International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
Multimodal Parkinson’s Disease Detection Using Handwriting and Voice Analysis with Auto-Weight Fusion
Yash Jain1 , Tejaswi2 , Swarnit Kumar3 , Shubham Kumar4, Dr. Savitha G5
1 Student, Department of Computer Science Engineering R V Institute of Technology and Management, Bengaluru, India
2 Student, Department of Computer Science Engineering R V Institute of Technology and Management, Bengaluru, India
3 Student, Department of Computer Science Engineering R V Institute of Technology and Management, Bengaluru, India
4 Student, Department of Computer Science Engineering R V Institute of Technology and Management, Bengaluru, India
5 Associate Professor, Department of Computer Science Engineering R V Institute of Technology and Management, Bengaluru– 560076, Karnataka, India ***
Abstract - Parkinson’s disease constitutes a progressive neurodegenerative disorder that persists long-term, which impairs both motor control and coordination and speech production abilities. The process of achieving effective treatment requires early diagnosis and precise diagnosis to enhance patient outcomes. The basic diagnostic methods depend on clinical assessments and neurological tests, which requiresignificanttime fortheirsubjectivenature toproduce results. Automatic systems have been developed to perform more dependable detection work. This study introduces a multimodal machine learning framework that detects Parkinson’s disease by analyzing handwriting images and voice-based acoustic features. The handwriting modality identifies motor disabilities, which researchers investigate throughaMobileNetV2transferlearningmethodthatenables accurate feature extraction. The voice modality employs acoustic attributes that the XGBoost model classifies after applying preprocessing operations to scale features and eliminate unnecessary elements. The system uses an autoweight optimizationfusionmethod tomerge both modalities, which calculates the optimal model contribution through classification success measurement. The handwriting model reached87.6percentaccuracy,whilethevoicemodelachieved 95.8percentaccuracyaccordingtoexperimentalfindings.The fusion model reached an overall accuracy of 89.3 percent, which showed better stability and equal performance across different tests. The proposed system delivers a non-invasive method that can detect Parkinson’s disease early while providing decision support for clinical purposes through its economical and expandable design.
Key Words: Parkinson, Multimodal, Handwriting, Voice Signal, MobileNetV2, XGBoost, Fusion of Features.
1. INTRODUCTION
Parkinson disease is a chronic and progressive neurodegenerative disorder, which involves the central nervous system and causes severe motor control,
coordination, and speech production dysfunction. It is mainlybroughtaboutbythedeteriorationofneuronsthat produce dopamine in the brain, which interferes with the normal operation of motor pathways. A variety of motor symptomsareassociatedwiththedisease,suchastremor, rigidity,bradykinesia(slownessofmovement),andpostural instability. Along with these motor impairments are nonmotor symptoms that are common in patients like voice difficulties,lackofspeechclarity,andcognitivechanges.The symptomsincreasewiththeprogressionofthediseaseand thus early detection and intervention are important in preventingdiseaseprogressionandimprovingtheoutcomes ofthepatient.
Conventional diagnostic methods of Parkinson disease are based on clinical observation, neurological examinationandpatienthistory.Althoughthesetechniques arecommonplace,theytendtobesubjectiveinnatureand greatly rely on the knowledge and experience of medical practitioners.Addition ally,thesymptomsofthediseaseat the initial stages can be insidious and hard to detect, resulting in a delay in diagnosis and treatment. This constraint explains why automated, objective and databased systems of diagnostic should be developed to help cliniciansmakemoreappropriateandtimelydecisions.
Over the past few years, machine learning and artificialintelligencehaveadvanced,whichhascreatednew opportunitiesin thecreationof suchdiagnostic tools.The methods facilitate the study of intricate biological signals and patterns that could be difficult to identify using conventionaltechniques.Oneofthedatamodalities,which havereceivedalotoffocus,ishandwritingandvoicesignals, as they are highly correlated with the symptoms of Parkinsondisease.Theanalysisofhandwritingcanalsohelp identifymotorimpairment,asitiscommontofindthatthe patient has tremors, deviant stroke patterns and slow writing. Likewise, voice analysis can detect vocal impairmentslikeadecreaseinpitchvariation,hoarseness,
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
and articulation, which are typical in patients with Parkinson’sdisease.
Though a number of studies have shown the usefulnessofeitherhandwritingorvoicedataalone,theuse of a single modality might not be able to capture the numerousmanifestationsofthedisease.Thesymptomsof Parkinsonarediverseamongdifferentpeopleandunimodal approachmightnotidentifysomepatternandhencethelow reliabilityindiagnosticresults.Thishasbeenthereasonthat has led to the exploration of multimodal frameworks that bringtogetherseveralsourcesofinformationtogiveamore holisticevaluation.
Here, the current work suggests a multimodal machinelearningmodel,whichintegrateshandwritingand voicedatatodetectParkinsondisease.Thetransferlearning method with MobileNetV2 is used to analyze the handwritingmodalityasitfacilitatestheextractionofhighlevelspatialfeaturesofimagedataefficiently.Inthevoice modality, an XGBoost classifier will be used to model the acousticfeatures,whichtakeadvantageofitscapabilityto usecomplexrelationshipsinstructureddata.Moreover,an auto-weight optimization fusion method is proposed to mergeforecastsofthetwomodels.Thisisdoneinadynamic waysuchthattheoptimalcontributionofeachmodalityis establishedresultinginabalancedandrobustclassification result.
The main goal of this paper is to improve the accuracy, reliability and strength of the detection of Parkinsondiseasebasedoncomplementaryinformationof various modalities. The proposed system will be used to offer an affordable, scalable, and non-invasive motor and vocalbiomarkersystemcapableofaidinginearlyscreening andhelpingmedicalspecialiststomakeclinicaldecisions.
2. LITERATURE SURVEY
Withthehighrateofdevelopmentofthecomputational intelligence and medical technologies, the process of diagnosis of neurodegenerative diseases, especially ParkinsonDisease(PD)hasundergoneaprofoundchange. PDisaprogressiveneurodegenerativediseasethatresultsin motorandnon-motorsymptoms(e.g.cognitiveimpairment andspeechdifficulties)duetodegenerationofdopaminergic neurons causing motor symptoms in the form of tremors, rigidity,andgaitimpairment.Earlydiagnosisisacomplicated issue because of its non-homogenous character and slow development. This has prompted many studies aimed at creating automated, data driven solutions to proper and promptdetection.
EarlystudiesonPDdetectionwerebasedontheuseof conventionalmachinelearningalgorithmsandhand-crafted features.Physiologicalcues,includinggait,speechpatterns, and handwriting, were the main cues analyzed by these methods.
Gait analysis is a diagnostic measure that has been extensively explored in sensor-based time-series measurements.Walkingpatternswerecapturedbyapplying models like Artificial Neural Networks (ANN), linear regressionandgradientboostingtechniques.AlthoughANN and the boosting-based models were found to be more accurateinprediction,simplermodelswerefoundtobemore explainableandtherewasatrade-offbetweenaccuracyand explainability.
Likewise,speechanalysishasalsobeenwidelyemployed indetectingPD.Acousticfeaturesthatwereextractedinthe voicerecordingsincludedpitchvariation,jitter,shimmerand harmonic-to-noiseratio.AlgorithmssuchasSupportVector Machines(SVM)andRandomForestswereusedtoclassify these features. These methods were rather moderate in termsofaccuracybuttheperformancewashighlydependent onfeatureselection,whichconstrainedtheirrobustnessand generalizability.Anothermethodthatisfrequentlyemployed is handwriting analysis since PD patients tend to have micrographia, or a small writing size and non-uniform patterns. The use of machine learning to recognize early symptoms in spiral and wave drawings has demonstrated good results. Nonetheless, these methods tend to have problemswithvariabledataandtheyarenotscalabletorealworldsituations.
With the advent of deep learning, PD detection has greatlyenhancedwiththeautomaticextractionoffeaturesin raw data. In contrast to the classical approaches, deep learningmodelsdonotrequiremanualfeatureengineering andhavethecapabilitytoidentifyintricatepatternsindata.
Convolutional Neural Networks (CNNs) have been popularly used in medical image, spectrogram, and handwriting analysis. These models are able to capture spatialandtemporaldependenciesandhencecanbeusedto detect subtle abnormalitiesrelatedto PD.In speech-based detection,voicesignals(spectrogram representations)have beenpresentedtoCNNmodels.Suchrepresentationsretain both frequency and time information, enabling models to detectsubtledeficienciesinthevoice.Thecurrentliterature with pre-trained CNN structures, including DenseNet, MobileNet,andShuffleNethasshownhighclassificationrates withamajoritybeingabove95percent,whichillustratesthe powerofdeeplearninginspeechrecognition.
Handwriting-based diagnosis has also been effectively implemented with deep learning. The high accuracy of transferlearningmethodsbasedonpre-trainedmodelslike ResNet and GoogLeNet take advantage of pre-learned features.Theperformancehasbeenfurtherenhancedbythe useofhybridmodelsthatcombineseveralarchitecturesby exploitingcomplementaryfeaturesandtheeffectivenessof deepfusiontechniquesisproved.
Single-modalitymethodshavedemonstratedencouraging outcomes, but they tend to be inadequate to describe the
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
whole complexity of the Parkinson Disease. The recent studies have thus been on multimodal techniques that combinevariousdatasetslikeclinicaldata,imaging,speech andgeneticdata.
To model relationships between patients and predict dis ease progression, graph-based models have been presented.Themodelscombineseveraldatamodalitiessuch asMRIscansandclinicalteststomakeindividualpredictions. These methods have demonstrated better capability to capture the heterogeneity and progression patterns of diseases.
Thecombinationofdeeplearningandclassicalmachine learningmodelshasbeenconsideredaswell.Deeplearning models are employed in feature extraction and classical algorithmsinclassificationinthesesystems.Thepurposeof this combination is to obtain high accuracy and better interpretability which is one of the greatest limitations of deeplearning.
Oneofthemajorflawsofsinglemodelsisthefactthat theycannotdescribeallthepertinentaspectsofPD.Inorder to deal with this, feature fusion methods and ensemble learningmethodshavebeenwidelyused.Infeaturefusion, theoutputofseveralmodelsisusedtocomeupwithamore detailed model of the data. This methodology allows to combine the complementary information of various architectures, which results in better generalization and strength.
Ensemblelearningalgorithmslikebagging,boostingand stacking also improve the performance of the model by combining the results of multiple models. XGBoost and gradient boosting algorithms have demonstrated good performancetomodeldetailedpatternsofgaitandspeech data. These methods enhance predictive accuracy and minimizemodelvariance.
AlthoughthePDdetectionhasachievedalot,thereare still several challenges. The first and one of the biggest challenges is the heterogeneity of data because symptoms differ greatly among patients. This inconsistency causes modelstobeunabletogeneralizeacrossdatasets.Theother problem that is critical is the lack of big and quality data. Smalldatasetsareusedinmanystudies,whichcanleadto overfittinganddecreasemodelreliability.Dataaugmentation anddatatransferlearningareeffectivemethodstoalleviate thisproblembuttheyarenotthesolutiontoit.
Also,deeplearningmodelstendnottobeinterpretable. Althoughtheyareveryprecise,theirdecisionmakingishard tofollowandhencetheycannotbeusedinclinicalsettings wheretransparencyisavitalrequirement.
Moreover,muchofthecurrentsolutionsconcentrateon only one type of data, which is not the comprehensive measureofPD.Whereasmultimodalapproachesresolvethis drawback, they need more complicated architectures and increasedcomputationresources.Trendsinrecentresearch
point towards the change toward personalized and noninvasive methods of diagnosis. The modeling of patientspecific disease progression is becoming more commonly done using graph-based models and deep learning architectures. Voice analysis and wearable sensors are increasinglyusedastheyarenon-invasiveandcost-effective. Wearables can be constantly used in monitoring motor symptoms, which is useful in early detection and disease management.PDdetectionsystemsarealsobeingequipped with explain able Artificial Intelligence (XAI) methods to enhance interpretability. Attention mechanisms and visualization techniques are some of the methods used to identify significant factors that contribute to model predictions.
Also, the transfer learning is still having a vital role in enhancing the performance of models particularly with limited data. Pre-trained models can be used to extract featuresefficientlyandenhanceaccuracywithouttheneedto usebigdatasets.
Inshort,theliteraturepresentsashiftinthemethodsof traditionalmachinelearningtowardsthesophisticateddeep learning and multimodal systems of detecting Parkinson’s Disease. Although the initial approaches were based on manualfeatureextraction,recentapproachestakeadvantage of automated feature learning and data integration. Multimodal, ensemble learning, and feature fusion have become important performance improvement strategies. Nevertheless, data variability, small datasets, and noninterpretabilityareissuesthatshouldbeovercometomakeit possibletoapplytheapproachinpractice.
3. PROPOSED METHODOLOGY
The given system is created asa multimodal model of the ParkinsonDisease(PD)detectionthatintegratestheanalysis ofimagesofhandwritingandtheclassificationofthevoicebased acoustic features. The system is a combination of concepts of both deep learning and machine learning to capture motor and vocal impairments that are associated withPD.
3.1 System Overview
Theoverallarchitecturewillhaveaparallelprocessing pipeline where handwriting and voice data is processed in dependently and the results combined using a fusion strategy.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
3.2 Handwriting-Based Detection
Thehandwritingmodalitycapturesmotorimpairments such as tremors, irregular strokes, and reduced writing precision.
1) Data Preprocessing: Input images are resized to 224×224 pixels and normalized. Grayscale images are con verted to three-channel format to match the requirementsofpre-trainedconvolutionalneuralnetworks.
2)ModelArchitecture:AbasemodelisMobileNetV2and atransferlearningtechniqueisemployed.Convolutedlayers arethenfrozentostorelearnedcharacteristicsandbinary classification custom classifier is included. This model is trained with the Binary Cross Entropy loss and optimized withAdamoptimizer.
3.3 Voice-Based Detection
Thevoicemodalityrecordstheacousticdeviationslike changeinthepitchandamplitude.
1)DataProcessing:Thedatasetcontainsvoicefeatures that researchers extracted from speech recordings. The dataset removes unnecessary attributes which include subjectIDandduplicatevariables.Abinarylabeliscreated basedathresholdwhichresearchersappliedtotheUPDRS score.
2) Feature Scaling: Standardization is done to ensure every feature of the data has zero mean and unit variance,consequentlyincreasingmodelstability.
3)ModelArchitecture:AnXGBoostclassifierisused due to its ability to model complex relationships in structureddata.
3.4Fusion Strategy
Theoutputsofbothmodelsarecombinedusinga probability-basedfusionapproach.
Pfusion =w·Phandwriting +(1−w)·Pvoice
(1) where w represents the weight assigned to the handwritingmodel.1)WeightOptimization:Inordertoset the weight that figures in an optimal way, the grid search
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
mechanism in the eval setup automatically chooses that valuewhichcoulddeterminethebestaccuracy.
2) Final Decision: The congruent probability is compared againstathresholdtoproducethefinalclassificationoutput.
4. RESULTS
Thissectionpresentstheperformanceevaluationof the proposed multimodal Parkinson’s Disease detection system.
4.1
Performance Metrics
The evaluation metrics used include accuracy, precision,recall,andF1-score.
4.2 Quantitative Results
Theresultsindicatethatthevoicemodelachieves thehighestaccuracyandrecall,whilethehandwritingmodel shows higher precision. The fusion model provides a balancedperformance.
4.3 Handwriting Model Results
The confusion matrix for the handwriting model showsthat316healthysampleswerecorrectlyclassifiedas truenegativesand256Parkinson’ssampleswerecorrectly identified as true positives which demonstrate strong classificationcapability.Thestudyfound11healthysamples which were mistakenly identified as Parkinson’s patients throughfalsepositiveswhile70Parkinson’ssampleswere incorrectlyidentifiedashealthythroughfalsenegatives.The
higher number of false negatives suggests that the model might overlook some Parkinson’s cases which results in decreased recall performance. The small number of false positives demonstrates high precision because the model successfully prevents erroneous identification of healthy individuals. The results show that the handwriting model can accurately detect normal cases but it needs further enhancements to improve its capabilities in detecting all Parkinson’sinstances.
4.4 Voice Model Results
The voice-based model achieved successful prediction results through its capacity to identify 228 healthy samples as true negatives together with 898 Parkinson’ssampleswhichitidentifiedastruepositives.The studyfoundthattherewere27healthysampleswhichwere wrongly identified as Parkinson’s disease (incorrect positiveresults)while22Parkinson’ssampleswerefalsely identifiedashealthy(incorrectnegativeresults).Themodel demonstrated high effectiveness in detecting Parkinson’s diseasethroughitsabilitytomaintainanexceptionallylow false negative rate which serves as a vital component for early disease identification. The model demonstrates high precisionthroughitscapacitytocorrectlyidentifyhealthy individuals without making false positive errors. The confusionmatrixvaluedistributiondemonstratesthatthe model performs well because it achieves both high sensitivityandhighspecificitywhichmakesitsuitablefor actualdiagnosticuse.
4.5 Fusion Model Performance
Theperformancecomparisongraphillustratesthe classificationaccuracyachievedbythehandwriting,voice, andfusionmodels. Thevoicemodel reaches itsmaximum individualaccuracybecauseitsuccessfullycapturesdefined acoustic elements. The handwriting model shows slightly
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
lower accuracy because its motor patterns demonstrate variableperformanceandithasahigherfalsenegativerate. Thefusionmodel,althoughmarginallylowerthanthevoice model in terms of raw accuracy, demonstrates a more balanced performance across all evaluation metrics. The fusionmethodsuccessfullymergesinformationfrom both modes because it diminishes the limitations found in separatemodels.Thefusionmodelusesbothmotor-based andvoice-basedfeaturestocreateasystemthatperforms better in real-world diagnostic situations which need consistentperformanceinsteadofmaximumaccuracy.
4.6 ROC Curve Analysis
TheReceiverOperatingCharacteristic(ROC)curve displays how true positive rate (sensitivity) and false positiveraterespondtodifferentclassificationthresholds. The model demonstrates strong discriminative ability because itscurve extendssignificantlyabovethediagonal baseline that represents random classification. The classification system achieves effective results because its curve reaches the top-left corner of the plot which demonstrates higher sensitivity with lower false positive rates.Thehighareaunderthecurve(AUC)valueshowsthat the model can accurately differentiate between healthy individuals and Parkinson’s patients. The model shows stable performance because it handles different threshold settingswhichmakesitsuitableforuseinclinicaldecisionmakingsituations
4.7
Discussion
Themultimodalapproachimprovesrobustnessby combiningmotorandvocalfeatures.Thehandwritingmodel capturesmotorimpairments,whilethevoicemodelcaptures speechabnormalities.Thefusionmodelintegratesbothfor improved reliability. H. Summary The proposed system demonstrates strong performance and highlights the effectiveness of multimodal learning for Parkinson’s Diseasedetection.
4.8 Summary
The proposed system demonstrates strong performanceandhighlightstheeffectivenessofmultimodal learningforParkinson’sDiseasedetection.
5. ADVANTAGES AND DISADVANTAGES
5.1 Advantages
• Improved Diagnostic Accuracy: The advanced multimodal system improves Parkinson’s Disease detectionaccuracybecauseitusesbothhandwritingand voice data. The system achieves better prediction accuracy through its dual operation which uses both motor and vocal biometric data to assess patient symptoms.
• Complementary Feature Utilization: The system successfullyutilizesadditionalinformationwhichcomes from various different sources. Handwriting analysis detectsmotorimpairmentsthroughitsabilitytoidentify tremors and irregular stroke patterns, while voice analysis detects vocal abnormalities through its detectionofpitchinstabilityandarticulationissues.The combination of these features provides a complete evaluationofthepatient’smedicalcondition.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
•EfficientUseofTransferLearning:Theapplication of transfer learning with MobileNetV2 enables researcherstoextracthandwritingimagefeatures without needing extensive training data requirements. The system achieves operational efficiencythroughdecreasedcomputationalneeds and reduced training durations while delivering exceptionalperformance,whichmakesthesystem appropriateforpracticaluse.
•RobustClassificationwithXGBoost:TheXGBoost classifier uses voice data to handle structured features with strong performance capabilities. XGBoostimprovesmodelperformancethroughits complex relationship modeling abilities which include overfitting prevention methods and regularizationtechniques.
• Scalability and Practical Applicability: The frameworkallowsforstraightforwardexpansionto additional modalities which include gait analysis and sensor data. The system’s modular structure enables healthcare facilities to implement it for ongoing patient surveillance and quick illness identification.
5.2
Disadvantages
•ModerateRecallinHandwritingModel:Although the handwriting model achieves high precision, it mayexhibitrelativelylowerrecall,indicatingthat some Parkinson’s cases may be misclassified as healthy. This limitation can affect early detection, whereidentifyingallpositivecasesiscritical.
•ComputationalComplexityinFusion:Thefusion process,particularlytheauto-weightoptimization step,introducesadditionalcomputationaloverhead. Searchingforoptimalweightsincreasesprocessing time,especiallywhendealingwithlargedatasets.
• Dataset Dependency and Generalization Issues: The model performance is dependent on the datasets used for training. Differences in data distribution across populations may affect generalization, requiring further validation on diversedatasets
6. CONCLUSIONS
We developed a complete multimodal machine learningsystemwhichdetectsParkinson’sDiseasethrough its two components. The primary motivation behind this workistodevelopareliable,non-invasive,andautomated diagnosticapproachwhichhelpswithearlydetectionthat remainsacriticalchallengeinclinicalpractice.Theproposed system uses multiple modalities to detect various disease symptoms which single-modality systems cannot fully observe.
The framework effectively utilizes both deep learning and traditional machine learning techniques to
exploit the strengths of each domain. MobileNetV2, implementedthroughatrans ferlearningstrategy,isused toextractimportantspatialdatafromhandwritingimages which helps identify motor related changes like tremors andstrokeformationproblemsandwritingstabilityissues. The XGBoost classifier analyzes voice-based features to identifyvocalimpairmentswhichincludechangesinpitch and articulation and speech dynamics. The dual-modality approach allows the model to accurately represent both motor and vocal symptoms which occur in Parkinson’s Diseasepatients.
The primary achievement of this research work introduces an automatic weight adjustment system which merges hand writing and voice model results through decision-levelprocessing.Theproposedmethodestablishes its optimal performance through dynamic assessment of each modal element’s contribution instead of using fixed weightdistribution whichis commonintraditional fusion methods. The system achieves its adaptive fusion capabilitiessinceitusesadaptivefusionwhichprotectsits functional ability when data quality changes and different modalitiesshowtheiruniquestrengths.
The system demonstrates its capability to deliver strong outcomes through various testing methods which measure its performance at different points including accuracyandprecisionandrecallandF1-score.Thevoicebasedmodelreacheshigheraccuracybecauseitprocesses acoustic features through structured methods but the handwriting model enables early detection because it reveals essential insights about motor impairments. The fusion model effectively integrates these complementary aspects, resulting in improved overall performance and better balance between sensitivity and specificity. The research results demonstrate that multimodal methods effectively handle the complex symptoms and variable manifestationsofParkinson’sDisease.
Theframework providesbetterdiagnostic results throughitsnewsystemwhichbringsmultipleoperational benefits.Thesystemachieveseconomicalscreeningcapacity through its use of non-invasive handwriting and voice recording data sources which enable widespread testing throughout populations. The system’s modular structure supportsupcomingresearchrequirementsandclinicalneeds throughitsabilitytoincorporatenewtestingmethods. The study has specific restrictions which researchersneedtoknowabout.Thesystemrequireshighquality consistent input data but voice recordings suffer from interrupted recording conditionsand environmental noise problems. The hospital environments require transparent systems which use deep learning models becausethesemodelscreatechallengesthatpreventpeople fromunderstandingthem.Theexistingsystemsneedbetter twoessentialfunctionswhichincludeunderstandingsystem operationandprotectingagainstunexpectedproblems.
International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 13 Issue: 04 | Apr 2026 www.irjet.net p-ISSN: 2395-0072
Systemoperationandprotectingagainstunexpected problems. Researchers should investigate additional data sources which include gait analysis and wearable sensor dataandmedicalimaging.Explainableartificialintelligence methods will prove beneficial because they make models moreunderstandabletouserswhileenablingtheirusagein healthcare settings. The proposed framework needs additionaltestingonlargerandmorevarieddatasetsaswell as the creation of systems that monitor processes in real timetoenhanceitsoperationaleffectivenessanddependable performance.
The multimodal system which researchers developed shows potential because it combines different data sources to achieve precise and strong detection of Parkinson’sDisease.Theframeworkusesanadaptivefusion strategy to combine motor and vocal biomarkers which makesitascalablesolutionthathelpsdoctorsmakebetter decisionsandenablesearlydiseaseidentification.
REFERENCES
[1] M.A.Islametal.,“Areviewofmachinelearninganddeep learning algorithms for Parkinson’s disease diagnosis,” Heliyon,vol.10,2024.:contentReference[oaicite:0]index=0
[2] D. F. Santos, “Parkinson’s Disease Detection using XGBoost and Machine Learning,” medRxiv, 2023. :contentReference[oaicite:1]index=1
[3]P.PradeepandJ.Kamalakannan,“Automateddetection ofParkinson’sdiseaseusingimprovedLinkNet-GhostNet modelbasedonhandwritingimages,”ScientificReports,vol. 15,2025.:contentRefer ence[oaicite:2]index=2
[4] H. Sedigh Malekroodi et al., “Voice-Based Detection of Parkinson’sDiseaseUsingDeepLearning,”Bioengineering, 2025.:contentRefer ence[oaicite:3]index=3
[5]Y.A.Babuietal.,“Parkinson’sDiseasePredictionUsing Handwritten and Voice Data,” Atlantis Press, 2025. :contentRefer ence[oaicite:4]index=4
[6]R.S.Ashoketal.,“Machinelearning-basedearlydetection ofParkinson’sdisease,”JournalofResearchinEngineering andManagement,2025.:contentReference[oaicite:5]index=5
[7] T. H. H. Aldhyani et al., “Modeling and diagnosis of Parkinsondiseaseusinghanddrawinganddeeplearning,” AIMS Mathematics, vol. 9, no. 3, 2024. :contentReference[oaicite:6]index=6
[8]RSIInternational,“DetectingParkinson’sdiseasethrough handwritingpatterns,”2025. :contentReference[oaicite:7]index=7
[9] Semantic Scholar, “Parkinson’s Disease Identification using Voice and Spiral Drawings,” 2023. :contentReference[oaicite:8]index=8
[10]B.Egboetal.,“Explainablemachinelearningforearly detection of Parkinson’s disease,” PMC, 2025. :contentReference[oaicite:9]index=9
[11]J.Shinetal.,“ParkinsonDiseaseDetectionBasedonInair Dynamics Feature Extraction,” arXiv, 2024. :contentReference[oaicite:10]index=10
[12]R.Tyagietal.,“Machinelearningforearlydiagnosisof Parkinson’s disease using handwriting,” arXiv, 2021. :contentReference[oaicite:11]index=11
[13]W.Rahmanetal.,“DetectingParkinson’sdiseasefrom speechsignals,”arXiv,2020.contentReference[oaicite:12]inde x=12
[14] R. Castrillon et al., “Characterization of handwriting skills as a biomarker for Parkinson disease,” arXiv, 2019. :contentRefer ence[oaicite:13]index=13
[15]S. Taleb etal.,“DetectionofParkinson’s diseasefrom handwritingusingdeeplearning,”EvolutionaryIntelligence, 2023.
[16]Z.Lietal.,“EarlydiagnosisofParkinson’sdiseaseusing convolutional neural networks,” Journal of Biomedical Informatics,2022.
[17]M.Garc´ıa-Ordasetal.,“Multi-taskneuralnetworksfor Parkinson’s´diseaseseveritydetection,”MultimediaTools andApplications,2024.
[18] Q. Du et al., “Machine learning-based Parkinson’s diseasedetectionusingclassificationboundaries,”Discover AppliedSciences,2024.
[19]H.Ahmadietal.,“Hybridmachinelearningmodelsfor early Parkinson’s disease diagnosis,” Biocybernetics and BiomedicalEngineering,2024.
[20] N. Mate et al., “Deep learning frameworks for multimodal Parkinson’s disease detection,” Complex & IntelligentSystems,2024