Efficient Handwriting-Based Alzheimer,s Disease Diagnosis Using a Low-Rank Mixture of Experts Deep Learning Framework

Wu Wang; Yuang Cheng; Fouzi Harrou; Ying Sun

cs.LGApr 14, 2026

Efficient Handwriting-Based Alzheimer,s Disease Diagnosis Using a Low-Rank Mixture of Experts Deep Learning Framework

Authors: Wu Wang, Yuang Cheng, Fouzi Harrou, Ying Sun

Organizations: Center for Applied Statistics and School of Statistics Renmin University of China Beijing 100872, China · King Abdullah University of Science and Technology (KAUST) Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division Thuwal 23955-6900, Saudi Arabia

Abstract

Early and reliable detection of Alzheimer's disease (AD) is crucial for timely clinical intervention and improved patient management. It also supports the evaluation of emerging therapeutic strategies. In this paper, we propose a Low-Rank Mixture of Experts (LoRA-MoE) deep learning framework for Alzheimer's disease diagnosis based on handwriting analysis. Handwriting signals provide a non-invasive and scalable digital biomarker that captures subtle cognitive-motor impairments associated with early AD progression. The proposed architecture allows multiple experts to specialize in different handwriting patterns while sharing a common base network. This design enables efficient learning of general representations while reducing interference between experts. Each expert is equipped with lightweight low-rank adapters. This mechanism significantly reduces the number of trainable parameters compared with standard Mixture of Experts (MoE) models and improves training stability. The proposed framework is evaluated on the Diagnosis AlzheimeR WIth haNdwriting (DARWIN) dataset. Extensive experiments are conducted, including ablation studies on key architectural parameters such as hidden dimension size, number of experts, and LoRA rank. The method is compared with multilayer perceptron (MLP) and conventional MoE architectures. In addition, stacking ensemble strategies (StackMean and StackMax) are investigated to improve robustness and predictive performance. Experimental results show that the LoRA-MoE framework achieves powerful diagnostic performance while activating significantly fewer parameters during inference. These results highlight the potential of the proposed approach as an accurate and computationally efficient solution for handwriting-based Alzheimer's disease screening and digital health applications.

Figures & tables

Figure 1: Overview of the proposed LoRA-MoE architecture for handwriting-based Alzheimer’s disease diagnosis.

Figure 2: The end-to-end workflow of the proposed LoRA-MoE architecture for Alzheimer’s disease diagnosis using handwriting dynamics

Figure 3: Handwriting data acquisition setup using a graphic tablet to capture pen movements in real time Cilia et al. (2022) .

Category	Description
Time-related features
total_time	Total duration required to complete the task.
air_time	Time spent performing in-air pen movements.
paper_time	Time spent performing on-paper pen movements.
Movement-related features
mean_speed_on_paper	Average speed of on-paper movements.

Table 1: Summary of features extracted from handwriting data.

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BaseLearner_1	0.8457	0.8500	0.8412	0.9239	0.8537	0.8446	0.69
	BaseLearner_2	0.8657	0.8667	0.8647	0.9250	0.8752	0.8647	0.67
	BaseLearner_3	0.8457	0.8500	0.8412	0.9237	0.8545	0.8447	0.70
	BaseLearner_4	0.8629	0.8556	0.8706	0.9221	0.8719	0.8617	0.72
	BaseLearner_5	0.8457	0.8611	0.8294	0.9314	0.8558	0.8443	0.74
	BaseLearner_6	0.8429	0.8444	0.8412	0.9255	0.8516	0.8417	0.96

Table 2: Model Comparison with Different Hidden Dimensions

Figure 4: Comparison of (a) training time, (b) base-learner accuracy distribution, and (c) efficiency index (performance/time) for LoRA-MoE, MoE, and MLP architectures.

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BL_1	0.8514	0.8667	0.8353	0.9320	0.8545	0.8511	0.66
	BL_2	0.8514	0.8667	0.8353	0.9369	0.8538	0.8512	0.77
	BL_3	0.8543	0.8722	0.8353	0.9335	0.8590	0.8539	0.87
	BL_4	0.8686	0.8722	0.8647	0.9310	0.8714	0.8684	0.98
	BL_5	0.8714	0.8833	0.8588	0.9361	0.8751	0.8711	1.07
	BL_6	0.8486	0.8556	0.8412	0.9243	0.8525	0.8482	1.15

Table 3: Model Comparison with Different Number of Experts

Figure 5: Comparison of (a) training time, (b) base-learner accuracy distribution, and (c) efficiency index (performance/time) for LoRA-MoE, MoE, and MLP under different numbers of experts.

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BL_1	0.8086	0.7722	0.8471	0.9160	0.8187	0.8075	0.97
	BL_2	0.8514	0.8389	0.8647	0.9252	0.8561	0.8511	1.09
	BL_3	0.8486	0.8278	0.8706	0.9257	0.8551	0.8480	1.10
	BL_4	0.8314	0.7944	0.8706	0.9240	0.8398	0.8307	1.14
	BL_5	0.8229	0.8000	0.8471	0.9224	0.8287	0.8223	1.16
	BL_6	0.8371	0.8111	0.8647	0.9224	0.8424	0.8368	1.16

Table 4: Model Comparison with Different Rank of LoRA

Figure 6: Comparison of (a) training time, (b) base-learner accuracy distribution, and (c) efficiency index (performance/time) for LoRA-MoE, MoE, and MLP under different ranks of LoRA.

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BL_1	0.6279	0.5854	0.6725	0.6803	0.6315	0.6264	13.27
	BL_2	0.6577	0.6301	0.6866	0.7112	0.6598	0.6571	13.23
	BL_3	0.6626	0.6148	0.7127	0.7178	0.6667	0.6615	13.47
	BL_4	0.6693	0.6189	0.7221	0.7293	0.6741	0.6679	13.14
	BL_5	0.6784	0.6283	0.7308	0.7321	0.6826	0.6774	14.02
	BL_6	0.6706	0.6335	0.7094	0.7328	0.6737	0.6698	13.91

Table 5: Performance Comparison of Models under Independent Task-Level Prediction.

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BL_1	0.8314	0.8278	0.8353	0.9162	0.8358	0.8310	1.72
	BL_2	0.8143	0.8167	0.8118	0.9041	0.8203	0.8136	1.77
	BL_3	0.8457	0.8500	0.8412	0.8938	0.8506	0.8452	1.98
	BL_4	0.8371	0.8333	0.8412	0.8946	0.8405	0.8368	2.12
	BL_5	0.8286	0.8389	0.8176	0.8923	0.8305	0.8283	2.29
	BL_6	0.8371	0.8722	0.8000	0.8869	0.8411	0.8365	2.92

Table 6: Model Comparison with Different Hidden Dimensions(5 Layers)

Arch	Model	Accuracy	Sensitivity	Specificity	AUC	Precision	F1 Score	Time(s)
LoRA-MoE	BL_1	0.8386	0.8556	0.8206	0.8955	0.8458	0.8374	4.04
	BL_2	0.8429	0.8444	0.8412	0.8856	0.8518	0.8410	4.33
	BL_3	0.8543	0.8861	0.8206	0.8907	0.8630	0.8531	4.90
	BL_4	0.8429	0.8667	0.8176	0.8796	0.8514	0.8417	5.24
	BL_5	0.8400	0.8611	0.8176	0.8834	0.8457	0.8392	5.79
	BL_6	0.8329	0.8833	0.7794	0.8852	0.8430	0.8312	7.60

Table 7: Model Comparison with Different Hidden Dimensions(8 Layers)

Exp	LM-5	LM-10	LM-15	LM-20	LM-25	LM-Max	LM-Mean
1	60.43	62.14	64.00	62.29	61.14	61.71	63.71
2	63.14	68.14	66.71	69.00	68.86	68.29	69.86
3	61.29	67.29	67.86	68.86	69.29	70.00	69.57
4	63.71	66.29	65.57	66.29	65.86	66.00	66.57
5	66.57	69.43	72.00	72.71	70.00	71.57	72.00
6	67.57	69.29	69.14	69.71	67.86	70.86	70.57

Table 8: Accuracy (%) of LM(LoRA-MoE) configurations on each experiment.

Configuration	Accuracy (%)	Sensitivity (%)	Specificity (%)
MLP
MLP-5	80.14 ± 6.61	81.67 ± 14.07	78.53 ± 9.71
MLP-10	81.71 ± 6.48	79.17 ± 10.23	84.41 ± 7.03
MLP-15	82.00 ± 7.18	76.11 ± 11.12	88.24 ± 5.26
MLP-20	84.71 ± 5.21	80.28 ± 9.21	89.41 ± 5.13
MLP-25	87.43 ± 4.64	84.17 ± 7.51	90.88 ± 5.09

Table 9: Performance of Voting Ensemble configurations (mean ± std, %).

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Exp	MLP-5	MLP-10	MLP-15	MLP-20	MLP-25	MLP-Max	MLP-Mean
1	57.57	61.86	63.86	63.29	63.00	63.14	64.14
2	64.29	68.00	67.14	69.14	70.14	69.57	69.14
3	65.71	67.43	68.57	67.29	68.71	68.00	67.43
4	62.43	66.43	66.29	66.29	67.71	66.86	66.43
5	64.14	72.00	72.14	70.71	70.00	71.86	72.14
6	65.14	71.43	70.43	72.14	71.43	71.43	71.43

Appendix

Table 10: Accuracy (%) of MLP configurations on each experiment.

Exp	MoE-5	MoE-10	MoE-15	MoE-20	MoE-25	MoE-Max	MoE-Mean
1	57.57	61.57	60.71	61.00	62.57	60.29	60.29
2	60.57	65.14	66.57	68.57	67.14	68.71	68.71
3	59.71	65.29	66.14	66.71	67.00	68.86	68.86
4	60.29	65.71	65.00	63.86	65.43	65.14	65.14
5	66.14	62.71	67.43	69.57	68.00	70.86	70.86
6	60.29	66.43	69.00	69.43	68.00	72.43	72.43

Appendix

Table 11: Accuracy (%) of MoE configurations on each experiment.

Exp	LM-5	LM-10	LM-15	LM-20	LM-25	LM-Max	LM-Mean
1	60.83	60.83	59.17	57.22	57.78	56.94	56.94
2	72.78	63.06	66.67	70.56	69.72	69.72	68.89
3	68.61	66.67	67.22	69.72	67.78	69.72	68.06
4	67.22	66.94	62.22	66.67	67.50	67.50	66.94
5	64.44	62.22	63.89	63.61	62.22	64.17	63.61
6	68.61	59.17	60.28	64.44	61.67	63.89	62.78

Appendix

Table 12: Sensitivity of LM (LoRA-MoE) configurations on each experiment.

Exp	MLP-5	MLP-10	MLP-15	MLP-20	MLP-25	MLP-Max	MLP-Mean
1	50.00	59.72	60.28	58.33	56.94	57.22	57.50
2	70.00	65.00	64.44	66.11	67.78	67.50	65.83
3	63.89	68.61	69.44	68.61	68.33	70.00	69.72
4	63.89	65.56	65.83	65.00	65.28	66.39	66.67
5	59.72	60.83	60.56	60.83	60.28	60.56	60.83
6	60.28	60.28	62.22	63.06	64.72	63.06	62.50

Appendix

Table 13: Sensitivity of MLP configurations on each experiment.

Exp	MoE-5	MoE-10	MoE-15	MoE-20	MoE-25	MoE-Max	MoE-Mean
1	56.11	58.06	56.94	61.39	61.11	58.33	58.06
2	65.00	61.94	65.83	64.72	63.89	68.06	66.94
3	58.89	60.28	58.33	60.00	62.50	65.83	65.00
4	57.22	60.83	58.06	56.67	59.17	58.06	57.78
5	62.50	61.39	62.50	60.28	63.06	65.28	64.44
6	54.17	57.78	61.39	64.44	62.50	65.83	64.44

Appendix

Table 14: Sensitivity of MoE configurations on each experiment.

Exp	LM-5	LM-10	LM-15	LM-20	LM-25	LM-Max	LM-Mean
1	60.00	63.53	69.12	67.65	64.71	66.76	70.88
2	63.24	75.00	71.18	70.59	70.88	72.35	72.94
3	65.29	65.29	66.47	66.18	69.41	67.94	67.35
4	58.24	67.94	74.12	70.29	75.29	73.53	73.53
5	64.12	74.41	76.47	73.53	74.12	75.88	76.47
6	66.47	80.00	78.53	75.29	74.41	78.24	78.82

Appendix

Table 15: Specificity of LM (LoRA-MoE) configurations on each experiment.

Exp	MLP-5	MLP-10	MLP-15	MLP-20	MLP-25	MLP-Max	MLP-Mean
1	65.59	64.12	67.65	68.53	69.41	69.41	71.18
2	62.35	75.29	71.76	72.06	73.24	71.76	72.65
3	67.94	72.65	74.71	74.41	75.59	73.53	74.12
4	51.47	72.06	74.71	77.65	77.24	72.06	72.06
5	55.59	72.65	78.24	76.76	78.53	77.94	79.41
6	70.29	83.24	79.12	81.76	78.53	80.29	80.88

Appendix

Table 16: Specificity of MLP configurations on each experiment.

Exp	MoE-5	MoE-10	MoE-15	MoE-20	MoE-25	MoE-Max	MoE-Mean
1	59.12	64.71	60.59	64.12	61.18	62.35	65.29
2	55.88	68.53	67.35	72.65	70.59	69.41	72.35
3	60.59	70.59	69.12	71.76	71.76	72.06	72.06
4	63.53	70.88	72.35	71.47	72.06	72.65	75.88
5	70.00	64.12	72.65	79.41	73.24	76.76	78.24
6	66.76	75.59	77.06	74.71	73.82	79.41	80.88

Appendix

Table 17: Specificity of MoE configurations on each experiment.

Alzheimer's Disease Classification

Alzheimer’s disease classification uses clinical, behavioral, and brain-imaging data to distinguish disease stages and predict progression.

See all 92 papers

Nov 8, 2025

Cross-Task Generalization in Handwriting-Based Alzheimer's Screening via Vision Language Adaptation

Alzheimer's disease (AD) is a prevalent neurodegenerative disorder for which early detection is critical. Handwriting, which can be disrupted by subtle motor and cognitive decline, provides a non-invasive and cost-effective window for AD screening. Existing handwriting-based AD studies mostly rely on online trajectories and hand-crafted features, while the influence of handwriting task type on diagnostic performance and cross-task generalization remains underexplored. Meanwhile, large-scale vision--language models have demonstrated strong transfer and adaptation ability in natural-image anomaly detection and several medical modalities, such as chest X-ray and brain MRI. However, handwriting-based disease detection remains unexplored within this paradigm. To address this gap, we introduce a lightweight Cross-Layer Fusion Adapter (CLFA) framework that repurposes Contrastive Language--Image Pre-training (CLIP) for handwriting-based AD screening. CLFA inserts multi-level adapters into a frozen visual encoder, combining cross-layer feature fusion with depthwise 2D convolution on patch grids to capture both local stroke irregularities and higher-level handwriting structure. This design progressively aligns pretrained vision--language representations with AD-related handwriting cues and supports transfer from supervised source tasks to task-disjoint unseen target tasks. On the Darwin dataset, under the subject-disjoint cross-task protocol, averaged over all 600 task-disjoint source-target pairs, CLFA achieves 74.63% AUC, 74.85% accuracy, and 73.72% F1 score, outperforming the best competing model by 2.15, 1.79, and 1.87 percentage points, respectively.

Changqing Gong, Huafeng Qin, Mounim A. El-Yacoubi

Telecom SudParis, Institut Polytechnique de Paris, 91120 Palaiseau, France · School of Computer Science and Information Engineering, Chongqing Technology and Business University, Chongqing 400067, China

Alzheimer's Disease VLM Adaptation Medical VLMs Alzheimer's Disease Classification

Ask PDF

Sep 24, 2026

Segment-Level Risk Discovery in Online Handwriting for Alzheimer's Disease Detection

Online handwriting provides a non-invasive and low-cost behavioral biomarker for Alzheimer's disease (AD) detection, as it reflects both cognitive planning and fine motor control. Existing handwriting-based AD detection methods usually rely on global trajectory features or whole-sample representations, which can be strongly affected by individual writing style, task-specific variation, and acquisition noise. In this paper, we propose NormPaST-Risk, a healthy-normative Paper-Air selective trajectory state-space risk network for interpretable AD detection from online handwriting. Instead of treating the entire trajectory as a single holistic representation, our method reformulates AD handwriting detection as local disease-relevant segment discovery. Specifically, a multi-scale temporal encoder captures stroke dynamics at different temporal resolutions, while a selective Paper-Air state-space encoder models long-range handwriting progression and distinguishes on-paper motor execution from in-air planning and transition behaviors. To explicitly characterize abnormal deviations, a healthy normative branch learns normal handwriting dynamics from healthy controls, and a task-aware multi-expert segment-risk module estimates segment-level AD risk calibrated by hidden-state changes and normative deviations. A weakly supervised segment-level objective further enables high-risk segment discovery without manual segment annotations. Experiments on the DARWIN benchmark demonstrate that the proposed framework achieves superior AD/HC classification performance compared with existing methods. Moreover, the discovered high-risk segments can be projected back to the original handwriting trajectory, providing interpretable evidence associated with AD-related handwriting variations.

Changqing Gong, Huafeng Qin, Mounîm A. El-Yacoubi

Alzheimer's Disease Classification

Ask PDF

Aug 27, 2026

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease

Biological heterogeneity in Alzheimer's Disease (AD) poses a critical diagnostic challenge, particularly for traditional linear methods that fail to capture non-linear neural dynamics. To address this, we propose a diagnostic framework utilizing the Large Brain Model (LaBraM), pretrained on over 2,500 hours of EEG data. By integrating these high-dimensional latent embeddings with a non-linear Random Forest classifier, our approach effectively isolates robust disease markers. Under a rigorous subject-independent 5-fold cross-validation protocol, the method achieves an ROC-AUC of 89.36% +/- 3.49%, PR AUC of 81.45% +/- 4.43%, and Balanced Accuracy of 82.44% +/- 4.34% in distinguishing dementia patients from healthy controls. Notably, this performance uses only 8-second EEG segments, surpassing traditional spectral baselines, including band-power and parameterized oscillatory features (FOOOF). Post-hoc occlusion analysis confirms the model captures clinically validated biomarkers, specifically occipital-frontal Alpha and Theta rhythm degradation. Additional neurophysiological alignment analysis demonstrated that higher LaBraM-predicted dementia probability significantly correlated with worse cognitive performance, greater clinical severity, increased theta and alpha relative power, and higher aperiodic exponent. These findings demonstrate that deep latent representations extract clinically relevant signatures from noisy signals, enabling precise, rapid, and data-efficient diagnosis.

Maggie Lin, Chung-Lin Hou, Tzyy-Ping Jung

Department of Bioengineering University of California, San Diego La Jolla, CA 92037, USA · Research and Development HippoScreen Neurotech Corp. Taipei, Taiwan · Swartz Center for Computational Neuroscience University of California, San Diego La Jolla, CA 92037, USA

Alzheimer's Disease Classification Electroencephalography EEG Foundation Models

Ask PDF

Abstract

Alzheimer's Disease Classification

Explore similar work

Cross-Task Generalization in Handwriting-Based Alzheimer's Screening via Vision Language Adaptation

Segment-Level Risk Discovery in Online Handwriting for Alzheimer's Disease Detection

Leveraging a Foundation Model for the EEG-Based Diagnosis of Alzheimer's Disease