The architecture of abnormal reward behaviour in dementia: multimodal hedonic phenotypes and brain substrate

Abstract Abnormal reward processing is a hallmark of neurodegenerative diseases, most strikingly in frontotemporal dementia. However, the phenotypic repertoire and neuroanatomical substrates of abnormal reward behaviour in these diseases remain incompletely characterized and poorly understood. Here we addressed these issues in a large, intensively phenotyped patient cohort representing all major syndromes of sporadic frontotemporal dementia and Alzheimer’s disease. We studied 27 patients with behavioural variant frontotemporal dementia, 58 with primary progressive aphasia (22 semantic variant, 24 non-fluent/agrammatic variant and 12 logopenic) and 34 with typical amnestic Alzheimer’s disease, in relation to 42 healthy older individuals. Changes in behavioural responsiveness were assessed for canonical primary rewards (appetite, sweet tooth, sexual activity) and non-primary rewards (music, religion, art, colours), using a semi-structured survey completed by patients’ primary caregivers. Changes in more general socio-emotional behaviours were also recorded. We applied multiple correspondence analysis and k-means clustering to map relationships between hedonic domains and extract core factors defining aberrant hedonic phenotypes. Neuroanatomical associations were assessed using voxel-based morphometry of brain MRI images across the combined patient cohort. Altered (increased and/or decreased) reward responsiveness was exhibited by most patients in the behavioural and semantic variants of frontotemporal dementia and around two-thirds of patients in other dementia groups, significantly (P < 0.05) more frequently than in healthy controls. While food-directed changes were most prevalent across the patient cohort, behavioural changes directed toward non-primary rewards occurred significantly more frequently (P < 0.05) in the behavioural and semantic variants of frontotemporal dementia than in other patient groups. Hedonic behavioural changes across the patient cohort were underpinned by two principal factors: a ‘gating’ factor determining the emergence of altered reward behaviour and a ‘modulatory’ factor determining how that behaviour is directed. These factors were expressed jointly in a set of four core, trans-diagnostic and multimodal hedonic phenotypes: ‘reward-seeking’, ‘reward-restricted’, ‘eating-predominant’ and ‘control-like’—variably represented across the cohort and associated with more pervasive socio-emotional behavioural abnormalities. The principal gating factor was associated (P < 0.05 after correction for multiple voxel-wise comparisons over the whole brain) with a common profile of grey matter atrophy in anterior cingulate, bilateral temporal poles, right middle frontal and fusiform gyri: the cortical circuitry that mediates behavioural salience and semantic and affective appraisal of sensory stimuli. Our findings define a multi-domain phenotypic architecture for aberrant reward behaviours in major dementias, with novel implications for the neurobiological understanding and clinical management of these diseases.

between them -a measure of the quality of representation of the feature by that factor, or how well the feature (presence vs. absence) is discriminated by the factor. Higher squared cosine values correspond to greater discriminatory power. The sum of squared cosine values across all retained factors denotes how well each feature is represented by the retained factors (normalised between 0 and 1, a value closer to 1 signifying that the feature is well represented).
Another useful property of MCA is that it allows additional data not included explicitly in the original analysis to be mapped as "supplementary data" onto the same dimensions as the original model. In other words, MCA allows us to map new data points, whether a new observation or a new feature (here, diagnostic groups), into the common, derived factor space. We can derive the factor values for the supplementary feature by exploiting the information derived from singular vector decomposition. If the original features are represented in the matrix f(COL) x f(ROW), a supplementary feature, c_1, is mapped into the column factor space as: f(COL+ c_1) x f(ROW); the additional information will be f(c_1) x f(ROW)). Such supplementary features do not affect the factor analysis of the original data, however, this process allows them to be assessed (and visualised) in a common space with the original features.

Cluster stability analysis: background
Cluster stability analysis 1 is a form of sensitivity analysis used for evaluating the performance of clustering algorithms. Here, we employed a bootstrapping technique that subsamples a designated proportion (here, 80%) of data from the whole dataset, with replacement after each iteration. The kmeans clustering algorithm was applied on all subsampled data. In each iteration, the similarity of the clustering result on the subsample with the entire original dataset was determined by calculating the mean percentage of participants in each cluster who belonged to the same cluster in the original analysis, across all clusters (see Supplementary Figure 3). A similarity of 100% would signify all participants in each cluster in the subsample were included in the same clusters in the original analysis on the entire dataset. A cluster stability index was derived by averaging the percentage similarity of every subsample over the assigned number of iterations.

Brain imaging acquisition and pre-processing
For each patient, a sagittal 3D magnetization-prepared rapid-gradient echo T1-weighted volumetric brain MRI sequence (TE/TR/TI 2.9/2200/900ms, dimensions 256 × 256 × 208, voxel volume of 1.1 × 1.1 × 1.1mm) was acquired on a Siemens Prisma 3T MRI scanner using a 32-channel phased-array head-coil and pre-processed using standard procedures in SPM12 (www.fil.ion.ucl.ac.uk/spm,details in Supplementary material). Ninety-six volumetric brain MRI scans from the patient cohort (22 bvFTD, 23 AD, 20 nfvPPA, 20 svPPA, 11 lvPPA) were included in the VBM analysis. Twenty-three patients were excluded either because their scan was unavailable or of inadequate quality. Pre-processing of brain images was performed using the New Segment and Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) toolboxes on SPM12, following an optimised protocol.
Normalisation, segmentation, and modulation of grey and white matter images were performed using default parameter settings. Grey matter images were smoothed using a 6 mm full-width-at-halfmaximum Gaussian kernel. A study-specific template brain image was created by warping all biascorrected native space brain images to the final DARTEL template and calculating the average of the warped brain images. Total intracranial volume for each participant was calculated by summing grey matter, white matter and cerebrospinal fluid volumes.

Supplementary Table 1. Neuropsychological and general behavioural characteristics of participant groups
Counts (percentage of group) are shown for behavioural change data; and mean (standard deviation) or median (interquartile range) scores are shown for neuropsychological tests (with maximum scores in parentheses). Differences between diagnostic groups were assessed using ANOVA, Kruskal-Wallis test and chi-square test with post-hoc correction. Significant differences between patient groups and healthy controls are in bold; significant differences between patient groups are coded as follows: 1 significantly different from AD, 2 significantly different from lvPPA, 3 significantly different from bvFTD, 4 significantly different from svPPA, 5 significantly different from nfvPPA (all pFDR < 0.05). AD, patient group with typical Alzheimer's disease; BPVS, British Picture Vocabulary Scale (Dunn, Dunn and Whetton, 1982); bvFTD, patient group with behavioural variant frontotemporal dementia; Controls, healthy control group ; D-KEFS, Delis Kaplan Executive System (Delis et al., 2001); DS, Digit Span; GDA, Graded Difficulty Arithmetic test (Jackson and Warrington, 1986); GNT, Graded Naming Test (McKenna and Warrington, 1983); lvPPA, patient group with   , 1997). A reduced number of participants completed certain tests, as follows: a n-1, b n-2, c n-3, d n-4, e n-5, f n-6, g n-7, h n-8, m n-13.

Supplementary Table 2. Symptom survey used to record changes in reward behaviour
Respondents were patients' primary caregivers or healthy control participants. Prior to completing the survey, caregivers were instructed that a relevant behavioural 'change' in a particular reward domain might comprise an evident alteration in liking, enjoyment and/or interest (e.g., seeking or avoidance of the relevant item) that the caregiver had observed in the person with dementia. This table displays the squared cosine value for each reward feature with the two principal factors (factor 1 and factor 2) identified from the multiple correspondence analysis. The sum of the squared cosines from factor 1 and factor 2 for each feature is shown in the column 'Sum'. Features with a high sum of squared cosine values (sum close to 1) are well-represented by the two principal factors. Dec, decreased; Inc, increased.