SQUIRREL SEARCH GRADIENT OPTIMIZED DEEP BELIEF NETWORK CLASSIFIER FOR THYROID DISEASE PREDICTION

Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4268 SQUIRREL SEARCH GRADIENT OPTIMIZED DEEP BELIEF NETWORK CLASSIFIER FOR THYROID DISEASE PREDICTION R.VANITHA1, D .K. PERUMAL2 1 Resea ch Schola , Depa men o Compu e Applica ions, Madu ai Kama aj Uni e si y, Madu ai, India 2 P o esso , Depa men o Compu e Applica ions, Madu ai Kama aj Uni e si y, Madu ai, India E-mail: 1 ani hachezian200[email p o ec ed], 2 pe uma[email p o ec ed]m ABSTRACT Thy oid disease is a ange o diso de s ha a ec he hy oid gland, a bu e ly-shaped o gan loca ed in he neck esponsible o p oducing ho mones ha egula e me abolism, ene gy le els, and o e all bodily unc ions. Ea ly de ec ion and managemen o hy oid disease a e c ucial, as un ea ed condi ions leads o se e e complica ions, including ca dio ascula issues, in e ili y, and me abolic diso de s. Ad anced diagnos ic me hods, including machine lea ning and deep lea ning echniques, a e inc easingly used o imp o e he accu acy and imeliness o hy oid disease de ec ion, acili a ing be e ea men ou comes. Bu , se e i y o hy oid disease p edic ion accu acy wi h minimal ime is majo challenging issues. In o de o imp o e he accu acy o hy oid disease p edic ion, a no el Squi el Sea ch G adien Op imized Deep Belie Neu al Classi ie (SSGODBNC) model is de eloped wi h minimal ime consump ion. The p oposed Deep Belie Ne wo k (DBN) is a ully connec ed a i icial eed- o wa d deep lea ning me hod comp ising wo isible laye s such as he inpu and ou pu laye and mul iple hidden laye s o p ocessing he gi en inpu . In he laye -by-laye p ocess, he i s hidden laye ecei es weigh ed inpu and pe o ms da a p ep ocessing. Then ex ac ing signi ican ea u es and elimina es he insigni ican ea u es om he da ase using he Spa se Au oencode model. These selec ed signi ican ea u es a e u ilized o classi y he se e i y le el o hy oid disease using Sokal–Michene ’s simple ma ching me hod. Du ing ine- uning, e o back-p opaga ion algo i hms adjus he hype pa ame e s using Squi el Sea ch G adien Op imiza ion o inc ease he accu acy o hy oid disease classi ica ion. This op imized ine- uning p ocess signi ican ly enhances he pe o mance o he deep belie ne wo k and imp o es o e all lea ning e iciency in classi ica ion asks. Finally, he accu a e hy oid disease se e i y p edic ion esul s wi h minimal e o a e ob ained a he ou pu laye . Expe imen al assessmen is conduc ed wi h di e en e alua ion me ics such as Accu acy, P ecision, Recall, F1-sco e, speci ici y and Thy oid disease p edic ion ime. The obse ed esul shows he e ec i eness o he p oposed SSGODBNC model wi h highe accu acy in hy oid disease p edic ion wi h minimum ime han he exis ing me hods. Keywo ds: Thy oid Disease P edic ion, Deep Belie Ne wo k, Fine-Tuning, Adap i e G adien Me hod, Squi el Sea ch G adien Op imiza ion, Sokal–Michene ’s Simple Ma ching Me hod. 1. INTRODUCTION Thy oid disease poses a subs an ial heal h isk, nega i ely impac ing an indi idual's quali y o li e while also leading o inc eased medical expenses o diagnosis and ea men . Iden i ying hy oid disease pa icula ly challenging, especially o less expe ienced heal hca e p o essionals, as i s symp oms o en o e lap wi h hose o o he condi ions. Recen ad ancemen s in medical esea ch ha e highligh ed he po en ial o machine lea ning echniques as e ec i e ools o diagnosing diseases. By analyzing pa e ns in clinical da a, machine lea ning models suppo s p ac i ione s in making accu a e and imely diagnoses, he eby imp o ing pa ien ou comes and educing he bu den on heal hca e sys ems. The scopes o he DBNs sugges he po en ial pa h o p edic ing hy oid diseases, wi h he aid o ea ly diagnosis and modi ied ea men . I s con ex ex ends o enhance he accu acy and e iciency in iden i ying di e en hy oid condi ions. The DBNs can be ained o o ecas he en i y pa ien esponses o a ious ea men s, allowing o u he pe sonalized and managemen o hy oid diso de and u ilized o de e mine he medical images such as, X- ays and CT scans o hy oid abno mali ies. I Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4269 classi ies and diagnoses hy oid condi ions based on hy oid scin ig aphy images, o e ing ano he ool o diagnosis. A Dynamic Selec ion Hyb id Model (DSHM) was p oposed in [1] o p edic hy oid disease and imp o e accu acy h ough obus ea u e selec ion. Howe e , he DSHM model equi es highe compu a ional ime o hy oid disease p edic ion. A S acked Ensemble wi h IG ea u e selec ion model was designed in [2] wi h he aiming o enhance hy oid disease de ec ion and educe sc eening ime and cos s conside ing ew clinical a ibu es. Bu i ailed o p edic he se e i y le el o hy oid disease p edic ion. Di e en machine lea ning models we e de eloped in [3] o de ec ing hy oid disease, inco po a ing a di e en ial e olu ion (DE)-based op imiza ion algo i hm o ine- une pa ame e s and minimize e o s. Bu , i did no add ess inc easing he da ase size, limi ing he abili y o u he analyze he pe o mance o deep lea ning models. A gene alized deep lea ning-based decision suppo sys em was p oposed in [4] o imp o e hy oid cance diagnosis and enhance o e all diagnos ic pe o mance. Howe e , challenges ela ed o p ecision and ecalls in hy oid cance de ec ion emain un esol ed. To enhance he pe o mance o p ecision and ecall, a no el andom o es -based sel -s acking classi ie model was de eloped in [5] o e icien hy oid disease de ec ion. Howe e , he ime complexi y o hy oid cance p edic ion emained unadd essed. Va ious machine lea ning app oaches we e de eloped in [6] o p edic ing papilla y hy oid cance . Howe e , deep lea ning models we e no u ilized o enhance he accu acy o cance p edic ion while minimizing ime consump ion. A Quan um Suppo Vec o Machine classi ie model was de eloped in [7] o mo e accu a e classi ica ion o hy oid cance by selec ing signi ican ea u es using he Quan um Pa icle Swa m Op imiza ion me hod. Howe e , i ailed o apply e ec i e ea u e selec ion and classi ica ion algo i hms o imp o e he pe o mance o hy oid disease p edic ion and achie e be e accu acy a es. Se e al machine lea ning echniques we e p oposed in [8] o classi ying hy oid disease p edic ions, which include da a p epa a ion, ea u e selec ion, and hype pa ame e uning. Howe e , hese me hods did no add ess he educ ion o ime complexi y in hy oid disease p edic ion. A obus and e ec i e machine lea ning- based me hod was de eloped in [9] o p edic ing hy oid disease by add essing class imbalance and pe o ming ea u e selec ion. Howe e , a ious ea u e selec ion echniques and obus handling o missing da a we e no adequa ely add essed. A ine- uned Ligh G adien Boos ing Machine (LGBM) model was de eloped in [10] o achie e high accu acy in hy oid disease p edic ion. Howe e , i did no inco po a e deep lea ning models o u he enhance he diagnos ic pe o mance o hy oid disease. Di e en machine lea ning algo i hms we e designed in [11] o p edic hypo hy oidism and hype hy oidism by iden i ying he mos signi ican ea u es o dis inguish hy oid diseases mo e accu a ely. Howe e , i ailed o de elop a mo e e ec i e ea u e selec ion scheme o u he imp o e he esul s. An ensemble lea ning model was de eloped in [12] o au oma ic, eliable, and accu a e hy oid ecogni ion wi h he aim o imp o ing p edic ion accu acy. Howe e , i ailed o cons uc a mul iclass hy oid classi ica ion model. A h ee-s age hyb id classi ie (3SHC) model was de eloped in [13] o disease p edic ion by educing he da ase dimension and pe o ming ea u e selec ion. Bu , he designed classi ie model equi es highe compu a ional esou ces and esul s in inc eased compu a ional cos s. An op imized ex eme g adien boos ing mul iclass classi ie model was in oduced in [14] o classi y pa ien s wi h di e en ypes o hy oid disease. Howe e , sophis ica ed deep lea ning models ha e no been applied o achie e e en mo e accu a e and e ec i e ou comes. The eg esso and classi ie model de eloped in [15] aimed o p edic he occu ence o hypo hy oidism by analyzing he ea u es equi ed o classi ica ion. Howe e , i ailed o op imize he model hype pa ame e s o minimize he s a is ical loss unc ions. 1.1 A No el Con ibu ion o The SSGODBNC Me hod The majo con ibu ions o he SSGODBNC model is lis ed as ollows,  To enhance he accu acy o hy oid disease p edic ion, he SSGODBNC model has been de eloped, inco po a ing p ep ocessing, da a ea u e selec ion, and classi ica ion.  A no el y o deep lea ning model pe o ms da a p ep ocessing and ea u e selec ion in he hidden laye using SSGODBNC model designed in o minimize he aining ime o hy oid disease p edic ion Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4270  A no el me hod o Sokal–Michene ’s simple ma ching echnique is de eloped o analyzing he aining and es ing da a samples and p o ides he mul i class classi ica ion ou come o enhance he accu acy wi h minimum e o .  A no el y o squi el Sea ch Algo i hm is ob ained in ine- uning p ocess o op imize he e o a e and imp o e he accu acy o hy oid disease p edic ion.  Finally, an expe imen al e alua ion is ca ied ou o es ima e he pe o mance o he SSGODBNC model using a ious me ics and compa ing i o o he classi ica ion me hods. 1.2 P oblem S a emen The hy oid disease p edic ion is imp o ing he occu ence o hy oid diso de s and equi e o co ec and ea ly de ec ion o enhance he pa ien esul s and minimum heal hca e cos s. The hy oid disease p edic ion [1] designed o enhance accu acy by obus ea u e selec ion. Bu , he DSHM model was no educing he compu a ional ime. The ea u e selec ion model [2] in oduced wi h imp o ed hy oid disease de ec ion wi h lesse sc eening ime and cos s. Howe e , he se e i y le el o hy oid disease p edic ion was no de e mined. The di e en machine lea ning me hods a e designed in [6] o hy oid cance . Bu , i ailed o imp o e accu acy o cance p edic ion wi h educed ime consump ion. To o e come his issue, he p oposed SSGODBNC model achie ed wi h be e accu acy in hy oid disease p edic ion wi h lesse ime han he exis ing me hods. 1.3 O ganiza ion The pape is s uc u ed as ollows: Sec ion 2 p o ides a e iew o ela ed wo ks in he ield, highligh ing issues. Sec ion 3 in oduces he p oposed SSGODBNC model, o e ing a de ailed explana ion along wi h a diag am o be e unde s anding. Sec ion 4 ou lines he expe imen al se up and p o ides a desc ip ion o he da ase used o e alua ion. In Sec ion 5, he pe o mance o he p oposed model is compa ed wi h exis ing me hods, conside ing a ious pa ame e con igu a ions. Finally, Sec ion 6 p esen s he conclusion. 2. RELATED WORKS A Ligh G adien Boos ing Classi ie model was de eloped in [16] o achie e high pe o mance in hy oid cance diagnosis. Howe e , he designed model ailed o be e ec i ely applied in clinical p ac ice o imp o e i s p edic i e accu acy. A new combina ion o K-Neighbo s (KN) and Random Fo es (RF) classi ie models was de eloped in [17] o he e ec i e iden i ica ion o hy oid synd ome. Howe e , i did no inco po a e mo e ad anced neu al ne wo k-based app oaches o u he enhance he pe o mance sco es o hy oid synd ome de ec ion. An ensemble machine lea ning classi ie model was in oduced in [18] o imp o e classi ica ion pe o mance wi h highe speci ici y. Bu , i ailed o ex end he classi ica ion o hy oid disease using an explainable machine lea ning app oach, which could enhance accu acy, anspa ency, and ou comes. A machine lea ning (ML) in eg a ion was de eloped in [19] by applying mul i-c i e ia decision-making o hy oid p edic ion. Howe e , he ime complexi y o he hy oid p edic ion was highe . A hyb id model combining ensemble s acking and an ad anced ea u e selec ion echnique was de eloped in [20] o enhance he accu acy o hy oid diso de de ec ion. Howe e , he pe o mance o sensi i i y analysis in hy oid diso de de ec ion was no add essed. An in e p e able hy oid ca ego iza ion app oach was in oduced in [21] using explainable AI, achie ing he highes accu acy pe o mance. Howe e , i did no apply a mul iclass classi ica ion app oach. A andom o es model was de eloped in [22] o achie e imp o ed p edic ion pe o mance o hy oid papilla y cance . Howe e , he issue o ime consump ion in p edic ing hy oid papilla y cance emained un esol ed. A machine lea ning app oach was de eloped in [23] o p edic di e en ia ed hy oid cance based on hype pa ame e uning. Howe e , i ailed o explo e hese models on la ge and mo e di e se da ase s o alida e he hy oid cance p edic ion. In [24], machine lea ning algo i hms we e designed o p edic medulla y hy oid ca cinoma. Howe e , op imiza ion and addi ional p edic i e ac o s we e no inco po a ed in o he ca cinoma p edic ion. A con olu ional neu al ne wo k (CNN) p edic ion model was de eloped in [25] wi h he aim o de ec ing papilla y hy oid cance , achie ing high sensi i i y and speci ici y. Bu , an e icien op imiza ion model was no applied o u he enhance he hy oid cance p edic ion. A machine lea ning model using eXplainable A i icial In elligence (XAI) was in oduced in [26] o imp o e hy oid disease p edic ion. Howe e , mul i-label hy oid disease p edic ion emained unadd essed. Risk p edic ion models we e de eloped in [27] wi h he aim o p edic ing he ce ical lymph node in ol emen in papilla y hy oid ca cinoma. Howe e , he models Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4271 ailed o alida e he e icacy o hese p edic ion models. An e icien nomog am model was de eloped in [28] by u ilizing isualized mul ipopula ion da a o accu a ely classi y hy oid ca cinoma. Bu , a deep lea ning classi ie model was no applied o enhance he accu acy o hy oid ca cinoma p edic ion. A bina y logis ic eg ession and Lasso eg ession model was de eloped in [29] o a iable selec ion and isk ac o analysis in hy oid disease p edic ion. Howe e , he e o a e in he isk ac o analysis was no e ec i ely add essed. A no el condi ional gene a i e ad e sa ial ne wo k model was de eloped in [30] wi h he aim o de ec ing hy oid disease by ex ac ing mul i-scale ea u es. Bu , i ailed o pe o m an in-dep h analysis o hy oid disease p edic ion. 3. PROPOSAL METHODOLOGY Thy oid disease is a signi ican cause o mo ali y, highligh ing he impo ance o ea ly diagnosis o mi iga e i s impac . Howe e , exis ing me hods in heal hca e diagnosis ace challenges ega ding pe o mance consis ency and accu a e disease p edic ion wi hin minimal ime. This sec ion in oduces a no el me hodology called SSGODBNC, de eloped o accu a e hy oid disease p edic ion. The wo king me hodology o he SSGODBNC model is di ided in o ou p ima y p ocesses namely da a acquisi ion, da a p ep ocessing, ea u e selec ion, and classi ica ion. Figu e 1 p o ides an o e iew o he en i e wo king p ocess o he SSGODBNC model. Figu e 1: A chi ec u e Diag am o SSGODBNC Model Figu e 1 abo e illus a es he a chi ec u e o he p oposed SSGODBNC model, which aims o achie e accu a e hy oid disease p edic ion in medical da a p ocessing. The SSGODBNC model in eg a es a ious undamen al p ocesses ha wo k collabo a i ely o enhance he p edic ion accu acy and e iciency. These p ocesses include da a p ep ocessing, ea u e selec ion, and e alua ion, each playing a c ucial ole in e ining he da ase and imp o ing he pe o mance o he disease p edic ion model. Th ough applying op imiza ion, he model p o ides he be e p edic ion esul s. In he ollowing subsec ions, each o hese p ocesses is explained in de ail, highligh ing hei signi icance in he o e all amewo k o he p oposed model. 3.1 Da a Acquisi ion Da a acquisi ion is he c ucial s ep in he SSGODBNC model ha in ol es ga he ing ele an and eliable da a om he heal hca e da abases namely Thy oid disease da ase ex ac ed om h ps://www.kaggle.com/da ase s/emmanuel we / hy oid-disease-da a. This s ep ensu es ha su icien in o ma ion is a ailable o ain and alida e he model e ec i ely. In he con ex o hy oid disease p edic ion, da a acquisi ion ocuses on collec ing pa ien - ela ed in o ma ion, including clinical ea u es, es esul s, and demog aphic de ails, o c ea e a comp ehensi e da ase o u he p ocessing. Accu a e and high-quali y da a acquisi ion plays a i al ole in enhancing he eliabili y and pe o mance o he p oposed SSGODBNC model. The da ase includes 9172 ins ances o eco ds o da a samples and 31 a ibu es o ea u es o accu a e hy oid disease p edic ion. The 31 a ibu es a e lis ed as ollows, age o he pa ien , sex o pa ien , on_ hy oxine, que y on hy oxine, on an i hy oid meds, sick, p egnan , hy oid_su ge y, I131_ ea men , que y_hypo hy oid, que y_hype hy oid, li hium, goi e , umo , hypopi ui a y, psych, TSH_measu ed, TSH, T3_measu ed, T3 le el in blood om lab wo k ( loa ), TT4_measu ed in he blood, TT4 le el in blood, T4U_measu ed in he blood, T4U le el in blood, FTI_measu ed , FTI le el in blood, TBG_measu ed , TBG, e e al_sou ce, a ge , pa ien _id Le us conside he da ase ‘𝐷𝑆’ and samples as well as ea u es a e a anged in he o m o ma ix. The e o e, he inpu ma ix is o mula ed as gi en below, Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4272 𝐼𝑀 = ⎣ ⎢ ⎢ ⎢ ⎡ 𝐹𝐹… 𝐹 𝑆 𝑆 … 𝑆 𝑆 𝑆 … 𝑆 ⋮ ⋮ … ⋮ 𝑆 𝑆 … 𝑆 ⎦ ⎥ ⎥ ⎥ ⎤ (1) Whe e, 𝐼𝑀 indica es an inpu ma ix, each column indica es a numbe o ea u es𝐹 = {𝐹,𝐹,…,𝐹}, each ow indica es a numbe o samples o ins ances o eco ds ‘𝑆 = {𝑆,𝑆,…,𝑆}’ espec i ely. 3.2 P oposed Deep Belie Neu al Ne wo k The SSGODBNC model employs a Deep Belie Ne wo k (DBN), a specialized deep lea ning model designed o enhance he accu acy o hy oid diseases de ec ion while educing p ocessing ime. This app oach imp o es he ea u e selec ion and classi ica ion, especially o handling he la ge olume o sequen ial da a. A DBN a chi ec u e consis s o mul iple laye s o Res ic ed Bol zmann Machines (RBMs) a anged hie a chically, unc ioning as a gene a i e model. This laye ed a chi ec u e minimized he compu a ional complexi y while handling he la ge olume o da a samples. Addi ionally, he p oposed deep lea ning a chi ec u e e ec i ely minimizes e o s du ing aining, leading o imp o ed o e all pe o mance in he hy oid diseases p edic ion. Figu e 2 depic s he a chi ec u e o a Deep Belie Ne wo k o accu a e hy oid disease p edic ion. The lea ning p ocess is di ided in o wo p ima y p ocesses namely laye -by-laye aining and ine- uning. Du ing he laye -by-laye aining phase, each laye o he Deep Belie Ne wo k p ocesses weigh ed inpu da a samples and ans e ed in o he nex laye . In he ine- uning phase, e o back p opaga ion is Figu e 2 : Cons uc ion o Deep Belie Ne wo k employed o adjus he ne wo k’s hype pa ame e s, by applying a Squi el Sea ch G adien Op imiza ion me hod o enhanced pe o mance. In he laye -by-laye app oach, Deep Belie Ne wo ks u ilizes he Res ic ed Bol zmann Machines (RBMs), which a e s ochas ic neu al ne wo ks. Res ic ed Bol zmann Machines (RBMs) a e a ype o s ochas ic neu al ne wo k used o unsupe ised lea ning, pa icula ly in ea u e ex ac ion and dimensionali y educ ion asks. An RBM consis s o wo laye s such as a isible laye and a hidden laye . The isible laye ep esen s he inpu da a samples, while he hidden laye p ocesses he da a samples. The ou pu gene a ed by one RBM se es as he inpu o he isible laye o he nex RBM, as illus a ed in Figu e 2. As exposed in he igu e 2, he DBNs consis o aining se {𝑆,𝑌} whe e 𝑆 deno es a inpu da a samples 𝑆={𝑆,𝑆,𝑆,…𝑆}’ collec ed om he da ase and a label o ou pu ‘𝑌’ ep esen ing i s ca ego y which belongs o he di e en classes(𝑌∈ 1,2,3…𝑘). The inpu da a samples is associa ed o a weigh ‘ 𝜗,𝜗,…,𝜗’ and added wi h bias ‘𝐵’. The p obabili y o he neu on ac i a ion in he isible laye is gi en below, 𝑃 = 𝐹∑𝑆   ⬚∗𝜗+𝐵 (2) Whe e, 𝑃 deno es a neu on ac i a ion p obabili y in isible laye , 𝐹 symbolizes a sigmoid ac i a ion unc ion, ‘𝑆’ indica es an inpu pa ien da a samples, 𝜗 ep esen s a weigh s in isible laye 𝐵 indica es a bias o isible laye . I he neu on ac i a ion p obabili y 𝑃 = 1 , hen he inpu da a Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4273 samples a e sen in o he hidden laye . In ha laye , da a p ep ocessing is ca ied ou by signi ican ly imp o e he pe o mance o p edic i e models by add essing issues o missing alues in he gi en da ase . 3.2.1 Da a P ep ocessing Da a p ep ocessing is essen ial o ensu ing ha he machine lea ning model lea ns om high-quali y, s uc u ed, and ele an da ase , leading o mo e eliable and accu a e disease p edic ions. In he p ep ocessing s ep, he p oposed SSGODBNC model add esses he missing da a in he gi en da ase h ough he nea es neighbo impu a ion me hod. The i s s ep in ol es analyzing he da ase o ecognize he dis ibu ion o missing alues. A e ha , he nea es neighbo impu a ion me hod is hen applied o inding he missing alues. The alues o he nea es neighbo s a e a e aged o ill in he missing en ies. The missing da a impu a ion p ocess is exp essed as ollows, 𝑆 =∑   ∑   (3) Whe e, 𝑆 indica es a missing da a alues, 𝑆 deno es an obse ed neighbo ing known da a sample alues a ailable in da ase , 𝛿 designa es a weigh s assigned o he neighbo ing known da a sample alues. A e inding he missing da a, he de e mined alues a e e ined by applying a no maliza ion p ocess. I is used o educe he dimensionali y o he da ase and cap u e he mos signi ican alues ha explain he a iance in he da a. This e inemen p ocess s abilizes he e ec o bo h con inuous and ca ego ical a iables which cap u e he unde lying s uc u e and ela ionships wi hin he da a. In his s ep, he mean o each known alue is compu ed as ollows, 𝜇 =  ∑𝑆   (4) Whe e, 𝜇 deno es a mean o each alue, 𝐾 deno es a numbe o neighbo ing da a samples. A e ha , he no maliza ion me hod is applied o escales da a in o s anda d no mal dis ibu ion. 𝑆 = ()  (5) Whe e,𝑆 deno es a no maliza ion o he espec i e missing alues ‘𝑆’ and 𝜇 deno es a mean, 𝜎 deno es a s anda d de ia ion. The missing alues, impu ed wi h he mean alue o minimize de ia ion, ensu e ha he unde lying s uc u e o he da a is accu a ely e lec ed. Finally, hese impu ed alues e ine he da ase , enhancing he accu acy o disease p edic ion while minimizing ime consump ion. 3.2.2 Fea u e Selec ion A e he p ep ocessing, he ea u e selec ion p ocess is pe o med o educe i s dimensionali y. This s ep in ol es iden i ying and e aining he mos impo an ea u es while disca ding less ele an o edundan ones. By ocusing on he mos in o ma i e a ibu es, ea u e selec ion no only simpli ies he da ase bu also enhances he e iciency and accu acy o subsequen modeling asks. This p ocess ensu es ha he model ope a es on a e ined se o ea u es, educing compu a ional complexi y and imp o ing p edic i e pe o mance. The p oposed SSGODBNC model u ilizes he Con e gen P opaga ed Spa se Au o encode model o selec ing he signi ican ea u es by emo ing he o he ea u es. The Spa se Au o encode is a a ia ion o an au o encode neu al ne wo k ha helps o enhance he lea ning o inpu samples and p o ides he ou pu in e ms o compac and meaning ul ep esen a ion. A Spa se Au o encode pe o ms wo majo p ocesses namely o wa d p opaga ion and backwa d p opaga ion o minimize he dimensionali y o he inpu da ase . The p oposed au o encode model conside he p ep ocessing ou pu ‘𝑃𝑂’ as inpu o ea u e selec ion. Fo wa d p opaga ion is used o ind he mos signi ican ea u e ha maximizes he objec i e unc ion’ 𝐽’ as exp essed as ollows, 𝐹= 𝑎𝑟𝑔𝑚𝑎𝑥 𝐽𝑂𝑢𝑡+ 𝑃𝑂,𝑤ℎ𝑒𝑟𝑒 𝑃𝑂 ∈ 𝑂𝑢𝑡, 𝑃𝑂 ∉ 𝑂𝑢𝑡 (6) Whe e, 𝐹 ea u es selec ion a he o wa d p opaga ion, 𝑎𝑟𝑔𝑚𝑎𝑥 deno es a gumen o maximum unc ion, 𝑂𝑢𝑡 deno es a o wa d p opaga ion o he ea u es combined wi h he p ep ocessing ou pu ‘𝑃𝑂’ du ing he ea u e selec ion p ocess. A e he o wa d p opaga ion iden i ies signi ican ea u es, backwa d p opaga ion is employed o elimina e insigni ican ea u es om he selec ed se . This ensu es ha he ea u e se is e ined o e ain only he mos ele an and impac ul ea u es. Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4274 𝐹= 𝑎𝑟𝑔𝑚𝑎𝑥 𝐽𝑂𝑢𝑡− 𝑃𝑂,𝑤ℎ𝑒𝑟𝑒 𝑃𝑂 ∈ 𝑂𝑢𝑡,𝑃𝑂 ∉ 𝑂𝑢𝑡 (7) 𝐹𝑆 = (𝐹 ,𝐹) (8) Whe e, 𝐹 ea u es elimina ion a he backwa d p opaga ion, 𝑎𝑟𝑔𝑚𝑎𝑥 deno es a gumen o maximum unc ion, 𝑂𝑢𝑡 deno es a backwa d p opaga ion ou pu combined wi h he p ep ocessing ou pu ‘𝑃𝑂’ du ing he ea u e selec ion p ocess, 𝐽 deno es a objec i e unc ion, 𝐹𝑆 indica es a inal dimensionali y ea u e se ou pu which includes he signi ican ea u e selec ion and elimina ion. The p ocess con e ges when he o al numbe o ea u es in he hy oid disease da ase is ully e alua ed (bo h selec ed and elimina ed), ensu ing ha he inal ea u e se ‘𝐹𝑆’ con ains only he mos ele an ea u es. This app oach enhances he model's pe o mance by op imizing he ea u e se h ough i e a i e e inemen . The selec ed signi ican ea u es a e ans e ed in o he hi d hidden laye . 3.2.3 Classi ica ion A e he ea u e ex ac ion phase, he classi ica ion p ocess is pe o med o analyze he ex ac ed ea u es om he aining and es ing da ase s. This phase is c ucial o building and alida ing he p edic i e model. The Sokal– Michene ’s simple ma ching me hod is a s a is ical me hod which applied o analyze he aining and es ing da a samples based on he co ela ion measu e. I is ma hema ically compu ed as ollows, 𝑐𝑜𝑟𝑟 (𝑆,𝑆)= 1−| ∆ |  (9) 𝑌 = 𝑐𝑜𝑟𝑟 (𝑆,𝑆) (10) Whe e, 𝑌 deno es an analysis ou comes, 𝑐𝑜𝑟𝑟 (𝑆,𝑆)indica es a co ela ion be ween he es ing da a samples ‘𝑆’ and aining samples ‘𝑆’,’ 𝑛’ deno es a numbe o samples,𝑆 ∆ 𝑆 deno es a de ia ion be ween he samples. Based on he Sokal–Michene ’s simple ma ching me hod, he co ela ion p o ides he simila i y ou comes om ‘0’ o ‘1. The maximum co ela ion esul s p o ide he inal classi ica ion ou comes. 3.2.4 Fine Tuning Fine-Tuning is a i al p ocess in deep lea ning whe e a classi ie model is u he op imized o pe o m a speci ic ask. I is used o e ine he weigh s o he ne wo k o imp o ed classi ica ion pe o mance. In ine uning p ocess, he e o a e is measu ed based on squa ed di e ence be ween he ac ual and p edic ed classi ica ion ou pu as ollows, 𝐸𝑅 = 𝑌 −𝑌 (11) Whe e, ‘𝐸𝑅’ symbolizes he e o a e, 𝑌signi ies he ac ual classi ica ion ou pu , 𝑌 symbolizes he p edic ed classi ica ion ou pu . In o de o minimize he e o , he adap i e G adien me hod is employed o upda e he weigh . 𝜗 = 𝜗 − 𝜂 󰇣  󰇤 (12) Whe e, 𝜗 indica es a new weigh , 𝜗speci ies a cu en weigh , 𝜂 indica es a lea ning a e,  signi ies he i s -o de de i a i e o ind ou a local minimum o a unc ion (i.e. e o a e) by upda ing he cu en weigh ‘𝜗’ In o de o ind op imal weigh alue, Squi el Sea ch Op imiza ion algo i hm is employed o educe he e o and enhance he accu acy o hy oid disease p edic ion. Squi el Sea ch Op imiza ion (SSO) is a me a-heu is ic algo i hm inspi ed by he adap i e o aging beha io o squi els. Du ing wa m wea he , squi els ac i ely glide be ween ees in sea ch o ood esou ces, showcasing dynamic explo a ion pa e ns. In his op imiza ion algo i hm, squi els symbolizes weigh s, while ood esou ces ep esen ed by a i ness unc ion. In colde pe iods, hei ac i i y dec eases as hey conse e ene gy o mee hei basic needs. When he wea he becomes cons uc i e again, he squi els esume hei ac i e o aging and explo a ion beha io s. This cyclic pa e n con inues h oughou he li espan o he squi els, o ming he basis o he op imiza ion p ocess. Fi s , popula ions o squi els (i.e. weigh s) a e ini ialized in sea ch space, 𝜗= 𝜗,𝜗,𝜗,….𝜗 (13) Whe e, 𝜗 deno es a ‘𝑏’ numbe o upda ed weighs. Fo each squi el (i.e. weigh ), he i ness is measu ed based on he e o a e. 𝑓(𝜗) = 𝑎𝑟𝑔𝑚𝑖𝑛𝐸𝑅(14) Whe e,𝑓(𝜗) ep esen s a i ness o weigh , 𝑎𝑟𝑔𝑚𝑖𝑛 indica es an a gumen o minimum unc ion, Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4275 𝐸𝑅 indica es an e o a e o he classi ie model. Based on he i ness es ima ion, he cu en bes weigh is selec ed among he popula ion. Then execu es a di e en beha io s o he squi els as ollows,  New Loca ions Gene a ion Th ough Gliding In his beha io , new loca ions a e gene a ed by mimicking he gliding beha io o squi els, which e lec s hei na u al o aging habi s. This gliding mechanism enables he algo i hm o e icien ly explo e he solu ion space in de ec ion o op imal esul s. 𝑋 = 𝑋+𝐷𝐺∗ |𝑋−𝑋| (15) Whe e, 𝑋 indica es a new loca ion o he squi els, 𝑋 indica es old loca ion o he squi el, 𝐷 deno es a andom gliding dis ance, 𝐺 indica es a gliding cons an , |𝑋−𝑋| indica es a de ia ion be ween he cu en posi ion o squi el ‘𝑋’ and 𝑋’ indica es a bes posi ion o he squi el.  Ve i y Seasonal Moni o ing Condi ion The o aging pa e ns o lying squi els a e signi ican ly in luenced by seasonal a ia ions. To add ess his, a seasonal moni o ing mechanism is implemen ed, ensu ing he algo i hm a oids becoming s uck in local op ima. 𝑍𝑆 = (𝑋− 𝑋) (16) Whe e, 𝑍𝑆 deno es a seasonal cons an , 𝑋 indica es a cu en solu ion, 𝑋 designa es an bes posi ion o squi el. 𝑍𝑆 = [] 󰇡  󰇢∗. (17) Whe e𝑍𝑆indica es a minimum seasonal cons an , 𝐼𝑡𝑒𝑟indica es an i e a ion, 𝐼𝑡𝑒𝑟designa es a maximum i e a ion. When 𝑍𝑆 < 𝑍𝑆 indica ing he end o win e , lying squi els lose hei abili y o na iga e he o es e icien ly and ins ead begin andomly explo ing new loca ions in sea ch o ood. This cycle con inues un il he maximum numbe o i e a ions is achie ed. I no , he p ocess o gene a ing new posi ions and e alua ing seasonal moni o ing condi ions is epea ed. Figu e 3 : Flow Cha o Squi el Sea ch Op imiza ion Figu e 3 demons a es he low diag am o he squi el sea ch op imiza ion o selec ing he op imal weigh wi h minimum e o . As a esul , hen he op imally selec ed weigh s a e used o enhance he disease p edic ion. Finally, he disease p edic ion esul s a e ob ained a he ou pu laye as ollows, 𝑌 = 𝐹 ( 𝛿ℎ∗ ℎ) (18) Whe e 𝑌 indica es a mul iclass classi ica ion ou pu , 𝐹 indica es a so max ac i a ion unc ion, ℎ indica es an ou pu o he p e ious hidden laye , 𝛿 deno es a weigh be ween he hidden and ou pu laye . A so max ac i a ion unc ion ‘𝐹’ in he ou pu laye o mul i class classi ica ion ou pu is o mula ed as ollows. 𝐹 =() ∑()   (19) F om he abo e (19), he so max ac i a ion unc ion is used o make a mul iple classi ica ion esul s, 𝑌 deno es a aw ou pu o he 𝐾 class, 𝐶 deno es a o al numbe o classes. The algo i hmic p ocess o Squi el Sea ch G adien Op imized Deep Belie Neu al Classi ie model is gi en below. // Algo i hm 1: Squi el Sea ch G adien Op imized Deep Belie Neu al Classi ie model Inpu : da ase ‘ 𝐷𝑆 ’ , numbe o ea u es 𝐹 = { 𝐹  , 𝐹  , … , 𝐹  } , samples ‘ 𝑆 = { 𝑆  , 𝑆  , … , 𝑆  } ’ Ou pu : Inc ease he disease p edic ion accu acy Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4276 Begin 1. Collec numbe o ea u es 𝐹 = {𝐹,𝐹,…,𝐹} and samples ‘𝑆 = {𝑆,𝑆,…,𝑆}--- inpu laye 2. Fo each sample 𝑆 3. Fo mula e he neu on ac i a ion p obabili y using (1) 4. End Fo 5. P ep ocessing he da a samples using (3) (4) (5)–[hidden laye 1] 6. Fo each p ep ocessing samples ‘𝑷𝑶’ 7. Pe o m o wa d p opaga ion o ex ac signi ican ea u es as gi en in (6) 8. Pe o m backwa d p opaga ion p ocess o elimina e insigni ican ea u es (7) 9. Ob ain he ea u e se ‘FS’ using (8) 10. End o 11. End o 12. Fo each aining and es ing da a samples 13. Apply Sokal–Michene ’s simple ma ching me hod using (9) 14. Ob ain he classi ica ion esul s (10) 15. End o 16. Fo each classi ica ion ou comes-- 17. Compu e he e o a e ‘𝐸𝑅’ using (11) 18. Apply adap i e G adien me hod o upda e weigh using (12) 19. End o 20. Ini ialize he popula ion o he weigh s 𝜗= 𝜗,𝜗,𝜗,….𝜗 21. Fo each weigh in popula ions 22. Compu e he i ness ‘𝐹’ using (14) 23. While (𝐼𝑡𝑒𝑟<𝐼𝑡𝑒𝑟) 24. Selec he cu en bes using 25. Gene a e new loca ion using (15) 26. Ve i y Seasonal Moni o ing Condi ion using (16) (17) 27. 𝒊𝒇(𝑍𝑆 < 𝑍𝑆) hen 28. Reloca e he sea ch space 29. 𝐼𝑡𝑒𝑟 = 𝐼𝑡𝑒𝑟 +1 30. go o s ep 23 31. Else 32. Find he op imal weigh 33. End i 34. End while 35. Ob ain he inal classi ica ion esul s using (18) (19) wi h so max ac i a ion unc ion a ou pu laye End Algo i hm 1 ou lines he p ocess o p edic ing di e en ypes o hy oid diseases wi h minimal ime consump ion. Fo each inpu da a sample, weigh s and biases a e assigned o he isible laye o he deep belie ne wo k (DBN) a chi ec u e. The inpu is hen ans e ed o he neu ons in he hidden laye , whe e da a p ep ocessing is pe o med o handle missing alues wi hin he da ase . Nex , signi ican ea u es a e selec ed in he subsequen hidden laye . Classi ica ion is ca ied ou in hi d hidden laye using he Sokal–Michene 's simple ma ching me hod o compa e aining and es ing da a samples. Based on his simila i y, di e en ypes o hy oid diseases a e classi ied. A e classi ica ion, a ine- uning p ocess is pe o med using he Squi el Sea ch op imiza ion algo i hm. Ini ially, he numbe o weigh s is de e mined, and a popula ion o squi els ( ep esen ing he weigh s) is ini ialized wi hin he sea ch space. The i ness o each squi el is compu ed based on he classi ica ion e o . The posi ion o each squi el is hen upda ed i e a i ely. This p ocess con inues un il he maximum numbe o i e a ions is eached. Th ough his i e a i e app oach, he Squi el Sea ch algo i hm iden i ies he op imal weigh alues ha minimize he classi ica ion e o . Finally, he p edic ion ou comes a e de e mined by minimizing he classi ica ion e o a he ou pu laye . 4. Expe imen al Se ings In his sec ion, expe imen al e alua ion o he p oposed, SSGODBNC and wo exis ing me hods DSHM [1] and S acked Ensemble wi h IG ea u e selec ion [2] a e implemen ed in Py hon high-le el gene al-pu pose p og amming language. In o de o conduc he expe imen , Thy oid disease da ase is applied and i aken om he h ps://www.kaggle.com/da ase s/emmanuel we / hy oid-disease-da a. This s ep ensu es ha su icien in o ma ion is a ailable o ain and alida e he model Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4283 2840 – 2855. DOI: 10.1109/TAI.2023.3327981 [14] Mona Alnagga , Mohamed Handosa, Tame Medha , M. Z. Rashad, “Thy oid Disease Mul i-class Classi ica ion based on Op imized G adien Boos ing Model”, Egyp ian Jou nal o A i icial In elligence, Volume 2, Issue 1, 2023, Pages 1- 13. DOI:10.21608/ejai.2023.205554.100 8 [15] Munisamy Shyamala De i, Venka esan Dhilip Kuma , Ad ian B ezulianu, Oana Geman and Muhammad A i , “A No el Blunge Calib a ion In elligen Fea u e Classi ica ion Model o he P edic ion o Hypo hy oid Disease”, Senso s, Volume 23, Issue 3, 2023, Pages 1- 30.h ps://doi.o g/10.3390/s23031128 [16] Wojciech Ksiazek, “Explainable Thy oid Cance Diagnosis Th ough Two-Le el Machine Lea ning Op imiza ion wi h an Imp o ed Naked Mole-Ra Algo i hm”, Cance s, Volume 16, Issue 24, 2024, Pages 1-25. h ps://doi.o g/10.3390/cance s16244128 [17] Muhammad Asad Abbas,Kashi Muni ,Ali Raza,Madiha Amjad,Nagwan Abdel Samee ,Mona M. Jamjoom,Zahid Ullah, “A no el me a lea ning based s acked app oach o diagnosis o hy oid synd ome”, PLoS ONE, Volume 19, Issue 11, 2024, Pages 1- 20.h ps://doi.o g/10.1371/jou nal.pone.0 312313 [18] Khandake Mohammad Mohi Uddina, Abdullah Al Mamun, Anamika Chak aba i, Ra id Mos a iz, “An ensemble machine lea ning-based app oach o p edic hy oid disease using hyb id ea u e selec ion”, Biomedical Analysis, Else ie , Volume 1, Issue 3, Sep embe 2024, Pages 229-239. h ps://doi.o g/10.1016/j.bioana.2024.08. 001 [19] Ahmed M. Ali, and Said B oumi, “Machine Lea ning wi h Mul i-C i e ia Decision Making Model o Thy oid Disease P edic ion and Analysis”, Mul ic i e ia Algo i hms wi h Applica ions, Volume 2, 2024, Pages 80–88. h ps://doi.o g/10.61356/j.mawa.2024.26 961 [20] Muhammad A mghan La i , Zohaib Mush aq, Saad A i , Sa a Rehman, Muhammad Fa ukh Qu eshi, Nagwan Abdel Samee, Maali Alabdulha i h, Yeong Hyeon Gu and Mohammed A. Al-masni, “Imp o ing Thy oid Diso de Diagnosis ia Ensemble S acking and Bidi ec ional Fea u e Selec ion”, Compu e s, Ma e ials, & con inua, Volume 78, Issue 3, 2024, Pages 4225-4241.DOI: 10.32604/cmc.2024.047621 [21] Ananda Su adha , Sha min Ak e , F M Ja ed Mehedi Sham a , P onab Ghosh, Xujuan Zhou, Mohd Yamani Idna Bin Id is, Kawsa Ahmed, Mohammad Ali Moni, “Ad ancing hy oid ca e: An accu a e us wo hy diagnos ics sys em wi h in e p e able AI and hyb id machine lea ning echniques”, Heliyon, Else ie , Volume 10, Issue 17, 2024, Pages 1- 17.h ps://doi.o g/10.1016/j.heliyon.2024.e3 6556 [22] Hongxi Wang, Chao Zhang, Qian ui Li, Tian Tian, Rui Huang, Jiajun Qiu & Rong Tian, “De elopmen and alida ion o p edic ion models o papilla y hy oid cance s uc u al ecu ence using machine lea ning app oaches”, BMC Cance , Sp inge , Volume 24, 2024, Pages 1- 12.h ps://doi.o g/10.1186/s12885-024- 12146-4 [23] Elizabe h Cla k, Saman ha P ice, The esa Lucena, Bailey Habe lein , Abdullah Wahbeh and Raed See an, “P edic i e Analy ics o Thy oid Cance Recu ence: A Machine Lea ning App oach”, Knowledge, Volume 4, Issue 4, 2024, Pages 1- 14.h ps://doi.o g/10.3390/knowledge40400 29 [24] Zhen-Tian Guo, Kun Tian, Xi-Yuan Xie, Yu- Hang Zhang, and De-Bao Fang, “Machine Lea ning o P edic ing Dis an Me as asis o Medulla y Thy oid Ca cinoma Using he SEER Da abase”, In e na ional Jou nal o Endoc inology, Hindawi, Volume 2023, Decembe 2023, Pages 1-10. h ps://doi.o g/10.1155/2023/9965578 [25] Zhongzhi Wang, Limeng Qu, Qi ong Chen, Yong Zhou, Hong ao Duan, Bai eng Li, Yao Weng, Juan Su and Wenjun Yi, “Deep lea ning-based mul i ea u e in eg a ion obus ly p edic s cen al lymph node me as asis in papilla y hy oid cance ”, BMC Cance , Sp inge , Volume 23, 2023, Pages Jou nal o Theo e ical and Applied In o ma ion Technology 31s May 2025. Vol.103. No.10 © Li le Lion Scien i ic ISSN: 1992-8645 www.ja i .o g E-ISSN: 1817-3195 4284 1-17.h ps://doi.o g/10.1186/s12885- 023-10598-8 [26] Sumya Ak e , Hossen A. Mus a a, “Analysis and in e p e abili y o machine lea ning models o classi y hy oid disease”, PLoS ONE, Volume 19, Issue 5, 2024, Pages 1- 30.h ps://doi.o g/10.1371/jou nal.pone.0 300670 [27] Qiong Chen, Xiao en Ye, Kangjian Wang and Haolin Shen, “P edic ion o papilla y hy oid me as ases o he cen al compa men : p oposal o a model aking in o conside a ion o he hy oid condi ions”, F on ie s in Endoc inology, Volume 14, 2023, Pages 1-8. h ps://doi.o g/10.3389/ endo.2023.1299 290 [28] Jiaqiang Dan, Jingya Tan, Yao Guo, Yang Xu, Lin Zhou, Junhua Huang, Zhiying Yuan, Xiang Ai, Junyan Li, “Cons uc ion and alida ion o a nomog am o p edic ing la e al lymph node me as asis in pedia ic and adolescen wi h di e en ia ed hy oid ca cinoma”, Endoc ine, Sp inge , Volume 84, 2024, Pages 1088–1096. h ps://doi.o g/10.1007/s12020-024- 03730-6 [29] Jianning Liu, Zhuoying Feng, Ru Gao, Peng Liu, Fangang Meng, Lijun Fan, Lixiang Liu and Yang Du, “Analysis o isk ac o s o papilla y hy oid ca cinoma and he associa ion wi h hy oid unc ion indica o s”, F on ie s in Endoc inology, Volume 15, 2024, Pages 1- 12. h ps://doi.o g/10.3389/ endo.2024.1 429932 [30] Anping Song, Tianyi Li, Xuehai Ding, Mingye Wu and Ren Wang, “TSE-GAN: s ain elas og aphy using gene a i e ad e sa ial ne wo k o hy oid disease diagnosis”, F on ie s in Bioenginee ing and Bio echnology, Volume 12, 2024, h ps://doi.o g/10.3389/ bioe.2024.13307 13