X-INCEPD: Enhancing Incep ion-based
Model o Pa kinson’s Disease P edic ion
wi h Explainable AI
Nikos Tsolakis
Ch is oniki Maga-N e e
S e anos V ochidis
In o ma ion Technologies
In o ma ion Technologies
In o ma ion Technologies
Ins i u e
Ins i u e
Ins i u e
Cen e o Resea ch &
Cen e o Resea ch &
Cen e o Resea ch &
Technology Hellas
Technology Hellas
Technology Hellas
Thessaloniki, G eece
Thessaloniki, G eece
Thessaloniki, G eece
School o In o ma ics
A is o le Uni e si y o
[email p o ec ed]
[email p o ec ed]
Thessaloniki
Thessaloniki, G eece
[email p o ec ed]
Abs ac — Accu a e and in e p e able p edic ion o Pa kinson’s Disease (PD) p og ession is essen ial o
e ec i e clinical decision-making and pe sonalized ca e. Building on IncePD, a deep lea ning amewo k
based on Incep ionTime a chi ec u e, which demons a ed high p edic i e accu acy using wea able senso
da a and clinical assessmen sco e, we p o ide an explainable amewo k. In his pape , we p esen a
signi ican ad ancemen o ha model by in eg a ing explainable a i icial in elligence (XAI) echniques in o
he p edic ion pipeline. Ou enhanced amewo k, X-IncePD, p ese es he empo al modelling s eng h o
he o iginal a chi ec u e, while o e ing in e p e able insigh s in o model p edic ions. We implemen an
enhanced LIME explaine o iden i y key ea u es in model p edic ions. Expe imen al e alua ions show ha
he model no only main ains g ea pe o mance, bu also p o ides anspa en jus i ica ions combined wi h
an easy- o-use p edic ion in e ace.
Keywo ds—XAI, Deep Lea ning, Pa kinson’s Disease, Heal hca e, Explainabili y, AI
I. INTRODUCTION
Pa kinson’s Disease (PD) is a p og essi e neu odegene a i e diso de cha ac e ized by mo o
and non-mo o symp oms ha signi ican ly a ec pa ien s’ quali y o li e [1]. The se e i y o
PD is ypically e alua ed using wo well-es ablished clinical ins umen s: he Mo emen
Diso de Socie y – Uni ied Pa kinson’s Disease Ra ing Scale (MDS-UPDRS) and he
Pa kinson’s Disease Ques ionnai e-9 (PDQ-8) [2,3]. The MDS-UPDRS, an upda ed e sion o
he o iginal UPDRS, was de eloped o add ess limi a ions in he ea lie scale. I is di ided in o
ou sec ions which collec i ely assess a wide spec um o PD symp oms, inscluding non-mo o
expe iences, mo o abili ies and o e all impac on daily unc ioning and emo ional heal h. In
pa allel, he PDQ-8, a concise al e na i e o PDQ-39, cap u es he pa ien ’s sel - epo ed
quali y o li e h ough eigh a ge ed ques ions co e ing mood, physical limi a ions, men al
s a e and daily ac i i ies. Highe cumula i e sco es on he PDQ-8 indica e mo e se e e
impai men . Toge he , hese ools o e obus , alida ed me ics and a e widely u ilized in bo h
clinical e alua ions and PD- ela ed esea ch. Ea ly de ec ion and accu a e moni o ing o
symp om p og ession a e c i ical o imely in e en ion and pe sonalized ea men planning.
Nick Bassiliades
Geo gios Medi skos
School o In o ma ics
School o In o ma ics
A is o le Uni e si y o Thessaloniki
A is o le Uni e si y o Thessaloniki
Thessaloniki, G eece
Thessaloniki, G eece
[email p o ec ed] h.g
[email p o ec ed] h.g
Recen ad ancemen s in Machnie Lea ning (ML) and wea able echnologies ha e anabled he
de elopmen o au oma ed sys ems o PD p edic ion, o e ing p omising a enues o emo e
moni o ing and clinical suppo [4]. In ou p io wo k, we in oduced ncePD, a Deep Lea ning
(DL) model based on he Incep ionTime a chi ec u e, designed o p edic PD se e i y using
ime se ies da a om wea able senso s combined wi h clinical assessmen sco es such as MDS-
UPDRS and PDQ-8 [5,6]. While IncePD achie ed g ea and compe i i e accu ace in cap u ing
disease p og ession pa e ns, i s in e p e abili y emain limi ed – a challenge common o mos
deep lea ning-based medical applica ions. In clinical se ings, black-box models pose
signi ican ba ie s o adop ion, as heal hca e p o essionals equi e anspa en easoning o
us and ac on algo i hmic decisions. This has led o a g owing demand o explainable
a i icial in elligence (XAI) echniques ha can unco e he a ionale behind p edic ions,
ensu ing models a e no only accu a e bu also in e p e able and cilinically aligned.
In his s udy, we in oduce X-IncePD, an enhanced e sion o ou p e ious model ha
embeds explainabili y in o i s a chi ec u e and p edic ion pipeline. By in eg a ing s a e-o - he-
a XAI me hods such as LIME, X-IncePD o e s insigh in o indi idualized decision pa hways.
These capabili ies no only enhance us and anspa ency, bu also acili a e clinical alida ion
o model beha io . We e alua e ou app oach on a mul i-sou ce da ase including wea able da a
and clinical sco es, demons a ing ha X-IncePD main ains high p edic i e pe o mance while
p o iding meaning ul explana ions ha align wi h expe medical unde s anding.
The es o his pape is o ganized as ollows. In Sec ion 2, a li e a u e e iew on
p e ious ela ed wo ks on PD p edic ion and explainabili y amewo ks is p esen ed. In Sec ion
3, he expe imen al se up and e alua ion me hods a e desc ibed. In Sec ion 5, he esul s and
he explainabili y amewo k a e discussed. Finally, Sec ion 6 p esen s he conclusions and
u u e di ec ions.
II. RELATED WORK
Recen esea ch e o s ha e inc easingly ocused on in eg a ing machine lea ning and
XAI o enhance he diagnosis and unde s anding o PD. These me hods no only imp o e
p edic i e pe o mance bu also o e c i ical insigh s in o he unde lying bioma ke s and
disease mechanisms. Bhanda i e al.conduc ed an in eg a i e gene exp ession analysis u ilizng
machine lea ning and XAI o iden i y signi ican genomic ea u es associa ed wi h ea ly PD
diagnosis [7]. A e sano e al. [8] p oposed an XAI-d i en app oach o ea ly PD diagnosis ha
cen e s on mo o symp oms. Thei sys em u ilized mo o beha io da a o p edic disease onse
and moni o pa ien p og ession, wi h an emphasis on model anspa ency o suppo clinical
decision-making. An XGBoos wi h SHAP (Shapley Addi i e exPlana ions) o enhance PD
de ec ion accu acy was in oduced by Roy e al. [9]. Thei wo k highligh ed how ea u e
a ibu ion echniques could inco e he mos in luen ial a iables enabling bo h high accu acy
and in e p e abili y. In a b oade con ex , P iyada shini e al. de eloped a comp ehensi e
amewo k o PD diagnosis ha employs mul iple machine lea ning echniques powe ed by
XAI o analyze MRI imaging da a [10]. Sa a anan e al. de eloped a hyb id deep lea ning
model combining VGG19 and GoogleNe o classi ying spi al and wa e d awings, achie ing
high accu acy in PD p edic ion [11]. Thei in eg a ion o XAI echniques enabled local
in e p e abili y, o e ing aluable insigh s in o how speci ic d awing ea u es in luenced model
p edic ions. An AI-enabled sys em ha le e ages noc u nal b ea hing signals o de ec ing PD
and acking i s p og ession p oposed by Yang e al [12]. The non-in asi e na u e o he da a
and he model’s longi udinal p edic ion capaci y ma k a signi ican s ep owa d passi e
moni o ing solu ions. Reddy e al. e iewed ea ly diagnos ic app oaches using AI and ML,
highligh ing hei capabili y o ex ac sub le bioma ke s missed by con en ional clinical
me hods [13]. Simila ly, Dixi e al. conduc ed a comp ehensi e su ey o AI models o PD
diagnosis, emphasizing he supe io i y o hese me hods in handling complex, non-linea
clinical da a [14]. Finally, Velázquez e al. in oduced an app oach using X- ec o s ex ac ed
om speech o au oma ic PD de ec ion, showing high p edic i e accu acy wi h minimal
ea u e enginee ing. This wo k suppo s he po en ial o oice as a iable bioma ke [15].
III. METHODOLOGY
Recen ad ancemen s in ML and DL ha e demons a ed signi ican p omise in he
p edic ion and diagnosis o a ious medical condi ions, including PD. In his s udy, we le e age
Incep ionTime o de elop models capable o assessing PD se e i y.
A. Da a Acquisi ion and P ep ocessing
This s udy employs da a sou ced om he mPowe Public Resea ch Po al [16], a la ge-
scale clinical obse a ional ini ia i e ocused on PD. The Mpowe s udy ga he ed a
combina ion o senso -de i ed da a and sel - epo ed su ey esponses om a b oad pa icipan
base, all h ough a mobile applica ion pla o m. The s udy was s uc u ed a ound se en dis inc
asks, including physical ac i i ies (e.g., walking, apping, oice, and memo y) as well as
su eys (demog aphics, MDS-UPDRS, and PDQ-8). Guided by medical expe
ecommenda ions, we selec ed ou key componen s o ou analysis: he walking ask, he
demog aphic su ey, and he MDS-UPDRS and PDQ-8 ques ionnai es. The walking ask
comp ises h ee segmen s—ou bound walk, s a iona y es , and e u n walk—du ing which he
sma phone’s buil -in accele ome e and gy oscope eco d h ee-dimensional mo ion and
angula eloci y. These mo ion signals a e le e aged o de ec PD- ela ed mo o impai men s
and o di e en ia e indi iduals wi h PD om heal hy con ols, while also es ima ing disease
se e i y h ough p edic i e modeling o ques ionnai e sco es.
Ex ensi e esea ch has highligh ed he impo ance o da a p ep ocessing o imp o e
he p edic i e accu acy o models es ima ing MDS-UPDRS and PDQ-8 sco es [17, 18]. Du ing
his s age, issues such as missing en ies, noise, and da a inconsis encies a e sys ema ically
add essed. The p ocess began wi h iden i ying and isola ing he mos ele an subse s o he
da ase aligned wi h he objec i es o his s udy. Pa icula a en ion was gi en o he ea men
o missing alues, which o en p esen a majo challenge o Deep Lea ning sys ems. To
enhance e iciency and model in e p e abili y, ea u e selec ion was pe o med, helping educe
dimensionali y and accele a e he aining p ocess. Fo his s udy, we ocused on pa icipan s
who had comple ed bo h ele an su eys and he walking ask, as iden i ied h ough guidance
om he clinical collabo a o s in ol ed. A c i ical aspec o p ep ocessing in ol ed
ans o ming aw senso da a in o a s uc u ed o ma compa ible wi h neu al ne wo k inpu s.
In addi ion, ha monizing he da ase equi ed es ablishing common indexing keys o handle
o e lapping and edundan en ies ac oss di e en da a s eams. Subsequen ly, ime-se ies da a
was segmen ed in o sho e windows using a sliding window echnique—each segmen
spanning 5 seconds (500 samples a a 100 Hz sampling a e) wi h a 50% o e lap be ween
consecu i e windows. Finally, he da ase was di ided in o aining and es ing subse s, wi h
80% alloca ed o aining he models and he emaining 20% ese ed o e alua ion.
B. Model A chi ec u e
The p oposed Ince-PD in [5] model in oduces se e al key a chi ec u al modi ica ions
o he o iginal Incep ionTime amewo k o enhance i s e ec i eness in p edic ing Pa kinson’s
Disease se e i y. The i s majo adjus men in ol es he emo al o he Bo leneck laye wi hin
he incep ion modules. Empi ical indings indica ed ha excluding his laye led o imp o ed
model e iciency wi hou sac i icing accu acy. In con as , esidual connec ions—which link
e e y hi d incep ion block—we e e ained om he o iginal a chi ec u e, as hey we e shown
o suppo be e g adien low and o e all model op imiza ion. Beyond changes o he incep ion
modules hemsel es, he b oade ne wo k a chi ec u e was e ined. Addi ional laye s such as
Ba ch No maliza ion and D opou ( a e = 0.5) we e inco po a ed o imp o e aining s abili y
and p e en o e i ing, espec i ely. Fu he mo e, Rec i ied Linea Uni s (ReLU) we e
adop ed as he ac i a ion unc ion ollowing he ba ch no maliza ion laye s o in oduce non-
linea i y and accele a e lea ning. No ably, ins ead o ending he ne wo k wi h a So max laye ,
he inal ou pu laye also u ilizes a ReLU ac i a ion, which acili a ed as e con e gence and
yielded imp o ed pe o mance in eg ession asks.
To u he enhance gene aliza ion, se e al hype pa ame e s we e ca e ully uned,
including he con olu ional ke nel sizes, ne wo k dep h, and numbe o il e s. The sequence
o ope a ions—modi ied incep ion modules ollowed by no maliza ion, ac i a ion, and
pooling—was capped wi h a 1D Global A e age Pooling laye , educing ea u e dimensionali y
be o e he ou pu s age. Ul ima ely, he ained models we e used o p edic agg ega e sco es
o MDS-UPDRS Pa s I & II and he PDQ-8 ques ionnai e, based on he p ep ocessed senso
da a. A isual ep esen a ion o he p oposed a chi ec u e is p esen ed in Figu e 1.
Fig. 1. IncePD a chi ec u e as p esen ed in [5].
C. Model Pe o mance E alua ion
To assess he p edic i e pe o mance o he p oposed Incep ion-based model, we employed
wo s anda d e alua ion me ics: Mean Absolu e E o (MAE) and Mean Squa ed E o
(MSE). These me ics we e compu ed on bo h a pe -window and pe -pa ien basis o p o ide
g anula and o e all insigh s in o model accu acy. The ma hema ical de ini ions o hese
me ics a e shown below, whe e (y_i ) , y_i a e he p edic ed alue and he ac ual alue.
MAE = 1
𝑛∑|𝑦𝑖− 𝑦𝑖|
𝑛
𝑖=1 (1)
MSE = ∑(𝑦
𝑖−𝑦𝑖)2
𝑛
𝑛
𝑖=1 (2)
To u he alida e ou app oach, we compa ed he pe o mance o he p oposed model agains
a ange o exis ing me hods epo ed in he li e a u e o p edic ing MDS-UPDRS I & II and
PDQ-8 sco es. All models we e e alua ed using he same da ase and a ge labels o ensu e
consis ency ac oss compa isons. De ails o he baseline a chi ec u es used o benchma king
a e p esen ed in he subsequen sec ions.
D. Expe imen al Se up
The Incep ion-based model in oduced in his s udy was de eloped using Py hon 3.8 and
execu ed on a wo ks a ion equipped wi h an In el(R) Xeon(R) Sil e 4210 CPU @ 2.20 GHz.
The neu al ne wo k was implemen ed wi h he Tenso Flow amewo k [19], and aining was
conduc ed o e 10 epochs, allowing he model su icien oppo uni y o lea n meaning ul
pa e ns om he da a and imp o e pe o mance. Model aining was ca ied ou using he i
unc ion, applied o bo h inpu ea u es and hei co esponding labels. Fo op imiza ion, we
employed he Adam op imize , a widely-used adap i e lea ning a e me hod known o i s
balance o e iciency and con e gence speed. Hype pa ame e uning was pe o med h ough
heu is ic expe imen a ion, suppo ed by Tenso Flow’s HPa ams lib a y, allowing o dynamic
adjus men o he model’s con igu a ion.
To e alua e and ine- une he model's pe o mance, we used MAE and MSE as p ima y
e alua ion me ics. These guided he op imiza ion p ocess by helping o iden i y he mos
e ec i e model se up. The objec i e was o achie e he lowes possible e o a es while
main aining a balance be ween compu a ional e iciency and model complexi y. Ou p oposed
con igu a ion was u he benchma ked agains op imized models ained on he same da ase
o ensu e obus compa ison.
IV. RESULTS
The pe o mance o he ini ial Ince-PD model was i s assessed by compa ing i agains se e al
baseline a chi ec u es in oduced in p e ious wo k. Thanks o he combina ion o Incep ion
modules, ReLU ac i a ions, esidual connec ions, and d opou egula iza ion, Ince-PD o e s
imp o ed lea ning e iciency and as e con e gence while main aining lowe compu a ional
o e head. Despi e ca e ul uning o he compe ing models, Ince-PD consis en ly ou pe o med
hem ac oss bo h e alua ion me ics—Mean Absolu e E o (MAE) pe window and pe
pa ien — o bo h he o al MDS-UPDRS I & II and PDQ-8 sco es. The model achie ed a MAE
o 1.97 (window) and 2.27 (pa ien ) o MDS-UPDRS I & II, and 2.17 (window) and 2.96
(pa ien ) o PDQ-8. These esul s e lec a signi ican imp o emen compa ed o s anda d
CNN, LSTM, and hyb id a chi ec u es.
IV. X-INCEPD EXPLAINABILITY FRAMEWORK
The anspa ency and clinical in e p e abili y in deep lea ning–based Pa kinson’s Disease
assessmen is an impo an aspec in mode n day clinical se ings. We in oduce X-IncePD—
an explainabili y amewo k buil a ound he IncePD model. X-IncePD in eg a es
complemen a y echniques o p o ide meaning ul insigh s in o how p edic ions a e made,
suppo ing bo h scien i ic alida ion and po en ial clinical deploymen . In his chap e , we
p esen he key componen s o X-IncePD, including:
• LIME-based local explana ions wi h ime-axis ea u e mapping,
• Saliency maps highligh ing inpu sensi i i y ac oss ime and axes,
• Combined isualiza ions me ging a ibu ion and g adien insigh s,
• An in e ac i e use in e ace o eal- ime explo a ion o p edic ions.
A. Co e Me hodology And Resul s
To enhance in e p e abili y in o ou deep lea ning amewo k, we applied he Local
In e p e able Model-Agnos ic Explana ions (LIME) me hodology o explain he p edic ions
made by he IncePD model [20]. LIME is a pos -hoc explainabili y echnique ha app oxima es
he local beha io o complex models by i ing an in e p e able su oga e model— ypically a
linea eg esso —a ound a speci ic p edic ion ins ance. Gi en ha ou model ope a es on
mul i a ia e ime-se ies da a wi h a shape o (500, 3), co esponding o 5-second windows o
i-axial accele ome e signals (a 100 Hz), he aw inpu was i s eshaped in o a 2D o ma
compa ible wi h LIME. Each ins ance was la ened in o a single ec o o 1500 ea u es,
enabling he use o he LimeTabula Explaine designed o abula da a. Α cus om p edic ion
unc ion o eshape he la inpu ec o s back in o he o iginal h ee-dimensional o m expec ed
by he model was de ined, allowing LIME o p obe he model’s in e nal decision unc ion
wi hou al e ing i s a chi ec u e. The explaine was ini ialized using a ep esen a i e subse o
he la ened es da ase and con igu ed in eg ession mode, app op ia e o he con inuous
ou pu s o MDS-UPDRS I & II and PDQ-8 sco e p edic ions.
Fo each p edic ion ins ance, LIME gene a ed a se o pe u bed samples a ound he da a poin
o in e es and eco ded he model’s esponse o hese a ia ions. Using hese obse a ions, i
ained a local linea model o es ima e he impo ance o each ea u e. This p o ided ea u e
a ibu ion sco es ha indica ed he ela i e con ibu ion—posi i e o nega i e—o each ime-
se ies componen o he model’s inal ou pu . No ably, we isualized hese esul s o iden i y
which senso channels and empo al segmen s in luenced he p edic ion, p o iding clinically
meaning ul in e p e a ions aligned wi h known symp om pa e ns in Pa kinson’s Disease.
Figu e 2 p esen s he esul s o he explaine , p o iding meaning ul explana ions o he IncePD
model.
Fig. 2. Fea u e impo ance in he IncePD model.
Table 1 p esen s he ea u e impo ance summa y o a single p edic ion ins ance, as gene a ed
by he LIME explaine . Each ow co esponds o a ea u e ( om he la ened ime-se ies inpu ),
wi h i s associa ed alue used in he p edic ion. Fea u es highligh ed in blue indica e nega i e
con ibu ions, meaning hey dec eased he p edic ed se e i y sco e, while ea u es in o ange
ep esen posi i e con ibu ions, sugges ing an upwa d in luence on he ou pu . Fo example,
ea u e1165 and ea u e408 exhibi s ong nega i e in luence wi h alues o −0.05 and −1.01
espec i ely, while ea u e583, wi h a ela i ely small posi i e alue o 0.12, sligh ly inc eased
he inal p edic ion.
TABLE I. FEATURES AND VALUES UTILIZING X-INCEPD
Fea u e
Value
Fea u e1165
-0.05
Fea u e1203
-1.01
Fea u e492
-0.95
Fea u e408
-1.01
Fea u e583
0.12
B. T acing Back Fea u es in X-IncePD
A co e ad an age o X-IncePD is i s abili y o b idge deep model easoning wi h human-
in e p e able mo emen pa e ns by acing explana ions back o he aw senso signals cap u ed
du ing he walking asks o he mPowe s udy. Each model inpu is a 500×3 ma ix, ep esen ing
500 ime s eps o i-axial accele ome e da a (x, y, z) collec ed om pa icipan s’ sma phones.
Fo compa ibili y wi h local su oga e explaine s, his ma ix is la ened in o a 1500-
dimensional ec o . Howe e , we main ain a p ecise and in e p e able mapping om each
la ened ea u e index back o i s o iginal senso con ex . This mapping allows us o in e p e
any iden i ied in luen ial ea u e (e.g., ea u e1348) in e ms o a speci ic momen in ime and
a speci ic axis o mo ion (e.g., ime s ep 449, y-axis), hus econnec ing abs ac model in e nals
o eal, physical mo emen .
C. Saliency-based Enhancing o Explainabili y
To u he in e p e he model’s in e nal beha io , we employed saliency map analysis, a
g adien -based me hod ha quan i ies he impo ance o each inpu alue wi h espec o he
model’s p edic ion. By compu ing he g adien o he p edic ed ou pu wi h espec o he inpu
ea u es, we gene a ed a 2D saliency hea map ha e eals he mos in luen ial ime s eps and
senso channels in a gi en inpu window.
Figu e 3 p esen s he saliency hea map o a single sample (Sample 0) om he es se , co e ing
a 5-second ime window o i-axial accele ome e signals (500 ime s eps a 100 Hz). The y-
axis co esponds o he h ee senso axes: 0 = X, 1 = Y, and 2 = Z, while he x-axis ep esen s
he empo al dimension. Each pixel's in ensi y e lec s he absolu e g adien magni ude (i.e.,
|∂y/∂x|), whe e b igh e (yellow/whi e) alues indica e highe sensi i i y—i.e., inpu egions
ha had a g ea e impac on he model's p edic ion. The hea map clea ly shows disc e e bu s s
o high saliency, pa icula ly concen a ed in he X and Y axes wi hin he i s 200 ime s eps.
This sugges s ha ea ly pa s o he signal—likely ep esen ing he ini ial mo emen o gai
ansi ion in he walking ask—play a pi o al ole in he model’s decision. In con as , he Z-
axis gene ally con ibu es less ac oss mos o he sequence, which may indica e lowe ele ance
o e ical mo ion in dis inguishing Pa kinsonian pa e ns.
Fig. 3. Saliency hea map o a single sample.
D. Fea u e A ibu ion Visualiza ion
To con ex ualize he mos in luen ial ea u es in he p edic ion p ocess, we isualized he aw
accele ome e signal co esponding o a 5-second window (500 ime s eps) om Sample 0,
highligh ing he impac ul ea u es. As shown in Figu e 4, he plo displays all h ee senso
axes: X (blue), Y (g een), and Z (o ange). These signals ep esen he pa icipan ’s linea mo ion
as cap u ed by he i-axial accele ome e du ing he walking ask.
Key poin s o in luence, as iden i ied by X-IncePD, we e o e laid as dis inc sca e ma ke s o
indica e he op 10 ea u es con ibu ing mos signi ican ly o he model's ou pu o his
ins ance. No ably, hese poin s align wi h dis inc luc ua ions o ansi ions in he senso
eadings—o en co esponding o gai e en s such as oo s ikes, di ec ional changes, o
changes in mo emen in ensi y.
Fig. 4. Raw i-axial accele ome e signals o Sample 0, showing X (blue), Y (g een), and Z (o ange) axes ac oss 500 ime s eps..
E. Uni ied In e p e abili y Visualiza ion
F.
To o e a holis ic iew o model in e p e abili y, we de eloped a combined isualiza ion ha
o e lays bo h LIME ea u e a ibu ions and saliency-based ele ance sco es on o he o iginal
i-axial accele ome e signals. This app oach emphasizes he deg ee o which hose alues
a ec ed he p edic ion. The isualiza ion enables side-by-side alida ion o insigh s de i ed
om local pe u ba ion-based (LIME) and g adien -based (saliency) me hods. As shown in
Figu e 5, he o iginal senso signals (X: ed, Y: g een, Z: blue) o a ep esen a i e sample a e
plo ed o e 500 ime s eps (co esponding o a 5-second walking ask). The backg ound
shading ep esen s saliency magni ude: mo e sa u a ed e ical bands indica e ime-axis egions
whe e he model is mos sensi i e o changes in inpu , calcula ed as he absolu e g adien o he
ou pu wi h espec o each ime–senso alue. O e laid on his a e he op 10 LIME ea u es,
depic ed as sca e poin s. These poin s ep esen he mos impac ul inpu ea u es iden i ied
h ough LIME’s locally linea su oga e model. Each ea u e is mapped back om he la ened
inpu o i s o iginal ime s ep and axis, allowing o di ec isualiza ion wi hin he empo al
signal con ex .
Fig. 5. Combined in e p e abili y iew o Sample 0. Raw i-axial accele ome e signals a e shown in ed (X), g een (Y), and
blue (Z). Saliency in o ma ion is ep esen ed by backg ound colo bands, whe e da ke hues indica e highe model sensi i i y o
inpu a ia ion.
G. In e ac i e Local In e p e abili y Tool
To enhance model anspa ency and acili a e human-cen e ed e alua ion, a isualiza ion
ool was de eloped. Buil upon he LIME amewo k, his ool allows use s o explo e how
indi idual inpu ea u es in luence he p edic ed MDS-UPDRS sco es gene a ed by he IncePD
model. As illus a ed in Figu e 6, use s can selec a speci ic es sample ia an in ui i e slide
and ecei e an immedia e p edic ion along wi h a co esponding local explana ion. Fo he
selec ed sample (e.g., index 17), he model p edic ed a UPDRS sco e o 12.23. The adjacen
ba cha displays he mos in luen ial ea u es a ec ing his p edic ion, sepa a ed in o posi i e
con ibu o s (g een ba s) and nega i e con ibu o s ( ed ba s). These con ibu ions a e
exp essed as ea u e condi ions (e.g., ea u e1383 > 0.92), along wi h hei impac magni ude.
The esul s o such explana ions can be he inpu in de eloping e alua ion amewo ks as
p esen ed in [21].
Fig. 6. The X-IncePD Explaine in e ace showcasing a sample p edic ion and i s co esponding explana ion. The ool allows
use s o selec a sample, iew he p edic ed MDS-UPDRS sco e and in e p e which inpu ea u es (shown as condi ional ules)
mos in luenced he p edic ion ou come.
Figu e 7 illus a es he ea u e impo ance a e iden i ying he ea u es based on he
me hodology p esen ed in sec ion B o his chap e . In his example (Sample 42), we u he
enhanced in e p e abili y by acing each LIME-iden i ied ea u e back o i s co esponding
ime s ep and senso axis, allowing us o p esen he explana ion in a mo e meaning ul o ma .
Ins ead o abs ac ea u e indices, he explana ion is now exp essed in e ms o eal signal
componen s (e.g., "Time 426, Axis X"), helping o di ec ly ela e in luen ial segmen s o
physical mo emen pa e ns cap u ed du ing he walking ask. This mapping suppo s mo e
in ui i e unde s anding and acili a es clinical in e p e a ion.
Fig. 7. In e p e able Explana ion wi h Time-Axis Mapping o Sample 42.
V. CONCLUSION
In his wo k, he IncePD model, a deep lea ning model based on he Incep ionTime a chi ec u e,
designed o he p edic ion o Pa kinson’s Disease se e i y using i-axial accele ome e signals
was enhanced wi h an explainabili y amewo k. The ini ial model was ained o es ima e
clinically alida ed ou comes including o al MDS-UPDRS Pa s I & II and PDQ-8 sco es and
demons a ed g ea esul s. To add ess he c i ical need o anspa ency in medical AI sys ems,
we de eloped X-IncePD, an explainabili y amewo k ha su ounds he IncePD model wi h
obus in e p e abili y ools. X-IncePD le e ages bo h LIME and g adien -based saliency maps
o p o ide ea u e-le el insigh s in o he model’s p edic ions. These echniques expose which
pa s o he inpu signal (in ime and senso axis) we e mos in luen ial, enabling a laye o pos -
hoc in e p e abili y ha b idges he gap be ween black-box AI and clinical easoning. In
addi ion, X-IncePD includes a use - iendly, in e ac i e isualiza ion in e ace, allowing
clinicians and esea che s o explo e indi idual p edic ions and hei co esponding
explana ions. Buil wi h G adio, his ool p esen s local LIME explana ions and p edic ed sco es
in eal ime, making model ou pu s accessible and ac ionable o heal hca e p o essionals—
e en hose wi hou echnical AI expe ise. This ease o use makes X-IncePD a aluable asse
in acili a ing us and adop ion o AI ools in clinical p ac ice.
While he p edic i e pe o mance o IncePD is well alida ed, he e alua ion o he X-IncePD
amewo k i sel —in e ms o clinical u ili y, explana ion quali y, and us wo hiness—
emains a key di ec ion o u u e wo k. Fo mal s udies in ol ing clinical expe s, quali a i e
assessmen s o in e p e abili y, and he inco po a ion o addi ional explana ion me ics will be
essen ial o assess and e ine i s e ec i eness in eal-wo ld deploymen s. In conclusion, IncePD
p o ides an accu a e and e icien model o Pa kinson’s Disease assessmen , and X-IncePD
enhances i s anspa ency h ough ac ionable, human-cen e ed explana ions. Toge he , hey
ep esen a p omising s ep owa d in e p e able and deployable AI ools in he neu ological
domain.
