Smartphone-based behavioral profiling for distinguishing Dementia with Lewy bodies from Alzheimer’s Disease
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background and Objective
Dementia with Lewy bodies (DLB) is frequently misdiagnosed as Alzheimer’s disease (AD) due to overlapping clinical presentations. In this study, we evaluated a smartphone-based digital platform that leverages behavioral markers to accurately differentiate between DLB and AD.
Methods
We conducted a cross-sectional study on 81 participants (Healthy controls = 40, AD = 21, DLB = 20), administering MMSE, and a smartphone-based noise pareidolia test (NPT) with integrated eye-tracking and speech inputs. Ground truth for patients were established using nuclear imaging (DAT-SPECT, IMP-SPECT, and/or MIBG). Core behavioral features such as response efficiency, gaze parameters (fixation count, saccadic patterns, surface area coverage) and acoustic features (latency, spectral properties) were extracted. A two-tiered classification system involving deep learning and support vector machine model was used. Based on the behavioral output, probability scores were generated with a focus on explainability to differentiate DLB from AD.
Results and interpretation
DLB group showed more pareidolias than AD group on the digital app. DLB patients exhibited several short saccades, inconsistent scanning behavior and smaller fixation dispersion, suggesting impaired top-down modulation of visual attention. However, fixation duration did not differ between AD and DLB. Vocal responses (latency, spectral properties) lacked specificity for DLB classification. Response fluctuation analysis revealed lower variability in AD patients compared to DLB. The app outperformed paper-based methods with 87% sensitivity and 93% specificity, and overall accuracy of 90% for AD–DLB classification. Two AD patients were classified as healthy controls (probability scores: 15%, 45%), likely due to early disease stage or unmeasured neurocognitive factors.
Conclusions
By effectively capturing core symptoms, our smartphone-based modality provides a reliable, scalable, and non-invasive tool for classification of DLB from AD. The integration of machine learning enabled precise quantification of behavioral microfluctuations between AD and DLB subtypes, underscoring its significant diagnostic potential. Future research is needed in refining classifiers to address early-stage cases and for disease staging.
