactions between covariates and dependent variables were included in the model. We based the covariance structure (Toeplitz Heterogeneous) on the lowest Akaike Information Criterion (AIC), as a lower AIC indicates a better model fit. We included the random intercept as this model presented with a lower AIC. For converters, we calculated deltas of the standardized values for the [i] qualitative fluency measures and [ii] relevant neuropsychological tests between restructured time-point 1 in the six phenoconverters that had data 4 years before phenoconversion, or time-point 2 in the four phenoconverters that only had data 2 years before phenoconversion, and time-point 3 (phenoconversion). We then explored the association between the delta fluency measures and delta neuropsychological tests (corrected for age, sex and education) using partial correlations. Change over time maps were generated by subtracting the GM maps calculated at time-point 3 (phenoconversion) from the maps calculated at time-point 1 or 2 in SPM12. We then explored the relationship between the delta qualitative fluency measures and delta GM maps by means of multiple regression models. Age, sex, TIV, and head coil were entered as covariates. We set the statistical threshold at p < 0.05, adjusted for multiple comparisons with familywise error (FWE) correction. Lastly, to investigate classification abilities of these delta z-scores, we performed binary logistic regression analyses. Assumptions were checked (non-linearity, dependence of errors, outliers, multicollinearity). The models were selected with a forward stepwise method according to the likelihood ratio test and applying the standard p-values for variable inclusion (0.05) and exclusion (0.10). Goodness of fit was evaluated with the HL Χ² test, with Nagelkerke R² as measure of effect size. The analyses were adjusted for age, sex, education, and total number of words generated. All models were corrected for multiple comparisons (Bonferroni).

Demographic and clinical data are shown in Table 1. There were no differences between phenoconverters, non-converters, and controls in age [F(2,117) = 0.212, p = 0.809], sex [X(2) = 1.568, p = 0.457], gene [X(2) = 4.830, p = 0.089] or education level [F(2,117) = 1.290, p = 0.279]. There were no differences between MAPT and GRN phenoconverters in age [F(2,9) = 2.966, p = 0.123], sex [X(2) = 0.524, p = 0.262] or education level [F(2,9) = 0.022, p = 0.945]. We found no differences between phenoconverters, non-converters, and controls regarding baseline MMSE [F(2,117) = 0.229, p = 0.796], FAB [F(2,40) = 2.504, p = 0.095], BDI [F(2,116) = 0.607, p = 0.547] or NPI-Q [F(2,73) = 0.031, p = 0.969]. There were no differences between MAPT and GRN phenoconverters in these measures (all p > 0.05). There were no differences in neuropsychological test scores between group at study entry (all p > 0.05).

Two independent raters (LCJ, SAAML), blinded to participant’s clinical and genetic status, scored the number of clusters, cluster sizes, and number of switches (see Methods—Qualitative fluency measures, and Appendix 1). The interrater reliability of all three measures was considered ‘good’ (i.e., intraclass correlation coefficients 0.75–0.90) [25]: number of clusters 0.81 [95% CI 0.76–0.85], cluster sizes 0.85 [95% CI 0.81–0.87], and number of switches 0.85 [95% CI 0.81–0.88].

Appendix 2 shows the results of the principal component analysis on the five qualitative measures per time-point. Based on our criteria and visual inspection of the scree plot, we could extract two components. The first component explained between 39.5 and 49.7% of the total variance, and both LF and AoA had high loadings (r > 0.900). The second factor explained between 29.0 and 43.3% of the total variance, and both the number of clusters and the number of switches, and at some time-points also the cluster size, had high loadings (r > 0.59).

The inflection points and longitudinal trajectories of the fluency measures are shown in Table 2. The individual trajectories in phenoconverters are displayed in Fig. 2. Phenoconverters had declining total scores from at least 4 years pre-phenoconversion (p < 0.001), while there was no decline in non-converters or controls (p > 0.05). As can be noted from Fig. 2, the qualitative measures demonstrated more noise in their fluctuation than the total score. Phenoconverters had higher LF from 4 years pre-phenoconversion in comparison to controls (p = 0.016). Also the number of clusters (p < 0.001) and switches (p = 0.004) started to decline at this point. AoA declined from phenoconversion onwards (p = 0.050). No longitudinal change was found for cluster size (p > 0.05). There was no change in qualitative fluency measures in non-converters or controls (p > 0.05). Different inflection points and longitudinal trajectories were found for GRN and MAPT phenoconverters. At least 4 years pre-phenoconversion, GRN phenoconverters started producing fewer words in comparison to controls (p = 0.005). Moreover, they started producing fewer but larger clusters (both p < 0.001), and used fewer switches (p = 0.004). GRN phenoconverters had larger cluster sizes than MAPT phenoconverters (p < 0.001). LF and AoA did not change in GRN phenoconverters compared to controls (p > 0.05). Starting at least 4 years pre-phenoconversion, MAPT phenoconverters produced fewer words than controls (p < 0.001). Moreover, LF started to increase (p = 0.007), while AoA (p = 0.034), the number of clusters (p = 0.009), and cluster size (p = 0.010) declined. The number of switches did not change (p > 0.05).

Partial correlation coefficients between the fluency measures and relevant neuropsychological tests are shown in Table 3. In phenoconverters, decline on the SAT verbal correlated with an increase in LF (p = 0.031) and a decline in AoA (p = 0.037). Worse performance on TMT-B (p = 0.031) and Stroop-III (p = 0.026) correlated with an increase in cluster size, while worse performance on Stroop-III correlated with a decline in the number of switches (p = 0.031). In GRN phenoconverters, decline on the SAT verbal correlated with a decline in AoA (p = 0.020). Worse performance on Stroop-III correlated with a decline in the number of clusters (p = 0.022), and an increase in cluster size (p = 0.014), while decline on TMT-B correlated with a decline in the number of switches (p = 0.024). In MAPT phenoconverters, decline on the SAT verbal correlated with an increase in LF (p = 0.015) and a decline in AoA (p = 0.014).

The relationships between the qualitative fluency measures and GM volume loss are displayed in Fig. 3 and Appendix 3. In the total group of phenoconverters, worse performance on all five qualitative fluency measures was associated with GM volume loss in large overlapping areas spanning the frontal (e.g., middle frontal gyrus) and temporal lobes (e.g., middle and superior temporal gyrus), and the anterior insular and cingulate cortices (p_{FWE-corrected} < 0.05). In GRN phenoconverters, an increase in LF and a decline in AoA were associated with GM volume loss of the cerebellum and predominantly temporal areas (e.g., medial temporal lobe, inferior temporal gyrus). A decline in the number of clusters, cluster size, and the number of switches correlated with GM volume loss of the cerebellum, insular cortex, and putamen (p_{FWE-corrected} < 0.05). In MAPT phenoconverters, an increase in LF and a decline in AoA were associated with GM volume loss of the cerebellum and predominantly temporal areas (e.g., anterior temporal pole, inferior temporal gyrus). A decline in the number of clusters, cluster size, and the number of switches correlated with GM volume loss of the cerebellum and predominantly frontal areas (e.g., middle and superior frontal gyrus, frontal pole) (p_{FWE-corrected} < 0.05).

Decline in the total score differentiated between phenoconverters and non-converters [X²(1) = 7.669, p = 0.006] and controls [X²(1) = 7.643, p = 0.006]. This was primarily driven by MAPT phenoconverters, as it differentiated well between this group and non-converters [X²(1) = 5.919, p = 0.015] and controls [X²(1) = 5.902, p = 0.015], but not between GRN phenoconverters, non-converters, and controls (p > 0.05). A decline in switches was predictive of phenoconversion in GRN [X²(1) = 5.069, p = 0.024], correctly classifying 90.3% of cases; a decline in cluster size was predictive of phenoconversion in MAPT [X²(1) = 3.894, p = 0.048], correctly classifying 89.3% of cases.

Our pilot study shows that qualitative aspects of semantic fluency change in presymptomatic FTD, and shows different profiles and inflection points depending on the mutation involved. This could provide important insight into the mechanisms as to why the “traditional” total score is declining. Its brief and easy-to-apply nature makes the total score of the semantic fluency test a likely candidate cognitive biomarker for upcoming clinical trials for FTD, but more research with a larger sample of phenoconverters is needed to replicate our findings, and to explore the additional value of qualitative measures in identifying and tracking mutation carriers as they convert to the symptomatic stage.