Edited by: Changiz Geula, Northwestern University, United States
Reviewed by: David Bergeron, Université de Montréal, Canada; Daniel Llano, University of Illinois at Urbana-Champaign, United States; Hidenao Fukuyama, Kyoto University, Japan
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
Primary progressive aphasia (PPA) is a clinical syndrome characterized by neurodegeneration of language brain systems (Mesulam et al.,
In this regard, some studies have found a large percentage of patients not fulfilling the diagnostic criteria for a specific subtype, especially the logopenic variant (Sajjadi et al.,
Neurodegenerative diseases are determined by a relatively specific predilection of each disease for certain brain regions and networks (Cummings,
We hypothesized that performing clustering analyses of regional brain metabolism could allow an improvement in the classification of PPA patients. Thus, we aimed to study FDG-PET imaging data of a large consecutive case series of patients with PPA using unsupervised clustering algorithms in order to find out the optimal classification groups. We aimed to verify the standard three-groups classification of PPA types and, then, to discover subtype representations of the disease that could derive into different clinical course.
This study involved 150 FDG-PET imaging studies belonging to 91 patients with PPA (31 of them were scanned a second time during the follow-up, making a total of 122 scans) and 28 healthy controls (all of whom were scanned once). Participants were recruited consecutively in our center between November 2011 and May 2017, and they were followed-up until March 2018. Three cases with crossed aphasia (i.e., those patients with predominant right-hemisphere hypometabolism) were excluded. All patients met the current consensus criteria for PPA (Gorno-Tempini et al.,
All participants underwent a detailed neurological and neuropsychological assessment, together with FDG-PET. Language assessment was performed following current recommendations for PPA (Gorno-Tempini et al.,
PET images were acquired following European guidelines (Varrone et al.,
Images were preprocessed using
Clustering analysis (Hennig et al.,
HCA belongs to the special case of overlapping clustering algorithms. These iterative bottom-up classification methods create a sequence of partitions, which satisfy that
The process starts clustering two closer observations and a new cluster is merged in every step. The process builds a tree structure known as
This algorithm does not guarantee finding the optimal solution, but it has demonstrated to provide a good behavior. Moreover, although the computational cost associated with Hierarchical Clustering is higher than partitional Clustering, the dendrogram obtained allows us to explore different partitions, simply by changing the cut-off level as shown in the dendrogram representation. In this way, the problem of not knowing the
Ward's method works in terms of dissimilarities and it is based on the minimum variance method and the Error Sum of Squares (SSE). Ward's method estimates the proximity between clusters through their centroids. It measures the proximity between two clusters according to the increase of the SSE. Ward's method tries to minimize the sum of the squared distances for each point into the cluster, with respect to each cluster's centroid.
Dissimilarities between a cluster with
where
The following analysis was carried out using both
Davies-Bouldin (DB) index measures the similarity of clusters, and how compact a cluster is. DB depends both on the data and the algorithm. A minimum value represents a more compact cluster and therefore the clustering performed has higher homogeneity. DB is computed as
After this set of experiments, the next task was to apply the Ward Linkage method according to the number of potential clusters to explore, and in order to obtain a classification of patients for this number of clusters. This process was carried out using the
FDG-PET images of each obtained cluster were compared to an additional control group of 32 healthy subjects. Prior to statistical analysis, images had been spatially normalized and smoothed at 12 mm full-width at half maximum. Statistical Parametric Mapping version 8 was used for preprocessing and analysis. A two-sample T test was conducted to compare between groups, using age and gender as covariates. Statistical significance was set at
The sample included 122 FDG-PET imaging studies from 91 patients with PPA: 46 with the non-fluent, 15 with semantic, and 61 with logopenic variants. The mean age of the PPA group was 73.48 ± 7.79, and 63 (51.6%) were women. Mean age of onset of symptoms was 70.65 ± 10.67 years. In the PPA group, at the moment of FDG-PET imaging, Mini-Mental State Examination score was 23 [interquartile range 14–27] and Addenbrooke's Cognitive Examination was 51.22 ± 23.4. Mean Functional Activities Questionnaire score was 4 [0–12].
Clustering results are represented in Figure
Distribution of patients within each cluster. X-axis represents the number of clusters, while in the Y-axis we show the number of patients assigned to each cluster. The different colors of the bars indicate the clinical PPA diagnosis for each patient within the cluster (1 non-fluent/agrammatic, 2 semantic, 3 logopenic) and healthy controls.
Distribution of patients per cluster and clinical PPA diagnosis for
Figure
Dendrogram from the classification with
Dendrogram from the classification with
The application of the
Davies Bouldin values for the number of cluster explored (4, 5, 6, 7, and 8). Lower values of this metric mean a better fitting of the sample data to the number of clusters.
The following groups were found when classifying for 4 clusters. The first group k0 included 65 patients, and mainly comprised patients with non-fluent and semantic PPA. The second (k1,
In comparison to healthy controls, k0 showed lower metabolism mainly in the left frontal lobe and the anterior temporal lobe. k1 showed hypometabolism in two main clusters: the first one involving the left supramarginal, superior, middle and inferior temporal gyri and the inferior parietal lobule; and the second one including left middle and inferior frontal gyri and precentral gyrus. k2 showed lower metabolism in a main cluster in the left hemisphere involving the middle, superior, and inferior temporal gyri, as well as fusiform, angular, and parahippocampal gyri. There was an additional cluster in the right temporal lobe and the right angular gyrus (Figure
Clustering in four groups. Voxel-based brain mapping analysis showing regions with lower metabolism in the group k1 (logopenic PPA subtype 1, in red) and the group k2 (logopenic PPA subtype 2, in green) in comparison to healthy controls.
When available, amyloid biomarkers were positive in all cases classified into k1 and k2 clusters (18/18 and 15/15, respectively) and negative in 90% of cases classified into the k0 cluster (1/10 positive).
During follow-up, k0 evolved mainly to progressive supranuclear palsy (
In this analysis, groups mainly involving logopenic PPA and healthy controls remained unchanged. Conversely, the group including non-fluent and semantic variants was divided in two (in the case of 5 clusters classification) and three (in 6 clusters) groups: k0 (
In comparison to healthy controls, k0 showed lower metabolism in the left frontal lobe (superior, middle, medial and inferior frontal gyru, cingulate), insula, caudate and extended also to left inferior parietal lobule and middle temporal gyrus. k2 showed lower metabolism in left frontal lobe (precentral, cingulate, middle, medial, and inferior frontal gyri) and also in the right frontal lobe (medial, middle and superior frontal, and cingulate gyri) (Figure
Clustering in six groups. Voxel-based brain mapping analysis showing regions with lower metabolism in the group k0 (non-fluent PPA subtype 1, in blue) and the group k2 (non-fluent PPA subtype 2, in yellow) in comparison to healthy controls.
Clustering in six or eight groups. Voxel-based brain mapping analysis showing regions with lower metabolism in the group k3 (semantic PPA, in green) in comparison to healthy controls.
Amyloid imaging was negative in all cases in k0 (
During follow-up, patients in the k0 group developed symptoms of progressive supranuclear palsy (
When 7-8 clusters were considered, the former k0 group in 6 clusters was further subdivided into two subgroups (k0,
In comparison to the healthy control group, k0 showed lower metabolism in the left frontal lobe (superior, middle and inferior frontal gyri, cingulate, insula), inferior and temporal gyri, left caudate, left inferior parietal lobule, and left rectal gyrus.
Furthermore, k5 showed lower metabolism in a large cluster involving inferior, middle and superior frontal gyri, anterior cingulate, insula and orbital gyri in the left hemisphere. k5 also showed additional regions of hypometabolism including left inferior parietal lobule and angular gyri, left inferior and middle temporal gyri, and some small clusters in right middle frontal gyrus, left thalamus, and left medial frontal gyrus (Figure
Clustering in eight groups. Voxel-based brain mapping analysis showing regions with lower metabolism in the group k0 (non-fluent PPA subtype 1A, in violet) and k5 (non-fluent PPA subtype 1B, in green) in comparison to healthy controls.
During follow-up, k5 evolved to dementia with (
Flowchart of patient distribution within clusters, taking into account predominant clinical variants according to consensus classification and second clinical syndromes. Non-fluent PPA is shown in blue, semantic PPA in green, logopenic PPA in yellow, and healthy controls in orange. ALS, amyotrophic lateral sclerosis; bvFTD, behavioral variant frontotemporal dementia; CBS, corticobasal syndrome; PPA, primary progressive aphasia; PSP, progressive supranuclear palsy.
Our study addresses an open issue in the field of PPA regarding how patients with PPA should be classified into different subtypes. This is a very relevant question because current classification into three clinical variants aims to predict underlying pathology. Classification of PPA patients should be useful for outcome prediction, and it may be crucial in the near future when disease-modifying therapies are available. Our results confirm the current classification in non-fluent, semantic, and logopenic variants, but also suggest that current categorization may be improved.
The analysis of the distribution of patients among 4 clusters indicates that the HCA method
Classification in 6 clusters divided the former group including several variants of PPA associated with frontotemporal degeneration in three subgroups: k0 (which could be called non-fluent subtype 1), k2 (non-fluent subtype 2), and k3 (which corresponds to the semantic variant). The second syndrome during the follow-up in k2 was more frequently progressive supranuclear palsy, which has been considered very specific for tauopathies 4R (Josephs et al.,
Our study suggests that FDG-PET may classify patients with PPA and, in turn, it could enable the prediction of the clinical course and possible underlying neuropathology. This is especially relevant considering some difficulties, limitations, and reduced availability of other PET tracers for amyloid and tau. In this regard, amyloid imaging may not be specific for AD, especially with aging, or it may be indicative of mixed pathology in cases associated with frontotemporal degeneration (Santos-Santos et al.,
According to Davies-Bouldin index, 6 subtypes of PPA (8 clusters) seem to be the a more optimized classification. We included a group of 31 patients with a second FDG-PET study during the follow-up and, in all cases, both studies were classified in the same cluster. Thus, none cluster seems to be a later stage of a previous one, and this supports the idea that each cluster represents a specific subtype of PPA. To our knowledge, this is the first study using computational analysis performed on the FDG-PET attributes in PPA, which may improve the classification of patients. Some previous studies have used data mining techniques applied to neuropsychological and language performance (Knibb et al.,
Our study has some limitations. First, although clusters are clearly defined, HCA methods, particularly
In conclusion, we found that unsupervised clustering analysis of FDG-PET data favored, based on the Davies-Bouldin index, the classification of PPA into six variants rather than three subtypes as currently recommended in consensus PPA criteria. These subtypes try to go beyond the current categorization in three variants, probably improving the prediction of clinical outcome. In this regard, we have identified three subtypes within non-fluent variant, two subtypes within logopenic PPA, and confirmed the semantic variant. These results also support the usefulness of FDG-PET in evaluating PPA and the possibility to improve the classification of patients with PPA using FDG-PET imaging exclusively. Furthermore, our study suggests the applicability of computational methods for clustering in the analysis of brain metabolism, which could provide new insights in neurodegenerative disorders. Future studies should evaluate clinical and language features, and longitudinal follow-up characteristics of each new subtype.
All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional research committee of the Hospital Clinico San Carlos and with the 1964 Helsinki declaration and its later amendments. Informed consent was obtained from all individual participants included in the study or their caregivers.
JAM-G and JA study concept and design. JM-G and JC study supervision. TM-R and VP literature search. JAM-G, MC-M, TM-R, and VP acquisition of data. JAM-G, JD-Á, JA, and JR interpretation of data. JD-Á, JA, and JR statistical analysis of data. JAM-G, JD-Á, and JA writing the manuscript. MC-M, JM-G, and JC critical revision of the manuscript for important intellectual content.
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
We acknowledge support from Spanish Ministry of Economy and Competitiveness and European Regional Development Fund (FEDER) under project TIN2017-85727-C4-4-P and projects IB16035, of the Regional Government of Extremadura, Department of Commerce and Economy, conceded by the European Regional Development Fund, A way to build Europe.
The Supplementary Material for this article can be found online at: