Toward a Multimodal Multitask Model for Neurodegenerative Diseases

Diagnosis and Progression Prediction

Soﬁa Lahrichi, Maryem Rhanoui, Mounia Mikram and Bouchra El Asri

IMS Team, ADMIR Laboratory, Rabat IT Center, ENSIAS, Mohammed V University, Rabat, Morocco

mrhanoui@esi.ac.ma, mmikram@esi.ac.ma

Keywords:

Alzheimer’s Disease, Multimodal Multitask Learning, Machine Learning, Deep Learning, Progression

Detection, Time Series.

Abstract:

Recent studies on modelling the progression of Alzheimer’s disease use a single modality for their predictions

while ignoring the time dimension. However, the nature of patient data is heterogeneous and time dependent

which requires models that value these factors in order to achieve a reliable diagnosis, as well as making it

possible to track and detect changes in the progression of patients’ condition at an early stage. This article

overviews various categories of models used for Alzheimer’s disease prediction with their respective learning

methods, by establishing a comparative study of early prediction and detection Alzheimer’s disease progres-

sion. Finally, a robust and precise detection model is proposed.

1 INTRODUCTION

Alzheimer’s disease (AD) is a progressive, irre-

versible neurodegenerative disease characterized by

an abnormal build-up of amyloid plaques and neu-

roﬁbrillary tangles in the brain, resulting in memory,

thinking and behavior issues. AD is the most common

form of dementia characterized by a slow and asymp-

tomatic progress of the disease.

AD is clinically very heterogeneous, varying from

patient to patient in terms of cognitive symptoms, test

results, and rate of progression. Indeed, several recent

therapeutic trials have shown variable efﬁcacy from

one subset of patients to another. Currently available

treatments only slow the progression of AD, and no

deﬁnitive cure has been developed yet. When it comes

to AD, patient data (Zhang et al., 2012) is heteroge-

neous, but complementary, of different types; mag-

netic resonance imaging (MRI), positron emission to-

mography (PET), genetics, cerebrospinal ﬂuid (CSF),

etc. The combination of multimodalities (Weiner

et al., 2013) facilitates the detection of changes in the

patient’s states and constitutes a reliable diagnosis.

Patient data is gathered from different visits and

from continuous patient monitoring. The state of the

disease at any given time is not independent of the

condition at a previous time. Therefore, AD data are

not only multimodal but could also be considered as

time series and longitudinal series.

In the practical diagnosis of AD (Jo et al., 2019),

the most widely used neuroimaging dataset comes

from the Alzheimer’s Disease Neuroimaging Initia-

tive (ADNI) which contains (Jack Jr et al., 2008)

socio-demographic data (gender, education level), the

APOE genotype and ﬁve neuropsychological test re-

sults: MMSE (mini-mental state examination), CDR-

SB (the sum of boxes of clinical dementia rating),

ADASCog (Alzheimer’s Disease Assessment Scale

cognitive sub-scale), LMT (Logical Memory Test )

and RAVLT (Rey Auditory Verbal Learning Test).

Several learning methods have been developed for

the classiﬁcation and prediction of MCI (Mild Cogni-

tive Impairment) to AD conversion, namely Machine

Learning and Deep Learning.

After extensive bibliographical and analytical

work on various articles, we will give an overview of

the categories of approaches used for the prediction of

Alzheimer’s disease, then we will study the different

learning techniques for each approach and establish

a comparative table of this last. This categorization

was essential to us because it will allow us at the end

of the study to propose a stable and precise detection

model.

The remainder of the article will consist of three

main parts; the ﬁrst will be the most important be-

cause through it, we will establish the bases of our

study by deﬁning the concepts of multimodal and

multitask learning. The second part will be devoted

322

Lahrichi, S., Rhanoui, M., Mikram, M. and El Asri, B.

Toward a Multimodal Multitask Model for Neurodegenerative Diseases Diagnosis and Progression Prediction.

DOI: 10.5220/0010600003220328

In Proceedings of the 10th International Conference on Data Science, Technology and Applications (DATA 2021), pages 322-328

ISBN: 978-989-758-521-0

to the comparative study of the different models for

predicting the progression of AD. And in the last part,

we will propose a model grouping together all the ad-

equate criteria for a better detection of the disease.

2 BACKGROUND

In this section, we introduce the main notions neces-

sary to understand the context of our proposal.

2.1 Multi-modal Learning

Multimodal learning combines (Zhang et al., 2020)

data from multiple sources that are semantically cor-

related and sometimes provide complementary infor-

mation to each other, resulting in more robust predic-

tions.

The multimodal learning model not only captures

the correlation structure between different modalities

but also allows to recover missing modalities given

those observed.

The multimodal learning model for example (Ba-

heti, 2020) can combine two deep Boltzmann ma-

chines, each corresponding to a modality. An addi-

tional hidden layer is placed on top of the two Boltz-

mann machines to give the common representation.

Multimodal learning is divided into four stages:

(1) Representing inputs and summarize data in a way

that expresses multiple modalities.

(2) Translating (map) the data from one modality to

another.

(3) Extracting important features from information

sources while creating models that best match the

type of data.

(4) Fusion and co-learning. Consists in combining

the information of two or more modalities to make

a prediction. This combination should be normally

weighted.

2.2 Multitask Learning

Multitask learning (MTL) aims to extract useful infor-

mation from multiple tasks and use their correlations

to help learn a more accurate pattern for each task.

Multitask learning has been used successfully in

all applications of machine learning, natural lan-

guage processing(Worsham and Kalita, 2020), speech

recognition(Pironkov et al., 2016), computer vision,

and drug discovery.

MTL (Zhang and Yang, 2018)can be categorized

into several categories, including supervised multi-

task learning(MTSL), unsupervised multitask learn-

ing, semi-supervised multitask learning...

MTSL aims to learn n functions for the n tasks

of the labeled training set, (Each task can be a clas-

siﬁcation or regression problem) after it uses ﬁ (•) to

predict the labels of the data instances i from the j the

task. There are three categories of MTSL patterns to

indicate the relationship between stains:

Feature-based MTSL. Which learns common fea-

tures of different tasks in order to share knowledge

and avoid using original representations directly. It

can easily be affected by outliers.

Parameter-based MTSL. Which ﬁnds how the

model parameters of different tasks are related, which

leads to ranking. It is more robust to outliers.

Instance-based MTSL. Which aims to ﬁnd in one

task the instances useful for other tasks. It is not used

too much.

In unsupervised multitask learning, each task, can

be a clustering problem, aims to identify useful pat-

terns contained in a training dataset composed only

of data instances.

In multitask semi-supervised learning, each task

aims to predict the labels of data instances from the

labeled and also unlabeled data.

3 COMPARATIVE STUDY

3.1 Models Categories

After an in-depth bibliographical study on the pro-

gression of AD, and after consulting a large number

of articles that seemed relevant to us, two main ideas

emerge: Research on modeling disease progression

has received special attention from researchers

around the world, and the modeling has been divided

into four categories.

Mono-modal Mono-task Learning. Where the

model optimizes only one objective function based

on a type of data. Through this model, neither the

correlation between the tasks, nor the collective

information between the modalities are explored.

Mono-modal Multitask Learning. Where the tasks

share a few training instances. The relationship be-

tween the spots is modeled assuming that they share

a common representation space or they share certain

parameters. However, they don’t take into account

the relationship between the different modalities of

the same task.

Multi-modal Mono-task Learning. Where several

modalities are taken into consideration in the predic-

tion of a state while ignoring information from other

tasks.

Multi-modal Multitask Learning. Where each

task of the problem has characteristics from several

Toward a Multimodal Multitask Model for Neurodegenerative Diseases Diagnosis and Progression Prediction

323

modalities and or several tasks are linked to each

other in a chronological sequence.

Artiﬁcial intelligence is advancing in understand-

ing the world around us and the operating model

that most closely matches our real environment and

provides greater precision is multitask multimodal

learning

3.2 Related Works

In our bibliographical study, we focused on the

following articles which seem to us the most relevant

and which meet our expectations.

Mono-modal Mono-task Learning. (Cui et al.,

2019) proposed a single-task, monomodal model

based on six-step MRI time series data for the

detection of AD progression. They use a stacked

CNN-BGRU (Convolutional Neural Network Bidi-

rectional Gated Recurrent Unit) pipeline. It achieves

an accuracy of 91.33% for AD vs CN (normal

cognitive), and 71.71% for pMCI (progressive MCI)

vs sMCI (stable MCI). However, relying on MRI

alone is insufﬁcient in the medical ﬁeld.

Mono-modal Multitask Learning. (Liu et al., 2018)

proposed a deep multi-channel multi-task convolu-

tional neural network for classiﬁcation (Multi-class

classiﬁcation) and regression using MRI data and BG

demographic information.

(Lopez-de Ipina et al., 2018) have developed a new

nonlinear multitasking (three tasks: animal naming

(AN), picture description (PD) and spontaneous

speech (SS)) approach based on automatic speech

analysis. They introduced linear features, perceptual

features, Castiglioni fractal dimension and Multiscale

Permutation Entropy into their analysis.

Based on the MRI data, they performed a clas-

siﬁcation, using Multilayer Perceptron (MLP) and

Deep Learning by means of Convolutional Neural

Networks (CNN) (biologically- inspired variants of

MLPs) which led to promising results.

Multi-modal Mono-task Learning. (Pan et al.,

2018) proposed a two-step deep learning framework

for the diagnosis of AD using both MRI and PET

data. in the ﬁrst step, they assign the corresponding

MRI data to the missing PET data using 3D-cGAN

(3D Cycle-consistent Generative Adversarial Net-

works) to capture their underlying relationship. In the

second step, they develop LM3IL (Landmark-based

Multi-modal Multi-Instance Learning Network)

which learns and merges the characteristics necessary

for the diagnosis of AD and the prediction of MCI.

Using the ADNI dataset, (Shi et al., 2017)

proposed a method of transforming nonlinear feature

space into more linearly separable data using SVM.

They chose TPS (Thin-platespline) as a geometric

model because of its power of representation. They

also adopted a feature fusion strategy based on a deep

network by SDAE (Denoising Sparse Auto-Encoder)

to merge the transverse and longitudinal features

estimated from MRI brain images.

(Lee et al., 2019) proposed a one-task multi-

modal deep learning approach by incorporating

multi-domain longitudinal data. They applied a GRU

(Gated Recurrent Unit) for each modality (4Gru) to

produce feature vectors of ﬁxed size which will be

concatenated to form an input for the ﬁnal prediction

where regularized logistic regression is used for

the classiﬁcation of MCI- C and MCI-NC (MCI

converter and MCI non-converter). The model on a

very small number of features (no addition of BG

data) has better prediction, it only optimizes binary

classiﬁcation tasks.

Multi-modal Multitask Learning. (Lahmiri and

Shmuel, 2019) proposed M3T multimodal multitask

model for the prediction of AD over two years. The

latter has two main stages: (1) Selecting multitask

features using the Lasso (2) Using an SVM (Support

Vector Machine) model for separate classiﬁcation and

regression.

Multi-source multitask learning (MSMT) simul-

taneously considers two types of prior knowledge. 1)

Source consistency 2) Slow temporal evolution

(Nie et al., 2016) proposed a linear MSMT model that

predicts the future disease status over 2 years (M06-

M48) of new patients, based on their health informa-

tion at the ﬁrst moment (Baseline).

(Tabarestani et al., 2020) propose a distributed

multitask multimodal model to predict MMSE cog-

nitive measures of AD progression, the latter individ-

ually exploits several multitask regression coefﬁcient

matrices for each modality, then It concatenates the

risk factors with the predicted y of each modality then

it goes through gradient boosting to group the results

of different modalities and reduce their prediction er-

ror.

(Nie et al., 2015) proposed an adaptive multi-

modal multitask linear learning model (aM2L) to reg-

ularize the modality agreement for the same task, the

temporal progression on the same modality and the

weight of the modalities.

(Nie et al., 2016) and (Nie et al., 2015) in their

predictions, don’t incorporate the follow-up observa-

tions. Example: To predict her condition at M24 (24

months after Baseline), they don’t merge the observa-

tions of Baseline, M06 AND M12.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

324

(Li et al., 2015) presented a robust deep learn-

ing system to identify the different stages of progres-

sion of patients with AD based on MRI and PET.

They used the dropout technique to improve classical

deep learning by preventing its weight co-adaptation,

which is a typical cause of deep learning overﬁt-

ting. They stacked multiple RBMs (Restricted Boltz-

mann machine) to build a robust deep learning frame-

work, which integrates stability selection and multi-

task learning strategy.

(El-Sappagh et al., 2020) proposed a robust en-

semble deep learning model based on a stacked con-

volutional neural network (CNN) and a bi-directional

long-term memory network (BiLSTM). This multi-

modal multitask model jointly predicts several vari-

ables based on the fusion of ﬁve types of multimodal

time series data plus a basic knowledge set (BG). The

predicted variables include the multi-class task and

four critical cognitive score regression tasks. This

model gave equal weights for the classiﬁcation and

regression tasks.

The table below 1 is a summary of some works

seen previously, it lists the advantages and limitations

of its different models.

3.3 Synthesis of Models

All the models seen previously for the study of the

prediction of the progression of AD, use the ADNI

database which can contain missing data. In order to

overcome this problem, among the methods that were

used:

DEL: We simply eliminated the subjects with either

missing sources or missing labels.

ZERO: We assigned zero value to any element that is

missing.

KNN: The k-nearest neighbor (KNN) method re-

placed the missing value in the data matrix with the

corresponding value from the nearest column ...

In recent years, machine learning algorithms (such as

SVM and random forest, SDAE) and deep learning

(CNN, GRU, LSTM, BGRU, BiLSTM,...) have been

used for the design of a predictive model of AD pro-

gression. And for the generalization of these models

to avoid overﬁtting ,among the regularizers that have

been used ,we ﬁnd: lasso ,sparce group lasso,l2,1

norm, ridge regression ,fused group lasso(convex and

non-convex),dropout ......

(Liu et al., 2013) and (Duchesne et al., 2009) used

regular machine learning techniques to study mul-

timodal single-task classiﬁcation and regression, re-

spectively.

To detect AD progression based on a multimodal

single-task deep learning model, we ﬁnd; (Spasov

et al., 2019) who proposed a classiﬁcation model

based on a CNN ,(Lee et al., 2019) who also proposed

a binary classiﬁcation model based on GRU and used

logistic regression for regularization.

In a real medical environment, many modalities

are analyzed and multiple clinical variables must be

predicted. Multimodal single-task models will not

provide in this case sufﬁcient information for the

study of AD progression, hence the interest of the

multimodal multitask model.

This model has been used by (Lahmiri and

Shmuel, 2019) for classiﬁcation and regression using

SVM and SVR respectively. In order to generalize

their model, they used the Lasso regularizer.

It was also used by el sappagh who added to their

model the time series constraint which is consistent

with the longitudinal nature of AD because the pa-

tient’s state at a given time is not independent of his

state at an earlier time.

They used the CNN-BILTSM for multiclass clas-

siﬁcation and for regression of 4 cognitive scores. For

the regularization of their model, they opted for the

dropout.

(El-Sappagh et al., 2020) consider that the modal-

ities have the same weight of importance which is not

always true in the medical ﬁeld.

4 PROPOSED MODEL

Alzheimer’s disease (AD) is a longitudinal disease,

that is to say the state of a patient at a time t depends

strongly on his state at t-1. AD cannot be triggered

spontaneously; there must be a prior appearance of

signs that predict the manifestation and evolution of

the disease. Most models studied in the previous sec-

tion do not take this notion into consideration.

Several factors can inﬂuence the diagnosis out-

come of AD, and even the most indirect symptoms

may have in some cases an effect on the detection

of AD, hence the interest of introducing Multitask

Learning. Multitask Learning will allow us to ﬁnd

a relationship between the features, the instances and

the parameters involved in the detection of the dis-

ease.

Moreover, patients’ data is considered multimodal

since they’re gathered from various sources such as

MRI, PET, CSD, ASD, NPD etc... These modalities

can also interact with each other, because the infor-

mation extracted from a modality can be complemen-

tary to other modalities, which opens the possibility

of improving the performance of the model for AD

progression prediction. Additionally, it is important

to note that the data retrieved via different modalities

Toward a Multimodal Multitask Model for Neurodegenerative Diseases Diagnosis and Progression Prediction

325

Table 1: Comparative table of models for predicting the progression of Alzheimer’s disease.

Study Data Modality Merging

Time-

series

Performance Model Task

(Lee

et al.,

2019)

ADNI

(Database,

2004)

Demographic,

MRI, CSD,

CSF(LCR)

YES (4

Steps)

Accuracy: 81% (MCI/AD) GRU Classiﬁcation

(Lahmiri

and

Shmuel,

2019)

ADNI

MRI, FDG-

PET, LCR

NO NO

Classiﬁcation Accu-

racy: 93,3%(CN/AD),

83,2%(CN/MCI), 73,9%

(sMCI/pMCI). Regression

Accuracy: 0,697(MMSE),

0,739(ADAS)

SVM

Classiﬁcation

and regres-

sion

(Nie

et al.,

2016)

ADNI

MRI,PET,

CSF,PROPT,

META

NO NO

MMSE:89,01 %ADAS-

Cog: 91,68%

aM2L Regression

(El-

Sappagh

et al.,

2020)

ADNI

MRI, PET,

CSD, ASD,

NPD

YES

YES(15

Steps)

ACC: 92.62%, PRE:

94.02%, F1: 92.56%, REC:

98.42. MAE: 0.107, 0.076,

0.075, and 0.085, (FAQ,

ADAS, CDR, MMSE)

StackedCNN-

BiLSTM

Classiﬁcation

and regres-

sion

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

326

may vary in terms of importance. For instance, MRI

and PET are considered a more reliable and relevant

data sources, hence the interest in introducing the idea

of a weighted model.

According to the results of the comparative study

seen in the previous section, which follows closely

the same direction mentioned previously, we can con-

clude that the most adequate, precise and stable model

is the Multimodal Multitask Model based on time se-

ries.

The research axis that can be developed from this

study, will center around a Multimodal Multitask sys-

tem based on time series, which will be able to si-

multaneously regularize the modality weighting, the

temporal progression as well as the modality agree-

ment, i.e. the status of the patient estimated by dif-

ferent modalities must be consistent. The model will

also be capable of taking into consideration the rela-

tionship between AD progression and the patient’s co-

morbidities (cardiovascular disease, depression, gen-

itourinary renal metabolism, endocrine, etc.). Such

model has not yet been addressed in the literature.

To advance on this research axis, we propose a

Machine Learning or a Deep Learning architecture,

multitask, multimodal (using MRI, PET, CSD, ASD,

NPD), trained on the ADNI database and the patients’

demographic data. This model predicts AD progres-

sion status of a through a multiclass classiﬁcation

task, and the values of two cognitive scores ADAS,

MMSE which will be implemented as two regression

tasks.

In order to meet the different expectations as well

as possible, two possible scenarios seem to open up at

the moment:

- The ﬁrst approach consists in extracting separately

the temporal characteristics of each modality (MRI,

PET, CSD, ASD, NPD), then merged with the de-

mographic data to extract the common characteris-

tics that respond to each task by applying a multitask

learning.

- Or by applying multitasking to each modality and

merging the initial results with demographic data,

which is assumed to be a time-invariant information.

This model has the ability to stop the propagation of

an error from one modality to another.

The loss function in both approaches will include

a ﬁrst term for classiﬁcation, a second term for regres-

sion and 2 last terms to regularize the weight and the

agreement of the modalities.

5 CONCLUSION

Alzheimer’s disease (Thung et al., 2017) is currently

calling the attention of many researchers. A consid-

erable amount of effort is being put to understand

the biological and physiological mechanisms of AD

as well as its monitoring and early detection. The

presented works in this paper have raised several ap-

proaches for the detection of AD, namely Monomodal

monotask, Monomodal multitask Multimodal mono-

task and Multimodal multitasking. Following our

comparative study of different works and papers, we

believe that a major research axis has been revealed

in which the Multimodal Multitask based on time se-

ries could at the same time to regularize the modal-

ity weighting, the temporal progression, the modality

agreement and which will even take into considera-

tion the relationship between the progression of AD

and the patient’s comorbidities. This model could

bring a considerable advance in the ﬁeld of medical

research.

REFERENCES

Baheti, P. (2020). Introduction to Multimodal Deep

Learning. https://heartbeat.fritz.ai/introduction-to-

multimodal-deep-learning-630b259f9291.

Cui, R., Liu, M., Initiative, A. D. N., et al. (2019).

Rnn-based longitudinal analysis for diagnosis of

alzheimer’s disease. Computerized Medical Imaging

and Graphics, 73:1–10.

Database (2004). Introduction to Multimodal Deep Learn-

ing. http://adni.loni.usc.edu.

Duchesne, S., Caroli, A., Geroldi, C., Collins, D. L., and

Frisoni, G. B. (2009). Relating one-year cognitive

change in mild cognitive impairment to baseline mri

features. Neuroimage, 47(4):1363–1370.

El-Sappagh, S., Abuhmed, T., Islam, S. R., and Kwak, K. S.

(2020). Multimodal multitask deep learning model

for alzheimer’s disease progression detection based on

time series data. Neurocomputing, 412:197–215.

Jack Jr, C. R., Bernstein, M. A., Fox, N. C., Thompson,

P., Alexander, G., Harvey, D., Borowski, B., Britson,

P. J., L. Whitwell, J., Ward, C., et al. (2008). The

alzheimer’s disease neuroimaging initiative (adni):

Mri methods. Journal of Magnetic Resonance Imag-

ing: An Ofﬁcial Journal of the International Society

for Magnetic Resonance in Medicine, 27(4):685–691.

Jo, T., Nho, K., and Saykin, A. J. (2019). Deep learning

in alzheimer’s disease: diagnostic classiﬁcation and

prognostic prediction using neuroimaging data. Fron-

tiers in aging neuroscience, 11:220.

Lahmiri, S. and Shmuel, A. (2019). Performance of ma-

chine learning methods applied to structural mri and

adas cognitive scores in diagnosing alzheimer’s dis-

Toward a Multimodal Multitask Model for Neurodegenerative Diseases Diagnosis and Progression Prediction

327

ease. Biomedical Signal Processing and Control,

52:414–419.

Lee, G., Nho, K., Kang, B., Sohn, K.-A., and Kim, D.

(2019). Predicting alzheimer’s disease progression us-

ing multi-modal deep learning approach. Scientiﬁc re-

ports, 9(1):1–12.

Li, F., Tran, L., Thung, K.-H., Ji, S., Shen, D., and Li, J.

(2015). A robust deep model for improved classiﬁca-

tion of ad/mci patients. IEEE journal of biomedical

and health informatics, 19(5):1610–1616.

Liu, F., Zhou, L., Shen, C., and Yin, J. (2013). Multiple ker-

nel learning in the primal for multimodal alzheimer’s

disease classiﬁcation. IEEE journal of biomedical and

health informatics, 18(3):984–990.

Liu, M., Zhang, J., Adeli, E., and Shen, D. (2018).

Joint classiﬁcation and regression via deep multi-task

multi-channel learning for alzheimer’s disease diag-

nosis. IEEE Transactions on Biomedical Engineering,

66(5):1195–1206.

Lopez-de Ipina, K., Martinez-de Lizarduy, U., Calvo, P. M.,

Mekyska, J., Beitia, B., Barroso, N., Estanga, A.,

Tainta, M., and Ecay-Torres, M. (2018). Advances

on automatic speech analysis for early detection of

alzheimer disease: a non-linear multi-task approach.

Current Alzheimer Research, 15(2):139–148.

Nie, L., Zhang, L., Meng, L., Song, X., Chang, X., and

Li, X. (2016). Modeling disease progression via

multisource multitask learners: A case study with

alzheimer’s disease. IEEE transactions on neural net-

works and learning systems, 28(7):1508–1519.

Nie, L., Zhang, L., Yang, Y., Wang, M., Hong, R., and

Chua, T.-S. (2015). Beyond doctors: Future health

prediction from multimedia and multimodal observa-

tions. In Proceedings of the 23rd ACM international

conference on Multimedia, pages 591–600.

Pan, Y., Liu, M., Lian, C., Zhou, T., Xia, Y., and Shen,

D. (2018). Synthesizing missing pet from mri with

cycle-consistent generative adversarial networks for

alzheimer’s disease diagnosis. In International Con-

ference on Medical Image Computing and Computer-

Assisted Intervention, pages 455–463. Springer.

Pironkov, G., Dupont, S., and Dutoit, T. (2016). Multi-

task learning for speech recognition: an overview. In

ESANN.

Shi, B., Chen, Y., Zhang, P., Smith, C. D., Liu, J., Initiative,

A. D. N., et al. (2017). Nonlinear feature transforma-

tion and deep fusion for alzheimer’s disease staging

analysis. Pattern recognition, 63:487–498.

Spasov, S., Passamonti, L., Duggento, A., Li

o, P., Toschi,

N., Initiative, A. D. N., et al. (2019). A parameter-

efﬁcient deep learning approach to predict conversion

from mild cognitive impairment to alzheimer’s dis-

ease. Neuroimage, 189:276–287.

Tabarestani, S., Aghili, M., Eslami, M., Cabrerizo, M., Bar-

reto, A., Rishe, N., Curiel, R. E., Loewenstein, D.,

Duara, R., and Adjouadi, M. (2020). A distributed

multitask multimodal approach for the prediction of

alzheimer’s disease in a longitudinal study. NeuroIm-

age, 206:116317.

Thung, K.-H., Yap, P.-T., and Shen, D. (2017). Multi-stage

diagnosis of alzheimer’s disease with incomplete mul-

timodal data via multi-task deep learning. In Deep

learning in medical image analysis and multimodal

learning for clinical decision support, pages 160–168.

Springer.

Weiner, M. W., Veitch, D. P., Aisen, P. S., Beckett, L. A.,

Cairns, N. J., Green, R. C., Harvey, D., Jack, C. R.,

Jagust, W., Liu, E., et al. (2013). The alzheimer’s dis-

ease neuroimaging initiative: a review of papers pub-

lished since its inception. Alzheimer’s & Dementia,

9(5):e111–e194.

Worsham, J. and Kalita, J. (2020). Multi-task learning for

natural language processing in the 2020s: Where are

we going? Pattern Recognition Letters, 136:120–126.

Zhang, D., Shen, D., Initiative, A. D. N., et al. (2012).

Multi-modal multi-task learning for joint prediction

of multiple regression and classiﬁcation variables in

alzheimer’s disease. NeuroImage, 59(2):895–907.

Zhang, L., Wang, M., Liu, M., and Zhang, D. (2020). A

survey on deep learning for neuroimaging-based brain

disorder analysis. Frontiers in neuroscience, 14.

Zhang, Y. and Yang, Q. (2018). An overview of multi-task

learning. National Science Review, 5(1):30–43.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

328