Beyond group classification: probabilistic differential diagnosis of frontotemporal dementia and Alzheimer’s disease with MRI and CSF biomarkers.

Abstract

Background
Neuroimaging and fluid biomarkers are used in clinics to differentiate frontotemporal dementia (FTD) from Alzheimer’s disease (AD) and other neurodegenerative and non-neurodegenerative disorders. We implemented a machine learning (ML) algorithm that provides individual probabilistic scores for these patients based on magnetic resonance imaging (MRI) and cerebrospinal fluid (CSF) data.

Methods
We used a calibrated classifier with a Support Vector Machine with MRI data. We obtained group classifications and individual probabilities associated with group correspondence. We used the individual probabilities to address the clinical problem of confidence in the diagnosis. We investigated whether combining MRI and CSF levels of Neurofilament light (NfL) and 14-3-3 could improve the diagnosis confidence.

Results
215 AD patients (65 ± 10 years, 137 women), 103 FTD patients (64 ± 8 years, 49 women), and 173 healthy controls (CTR) (59 ± 15 years, 106 women) were studied. With MRI data only, we obtained accuracies of 88% in the AD vs. healthy controls (CTR) classification, 87% for FTD vs. CTR, 82% for AD vs. FTD, and 80% when differentiating the three groups. A total of 74% of FTD and 73% of AD participants have a high (≥ 0.8) probability of accurate diagnosis in the FTD vs. AD comparison. Adding CSF-NfL and 14-3-3 levels slightly improved the accuracy and the number of patients in the high diagnosis confidence group.

Conclusion
We propose a ML algorithm that provides individual diagnostic probabilities, and we validate it using MRI and/or CSF data. Our solution holds promise towards clinical applications as support to clinical findings or in settings with limited access to expert diagnoses.

Full text

Page 1/21
Beyond group classication: probabilistic
differential diagnosis of frontotemporal dementia
and Alzheimer’s disease with MRI and CSF
biomarkers.
Agnès Pérez-Millan
Hospital Clínic de Barcelona
Bertrand Thirion
Inria Saclay - Île-de-France Research Centre
Neus Falgàs
Hospital Clínic de Barcelona
Sergi Borrego-Écija
Hospital Clínic de Barcelona
Beatriz Bosch
Hospital Clínic de Barcelona
Jordi Juncà-Parella
Hospital Clínic de Barcelona
Adrià Tort-Merino
Hospital Clínic de Barcelona
Jordi Sarto
Hospital Clínic de Barcelona
Josep Maria Augé
Hospital Clínic de Barcelona
Anna Antonell
Hospital Clínic de Barcelona
Nuria Bargalló
Hospital Clínic de Barcelona
Mircea Balasa
Hospital Clínic de Barcelona
Albert Lladó
Hospital Clínic de Barcelona
Raquel Sánchez-Valle
Hospital Clínic de Barcelona
Roser Sala-Llonch (  roser.sala@ub.edu )
University of Barcelona

Page 2/21
Research Article
Keywords: Alzheimer's disease, frontotemporal dementia, magnetic resonance imaging, machine learning,
CSF biomarkers, SVM, individual probability
Posted Date: December 2nd, 2023
DOI: https://doi.org/10.21203/rs.3.rs-3627150/v1
License:   This work is licensed under a Creative Commons Attribution 4.0 International License. 
Read Full License
Additional Declarations: No competing interests reported.

Page 3/21
Abstract
Background
Neuroimaging and uid biomarkers are used in clinics to differentiate frontotemporal dementia (FTD)
from Alzheimer’s disease (AD) and other neurodegenerative and non-neurodegenerative disorders. We
implemented a machine learning (ML) algorithm that provides individual probabilistic scores for these
patients based on magnetic resonance imaging (MRI) and cerebrospinal uid (CSF) data.
Methods
We used a calibrated classier with a Support Vector Machine with MRI data. We obtained group
classications and individual probabilities associated with group correspondence. We used the individual
probabilities to address the clinical problem of condence in the diagnosis. We investigated whether
combining MRI and CSF levels of Neurolament light (NfL) and 14-3-3 could improve the diagnosis
condence.
Results
215 AD patients (65 ± 10 years, 137 women), 103 FTD patients (64 ± 8 years, 49 women), and 173 healthy
controls (CTR) (59 ± 15 years, 106 women) were studied. With MRI data only, we obtained accuracies of
88% in the AD vs. healthy controls (CTR) classication, 87% for FTD vs. CTR, 82% for AD vs. FTD, and
80% when differentiating the three groups. A total of 74% of FTD and 73% of AD participants have a high
(≥ 0.8) probability of accurate diagnosis in the FTD vs. AD comparison. Adding CSF-NfL and 14-3-3
levels slightly improved the accuracy and the number of patients in the high diagnosis condence group.
Conclusion
We propose a ML algorithm that provides individual diagnostic probabilities, and we validate it using MRI
and/or CSF data. Our solution holds promise towards clinical applications as support to clinical ndings
or in settings with limited access to expert diagnoses.
1. INTRODUCTION
Alzheimer's Disease (AD) is the most frequent neurodegenerative disorder. Frontotemporal dementia
(FTD) is also frequent in people younger than 65 years old and is the main differential diagnosis with AD
in this age group. AD and FTD are characterized by prototypical clinical features and patterns of
progressive brain atrophy that constitute the disease's ngerprint or signature. An early and accurate
diagnosis is essential for treatment, prognosis, and genetic counseling. However, there is considerable

Page 4/21
individual variability in the clinical features, especially in the early stages. This limits the accuracy of the
clinical diagnosis at this stage.
During the last two decades, uid biomarker studies have substantially improved the diagnosis of
neurodegenerative dementias. The current clinical criteria for AD diagnosis include cerebrospinal uid
(CSF) biomarkers, such as the amyloid-beta protein 42 (Aβ42), the total tau (t-tau), and phosphorylated
tau (p-tau) (1, 2). However, currently, FTD criteria do not include biochemical markers. Neurolament light
chain (NfL) levels, a marker of neuroaxonal damage, and 14-3-3 protein levels, a marker of synaptic-
neuronal loss, have been both proposed as nonspecic neurodegeneration markers that could support the
diagnosis of FTD, although their levels are also increased in AD compared to controls (3–6).
Magnetic Resonance Imaging (MRI) is broadly used in the study of AD and FTD, both at the research and
the clinical levels. Visual evaluation of the atrophy pattern is mainly used in the clinical setting (7, 8).
Quantitative MRI studies have described patterns of cortical thickness and gray matter (GM) volume loss
in AD and FTD at the group level when compared separately with healthy populations (9–14). However,
quantitative MRI studies are only scarcely used in clinics due to technical diculties and limited accuracy
in performing the diagnosis at the individual level.
A growing body of evidence supports the role of machine learning (ML) techniques using brain MRI (15–
17) to support the clinical diagnosis of these two dementias (18–22). Many studies have shown that a
support vector machine (SVM) with neuroimaging data differentiates AD or FTD patients from healthy
controls (22–27). However, fewer studies exist on the differential diagnosis of these two dementias, even
though the clinical symptoms of FTD and AD can display a substantial overlap between them (28–30).
In this study, we aimed to develop a probabilistic computer-aided classication method for FTD and AD,
using MRI data assuming that there will be overlapping and differential brain patterns in these two
neurodegenerative disorders. Then, we addressed the clinical problem of diagnosis condence using
individual prediction probabilities. Finally, we proposed investigating whether combining MRI and CSF
biomarkers could lead to better differentiation of these two dementias and gain more condence in the
diagnosis.
2. MATERIALS AND METHODS
2.1. Participants
We recruited the participants from the Alzheimer's disease and other cognitive disorders unit of the
Hospital Clínic de Barcelona (HCB), Barcelona, Spain. All participants underwent a complete clinical and
cognitive evaluation and a 3T high-resolution structural MRI scan. Participants with a history of stroke,
traumatic brain injury, major psychiatric disorder, or alcohol abuse were excluded.
All AD participants fullled the criteria for mild dementia due to AD (1, 2) supported by the CSF
biomarkers prole suggesting underlying AD neuropathology according to National Institute on

Page 5/21
Aging/Alzheimer's Association Research Framework 2018 (31). The FTD participants fullled the
diagnostic criteria for either behavioral variant frontotemporal dementia (bvFTD) or FTD-related primary
progressive aphasia (PPA) phenotypes, including Semantic Variant Primary Progressive Aphasia (svPPA)
and Nonuent Variant Primary Progressive Aphasia (nfvPPA) (32, 33). All FTD patients included here
showed a CSF prole not suggestive of AD. Healthy adults (CTR) had cognitive performance within the
normative range (cutoff 1.5 SD from the normative mean).
The HCB Ethics Committee approved the study (HCB 2019/0105), and all participants gave written
informed consent.
2.2 Biochemical markers
We used commercially available single-analyte enzyme-linked immunosorbent assay (ELISA) kits to
determine levels of CSF NfL (IBL International, Hamburg, Germany) and CSF 14-3-3 (CircuLex, MBL
International Corporation, Woburn, MA) at the Alzheimer's disease and other cognitive disorders unit
laboratory, Barcelona, Spain.
2.3 MRI acquisition
We acquired a high-resolution 3D structural dataset (T1-weighted, MP-RAGE, repetition time = 2.300 ms,
echo time = 2.98 ms, 240 slices, eld-of-view = 256 mm, voxel size = 1 × 1 × 1 mm) for everyone at each
time point in a 3T Magnetom Trio Tim scanner (Siemens Medical Systems, Germany) upgraded to a 3T
Prisma scanner (Siemens Medical Systems, Germany) during the study.
2.4 MRI processing
We used the processing stream available in FreeSurfer version 6.0
(http://surfer.nmr.mgh.harvard.edu.sire.ub.edu/) to perform cortical reconstruction and volumetric
segmentation of the T1-weighted acquisitions. FreeSurfer allowed us to obtain cortical thickness (CTh)
maps and segment the subcortical structures (34,35). From reconstructed data, we got summary
measures of mean CTh and GM volumes across the left and right hemispheres and summary measures
of mean CTh in 68 cortical regions and GM volumes of 16 subcortical structures, all derived from atlases
available in FreeSurfer (36, 37). The estimated intracranial volume estimated with FreeSurfer was used to
normalize volume measures. All images and individual segmentations were visually inspected and
manually corrected if needed.
2.5 MRI-based individual probabilistic classication
algorithm
We used all CTh values, GM subcortical volumes, and the age of the participants to create our ML
algorithm. We introduced the regional measures of both hemispheres separately, leading to a total of 84
features per subject (see Supplementary Material).

Page 6/21
We rst converted MRI data to z-scores. We implemented a calibrated classier with an SVM as a base
estimator to predict these values. For each classier, we tted a regression that distributes the classier's
output to calibrate the probability between 0 and 1. We created classiers for each pair of diagnostic
groups (AD vs. CTR, FTD vs. CTR, and AD vs. FTD) and across the three groups (AD vs. FTD vs. CTR).
Then, we subdivided the FTD group into bvFTD and PPA, and we used them as independent groups in a
new set (AD vs. bvFTD, AD vs. PPA, CTR vs. bvFTD, CTR vs. PPA, and bvFTD vs. PPA). All the
comparisons were performed with a 5-fold cross-validation to evaluate the performance of the
classication. Then, we analyzed the importance of each region for the decision of the classication
through a permutation feature importance estimation (38) using the test data of each run. The higher the
weight, the larger the importance of the feature in the classication.
We obtained individual probabilities associated with group correspondences as output values for each
test data point given by the calibrated SVM. Notably, they had complementary values (i.e., the probability
of one group is equal to 1 minus the probability of the other in the classication between two
diagnostics), and they were directly associated with the output category (i.e., the nal classication was
the one with probability > 0.5). We conventionally set two levels of diagnosis condence: an individual
probability ≥ 0.8 (or ≤ 0.2) was considered to provide high diagnosis condence, while probabilities
between 0.2 and 0.8 were considered a “gray zone”, with lower or insucient diagnosis condence for the
clinical decision. Thus, we estimated the accuracy and the number of individuals with a high probability
of being from the group for each classication.
Finally, we aimed to explore if NfL and 14-3-3 levels could help diagnose the individuals of the gray zone
of the MRI diagnosis for the following comparisons due to the available data: AD vs. CTR, FTD vs. CTR,
and AD vs. FTD. Thus, we created a reduced dataset with subjects having MRI data, NfL, and 14-3-3
levels. We trained and tested the proposed algorithm in 3 situations: MRI-based algorithm, CSF-based
algorithm, and MRI and CSF-based algorithm to study if the individual probabilities towards the actual
class increased. We did not include Aβ42, t-tau, and p-tau levels to avoid circularity, as these markers were
used in the clinical diagnosis according to current criteria.
We implemented the ML algorithm in Python version 3.10.6 (www.python.org) with the Scikit-learn library
(39).
3. RESULTS
3.1. Sample demographics
The prospective study includes 491 subjects: 215 AD, 103 FTD (56 bvFTD, 24 svPPA, 21 nfvPPA, and 2
PPA), and 173 cognitively normal control (CTR) participants. A subset of the study participants had CSF
measures available: NfL (N = 365) and 14-3-3 (N = 182). Table1 shows demographic information, group
statistics, and biomarker levels. As expected, CSF biomarkers levels showed signicant differences
between groups (corrected p-value < 0.05). There were differences in age and sex. As expected, based on

Page 7/21
previous studies, AD and CTR groups had more women than men; meanwhile, the FTD groups were more
harmonized. Regarding age, healthy controls were younger than AD and FTD participants.
Table 1
Group summaries written as each measure's mean and standard deviation. We calculated differences
between groups using Fisher Test for sex or the Anova Test for the rest of the variables. We highlighted
the signicant group differences in bold. We measured pairwise differences with a Benjamini-Hochberg
correction p-value). CTR: healthy subjects, AD: Alzheimer’s disease, FTD: frontotemporal dementia, NfL:
neurolament light chain.
CTR AD FTD CTR-AD p-
values CTR-
FTD
p-
values
AD-FTD p-
values
N MRI 173 215 103 --- --- ---
Sex at MRI,
Men/Women 67/106 78/137 54/49 0.67 0.049 0.022
Age at MRI,
years (SD)
59.4
(15.0) 65.0 (9.9) 63.7 (8.3) 1.3e-5 0.0045 0.39
N CSF NfL 112 175 78 --- --- ---
CSF NfL,
pg/mL
(SD)
536.1
(312.6) 1134.7
(587.1) 2340.6
(1736.3) 1.2e-07 < 2e-
16 5.9e-06
N CSF 14-3-3 50 68 64 --- --- ---
CSF 14-3-3,
pg/mL
(SD)
2531.9
(748.2) 5727.3
(2303.5) 4234.9
(1869.1) < 2e-16 3.0e-
06 5.9e-06
3.2. MRI-based probabilistic classication algorithm
We estimated the accuracy performance of our algorithm as the mean accuracy obtained in each k-fold
of the test data. We got an accuracy of 88 ± 8% when discriminating AD patients from CTR, and 87 ± 4%
when determining FTD patients from CTR. When we tried classifying AD vs. FTD patients, the accuracy
was 82 ± 6%. Finally, we obtained an accuracy of 77 ± 6% when discriminating between the three groups
(AD vs. FTD vs. CTR) (Table2).

Page 8/21
Table 2
Classication performance of the different approaches and the percentage of
participants with a higher probability of 80% in the diagnosis grouped by
diagnosis.
AD vs CTR FTD vs CTR AD vs FTD
MRI all data
(N = 491)
Accuracy: 87.7%
AD: 73.4%
CTR: 64.5%
Accuracy: 86.9%
FTD: 74.2%
CTR: 73.3%
Accuracy: 81.8%
AD: 73.3%
FTD: 74.2%
MRI reduced data
(N = 178)
Accuracy: 88.5%
AD: 67.2%
CTR: 55.3%
Accuracy: 85.6%
FTD: 68.1%
CTR: 54.3%
Accuracy: 84.6%
AD: 53.2%
FTD: 54.4%
CSF data
(N = 178)
Accuracy: 93.0%
AD: 72.1%
CTR: 71.7%
Accuracy: 86.6%
FTD: 72.1%
CTR: 23.5%
Accuracy: 83.8%
AD: 40.6%
FTD: 45.9%
MRI and CSF data
(N = 178)
Accuracy: 90.3%
AD: 68.1%
CTR: 64.4%
Accuracy: 86.5%
FTD: 70.8%
CTR: 53.2%
Accuracy: 88.5%
AD: 60.7%
FTD: 55.1%
As can be seen in Fig.1, the resulting algorithms were well-calibrated, which allowed us to create
condence ranges in the algorithm classication. The comparison of AD vs. CTR showed that 73% of AD
participants and 65% of CTR participants presented a probability higher than 0.8. In the FTD vs. CTR
comparison, we found 74% FTD participants and 73% CTR participants with a probability ≥ 0.8. Finally,
when discriminating AD vs. FTD, we found 73% AD participants and 74% FTD participants with
probabilities above 0.8 for being classied as AD or FTD, respectively. Figure2 shows the density of the
individual probabilities and how the distribution between the clinical and the algorithm diagnosis is
distributed within the group with an individual probability ≥ 0.8. Notably, the algorithm diagnosis and the
clinical algorithm did not always coincide, also inside the higher probability of 0.8 (high condence).
Then, we aimed to study the FTD clinical subtypes separately. Due to limitations in sample size, we
merged svPPA and nfvPPA in the same group-PPA. We obtained 91 ± 2% accuracy for classifying bvFTD
patients vs. CTR and 93 ± 4% when discriminating PPA patients from CTR. In both cases, the accuracy
increased compared to the accuracy reported for all FTD together (87 ± 4%). Compared with AD, we
obtained 85 ± 3% for the bvFTD vs. AD comparison and 91 ± 3% for the PPA vs AD. Finally, we obtained
an accuracy of 68 ± 6% discriminating bvFTD from PPA.
3.3. Important MRI regions for classication

Page 9/21
Figure 3 shows the region weights associated with each comparison. In summary, when comparing AD
versus CTR, the GM volume of the hippocampus, putamen, and amygdala played the most crucial role.
For FTD vs. CTR, we found that occipital, parietal, and frontal regions emerged as the top regions for the
classication. Finally, when discriminating both dementias (AD vs. FTD), we found a widespread pattern
in which the CTh measures were generally more important than subcortical GM volumes, especially those
in the frontal lobe.
The results of the most crucial regions in the classications considering bvFTD and PPA participants are
shown in Fig.4. When discriminating bvFTD from CTR, the frontal and temporal lobes and the GM
volume of the ventricles were the most important areas. In contrast, when differentiating between CTR
and PPA participants, the top regions were GM volumes of the hippocampus, amygdala, and temporal
lobe. When discriminating AD from bvFTD, the most important areas were the temporal, parietal, and
occipital lobes. The frontal, parietal, and occipital lobes emerged in the PPA vs. AD discrimination. Finally,
when discriminating bvFTD vs. PPA, the regions which contributed the most were the frontal, temporal,
and occipital lobes.
3.4. Individual probabilities using MRI and CSF data
The group classication performance of the algorithm and the percentage of participants with an
individual probability ≥ 80% using MRI-only, CSF-only, and combined MRI and CSF data are presented in
Table2. Adding NfL and 14-3-3 data to the MRI data did not improve the results. However, some
conclusions could be derived from the results.
For comparing AD vs. CTR, adding CSF data to the MRI did not increase the classication rate (Table2,
Fig.5). However, CSF data (NfL and 14-3-3) alone was enough to discriminate between AD and CTR
participants.
Contrarily, in the comparison between FTD and CTR having MRI and CSF, more participants were
classied with high condence, with a probability ≥ 0.8 (see Table2 and Fig.5). The accuracy with only
CSF data and that of MRI + CSF data was very similar, 87 ± 8% and 86 ± 9%, respectively; however, when
adding CSF data, CTR subjects with a CTR-probability higher than 0.8 increased from 23–53%. Thus, in
this case, combining the MRI and CSF data reduced the participants in the gray zone of the diagnosis.
Finally, when we compared AD and FTD participants, combining MRI and CSF data increased both the
accuracy and the number of subjects with a probability ≥ 0.8 (see Table2 and Fig.5).
4. DISCUSSION
In this study, we implemented a machine learning algorithm that discriminates FTD and AD subjects
using data from structural MRI. In addition, our algorithm was able to differentiate subtypes of FTD with
good accuracy. Clinical diagnosis requires decisions at the individual level, and the degree of condence
in the diagnosis is key in managing the patient. We approach the clinical question of diagnosis

Page 10/21
condence using individual probabilities. Among our key results, we found that 74% FTD and 73% AD
participants showed an individual probability ≥ 0.8 of being well-classied by the algorithm in the FTD
vs. AD comparison. Adding CSF neurodegeneration markers (NfL and 14-3-3) levels did not signicantly
improve the diagnosis classication or the number of patients with high individual probability for the
diagnosis.
Previous ML algorithms using structural MRI data have reported accuracies between 76 and 97% for AD
vs. CTR, 72–88% for FTD versus CTR, 51–90% for AD versus FTD, and 54–70% in discriminating between
AD, FTD, and CTR (7, 11, 18, 21, 22, 40–47). These studies used different algorithms, with the SVM being
the most common.
We obtained accuracies that are in accordance with, or even outperformed, previously reported
algorithms, especially for AD vs. FTD (16, 17, 48–50). We have differentiated FTD expressions (bvFTD
and PPA) against AD or CTR, outperforming previously published works (11, 21, 46). First, regarding the
comparisons with CTR, for bvFTD, we obtained a 91% accuracy, and in the case of the PPA participants,
an accuracy of 93%. When classifying bvFTD and PPA separately against AD, we obtained accuracies up
to 90% for both cases. However, when we tried to classify bvFTD vs. PPA, we obtained an accuracy of
68%, which is lower than the accuracy reported by Kim et al. (41), probably due to differences in the
algorithm.
We depicted the patterns that drive accuracy for each classication setting, to obtain a comprehensive
explanation of structural changes in both dementias. The GM volume of the hippocampus, putamen, and
amygdala were essential in differentiating AD from CTR. By contrast, when differentiating FTD from CTR,
the cortical regions were the most important, especially the CTh of occipital, parietal, and frontal regions.
According to this, GM volumes of subcortical areas could help to identify AD patients, and the thickness
of the cortex could be the key to identify FTD participants. This is in agreement with ndings obtained
with more classical analysis methods (12, 51–57). Finally, regarding the FTD variants, frontal brain
regions emerged for the bvFTD, and hippocampus and temporal regions were the most important in PPA,
as reported before (51).
Besides reaching good accuracies, one of the main novelties of our work is that we obtained the
individual probabilities for each diagnosis in all comparisons. Notably, as we built our rst set of
algorithms uniquely with MRI data, these probabilities might reect each individual's brain atrophy
severity. Using these values, we could identify the participants with high diagnosis condence (with a
probability upper to 80%) and those who do not have that high condence that could be a candidate for
further evaluations. Notably, more than 70% of AD and FTD participants were classied with high
diagnosis condence in the FTD vs. AD comparison.
Other studies using multimodal information also reported high classication accuracy combining data
from different imaging modalities or other biological and clinical measures (24, 40, 49, 58). Even so, in
some cases, our scores with only structural MRI data showed better accuracy (18, 46, 59, 60). Here, we
evaluated if adding CSF data to the MRI could improve the accuracy or the number of participants with a

Page 11/21
high diagnosis condence. This could help in reducing the number of clinical tests that usually have to
take these patients. This approach could detect which participants need extra clinical tests and which
ones; another clinical examination would not provide new evidence. In our cohort, adding NfL and 14-3-3
CSF data to the MRI data provided was benecial for the accuracy of the group classication or the
number of participants with high individual diagnosis probability, especially in the comparison where FTD
participants were involved. The fact that both, NfL and 14-3-3 levels and MRI data reect
neurodegeneration and no other aspects of the pathophysiological processes in these diseases could
explain the reduced improvement by adding CSF data. We did not use Aβ42, t-tau, and p-tau to avoid
circularity, as they were used for the AD clinical diagnosis.
Overall, our study has several strengths. First, its good performance makes it suitable for potential
implementation in a clinical setting, especially in doubtful cases or locations with limited access to expert
opinion or additional biomarkers. The key to individual probabilities, thanks to the calibrated algorithm,
also might represent a step toward personalized medicine. For example, it could be used to identify the
best patients to receive the new drugs or for clinical trials. In addition, at the level of explainable ML, we
identied the most critical regions for classication, contributing to the denition of structural atrophy
patterns, and may be used for identifying target regions in further studies.
Our study also presents several limitations. First, it is unicentric. It has the advantage that all the
participants had the same MRI scanner protocol and clinical criteria for the diagnosis. In the case of
using the algorithm in other centers, the increased heterogeneity of the data could worsen the algorithm's
performance. Another limitation regarding the FTD participants is that, when looking at the different
clinical expressions, we reduced the sample size to approximately 50 participants for each group, and
svPPA and nfvPPA had to be studied together. This means that the results are subject to large sampling
variability. Future studies could further explore the subanalyses with the FTD phenotype subtypes in more
detail. Finally, only some participants had NfL and 14-3-3 data available, and the smaller sample size
might have impacted the results.
In conclusion, the proposed diagnosis algorithm has shown high accuracy classication scores with
structural MRI data to discriminate AD, FTD, and CTR. This approach also provided individual MRI-based
classication probability scores as an ancillary tool for studying the overlapping results between FTD and
AD and a surrogate estimation for the condence in the ML diagnosis.
Abbreviations
AD=Alzheimer’s disease
Aβ42=Amyloid-beta protein 42
bvFTD=Behavioral variant frontotemporal dementia
CSF=Cerebrospinal uid

Page 12/21
CTh=Cortical thickness
CTR=healthy subjects
FTD=Frontotemporal dementia
HCB=Hospital Clínic de Barcelona
ML=Machine Learning
MMSE=Mini-Mental State Examination
MRI=Magnetic Resonance Imaging
NfL=neurolament light chain
nfvPPA=Nonuent Variant Primary Progressive Aphasia
PPA=Primary progressive aphasia
SVM=Support Vector Machine
svPPA=Semantic Variant Primary Progressive Aphasia 
Declarations
Ethical Approval and Consent to participate
The Hospital Clínic de Barcelona Ethics Committee approved the study (HCB 2019/0105), and all
participants gave written informed consent.
Consent for publication
Not applicable.
Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author
on reasonable request.
Competing interests
The authors declare that they have no competing interests.
Funding

Page 13/21
A. Pérez-Millan is a recipient of the French Embassy in Spain / Institut français de España fellowship and
a travel fellowship from María de Maeztu Unit of Excellence (Institute of Neurosciences, University of
Barcelona) MDM-2017-0729. This study was partially funded by Instituto de Salud Carlos III, Spain (grant
no. PI20/0448 to Dr. R. Sanchez-Valle, PI19/00449 to Dr. A. Lladó and project PI19/00198 to Dr. M.
Balasa andco-funded bytheEuropeanUnion, “Una manera de hacer Europa” AGAUR, Generalitat de
Catalunya (SGR 2021- 01126 and 2021 SGR 00523) and by Spanish Ministry of Science and Innovation
(PID2020-118386RA-I00/ AEI/10.13039/501100011033 to Dr. R. Sala-Llonch). Dr. B. Thirion is supported
the KARAIB AI chair (ANR-20-CHIA-0025-01) and the H2020 Research Infrastructures Grant EBRAIN-Health
101058516. Dr. N. Falgàs is a recipient of the Joan Rodes fellowship from the Instituto de Salud Carlos III,
Spain. Dr. S. Borrego-Écijais a recipient of the Joan Rodés Josep Baselga grant from FBBVA. 
Authors' contributions
APM, BT, RSV, and RSL contributed to the study's design. APM, NF, SB, BB, JJP, ATM, JS, JMA, AA, NB, MB,
AL, and RSV contributed to the data acquisition. APM and BT contributed to the data analyses. APM, BT,
NF, AL, RSV, and RSL contributed to interpreting the data. APM, BT, RSV, and RSL contributed to the draft
of the article. NF, SB BB, JJP, ATM, JS, JMA, AA, NB, MB, and AL revised the manuscript critically for
important intellectual content and approved the nal version of the manuscript. All authors contributed to
the article and approved the submitted version.
Acknowledgments
The authors thank patients, their relatives, and healthy controls for participating in the research.
References
1. Albert MS, DeKosky ST, Dickson D, Dubois B, Feldman HH, Fox NC, et al. The diagnosis of mild
cognitive impairment due to Alzheimer’s disease: Recommendations from the National Institute on
Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease.
Alzheimer’s and Dementia. 2011;7(3):270–9.
2. McKhann GM, Knopman DS, Chertkow H, Hyman BT, Jack CR, Kawas CH, et al. The diagnosis of
dementia due to Alzheimer’s disease: Recommendations from the National Institute on Aging-
Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s
and Dementia. 2011;7(3):263–9.
3. Alcolea D, Vilaplana E, Suárez-Calvet M, Illán-Gala I, Blesa R, Clarimón J, et al. CSF sAPPβ, YKL-40,
and neurolament light in frontotemporal lobar degeneration. Neurology. 2017 Jul 11;89(2):178–88.
4. Antonell A, Tort-Merino A, Ríos J, Balasa M, Borrego-Écija S, Auge JM, et al. Synaptic, axonal damage
and inammatory cerebrospinal uid biomarkers in neurodegenerative dementias. Alzheimer’s &
Dementia [Internet]. 2019 Oct 24 [cited 2022 Dec 7]; Available from:
https://www.sciencedirect.com/science/article/pii/S1552526019353695

Page 14/21
5. McFerrin MB, Chi X, Cutter G, Yacoubian TA. Dysregulation of 14-3-3 proteins in neurodegenerative
diseases with Lewy body or Alzheimer pathology. Ann Clin Transl Neurol. 2017 Jul;4(7):466–77.
. Rohrer JD, Woollacott IOC, Dick KM, Brotherhood E, Gordon E, Fellows A, et al. Serum neurolament
light chain protein is a measure of disease intensity in frontotemporal dementia. Neurology. 2016
Sep 27;87(13):1329–36.
7. Davatzikos C, Resnick SM, Wu X, Parmpi P, Clark CM. Individual patient diagnosis of AD and FTD via
high-dimensional pattern classication of MRI. NeuroImage. 2008 Jul 15;41(4):1220–7.
. Du AT, Schuff N, Kramer JH, Rosen HJ, Gorno-Tempini ML, Rankin K, et al. Different regional patterns
of cortical thinning in Alzheimer’s disease and frontotemporal dementia. Brain. 2007 Apr
1;130(4):1159–66.
9. Bocchetta M, Todd EG, Peakman G, Cash DM, Convery RS, Russell LL, et al. Differential early
subcortical involvement in genetic FTD within the GENFI cohort. NeuroImage: Clinical.
2021;30:102646.
10. Borrego-Écija S, Sala-Llonch R, van Swieten J, Borroni B, Moreno F, Masellis M, et al. Disease-related
cortical thinning in presymptomatic granulin mutation carriers. NeuroImage: Clinical. 2021 Jan 1;29.
11. Canu E, Agosta F, Mandic-Stojmenovic G, Stojković T, Stefanova E, Inuggi A, et al. Multiparametric
MRI to distinguish early onset Alzheimer’s disease and behavioural variant of frontotemporal
dementia. NeuroImage: Clinical. 2017;15:428–38.
12. Contador J, Pérez-Millán A, Tort-Merino A, Balasa M, Falgàs N, Olives J, et al. Longitudinal brain
atrophy and CSF biomarkers in early-onset Alzheimer’s disease. NeuroImage: Clinical. 2021 Jan
1;32:102804.
13. Möller C, Hafkemeijer A, Pijnenburg YAL, Rombouts SARB, van der Grond J, Dopper E, et al. Joint
assessment of white matter integrity, cortical and subcortical atrophy to distinguish AD from
behavioral variant FTD: A two-center study. NeuroImage: Clinical. 2015 Jan 1;9:418–29.
14. Möller C, Vrenken H, Jiskoot L, Versteeg A, Barkhof F, Scheltens P, et al. Different patterns of gray
matter atrophy in early- and late-onset Alzheimer’s disease. Neurobiology of Aging. 2013 Aug
1;34(8):2014–22.
15. Abraham A, Pedregosa F, Eickenberg M, Gervais P, Mueller A, Kossai J, et al. Machine learning for
neuroimaging with scikit-learn. Frontiers in Neuroinformatics [Internet]. 2014 [cited 2023 Mar 13];8.
Available from: https://www.frontiersin.org/articles/10.3389/fninf.2014.00014
1. Frizzell TO, Glashutter M, Liu CC, Zeng A, Pan D, Hajra SG, et al. Articial intelligence in brain MRI
analysis of Alzheimer’s disease over the past 12 years: A systematic review. Ageing Research
Reviews. 2022 May 1;77:101614.
17. Mateos-Pérez JM, Dadar M, Lacalle-Aurioles M, Iturria-Medina Y, Zeighami Y, Evans AC. Structural
neuroimaging as clinical predictor: A review of machine learning applications. NeuroImage: Clinical.
2018 Jan 1;20:506–22.
1. Bron EE, Smits M, Papma JM, Steketee RME, Meijboom R, de Groot M, et al. Multiparametric
computer-aided differential diagnosis of Alzheimer’s disease and frontotemporal dementia using

Page 15/21
structural and advanced MRI. European Radiology. 2017 Aug 1;27(8):3372–82.
19. Chagué P, Marro B, Fadili S, Houot M, Morin A, Samper-González J, et al. Radiological classication
of dementia from anatomical MRI assisted by machine learning-derived maps. Journal of
Neuroradiology. 2021 Nov 1;48(6):412–8.
20. Klöppel S, Stonnington CM, Chu C, Draganski B, Scahill RI, Rohrer JD, et al. Automatic classication
of MR scans in Alzheimer’s disease. Brain. 2008 Mar;131(3):681–9.
21. Möller C, Pijnenburg YAL, Van Der Flier WM, Versteeg A, Tijms B, De Munck JC, et al. Alzheimer
disease and behavioral variant frontotemporal dementia: Automatic classication based on cortical
atrophy for single-subject diagnosis. Radiology. 2016 Jun 1;279(3):838–48.
22. Pérez-Millan A, Contador J, Juncà-Parella J, Bosch B, Borrell L, Tort-Merino A, et al. Classifying
Alzheimer’s disease and frontotemporal dementia using machine learning with cross-sectional and
longitudinal magnetic resonance imaging data. Human Brain Mapping [Internet]. 2023 [cited 2023
Jan 24]; Available from: https://onlinelibrary.wiley.com/doi/abs/10.1002/hbm.26205
23. Bisenius S, Mueller K, Diehl-Schmid J, Fassbender K, Grimmer T, Jessen F, et al. Predicting primary
progressive aphasias with support vector machine approaches in structural MRI data. NeuroImage:
Clinical. 2017 Jan 1;14:334–43.
24. Bron EE, Klein S, Papma JM, Jiskoot LC, Venkatraghavan V, Linders J, et al. Cross-cohort
generalizability of deep and conventional machine learning for MRI-based diagnosis and prediction
of Alzheimer’s disease. NeuroImage: Clinical. 2021 Jan 1;31:102712.
25. Cuingnet R, Gerardin E, Tessieras J, Auzias G, Lehéricy S, Habert MO, et al. Automatic classication
of patients with Alzheimer’s disease from structural MRI: A comparison of ten methods using the
ADNI database. NeuroImage. 2011 May 15;56(2):766–81.
2. Magnin B, Mesrob L, Kinkingnéhun S, Pélégrini-Issac M, Colliot O, Sarazin M, et al. Support vector
machine-based classication of Alzheimer’s disease from whole-brain anatomical MRI.
Neuroradiology. 2009 Feb 1;51(2):73–83.
27. Meyer S, Mueller K, Stuke K, Bisenius S, Diehl-Schmid J, Jessen F, et al. Predicting behavioral variant
frontotemporal dementia with pattern classication in multi-center structural MRI data. NeuroImage:
Clinical. 2017 Jan 1;14:656–62.
2. Mendez MF. The accurate diagnosis of early-onset dementia. International Journal of Psychiatry in
Medicine. 2006;36(4):401–12.
29. Wojtas A, Heggeli KA, Finch N, Baker M, DeJesus-Hernandez M, Younkin SG, et al. C9ORF72 repeat
expansions and other FTD gene mutations in a clinical AD patient series from Mayo Clinic. Am J
Neurodegener Dis. 2012 May 16;1(1):107–18.
30. Zee J van der, Sleegers K, Broeckhoven CV. Invited Article: The Alzheimer disease–frontotemporal
lobar degeneration spectrum. Neurology. 2008 Oct 7;71(15):1191–7.
31. Jack CR, Bennett DA, Blennow K, Carrillo MC, Dunn B, Haeberlein SB, et al. NIA-AA Research
Framework: Toward a biological denition of Alzheimer’s disease. Alzheimer’s and Dementia. 2018
Apr 1;14(4):535–62.

Page 16/21
32. Gorno-Tempini ML, Hillis AE, Weintraub S, Kertesz A, Mendez M, Cappa SF, et al. Classication of
primary progressive aphasia and its variants. Neurology. 2011 Mar;76:1006–14.
33. Rascovsky K, Hodges JR, Knopman D, Mendez MF, Kramer JH, Neuhaus J, et al. Sensitivity of revised
diagnostic criteria for the behavioural variant of frontotemporal dementia. Brain. 2011
Sep;134:2456–77.
34. Fischl B, Van Der Kouwe A, Destrieux C, Halgren E, Ségonne F, Salat DH, et al. Automatically
Parcellating the Human Cerebral Cortex. Cerebral Cortex. 2004 Jan;14(1):11–22.
35. Fischl B, Dale AM. Measuring the thickness of the human cerebral cortex from magnetic resonance
images. Proceedings of the National Academy of Sciences of the United States of America. 2000
Sep 26;97(20):11050–5.
3. Desikan RS, Ségonne F, Fischl B, Quinn BT, Dickerson BC, Blacker D, et al. An automated labeling
system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest.
NeuroImage. 2006 Jul 1;31(3):968–80.
37. Seidman LJ, Faraone SV, Goldstein JM, Goodman JM, Kremen WS, Matsuda G, et al. Reduced
subcortical brain volumes in nonpsychotic siblings of schizophrenic patients: A pilot magnetic
resonance imaging study. American Journal of Medical Genetics - Neuropsychiatric Genetics.
1997;74(5):507–14.
3. Breiman L. Random Forests. Machine Learning. 2001;45(1):5–32.
39. Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine
Learning in Python. Journal of Machine Learning Research. 2011;12(Oct):2825–30.
40. Dukart J, Mueller K, Horstmann A, Barthel H, Möller HE, Villringer A, et al. Combined Evaluation of
FDG-PET and MRI Improves Detection and Differentiation of Dementia. PLOS ONE. 2011 Mar
23;6(3):e18111.
41. Kim JP, Kim J, Park YH, Park SB, Lee JS, Yoo S, et al. Machine learning based hierarchical
classication of frontotemporal dementia and Alzheimer’s disease. NeuroImage: Clinical. 2019 Jan
1;23:101811.
42. Li B, Jang I, Riphagen J, Almaktoum R, Yochim KM, Ances BM, et al. Identifying individuals with
Alzheimer’s disease-like brains based on structural imaging in the Human Connectome Project Aging
cohort. Human Brain Mapping. 2021 Dec 1;42(17):5535–46.
43. Lin W, Tong T, Gao Q, Guo D, Du X, Yang Y, et al. Convolutional Neural Networks-Based MRI Image
Analysis for the Alzheimer’s Disease Prediction From Mild Cognitive Impairment. Frontiers in
Neuroscience [Internet]. 2018 [cited 2023 Apr 9];12. Available from:
https://www.frontiersin.org/articles/10.3389/fnins.2018.00777
44. Moore PJ, Lyons TJ, Gallacher J, Initiative for the ADN. Random forest prediction of Alzheimer’s
disease using pairwise selection from time series data. PLOS ONE. 2019 Feb 14;14(2):e0211558.
45. Salvatore C, Cerasa A, Battista P, Gilardi MC, Quattrone A, Castiglioni I. Magnetic resonance imaging
biomarkers for the early diagnosis of Alzheimer’s disease: a machine learning approach. Front
Neurosci. 2015 Sep 1;9:307.

Page 17/21
4. Wang J, Redmond SJ, Bertoux M, Hodges JR, Hornberger M. A Comparison of Magnetic Resonance
Imaging and Neuropsychological Examination in the Diagnostic Distinction of Alzheimer’s Disease
and Behavioral Variant Frontotemporal Dementia. Frontiers in Aging Neuroscience [Internet]. 2016
[cited 2023 Apr 9];8. Available from: https://www.frontiersin.org/articles/10.3389/fnagi.2016.00119
47. Wang SH, Phillips P, Sui Y, Liu B, Yang M, Cheng H. Classication of Alzheimer’s Disease Based on
Eight-Layer Convolutional Neural Network with Leaky Rectied Linear Unit and Max Pooling. J Med
Syst. 2018 Mar 26;42(5):85.
4. Basheera S, Sai Ram MS. Convolution neural network–based Alzheimer’s disease classication
using hybrid enhanced independent component analysis based segmented gray matter of T2
weighted magnetic resonance imaging with clinical valuation. Alzheimer’s & Dementia: Translational
Research & Clinical Interventions. 2019 Jan 1;5:974–86.
49. Dashtipour K, Taylor W, Ansari S, Zahid A, Gogate M, Ahmad J, et al. Detecting Alzheimer’s Disease
Using Machine Learning Methods. In: Ur Rehman M, Zoha A, editors. Body Area Networks Smart IoT
and Big Data for Intelligent Health Management. Cham: Springer International Publishing; 2022. p.
89–100. (Lecture Notes of the Institute for Computer Sciences, Social Informatics and
Telecommunications Engineering).
50. McCarthy J, Collins DL, Ducharme S. Morphometric MRI as a diagnostic biomarker of frontotemporal
dementia: A systematic review to determine clinical applicability. NeuroImage: Clinical. 2018 Jan
1;20:685–96.
51. Avants BB, Cook PA, Ungar L, Gee JC, Grossman M. Dementia induces correlated reductions in white
matter integrity and cortical thickness: A multivariate neuroimaging study with sparse canonical
correlation analysis. NeuroImage. 2010 Apr 15;50(3):1004–16.
52. Dickerson BC, Goncharova I, Sullivan MP, Forchetti C, Wilson RS, Bennett DA, et al. MRI-derived
entorhinal and hippocampal atrophy in incipient and very mild Alzheimer’s disease. Neurobiology of
Aging. 2001;22(5):747–54.
53. Frisoni GB, Fox NC, Jack CR, Scheltens P, Thompson PM. The clinical use of structural MRI in
Alzheimer disease. Nat Rev Neurol. 2010 Feb;6(2):67–77.
54. Gil-Navarro S, Lladó A, Rami L, Castellví M, Bosch B, Bargalló N, et al. Neuroimaging and Biochemical
Markers in the Three Variants of Primary Progressive Aphasia. DEM. 2013;35(1–2):106–17.
55. Gordon E, Rohrer JD, Fox NC. Advances in neuroimaging in frontotemporal dementia. Journal of
Neurochemistry. 2016;138(S1):193–210.
5. Hodges JR, Patterson K. Semantic dementia: a unique clinicopathological syndrome. The Lancet
Neurology. 2007 Nov 1;6(11):1004–14.
57. Jack CR, Petersen RC, Xu Y, O’Brien PC, Smith GE, Ivnik RJ, et al. Rates of hippocampal atrophy
correlate with change in clinical status in aging and AD. Neurology. 2000;55(4):484–9.
5. Chen Y, Sha M, Zhao X, Ma J, Ni H, Gao W, et al. Automated detection of pathologic white matter
alterations in Alzheimer’s disease using combined diffusivity and kurtosis method. Psychiatry
Research: Neuroimaging. 2017 Jun 30;264:35–45.

Page 18/21
59. Bouts MJRJ, Möller C, Hafkemeijer A, van Swieten JC, Dopper E, van der Flier WM, et al. Single
Subject Classication of Alzheimer’s Disease and Behavioral Variant Frontotemporal Dementia Using
Anatomical, Diffusion Tensor, and Resting-State Functional Magnetic Resonance Imaging. Journal
of Alzheimer’s Disease. 2018 Jan 1;62(4):1827–39.
0. Dyrba M, Grothe M, Kirste T, Teipel SJ. Multimodal analysis of functional and structural
disconnection in Alzheimer’s disease using multiple kernel SVM. Human Brain Mapping.
2015;36(6):2118–31.
Figures
Figure 1
We plot the calibrated probabilities versus the uncalibrated probabilities. Each point represents the
individual probability for each classication obtained with calibration and without calibration step. We
represent the test data together.

Page 19/21
Figure 2
Density plot to study the obtained individual probabilities with the MRI-based algorithm. The clinical
diagnosis is identied with triangles or circles, and the algorithm’s diagnoses are plotted with different
colors. CTR: healthy subjects, AD: Alzheimer’s disease, FTD: frontotemporal dementia.

Page 20/21
Figure 3
Cortical (top) and subcortical (bottom) patterns of the feature importance of each region associated with
AD and FTD. At a higher value major importance of that region for the classication.
Figure 4
Cortical (top) and subcortical (bottom) patterns of the feature importance of each region associated with
bvFTD and PPA. At a higher value major importance of that region for the classication.

Page 21/21
Figure 5
Density plot to study the obtained individual probabilities with the MRI-based, CSF-based, and MRI- and
CSF-based algorithms. It can be seen in the clinical diagnosis with triangles or circles and the algorithm’s
diagnosis, plotted with different colors. CTR: healthy subjects,
Supplementary Files
This is a list of supplementary les associated with this preprint. Click to download.
SupplemntaryMaterial.docx