Neu oXVocal: De ec ion and Explana ion o
Alzheime ’s Disease h ough Non-in asi e
Analysis o Pic u e-p omp ed Speech
Nikolaos N ampakis1,2[0009−0001−7546−4104], Kons an inos
Diaman a as1[0000−0003−1373−4022], Ioanna Chou a da3[0000−0001−8915−6658],
Magda Tsolaki4[0000−0002−2072−8010], Panagio is
Sa igianndis5,2[0000−0001−6042−0355], and Vasileios
A gy iou6[0000−0003−4679−8049]
1In e na ional Hellenic Uni e si y, Sindos, G eece
[email p o ec ed]
2Me aMind Inno a ions, Kozani, G eece
[email p o ec ed]
3A is o le Uni e si y o Thessaloniki, Thessaloniki, G eece
4G eek Associa ion o Alzheime ’s Disease & Rela ed Diso de s, Thessaloniki, G eece
5Uni e si y o Wes e n Macedonia, Kozani, G eece
6Kings on Uni e si y London, London, UK
Abs ac . The ea ly diagnosis o Alzheime ’s Disease (AD) h ough non
in asi e me hods emains a signi ican heal hca e challenge. We p esen
Neu oXVocal, he i s end- o-end explainable AD classi ica ion sys em
ha achie es s a e-o - he-a pe o mance while p o iding clinically in-
e p e able explana ions. Ou no el dual-componen a chi ec u e con-
sis s o : (1) Neu o, a mul imodal classi ie implemen ing a unique ans-
o me based usion s a egy ha p ojec s acous ic, ex ual, and speech
embeddings in o a common dimensional space o complex c oss-modal
in e ac ions; and (2) XVocal, a specialized RAG-based explaine ha
e ie es ele an clinical li e a u e o gene a e e idence-based explana-
ions. Unlike p e ious app oaches using la e usion o simple conca e-
na ion, ou a chi ec u e enables bo h obus classi ica ion and meaning-
ul clinical insigh s. Using he IS2021 ADReSSo Challenge benchma k
da ase , Neu oXVocal achie ed 95.77% accu acy, signi ican ly ou pe -
o ming p e ious s a e-o - he-a . Medical p o essionals alida ed he
clinical ele ance o XVocal’s explana ions h ough s uc u ed e alua-
ion. This wo k ad ances beyond pu e classi ica ion o b idge he gap
be ween machine lea ning p edic ions and clinical decision-making. Code
a ailable a :
h ps://gi hub.com/NN amp/Neu oXVocal.
Keywo ds: Alzheime ·Mul imodal ·Explainable Heal hca e AI.
1 In oduc ion
Alzheime ’s Disease (AD) has eme ged as a c i ical global heal h conce n, a -
ec ing o e 55 million people wo ldwide wi h nea ly 10 million new cases an-
2 N. N ampakis e al.
nually [1]. Ea ly de ec ion h ough non-in asi e me hods emains c ucial o
e ec i e in e en ion and ea men planning. While adi ional diagnos ic ap-
p oaches ely on neu oimaging o in asi e p ocedu es, ecen ad ances in a i i-
cial in elligence ha e opened new possibili ies o ea ly de ec ion h ough speech
analysis [2, 3]. This pape p esen s Neu oXVocal, a no el dual-componen sys-
em ha no only classi ies bu also explains i s diagnos ic p edic ions h ough
speech analysis o pa ien s desc ibing images, whe he hey a e iden i ied as ha -
ing Alzheime ’s disease o being cogni i ely heal hy. The ela ionship be ween
cogni i e decline and speech pa e ns has been ex ensi ely s udied using he
ADReSSo benchma k da ase [4]. Syed e al. achie ed signi ican esul s using
unc ionals o deep ex ual embeddings, epo ing 84.51% accu acy in AD de-
ec ion [5]. Shah e al. u he in es iga ed language-agnos ic speech ep esen a-
ions, demons a ing he e ec i eness o speech in elligibili y ea u es wi h 79.6%
accu acy [6]. Mo e ecen ly, Fu e al. p oposed a mul imodal usion me hod com-
bining acous ic and seman ic in o ma ion using ImageBind audio encode and
ELMo, achie ing 90.3% accu acy [7]. Li e al. demons a ed p omising esul s
using Whispe -based ans e lea ning, achie ing 84.51% accu acy and 84.50%
F1-sco e by inno a i ely using ull ansc ip s as p omp s du ing ine- uning [8].
The la es ad ancemen by Lee e al. in oduced a g aph neu al ne wo k le e -
aging image- ex simila i y om ision language models, achie ing 88.73% accu-
acy [9]. While hese app oaches ha e shown p omising esul s in classi ica ion,
he ield has seen limi ed p og ess in explaining he easoning behind diagnos ic
p edic ions. Recen wo k by Iqbal e al. employed Local In e p e able Model-
agnos ic Explana ions (LIME) and SHapley Addi i e exPlana ions (SHAP) o
p o ide insigh s in o linguis ic ma ke s o cogni i e decline [10]. Simila ly, Bang
e al. explo ed he use o LLMs o gene a ing e idence-based explana ions o
speech pa e ns, hough hei app oach was limi ed by he in e p e abili y o
he unde lying language model [11]. Howe e , hese s udies s ill ace challenges
in p o iding comp ehensi e, clinically-ac ionable explana ions ha b idge he
gap be ween machine lea ning p edic ions and medical decision-making. Build-
ing upon hese ounda ions, we p esen Neu oXVocal, which add esses hese
limi a ions h ough he ollowing key con ibu ions:
1. No el A chi ec u e: Fi s end- o-end amewo k seamlessly in eg a ing AD
classi ica ion wi h clinically in e p e able explana ions, whe e mul imodal
ea u es con ibu e o bo h diagnosis and explana ion gene a ion.
2. Ad anced Fusion S a egy: A ans o me -based a chi ec u e ha p ojec s
acous ic ea u es, ex ual ea u es, and speech embeddings in o a common
dimensional space be o e usion, enabling complex c oss-modal in e ac ions
supe io o exis ing la e- usion app oaches.
3. Clinical Explainabili y: In oduc ion o XVocal, a specialized RAG compo-
nen ha e ie es ele an AD esea ch o gene a e e idence-based expla-
na ions. Medical p o essionals alida ed i s clinical ele ance, con i ming i s
po en ial as a diagnos ic suppo ool.
Neu oXVocal 3
4. S a e-o - he-a Pe o mance: Achie emen o 95.77% accu acy on he ADReSSo
benchma k, signi ican ly ou pe o ming exis ing me hods while main aining
in e p e abili y.
2 Me hodology
Ou p oposed no el Neu oXVocal sys em consis s o wo p ima y componen s, as
in Fig. 1: (1) he Neu o classi ie o AD de ec ion h ough mul imodal analysis o
speech da a, and (2) he XVocal explaine o gene a ing clinically-in e p e able
jus i ica ions. The sys em p ocesses inpu audio samples h ough mul iple pa -
allel s eams o ex ac complemen a y ea u es be o e usion and classi ica ion.
Fig. 1. Neu oXVocal A chi ec u e
4 N. N ampakis e al.
2.1 Fea u e Ex ac ion and P ocessing
Le xbe an inpu audio sample. F om his inpu , we ex ac h ee dis inc ea u e
ep esen a ions. The acous ic ea u es a(x) = ϕa(x)∈R47 comp ise empo al
cha ac e is ics (speech/pause a ios), p osodic ea u es (pi ch, in ensi y), a ic-
ula ion me ics, spec al p ope ies, oice quali y indica o s (ji e , shimme ,
ha monics- o-noise a io), and 13 Mel F equency Ceps al Coe icien s(MFCC)
coe icien s wi h hei s anda d de ia ions. These ea u es (47 in o al) unde go
s anda diza ion and missing alue impu a ion. Fo speech embeddings, we em-
ploy Wa 2Vec2-base-960h [12] a e con e ing audio o mono and esampling
o 16kHz:
e(x) = Mean(Wa 2Vec2(P ep ocess(x))) ∈R768 (1)
whe e he embeddings a e s anda dized be o e u he p ocessing. The ex ual
ea u es a e ob ained using Whispe ASR [13] o ansc ip ion ollowed by
DeBERTa- 3-base [14] encoding:
(x) = DeBERTa(P ep ocess(Whispe (x))) ∈R768 (2)
whe e p ep ocess includes lowe case con e sion, special cha ac e emo al, and
space no maliza ion.
2.2 Neu o Classi ie
The classi ica ion componen implemen s a no el usion a chi ec u e. We i s
p ojec he acous ic and speech embedding ea u es o a common dimensional
space:
ha=Linea ( a(x)) ∈R768, he=Linea ( e(x)) ∈R768 (3)
The usion p ocess conca ena es hese p ojec ions wi h he ex embeddings in
he p ojec ed dimensions o 514 ×768:
H= [ha;he; (x)] ∈R514×768 (4)
A wo-laye ans o me encode p ocesses his ep esen a ion. Each laye im-
plemen s mul i-head a en ion wi h 8 heads, whe e he inpu His p ojec ed o
que ies (Q), keys (K), and alues (V):
A en ion(Q, K, V ) = Conca (head1, ..., head8)WO(5)
whe e WOis he ou pu p ojec ion ma ix, and each a en ion head is compu ed
as:
headi=so max(QW Q
i(KW K
i)T
√dk
)V W V
i(6)
whe e WQ
i, W K
i, W V
i∈R768×96 a e lea ned pa ame e ma ices, and dk= 96
is he head dimension. This is ollowed by a eed- o wa d ne wo k wi h GELU
ac i a ion:
Z=FFN(A en ion(H)) ∈R514×768 (7)
Neu oXVocal 5
The inal classi ica ion uses a wo-laye classi ie :
p(y|x) = σ(Dense(z0)) (8)
whe e σis he sigmoid ac i a ion unc ion o bina y classi ica ion.
2.3 XVocal Explaine
The no el explainabili y componen implemen s a RAG app oach ha p ocesses
he ex ac ed ea u es along wi h he Neu o classi ie ’s p edic ion. The explana-
ion gene a ion begins by cons uc ing a s uc u ed p omp que y q(a ailable
in p ojec ’s Gi Hub eposi o y) h ough a empla e:
q={class(p(y|x)) ⊕ ea u es( a(x)) ⊕speech( e(x)) ⊕ ansc ip ( (x))}(9)
The ele an li e a u e co pus Lis p ep ocessed in o seman ic chunks by spli ing
each documen in o pa ag aphs and hen in o indi idual sen ences o c ea e a
ine-g ained con ex pool {c1, ..., cn}. Using all-MiniLM-L6- 2 [16], we cons uc
a dense ec o index:
Ec={MiniLM(ci)∈R384|ci∈ L} (10)
whe e each chunk is encoded in o a 384-dimensional embedding space. These
embeddings a e indexed using FAISS [15] L2 dis ance me ic:
I=FAISSL2(Ec)(11)
Fo e ie al, he que y qis encoded in he same embedding space and he op
5 mos ele an chunks a e e ie ed using nea es neighbo sea ch:
L ={I.sea ch(MiniLM(q), k = 5)}(12)
The inal explana ion is gene a ed using FLAN-T5-XL [17]:
E=FLAN-T5(q⊕L ;τ, p)(13)
whe e τand pa e he empe a u e and op-p sampling pa ame e s espec i ely,
con olling he gene a ion cohe ence.
3 Expe imen s and Resul s
3.1 Implemen a ion De ails
All expe imen s we e conduc ed on Ubun u using 8xNVIDIA A16 GPUs wi h
126GB sys em RAM. The Neu o classi ie ained o a maximum o 200 epochs.
Fo he XVocal componen , we used FAISS (CPU) o e ie al and deployed
using 4-bi quan iza ion. Each aining ound was comple ed on an a e age o 9
hou s in ou se up.
6 N. N ampakis e al.
3.2 Da ase
We u ilised he ADReSSo Challenge da ase [4] o he p obable AD p edic ion
ask. The da a is o ganized in he diagnosis olde , wi h 166 pa ien s in he ain-
ing se (79 cogni i ely no mal [cn], 87 p obable Alzheime ’s disease [ad]) and 71
pa ien s in he es se . The es se is kep independen o anspa en e alu-
a ion. The da ase is accessible h ough Demen iaBank membe ship, equi ing
egis a ion and adminis a o app o al. The comple e da ase documen a ion
is a ailable h ough ou p ojec eposi o y.
3.3 Resul s
Table 1. Pe o mance compa ison on ADReSSo da ase . A: Acous ic, T: Tex , S:
Speech embeddings
Me hodology Modali ies 5- old Accu acy(±s d%) Acc[%] F1-sco e[%]
Syed e al.(2021) [5] T 84.51% 84.45%
Shah e al.(2023) [6] A+T 79.60%
Fu e al.(2024) [7] A+T 90.3% 91.4%
Li e al.(2024) [8] T+S 84.51% 84.5%
Lee e al.(2025) [9] T+S 88.73% 88.23%
(Neu o)XVocal A+T+S 96.24% ±2.47% 95.77% 95.76%
Rega ding he esul s o he Neu o Classi ie inco po a ed in ou Neu oX-
Vocal me hodology, we compa ed wi h p ominen and ecen s a e-o - he-a
me hodologies as shown in Table 1. To e alua e he pe o mance, we u ilized
he widely adop ed accu acy and F1-sco e me ics. As demons a ed in Table 1,
ou Neu o classi ie achie ed obus pe o mance ac oss mul iple e alua ion sce-
na ios. In he 5- old c oss- alida ion se ing, we ob ained an a e age accu acy o
96.24% wi h a s anda d de ia ion o 2.47%. When ained on he ull aining se
and e alua ed on he independen es se , ou me hod achie ed 95.77% accu-
acy and 95.76% F1-sco e, subs an ially ou pe o ming all p e ious app oaches.
To assess he clinical ele ance and u ili y o XVocal’s explana ions, we con-
duc ed a comp ehensi e quali a i e e alua ion wi h medical expe s. Each expe
e alua ed explana ions o 20 pa ien cases (10 AD, 10 CN) using a s uc u ed
ques ionnai e wi h 10 c i e ia1, a ed on a 5-poin Like scale. Fo he knowl-
edge base o he RAG componen , we ha e inco po a ed a cu a ed co pus o
10 seminal publica ions [18–27] co e ing linguis ic ma ke s, spon aneous speech
analysis, and LLM applica ions in AD de ec ion.
The e alua ion esul s (Table 2) demons a e s ong pe o mance ac oss mul-
iple dimensions o clinical u ili y. XVocal achie ed no ably high sco es in AD
ma ke iden i ica ion (3.98) and explana ion cla i y (3.96), indica ing i s e ec-
i eness in highligh ing ele an diagnos ic ea u es. The sys em also pe o med
1Ques ionnai e a ailable a : h ps:// o ms.gle/ AFuC6ediUY qQz 8
Neu oXVocal 7
Table 2. C i e ia and expe e alua ion esul s o XVocal’s explana ions
Assessmen Focus Scale Mean Sco e
Clea jus i ica ion o diagno-
sis
1-No clea , 5-Ve y clea 3.96
Pe inence o iden i ied
ma ke s
1-No ele an , 5-Highly ele an 3.85
Consis ency wi h medical
knowledge
1-No alignmen , 5-High alignmen 3.63
Explana ion-based con i-
dence
1-No con iden , 5-Highly con iden 3.63
Recogni ion o disease indi-
ca o s
1-No ma ke s iden i ied, 5-Highly app o-
p ia e ma ke s iden i ied
3.98
U ili y o diagnosis 1-No use ul, 5-Highly use ul 3.70
Cohe ence and plausibili y 1-No sound, 5-Ve y sound 3.74
Expec ed consensus 1-Ve y unlikely, 5-Ve y likely 3.56
Robus ness o easoning 1-No a all plausible, 5-Highly plausible 3.77
Po en ial o misin e p e a-
ion
1-No misleading, 5-Highly misleading 2.38
well in iden i ying ele an linguis ic ea u es (3.85) and main aining logical
soundness (3.74), sugges ing eliable diagnos ic easoning. Pa icula ly no ewo -
hy is he low sco e o po en ially misleading aspec s (2.38), indica ing ha
expe s ound minimal isk o misin e p e a ion in XVocal’s explana ions. This
is c ucial o clinical applica ions whe e accu acy and eliabili y a e pa amoun .
The sys em also demons a ed good alignmen wi h clinical unde s anding (3.63)
and s ong u ili y o suppo ing diagnos ic decisions (3.70). XVocal success ully
iden i ied key speech ma ke s such as inc eased pause du a ions and educed
seman ic luency, connec ing hese ea u es o es ablished AD li e a u e.
4 Abla ion S udy
To assess he con ibu ion o each modali y, we conduc ed sys ema ic expe i-
men s by emo ing componen s and adap ing he ne wo k a chi ec u e acco d-
ingly. Fo each combina ion, we modi ied he dimensions o he inpu laye s o
ma ch he sizes o he ea u e ec o . The usion laye and a en ion mechanisms
we e adjus ed p opo ionally while main aining he co e a chi ec u e design.
Resul s, as shown in Table 3, demons a e he syne gis ic e ec o mul i-
modal usion, wi h ansc ip ion ea u es p o iding he s onges indi idual con-
ibu ion when combined wi h audio embeddings (91.30%). The ansc ip ion
ea u es p o e c ucial, as con igu a ions lacking his componen show educed
pe o mance (84.78%). Acous ic ea u es seems o be he weake modali y, sug-
ges ing hey cap u e complemen a y speech cha ac e is ics. The op imal pe -
o mance (95.77%) achie ed wi h all h ee modali ies indica es each componen
con ibu es unique disc imina i e in o ma ion essen ial o obus AD de ec ion.
8 N. N ampakis e al.
Table 3. Abla ion s udy esul s showing modali y combina ions
Audio Audio Tex
Embed. Fea . T ans. Accu acy[%] F1-sco e[%]
✓ ✓ ✓ 95.77 95.76
✓ ✓ 89.86 89.86
✓ ✓ 91.30 91.29
✓ ✓ 84.78 84.70
5 Conclusion
We p esen ed Neu oXVocal, a no el dual-componen sys em ha ad ances he
s a e-o - he-a in bo h AD de ec ion accu acy and clinical in e p e abili y. Ou
key con ibu ions include: (1) he i s end- o-end amewo k seamlessly in eg a -
ing high-accu acy classi ica ion (95.77%) wi h e idence-based explana ions, (2)
a ans o me -based a chi ec u e enabling supe io c oss-modal usion h ough
common dimensional space p ojec ion, and (3) a specialized RAG-based ex-
plaine alida ed by medical p o essionals o clinical ele ance. Unlike p e ious
app oaches ocusing solely on classi ica ion, Neu oXVocal b idges he c i ical
gap be ween machine lea ning p edic ions and clinical decision-making. Fu u e
wo k will ocus on de eloping a eal- ime in e ence pipeline and implemen ing
s eaming audio p ocessing o immedia e ea u e ex ac ion. We plan o ex end
he sys em wi h an in e ace o clinical deploymen , inco po a ing inc emen-
al lea ning capabili ies o adap o new da a pa e ns. Addi ionally, we aim
o expand he knowledge base wi h con inuous li e a u e upda es and enhance
he RAG componen wi h domain-speci ic p omp enginee ing o mo e a ge ed
explana ions. Fu he alida ion h ough la ge-scale clinical ials will help es-
ablish Neu oXVocal’s e icacy as a p ac ical diagnos ic suppo ool.
Acknowledgmen s. This p ojec has ecei ed unding om he Eu opean Union’s
Ho izon Eu ope esea ch and inno a ion p og amme (G.A. No. 101135800 - RAIDO)
and by UK Resea ch and Inno a ion (UKRI) unde he UK go e nmen ’s Ho izon
Eu ope unding gua an ee (G.N. 10099264). The au ho s ex end hei g a i ude o he
medical expe s eam o P o . Magda Tsolaki o hei in aluable con ibu ion o he
e alua ion p ocess and alida ion o he XVocal componen .
Disclosu e o In e es s. The au ho s ha e no compe ing in e es s o decla e ha
a e ele an o he con en o his a icle.
Re e ences
1. Wo ld Heal h O ganiza ion.: Global s a us epo on he public heal h esponse o
demen ia. WHO P ess, Gene a (2021)
2. Wong, K.K., e al.: Enhance ea ly diagnosis accu acy o Alzheime ’s disease by
elucida ing in e ac ions be ween amyloid cascade and au p opaga ion. In: MICCAI
2023, pp. 1–10. Sp inge , Vancou e (2023).
Neu oXVocal 9
3. Luz, S., Haide , F., de la Fuen e, S., e al.: Nonin asi e au oma ic de ec ion o
Alzheime ’s disease h ough speech analysis. In: F on ie s in Aging Neu oscience,
ol. 15, A icle 1224723 (2023)
4. Luz, S., Haide , F., de la Fuen e, S., F omm, D., MacWhinney, B. (2021).: ADReSSo
Challenge Da ase [Da a se ]. Demen iaBank. Re ie ed Janua y 9, 2025, om
h ps://demen ia. alkbank.o g/ADReSSo-2021/
5. Syed, Z.S., Shah, M.S., Lech, M., e al.: Tackling he ADRESSO Challenge 2021:
The MUET-RMIT Sys em o Alzheime ’s Demen ia Recogni ion om Spon aneous
Speech. In: In e speech 2021, pp. 3780–3784. ISCA, B no (2021)
6. Shah, Z., Sawalha, J., Tasnim, M., e al.: Explo ing language-agnos ic speech ep e-
sen a ions using domain knowledge o de ec ing Alzheime ’s demen ia. In: ICASSP
2023, pp. 1–5. IEEE P ess, Rhodes (2023)
7. Fu, Y., Xu, L., Zhang, Y., e al.: Classi ica ion and diagnosis model o Alzheime ’s
disease based on mul imodal da a usion. Medicine 103(52), (2024)
8. Li, J., Zhang, W.Q.: Whispe -Based T ans e Lea ning o Alzheime Disease Classi-
ica ion: Le e aging Speech Segmen s wi h Full T ansc ip s as P omp s. In: ICASSP
2024, pp. 11211–11215. IEEE P ess, Seoul (2024)
9. Lee, B., Bang, J.U., Song, H.J., e al.: Alzheime ’s disease ecogni ion using g aph
neu al ne wo k by le e aging image- ex simila i y om ision language model.
Scien i ic Repo s 15, 997 (2025)
10. Iqbal, F., Syed, Z.S., Syed, M.S.S., Syed, A.S.: An Explainable AI App oach o
Speech-Based Alzheime ’s De ec ion Using Linguis ic Fea u es. In: ISCA SMM
2024, pp. 1–6. ISCA P ess (2024)
11. Bang, J.-U., Han, S.-H., Kang, B.-O.: Alzheime ’s Disease Recogni ion om Spon-
aneous Speech Using La ge Language Models. In: ETRI Jou nal 46(1), pp. 1–10
(2024)
12. Bae ski, A., Zhou, H., Mohamed, A., Auli, M.: wa 2 ec 2.0: A amewo k o
sel -supe ised lea ning o speech ep esen a ions. Ad ances in Neu al In o ma ion
P ocessing Sys ems 33, 12449–12460 (2020)
13. Rad o d, A., Kim, J. W., Hallacy, C., e al.: Robus Speech Recogni ion ia La ge-
Scale Weak Supe ision. OpenAI (2022)
14. He, P., Gao, J., Chen, W.: DeBERTaV3: Imp o ing DeBERTa using ELECTRA-
S yle P e-T aining wi h G adien -Disen angled Embedding Sha ing. In e na ional
Con e ence on Lea ning Rep esen a ions (ICLR) (2023)
15. Douze, M., Guzh a, A., Deng, C., Johnson, J., Szil asy, G., e al.: The Faiss Lib a y.
IEEE T ansac ions on Big Da a 7(3), 535–547 (2024)
16. Wang, W., Wei, F., Dong, L., Bao, H., Yang, N., Zhou, M.: MiniLM: Deep Sel -
A en ion Dis illa ion o Task-Agnos ic Comp ession o P e-T ained T ans o me s.
Ad ances in Neu al In o ma ion P ocessing Sys ems 33, 5776–5788 (2020)
17. Chung, H. W., Hou, L., Longp e, S., e al.: Scaling Ins uc ion-Fine uned Language
Models. In: P oceedings o he Associa ion o Compu a ional Linguis ics (ACL),
pp. 1–12 (2023)
18. F ase , K.C., Mel ze , J.A., Rudzicz, F.: Linguis ic ea u es iden i y Alzheime ’s
disease in na a i e speech. Jou nal o Alzheime ’s Disease 49(2), 407–422 (2016)
19. Tó h, L., e al.: Speech ecogni ion-based classi ica ion o mild cogni i e impai -
men and demen ia. LNCS 11096, 89–98. Sp inge , Heidelbe g (2018)
20. Ahmed, S., Haigh, A.M., de Jage , C.A., Ga a d, P.: Connec ed speech as a ma ke
o disease p og ession in au opsy-p o en Alzheime ’s disease. B ain 136(12), 3727–
3737 (2013)
