A Pilot Study on Speech and Handwriting Multimodal Fusion Analysis in Prodromal Dementia with Lewy Bodies

A Pilot Study on Speech and Handwriting Multimodal Fusion Analysis in Prodromal Dementia with Lewy Bodies

Paper in proceedings outside WoS and Scopus

Prodromal dementia with Lewy bodies (preDLB) is difficult to detect because of heterogeneous clinical manifestations and substantial overlap with related neurodegenerative disorders. This pilot study investigates multimodal machine learning for non-invasive preDLB classification using speech and online handwriting digital biomarkers. The analysed cohort comprised 42 preDLB patients and 44 healthy controls. Features were extracted from two speech tasks and seven handwriting tasks and organized into domain-specific feature sets. Three machine learning models (logistic regression, support vector machine with a radial basis function kernel, and XGBoost) were evaluated using repeated stratified 10-fold cross-validation, Bayesian hyperparameter optimization, and iterative feature selection. Unimodal analyses, early fusion feature integration, and late fusion prediction aggregation strategies were systematically compared, with and without demographic variables. Among unimodal domains, temporal handwriting features achieved the highest performance (AUC=0.854, BACC=0.767). In the multimodal setting, early fusion with demographic variables yielded the best overall results, with XGBoost achieving AUC=0.913 and BACC=0.834. These findings support the feasibility of combining speech and handwriting biomarkers for multimodal preDLB assessment and highlight the potential of machine-learning-based digital biomarkers for non-invasive early screening.

Prodromal dementia with Lewy bodies (preDLB) is difficult to detect because of heterogeneous clinical manifestations and substantial overlap with related neurodegenerative disorders. This pilot study investigates multimodal machine learning for non-invasive preDLB classification using speech and online handwriting digital biomarkers. The analysed cohort comprised 42 preDLB patients and 44 healthy controls. Features were extracted from two speech tasks and seven handwriting tasks and organized into domain-specific feature sets. Three machine learning models (logistic regression, support vector machine with a radial basis function kernel, and XGBoost) were evaluated using repeated stratified 10-fold cross-validation, Bayesian hyperparameter optimization, and iterative feature selection. Unimodal analyses, early fusion feature integration, and late fusion prediction aggregation strategies were systematically compared, with and without demographic variables. Among unimodal domains, temporal handwriting features achieved the highest performance (AUC=0.854, BACC=0.767). In the multimodal setting, early fusion with demographic variables yielded the best overall results, with XGBoost achieving AUC=0.913 and BACC=0.834. These findings support the feasibility of combining speech and handwriting biomarkers for multimodal preDLB assessment and highlight the potential of machine-learning-based digital biomarkers for non-invasive early screening.

multimodal machine learning, early fusion, late fusion, speech processing, online handwriting

multimodal machine learning, early fusion, late fusion, speech processing, online handwriting

NOVOTNÝ, K.; LADISLAV, R.

28.04.2026

Brno University of Technology, Faculty of Electrical Engineering and Communication

Brno

978-80-214-6417-0

Proceedings I of the 32nd Conference STUDENT EEICT 2026 General Papers

292

296

5

https://www.eeict.cz/eeict_download/archiv/sborniky/EEICT_2026_sbornik_1.pdf