EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

An Integrated Multimodal Machine Learning Framework for Parkinson’s Disease Screening Using Behavioral and Physiological Biomarkers

Ashok R S

Journal of Intelligent Decision Making and Information Science · 09.09.2026 · 3 (18s) · 01-13

DOI: 10.17762/jidmis.v3.5714 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Journal of Intelligent Decision Making and Information Science, online erschienen am 09.09.2026. Verfasst von Ashok R S. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Parkinson’s disease is often recognized late because clinical assessment depends on visible motor signs and on examiner judgment. Low-cost behavioral and physiological measurements could support earlier screening, but most computational studies analyze one signal type at a time. This study presents an integrated multimodal machine learning framework that combines voice, spiral drawing, wave drawing, electrocardiogram with oxygen saturation, and hand-worn gait signals acquired with inexpensive embedded hardware, and asks whether multimodal evidence can improve on single-modality screening. Each modality was processed with its own preprocessing and feature pipeline, and about twelve classifiers per modality were compared using accuracy, precision, recall, F1-score, and log loss. The best single-modality models were XGBoost for voice (90.00%), CatBoost for spiral (87.50%) and wave drawings (89.50%), gradient boosting for the electrocardiogram (96.26%), and AdaBoost for gait (85.71%). The system-level figure of 89.79% reported for the original prototype is shown to be the arithmetic mean of these five accuracies rather than the accuracy of a fused classifier. Because the five datasets were collected independently and are not linked to the same participants, subject-level fusion could not be evaluated. Instead, two fusion designs (feature-level and calibrated decision-level) and a validation protocol covering ablation, calibration, and repeated subject-independent splits are specified, and Wilson confidence intervals are reported, showing wide uncertainty for the smaller test sets (gait: 48.7% to 97.4%). Tree-based ensembles performed best for every modality, and electrocardiogram features were the most discriminative in the available data. Claims of multimodal superiority require participant-aligned data; the framework and protocol provide a basis for a rigorous evaluation of low-cost multimodal screening.

Zitieren

APA

S, A. R. (2026). An Integrated Multimodal Machine Learning Framework for Parkinson’s Disease Screening Using Behavioral and Physiological Biomarkers. Journal of Intelligent Decision Making and Information Science, 3(18s), 01-13. https://doi.org/10.17762/jidmis.v3.5714

Vancouver

S AR. An Integrated Multimodal Machine Learning Framework for Parkinson’s Disease Screening Using Behavioral and Physiological Biomarkers. Journal of Intelligent Decision Making and Information Science. 2026;3(18s):01-13. doi:10.17762/jidmis.v3.5714

BibTeX
@article{S2026,
  author = {S, Ashok R},
  title = {An Integrated Multimodal Machine Learning Framework for Parkinson’s Disease Screening Using Behavioral and Physiological Biomarkers},
  journal = {Journal of Intelligent Decision Making and Information Science},
  year = {2026},
  volume = {3},
  number = {18s},
  pages = {01--13},
  doi = {10.17762/jidmis.v3.5714}
}
Kostenlos starten →