Title: Machine Learning's Diagnostic Performance Of Cognitive Decline Via Handwriting: A Systematic Review And Meta-analysis

Author Block: 
Ching-Wei Ye¹, Chien-Liang Liu², Fu-Ling Yang³, Nga-Weng Ieong⁴, Shu-Wei Chang⁵, Tien-Ni Wang¹, Hao-Ling Chen^1,6.
¹School of Occupational Therapy, National Taiwan University, College of Medicine, Taipei City, Taiwan, 
²Department of Neurology, Taipei City Hospital, Taipei City, Taiwan, 
³School of Nursing, National Taiwan University, College of Medicine, Taipei City, Taiwan, 
⁴Institute of Epidemiology and Preventive Medicine, National Taiwan University, Taipei City, Taiwan, 
⁵Department of Civil Engineering, National Taiwan University, Taipei City, Taiwan, 
⁶Health Science and Wellness Center, National Taiwan University, Taipei City, Taiwan.

Introduction: 

Cognitive decline—from mild cognitive impairment (MCI) to Alzheimer’s disease (AD)
—is a growing global health burden, yet neuroimaging and cerebrospinal fluid biomarkers are costly, invasive, and inaccessible in primary care.

Handwriting integrates fine motor control, visuospatial attention, and cognitive processing;
its kinematics (velocity, in-air time, fluency) are a promising non-invasive digital biomarker.

Machine learning (ML) enables automated HC / MCI / AD classification,
but reported accuracy varies widely. We synthesized the evidence and identified the methodological factors that drive diagnostic performance.

Methods: 

PRISMA-guided systematic review; PubMed & Cochrane Library (2004–2024), English only; 18 studies included.
Population: adults ≥60 with AD, MCI, or healthy controls (HC);
Index test: ML on digitized handwriting or drawing;
Reference: established clinical or neuropsychological criteria.
The risk of bias was assessed using the QUADAS-2.
The 2×2 confusion matrices were extracted or reconstructed for each study (TP/FP/TN/FN).
Bivariate random-effects meta-analysis; HSROC curves; meta-regression on handwriting task type, ML algorithm category.

Results: 
Common handwriting features:
mean/peak velocity, motor quality index (velocity peaks, jerks), pressure variability, pen-in-air time.

Pooled Diagnostic Accuracy:
Fig A. HC vs. AD (16 studies): sensitivity 88.1% (83.9–91.3), specificity 90.9% (86.8–93.7), AUC 0.947.
Fig B. HC vs. MCI (11 studies): sensitivity 77.8% (69.2–84.6), specificity 76.8% (71.1–81.1), AUC 0.83.

Meta-regression: task type and ML algorithm significantly drove heterogeneity.
Combining multiple tasks with ensemble or deep learning models achieved the highest accuracy.

Conclusion: 

Handwriting kinematics are promising behavioral markers for screening AD and MCI.

Combined writing-and-drawing protocols with ensemble or deep-learning models give the best diagnostic performance.

Trade-off: multi-task batteries raise patient burden; community screening may favor optimized single tasks.

Suggestion: explainable AI (XAI), early pre-symptomatic MCI detection, standardized multi-center datasets.

Future Work: include 2025-2026’s studies, identify the optimal single handwriting task,
and examine the smallest feature amount required to achieve high accuracy through further meta-regression.