Early diagnosis of Alzheimer's disease (AD) is crucial for providing timely treatment and care to patients. However, current diagnostic methods rely on clinical symptoms and biomarkers, which are often unreliable and invasive. Brain networks model the brain's structure and function in AD and other brain diseases. To fully capture their complexity, we need multi-modal models that combine diferent types of data, such as structural and functional connectivity, clinical and genetic information. This gives us a holistic view of the disease's many aspects. In this paper, we argue that brain networks and multi-modal data fusion can improve early diagnosis of AD by capturing the complex and heterogeneous nature of the disease. Using brain network modeling and multi-modal data fusion, we envisage a novel framework for detecting AD and its prodromal stages. The framework can simultaneously capture network properties from multi-modal as well as longitudinal datasets, which provide complementary information.
CEUR ceur-ws.org
the disease and predicting its progression. Various measurements have been developed and
assessed for detecting AD, typically encompassing physical health examinations,
neuropsychological assessments, and brain imaging techniques. Biomarkers are usually categorized into
physiological, cognitive, behavioral, and psychological domains. Among the non-cognitive tests,
neuroimaging techniques are commonly utilized, while the Mini Mental State Examination
(MMSE) [
However, current diagnostic methods rely on clinical symptoms and biomarkers, which are often unreliable and invasive. Clinical symptoms are subjective and vary across individuals and stages of the disease. Biomarkers, such as cerebrospinal fluid (CSF) and amyloid-beta (Aβ), require invasive procedures and expensive equipment. Moreover, both clinical symptoms and biomarkers are not sensitive enough to detect the early and prodromal stages of AD, such as mild cognitive impairment (MCI) and subjective cognitive decline (SCD).
To capture the various symptoms of AD, including subtle changes that occur throughout the progression of the disease, there’s widespread agreement that a reliable method for detecting early-stage Alzheimer’s disease cannot rely solely on measurements from one source. Instead, it should employ a multi-modal approach by combining diferent types of biomarkers. Each type of data reveals unique aspects of the condition, and integrating them all provides a more comprehensive understanding, ultimately improving diagnostic accuracy.
Over the past few years, the progress in Artificial Intelligence-based methods for analyzing
multi-modal data has fueled research seeking new approaches for early disease detection. In
Comito et al. [
The objective of this paper is to explore new avenues and methods to integrate brain networks within multi-modal learning architecture to improve early diagnosis of Alzheimer’s disease. To this purpose, the paper proposes a novel framework for detecting AD and its prodromal stages using brain network modeling and multi-modal data fusion. Unlike existing methods, our framework can simultaneously capture network properties from multi-modal as well as longitudinal datasets, which provide complementary information. We use network models to represent the structural and functional connectivity of the brain regions, and integrate multiple types of data, such as images, text, audio, etc., to capture the complex and heterogeneous nature of AD.
The paper is organized as follows. In Section 2, a summary of the current state-of-the-art
multi-modal approaches for AD prediction overviewed in [
This section summarizes the main data modalities and AI methods utilized in Alzheimer’s
disease research reviewed in [
As far as data modalities are concerned (see Figure 1), neuroimaging (i.e. diferent tipology of MRI and PET scans) stands out as the most prevalent data modality due to its non-invasive nature, the availability of large datasets, the advancement of robust AI tools capable of extracting significant features and classifying images, and the possibility of analyzing both structural and functional brain anomalies and changes in AD patients.
Following neuroimaging, biological and genetic markers (APOE-e4, SNPs, CSF) are predominantly used to identify individuals likely to develop AD. Actually, certain genes and CSF markers have been linked to an increased risk of the disease. Neuropsychological and cognitive assessment tests are the third most common, primarily serving as screening tools to pinpoint those requiring further evaluation. Lastly, demographic and clinical data, including blood markers, are less frequently employed in the classification and progression tracking of AD.
The complexity and diversity of the data involved in AD research are, in turn, reflected in the use of various AI techniques to classify the stages of the disease and predict its progression.
Traditional ML classifiers like SVM, DT, GB, RF, and LOR are widely used (see Figure 2), particularly when clinical, demographic, and cognitive data are adopted. On the other hand, DL methods, including NN, CNN, and RNN, show promise in medical image analysis, which is crucial for AD diagnosis and monitoring. However, there is a remarkable number of AI methods (referenced in Figure 2 as “OTHERS”) that are used only in one or two approaches. This highlights that the application of AI in AD prediction and progression is indeed a multifaceted ifeld.
The mixed outcomes of the research in this field suggest that there is no one-size-fits-all solution, and the choice of technique may depend on the specific dataset and task at hand. Moreover, the need for careful parameter tuning, data selection, and experimental settings cannot be overstated, especially when dealing with limited data availability, which is a common challenge in AD research. The exploration of ensemble neural networks and the comparison of various ML models underscore the ongoing eforts to refine predictive models for AD.
Overall, the continuous evolution of AI methods in AD prediction and progression demonstrates the dynamic nature of the field and the potential for AI to contribute to our understanding and management of this complex disease.
In the last decade, the study of the progression of Alzheimer’s disease and its early detection
has provided promising results from the joint use of advanced neuroimaging techniques and
complex network theory [
Graphs are a mathematical model widely used for studying complex systems. In the literature
on brain networks and neurodegenerative diseases in general, there are diferent ways to
formalize and model the brain through a graph. More formally, the entities of a brain and their
relationships can be represented with a brain network modeled as a graph = ( , , )
where is a set of objects, called nodes or vertices, ⊆ × is a set of links, called edges,
that connect two elements of , and ∶ × → is a function which assigns a weight to a
couple (, ) of nodes and , if there exists an edge connecting and , and 0 if an edge between
and does not exist. In almost all the AD-related studies, the nodes of the graph are usually
brain regions, while the edges may capture diferent relationships between regions [
• Anatomical/structural brain networks: constructed from structural magnetic
resonance imaging (MRI), the edges represent physical connections between regions (e.g., the
estimated white matter connection strength in terms of number of fibers between any
pair of brain regions [
functional-like edges representing correlations between regions. • Directed progression networks (DPNets): closely related to cortical thickness networks, they attempt to capture the temporal progression of the disease, more than the correlations between regions, similar to the epidemic network models where an edge represents the spreading of the disease. In this network formalism, the edges are directed and capture the degree to which one brain region thinning precedes the second region thinning.
To characterize network topology, thresholding sparsification techniques are usually applied to
structural or functional connectivity matrices to remove edges with noisy weights [
Even if these types of networks are clearly related, the comparison between them and their
joint analysis with the goal of having a systemic and multi-layer view of the disease progression
is not straightforward [
An important aspect that the study of brain networks has highlighted in recent years is the
detection of building blocks, i.e. the presence of overabundant small subgraphs sharing patterns
of interconnections, called network motifs, occurring with a frequency higher than that in a
random network [
A of a graph is defined as an unordered subset = { 1, … , } of nodes of presenting
a particular pattern of interconnections. Fig. 3 shows five types of motifs among three and four
nodes (Fig. 3(a)-Fig. 3(e)). Their labeling follows the same convention adopted in [
Motifs have been largely analyzed in brain networks, especially in structural and functional
networks. In [
RESEARCH | REPORTS (a) (b) (c) (d) (e) spoicotnrinioaplntriaoelngruaelglaruteligaoutniloantion sia1g(Ar.Fe,1iAMg.,11M–AM,1–M7M)1a–7)rMear)eare
7 uytrbsutstocrtusutcrrtuaulcrtahulurhablusbhsubs h1aAoptph,astph(asFt hi(gFs.i(g1F.Ai1g,A.,1A, eakotvwremkoomrtkiofmstifohstaihvfseavheave nortnikizosaontfioonnfentoweftonwrekotswrkosrks esralgylnaerrlygeemrleyamirneasminaasninasnan owuorrk motif M, our rmkomtioftMif ,Mo,ulaoryuerr. 8 JULY 2016 • VOL83J on another region. To calculate this pairwise value between brain regions, the measure of
Phase Transfer Entropy (PTE) is applied. Here, some 3 and 4 nodes motifs are over-expressed,
tornwereskotwrsktorsrutkcrutsutcrrtueumcsroteruesorveesr, motif-based clustering reveals a strong symmetry between the two hemispheres,
supporting the idea of a higher-order organization of the efective connectivity brain network.
oaizmrnagi-tazianotinizoaontfioocnofmco-mBcaot-tiston et al. [
network is built by forming a first structural network layer with DW-RMI data (Difusion ofsr.kFor each network FroeracehacnhentwetowrMkoargknetic RMI) and a second functional layer with rs-fMRI data. The motifs analyzed are ihnegr-order clusmtueltriipnlegx, which is statistically overabundant small subgraphs spanning over the two layers r-doerrdecrlucsltuesrtienrging
and jointly considered. Results show that motifs in whichFigl.i2n.Hkighser-boredetrwcluseteerinnthbeCr. aeleignansrneeugroinoal nnestwoarkt. [Stehe(e29).] have been proposed as pioneers
diicfh- means that dif- (A) The four-node bi-fan motif, which is overexpressed in neuronal networks neurons are known regulators of
emaneasnthsathtadtisftd-ruicft-ural layer also have a functional activity are frequent,(1b). uInttuittivhelye,trhiesmaotrifedesacrlibseosafcrooepeqrautivee nprotpamgatuionlotfiipnfolremxation exposestheirorganization.Theclu
dnes-are exposemdote,if-dshea-ving distant regions with strong functional dynamclusitecr sin.thFeCi.neleaglanlsyf,rontthal eneuaronnalanletywosrkisbassehdoon wthesmottihfine(A). three RME neurons, outgoing con
fromthe nodesonthe lefttothe nodesonthe right.(B)The besthigher-order intermediary of information proce
eareexpexopseodse,dd,ed- The cluster contains three ring motor neurons (RMEL, -V, and -R; cyan) with the largest total number of conn
existence of a reinforcement mechanism between the strumcantyuourtgaoilngaconnndecttiohnse,whficuh nsecrvetiaostnheasolurlcaeyof einfrorsmaftoionr;six lustrationofthehigher-orderclust
if. inner labial sensory neurons (IL2DL, -VR, -R, -DR, -VL, and -L; orange) with locations are the true two-dimen
which the probability of having a link is related to the intemnansyiintcoyminogfcontnhecetioncs,osenrvinngaescthteidoestninatiionnoftinhfoermaotiotnh;aendrfour Mostinformationflowsfromleft to
URA motor neurons (purple) acting as intermediaries. These RME neurons RIH serve as sources of informati
nrdfo.erddu.edu
• Step 2: Compute the spectral ordering s of in thenetwork,in practice,thealgorithmismuch tion) (
sither of nodes SRwecietnhtstudies have highlighted that network motif analythesinsodecs afronm ptheronovrmiadlizeed nmoetiwfLapmlaciaanrk efarstser. fBoyarnaltyhzineg16 real-world networks where higher
rf onfondoedseSswSiwtdhiiatghnosis and progression of Alzheimer’s disease. Friedman• Setetp a3:l.Fin[d1th4e]prwefixesreteofasbwlieth tthoe d2isbtililinong,wueisfohund the computational com- the su
matrix constructed via WM (15). the number of edges m ranges from 159,000 to knowl
Satsehould participate smallest motif conductance (the argument of the plexity to scaleasQ(m1.2).Moreover,thealgorithm when
huoldulpdaprtairctipicaiptAeaDtferom HC patients with directed motif analysis of theirmDinimPuNm), efotrmsa.llyI,Sn: =aadrgdmiintrifoM(Snr),,wthheree ecnanterasoilypbyepoarafllelized, and sampling tech- is not
Sr = {s1, …, sr}. niques can be used to further improve perform- also se
unldd, the set S sthheomuotlidfdistribution provides a new imaging marker for MForCtriIa.ngular motifs, the algorithm scales to ance (16). for the
, ethseetseStsShoshuolduIlndthe next section, a review of the most recent propostaaklesssesvetraulhdouyrsitonpgrocebssrgraapihns ofnsuechtswize.ordkirsectefdo,arndtwheigehtednetworks,aswellasmotifs higher
networkswithbillionsofedgesand,typically,only The framework can be applied to directed, un- (13). S
wenhich oFci gcu.r1s. wHhigenher-oFrdige.r 1n.eHtiwgohrekr-sotrrduecrtunreetswoaOnrnsmkadllernsettwtohrrkseuwithchhutndiuregdsrohfethoeussarnd-saon(1r3d)d.Moeretovrher, it encanealhsotbiewagpplohiedrteonkertw-orocks lrudbersof
hicohccoucrcsurwshwedhniaegnnosisFoifgFA.lizgh1e..im1He.ri’sHgdhiisgeeahrse-eoirsr-rodepreodrrteednr.entewtowrokrsksizet9(1rs3).uAtlthrcouugthtcuhewtroruest-carsseecomasputnatiaodnanl dtare hcomtemhon eihnsoiciahglneihtwgorekhs(rfriee-nodrve-rrsuodsforeedrmeathe
of edges, the algorithm can process motifs up to withpositiveandnegativesignsontheedges,which motifs
tmhea mo(tAif)aHreigihnetrh-oerder s(tAru)cHtuigrehsera-roerdcearpstutrruecdtubryesneatrwecomopclerxaikty pof mtthueaolrgoetritidhfmsfob.rtyrFianognurlaeremtoxw-aoor mtrrukst pvemrlseus ,doistartusiltflesd1ge.s3)aFndocmoertabneolicnxneaet-mcstpeepctr
aomtiof taifrearine itnhethe (A()AH)igHhigerh-eorr-doerdresrtrsutcrutucrteusreasrearceapcatuprteudrebdybnyentwetowrkormkomtiofsti.fsF.oFroerxeaxmapmlep,l
tifs is Q(m1.5), where m is the number of edges works (edges signifying activation versus inhibi- that fi
lyM, ,givenmaotmifs3oa.t1ri.ef NsMheto,wwonrkh-mbearoeste.ifd(sBaa)prpCerlsouhascothewernisngheorfea.(nBe)twColursk16t4ebra8iJUnsLY eg201d6o•VoOfL 3an53 ISnmSUEe6o2t95 wtifoMrk b.Faosread goinvemnomtifoMtif7M.F,oo
7
geinvean ma omtioftMif ,M, momtifostiafsreasrehoshwonwhnehree.r(eB. ()BC)luCsltuesrtienrginogf oafnaentwetowrkorbkabsaesdeodnomn omtiof tMif M.F7o.Froar
frtaomewfoinrdk aaAsimseptsarottfoofntohedNeesftiwnSodrtkahMasetedimtcinoienfaimnppoirzdoeaecsshemSs, tothhteiafnetctwomonrindk-iubmacstiezadenmsceemth,oofdtsiaf(rSce)oa,bnlwedthuoiccrethvaenwalceteh,edfeMfi(nSe),a7wshthiceh r
M
mweowrkorakimaismtnsoettwoork dfiinsrdufipntaidonsaseotsretohtfe osntrfoundctoeurdsaleScshatSnhgaetshtianmittiiamnteimdinbiiymztehiszeedmisseoamsteiofthtcrifooucngohdntuhdceutaapcpntlaiccneatc,ioefn,off(MS()S,w),hwichhicwhew
sheatt of nmodoetisfsSc)utth(afiltled trmianogtilfessccuutt()fitloledthterimanignliemsucmut)nutomtbheermofinnimoduems innuinmsbtaenrcoMefsnoofdtehseinminostitfai
methodologies proper of Network Science and Graph Theory, ofering practical techniques able
onfondoedseSs) St)hathtat momtifosticfsutcu(ftill(efidlletdriatrniagnlegslecsuct)utto) ttohethme imniminiummu mnunmubmebreorf onfondoedseins iinnsintasntac
this catsoeta,ctkhleertheeislimotinthaeitsiomcnsaoostfeift,hctehuceto.rn(esCoils)idToatnhedeebmhioimgohatriefkrce-ruostr[.d5](e.CrT)hneTeshetwemoehtrihgkohdcseluar-rseoterindrciernergansifnergatlwmyoerwkoclruks.tGeirvineng
attractingthinisttehrciesastscdeau,setehto,ettrhheeeirriseaboiislnitoeyntmoesomimtiopfltcifiuyfctt.hue(tCu.n()CdTe)rhsTteahnhdeiignhghigeofhr-tehorer-dmoeroddreenlirenngtweoftowtrhkoercklucslutesrtienrginfgraf
tgh,ematicalinEtnegrienseter(iinng,this casinet,eMres),tt(hine tfhraismceawseo,rMkf7o),rmthse aframmoetiwfaodrkjafcoernmcsy ammatoritxif (aWdja)cebnyccyomunattinri
eic3naa0ttli5cE,anUlgSEinAnteg.wei2nrDoeineepgnrai,onidrnigtdsmt,eehaeisssneatcarteionam-tboueiynlcrtrtiecpeetupslwerotrebro(tsiiioinntnln7og(goatiiinnhdcnaTietslahislnbecilvsseacet1loscaste-an[ho1,sec5cMe]tec.y,p7IuMno)er,fto7htfi)etnhb,hfertoeaahliflnnoermwanfieminrontawgstem,iotfwwa.rkeneAomwrcrneeokvodiereefowkoliigfnrtfgemhot.nehrsvmmeeoasmcstmtarooerotlmeiftvo.iaoffnAtatianfdwajoLaedrakcijgsapeeMclnaneccvnyieaccmnytoamtrrtaraoinxtfrsi(fxao
2
S,IsNAiUty.S,2AWD.eepsDaterLptamafraetymnetettnet, IN twotwnoondoedsecso-cooc-coucrcuinr iannainnsitnasntacnecoef otfhethmeomtiof.tiAf.nAneigeeignevnevcetocrtoorfoaf aLa
adjacency matrix isadthjaecnenccoymmpuatteridx.iTshtheeonrdceorminpgusteodf.tThheenoorddeesripnrgosvidoefdthbeyntohdeecsopmropvoidnee
p,eaWsrttemLsatefnLatya,efaStytteae,nttfIeNo,r dIN• Dragomir and Vrahatis [
d fMf(SMr()S,ris a
approaches and connectivity from MRI data on connectivity networks). In this context,
the benefits of monitoring the early stages of the disease by jointly considering genotype
and brain phenotype information as noninvasive biomarkers are discussed. Particular
emphasis is placed on methods that analyze the network substructure through methods
of complex network analysis since disruptions usually occur locally, in diferent brain
regions and molecular pathways, and at diferent rates. Finally, a framework
integrating knowledge from the two information levels, molecular omics-based data collected
from blood samples and brain connectivity obtained from neuroimaging techniques, is
proposed, showing that this multi-level solution can further improve diagnosis.
• In [
MRI images, and the clinical progression in nondemented subjects who have abnormal amyloid markers in the cerebrospinal fluid (CSF), that is a marker of predementia AD.
The study investigates if network structural measures like size, connectivity density,
degree, clustering coeficient, path length, betweenness centrality, and the small-world
property are somehow associated with the rate of progression to MCI or dementia, using
Cox proportional hazard models to assess associations between the structural measures
and time to clinical progression. Results indicate that when these measures have low
values there is an increased risk of fast progression to MCI. In particular, lower clustering
values, indicative of a more randomly organized network, in specific anatomical areas are
associated with clinical outcomes and fast clinical progression.
• In recent work, [
MCI, or HC, modeling the brains as functional graphs and exploiting graph theory-based features. The functional brain network is built by setting on each edge the Pearson’s correlation functional connectivity between ROIs. Then, graph embedding (node2vec) is used to transform graphs into a vector and applying machine learning techniques.
To classify the subjects into classes, diferent classification techniques including linear support vector machines (LVSM) and regularized extreme machine learning (RELM) are explored. The highest accuracy is obtained by combining LASSO with LSVM.
Brain networks are complex and intricate systems that reflect the brain’s structure and function in various brain diseases, such as AD. Multi-modal data is essential to capture their complexity. Multi-modal data integrates diferent types of data, such as structural and functional, connectivity, clinical, and genetic information, to form a holistic understanding of the disease’s multifaceted nature. Previous research has tried to fuse various modalities.
However, studies investigating the combination and the relationship of the brain connectome
with biomarkers and genetics are very few. Only recently, Yu et al. [
To bridge this gap we propose a novel multi-modal architecture that integrates brain network with biomarkers and genetics data. Integrating brain networks and multi-modal data within a learning framework for Alzheimer’s prediction is a promising approach that leverages diverse sources of information to improve the accuracy and reliability of predictive models. Brain networks provide a comprehensive representation of the brain’s structural and functional connectivity patterns, ofering valuable insights into the underlying neurobiology of Alzheimer’s disease. Multi-modal data, on the other hand, encompasses various types of information such as neuroimaging scans (e.g., MRI, PET), clinical assessments, genetic markers, and cognitive scores, each providing unique perspectives on the disease.
Figure 4 shows the proposed learning framework. The model is crafted for a multi-modal multitask objective, aiming to grasp Alzheimer’s disease progression and cognitive scores based on a variety of data. As illustrated in Figure 4, the model initially processes data from five modalities: neuroimaging, biomarkers, genomics, clinical, and demographics. Specifically, multimodal data including MRI scans, demographics, medical history, functional assessments, and neuropsychological test results, are used to develop learning models on various classification tasks. Local and temporal feature learning in the model is facilitated by exploiting machine and deep learning cutting-edge methods. The features from various modalities (cognitive scores, neuropsychological battery, MRI, PET, and assessment modalities) undergo preprocessing to enhance data quality. AI techniques are utilized for feature reduction, extracting principal components from high-dimensional MRI and PET data. For example, neuroimaging features from MRI and PET modalities can be extracted using FreeSurfer. Subsequently, deep features are independently learned from each modality using both ML and DL approaches like stacked CNN-BiLSTM models. Abstract deep features obtained from the previous step are then fused to extract common features from all modalities using a series of dense layers within a deep neural network. As the final stage of the learning framework, a classification is produced.
By combining brain networks with multi-modal data in a learning framework, we can capitalize on the complementary nature of these sources to enhance prediction accuracy and understand the complex progression of AD. Here are some key points to consider: • Feature Fusion. Integrating information from brain networks and multi-modal data involves fusing features extracted from diferent sources into a unified representation. This can be achieved through techniques such as feature concatenation, attention mechanisms, or graph-based fusion methods. • Learning Architecture. The learning framework should incorporate neural network architectures capable of processing multi-modal inputs and capturing complex relationships within the data. This may involve the use of deep learning models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), or graph neural networks (GNNs) tailored to handle multi-modal data. • Regularization and Adaptation. Given the high-dimensional and heterogeneous nature of multi-modal data, regularization techniques such as dropout, and batch normalization, can help prevent overfitting and improve generalization performance. Additionally, model adaptation strategies may be employed to adapt the learning process to diferent data modalities and patient cohorts. • Evaluation and Validation. Robust evaluation metrics and validation procedures are essential for assessing the performance of the learning framework. Cross-validation, hold-out validation, and external validation on independent datasets can help validate the generalizability of the predictive models. • Clinical Interpretability. Interpretable models are crucial for translating predictive insights into actionable clinical decisions. Techniques such as attention mechanisms, feature importance analysis, and visualization methods can provide insights into the contribution of diferent modalities to the prediction task and aid in clinical interpretation. • The rapid growth of new AI approaches in the last decade has largely overlooked the importance of computational eficiency in algorithm design and data generation. This has led to the widespread adoption of complex AI techniques with high computational costs and energy consumption. This is true also in medical applications, such as AD detection, where improvements in accuracy come at the cost of increased data availability. The development of novel algorithms able to deal with limited resources while maximizing the quality of the results obtained is a main objective of the emerging so-called Green AI methods. Considering the large volumes of multimodal data nowadays available for AD, the design and development of new energy-aware AI techniques with low computational needs and reduced data while reaching high predictive accuracy is a desirable objective in the near future of AD research.
Overall, integrating brain networks and multi-modal data within a learning framework holds great potential for advancing Alzheimer’s prediction research, leading to more accurate and reliable predictive models that can aid in early diagnosis and personalized treatment planning for patients.
In conclusion, early detection of Alzheimer’s disease is paramount for efective intervention and patient management. However, existing diagnostic methods often fall short, relying heavily on clinical symptoms and biomarkers that may not provide accurate or timely results. Brain networks ofer a promising avenue for understanding the complex structural and functional changes associated with AD. To fully grasp this complexity, a multi-modal approach is necessary, incorporating various data types such as structural and functional connectivity, clinical assessments, and genetic information.
In this paper, we advocate for the integration of brain networks and multi-modal data fusion to advance early AD diagnosis. By combining these approaches, we can better capture the diverse and nuanced characteristics of the disease. Our proposed framework seeks to leverage brain network modeling and multi-modal data fusion to develop a comprehensive understanding of AD and identify prodromal stages. This innovative approach aims to extract network features from diverse datasets, including longitudinal data, to enhance diagnostic accuracy and inform timely interventions for individuals at risk of AD.
This work has been partially supported by the project ALCMAEON (F/260014/01-03/X51) funded by MISE. We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), Spoke 9 - Green-aware AI, under the NRRP MUR program funded by the NextGenerationEU.