Agentic Cognitive Profiling: Realigning Automated Alzheimer's Disease Detection with Clinical Construct Validity

Jiawen Kang; Kun Li; Dongrui Han; Jinchao Li; Junan Li; Lingwei Meng; Xixin Wu; Helen Meng

cs.MAMar 18, 2026

Agentic Cognitive Profiling: Realigning Automated Alzheimer's Disease Detection with Clinical Construct Validity

Authors: Jiawen Kang, Kun Li, Dongrui Han, Jinchao Li, Junan Li, Lingwei Meng, Xixin Wu, Helen Meng

Organizations: The Chinese University of Hong Kong

Abstract

Automated Alzheimer's Disease (AD) screening has predominantly followed the inductive paradigm of pattern recognition, which directly maps the input signal to the outcome label. This paradigm sacrifices construct validity of clinical protocol for statistical shortcuts. This paper proposes Agentic Cognitive Profiling (ACP), an agentic framework that realigns automated screening with clinical protocol logic across multiple cognitive domains. Rather than learning opaque mappings from transcripts to labels, the framework decomposes standardized assessments into atomic cognitive tasks and orchestrates specialized LLM agents to extract verifiable scoring primitives. Central to our design is decoupling semantic understanding from measurement by delegating all quantification to deterministic function calling, thereby mitigating hallucination and restoring construct validity. Unlike popular datasets that typically comprise around a hundred participants under a single task, we evaluate on a clinically-annotated corpus of 402 participants across eight structured cognitive tasks spanning multiple cognitive domains. The framework achieves 90.5% score match rate in task examination and 85.3% accuracy in AD prediction, surpassing popular baselines while generating interpretable cognitive profiles grounded in behavioral evidence. This work demonstrates that construct validity and predictive performance need not be traded off, charting a path toward AD screening systems that explain rather than merely predict.

Figures & tables

Figure 1: The Conceptual Framework. We align AD detection with clinical construct validity by operationalizing the causal chain from Alzheimer’s Disease to Cognitive Deficits into an agentic workflow comprising Cognitive Tasks and Verifiable Metrics.

Figure 2: Overview of the Agentic Cognitive Assessment Framework. The workflow comprises three stages: (1) Administration: Collection of standardized task responses; (2) Examination: Multi-agent Workflow with Deterministic Function Calling and Verification Loop; (3) Inference: Aggregation of verified metrics for classification and explainable reporting.

Task	Scoring Primitives	Score
MoCA-SL (Assess language, attention, executive functions)
Picture Naming	Per-item correctness ( $\times$ 3)	3
Digit Span	Forward/backward correctness	2
Serial 7 Subtraction	# correct subtractions	3
Sentence Repetition	Per-sentence correctness ( $\times$ 2)	2
Animal Fluency	# valid animals	1

Table 1: Overview of Cognitive Tasks and Scoring Primitives. Scoring primitives define measurable performance units and their maximum scores.

Figure 3: Cognitive profile inference pipeline. Verified scoring primitives from each task are aggregated, normalized against demographic norms, and used for classification and report generation.

Figure 4: Demographic distribution of participants in datasets. Subplots depict age and years of education for Alzheimer’s (AD) and healthy control (HC) groups.

	Full		w/o Verifier		w/o Func. Call
Task	SMR	MAE	SMR	MAE	SMR	MAE
MoCA-SL
Picture Naming (/3)	97.0%	0.03	97.0%	0.03	92.6%	0.07
Digit Span (/2)	98.5%	0.01	98.5%	0.01	77.3%	0.24
Serial 7 Sub. (/3)	82.4%	0.19	70.6%	0.32	63.2%	0.63
Sentence Rep. (/2)	89.7%	0.10	89.7%	0.10	86.8%	0.13

Table 2: Task-level examination result. Score Match Rate (SMR) indicates the exact agreement percentage with manual scores; Mean Absolute Error (MAE) contextualizes error magnitude. Score range is annotated per task. Func. Call : Deterministic Function Calling.

System	Classifier	Accuracy	F1	Precision	Recall
Baseline
Handcrafted	MLP / SVM	68.2 / 70.6	55.7 / 67.8	58.3 / 56.8	54.4 / 84.0
BERT	MLP / SVM	73.7 / 79.4	59.5 / 72.0	68.3 / 72.0	52.8 / 72.0
RoBERTa	MLP / SVM	70.4 / 75.0	53.9 / 66.7	63.1 / 65.4	47.2 / 68.0
LLM-CoT	-	70.6	66.7	57.1	80.0
Proposed

Table 3: Comparison of Alzheimer’s disease screening performance (%) across baseline and proposed systems. Results exceeding 80% are highlighted in bold.

Figure 5: (a) Alignment of the full MoCA score and the subset MoCA-SL score; (b) Performance comparison of the Score Match Rate (SMR) with varying maximum retry limits.

Figure 6: Feature importance analysis using SHAP values for the supervised SVM classifier. Top-20 features are depicted.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Description	Cognitive Domain (DSM-5)	Scoring Primitives	Score
Alternating Trail Making	Draw a line to connect numbers and characters in alternating order	Exec. Function	Successful connection pattern	1
Copy Cube	Copy a 3-dimensional cube structure	Perceptual Motor	Correct 3D structure	1
Clock Drawing	Draw a clock face, numbers, and set time to 11:10	Perceptual Motor	Contour (1), Numbers (1), Hands (1)	3
*Picture Naming	Name three depicted animals (e.g., Lion, Rhino, Camel)	Language	Per-item correctness ( $\times$ 3)	3
*Digit Span	Repeat digits forward and backward	Complex Attention	Forward (1), Backward (1)	2
Vigilance (Tap)	Tap hand at each target digit (e.g., ’1’) in a sequence	Complex Attention	$\leq$ 1 error allowed	1

Appendix

Table 4: Detailed description and scoring criteria for the Montreal Cognitive Assessment (MoCA, Hong Kong version). Tasks marked with * belong to the spoken-language subset (MoCA-SL) used in this study

Trial / Task	Description	Cognitive Domain (DSM-5)	Scoring Primitives	Score
Acquisition & Learning (Form 1: Random Condition)
Trial 1	Listen to a list of 16 words (4 categories) presented orally, followed by immediate free recall.	Immediate Memory	# correctly recalled words	16
Trials 2 & 3	Listen to the same list repeated (same order), followed by free recall each time.	Learning Rate	# correctly recalled words (per trial)	16
Delayed Recall (Retention)
Trial 4 (10-min Delay)	Unannounced free recall of the 16 words after a 10-minute non-verbal interval.	Short-term Retention	# correctly recalled words	16
Trial 5 (30-min Delay)	Unannounced free recall of the 16 words after a 30-minute interval.	Long-term Retention	# correctly recalled words	16

Appendix

Table 5: Detailed description and scoring criteria for the Hong Kong List Learning Test (HKLLT)

ID	Feature Name	Description
L1	# words	The number of words
L2	% Stop words	# stop words / # words
L3	% Filled pauses	# filler words / # words
L4	% Lexical filler	# lexical filler / # words
L5	% Backchannel	# backchannel words / # words
L6	% Repetition	# repeated words / # words

Appendix

Table 6: Handcraft linguistic features used in this work. “%”: “ratio of”, “#”: “number of”, “dur”: “duration”.

ID	Feature Name	Description
HKLLT (Hong Kong List Learning Test)
C1	hkllt4_zScore	Trial 4 z-score
C2	hkllt5_zScore	Trial 5 z-score
C3	n_hkllt4_recall	Trial 4 recall count
C4	n_hkllt5_recall	Trial 5 recall count
C5	hkllt4_n_clustering	Trial 4 semantic clustering count

Appendix

Table 7: Cognitive scoring primitives used in this work. “#”: “number of”, “z-score”: “standardized score relative to norms”.

Approach	Accuracy	Precision	Recall	F1
1	0.818	0.813	0.786	0.796
2	0.824	0.814	0.802	0.802
3	0.818	0.813	0.786	0.796

Appendix

Table 8: Comparison of AD screening performance using different MoCA-SL normative estimation approaches.

Age	Education	N	Median	IQR	Percentile
(years)	(years)				16th	7th	2nd
65–69	0–3	64	9.1	1.7	7.4	6.1	3.9
	4–6	82	10.0	2.2	8.2	7.8	5.6
	7–9	74	10.4	2.2	9.1	8.2	6.9
	10–12	82	10.8	1.7	9.5	8.7	7.4
	$>$ 12	67	11.7	1.3	10.8	10.0	9.1

Appendix

Table 9: Normative Table for MoCA-SL scores using proportional rescaling.

Task	Metric	Value
MoCA-SL
Picture Naming	SMR (Exact)	97.1%
	SMR ( $\pm 1$ )	100.0%
	MAE (Score)	0.029
	RMSE (Score)	0.171
Digit Span	SMR (Exact)	98.5%

Appendix

Table 10: Full metric breakdown for MoCA-SL and HKLLT tasks. This table reports all available performance indicators, including exact score matching (SMR Exact), tolerance-based matching (SMR

\pm 1

), and granular error metrics (MAE, RMSE) for both final scores and intermediate counts (e.g., recall number, item accuracy).

Panel of Figure 7 — Figure 7: An example of the Examiner’s prompt with a verifier step for MoCA Abstraction task. The left column shows the original Cantonese prompt, and the right column shows the English translation.

Panel of Figure 8 — Figure 8: An example of the Examiner’s prompt with a verifier step for MoCA Animal Fluency task. The left column shows the original Cantonese prompt, and the right column shows the English translation.

Panel of Figure 9 — Figure 9: Clinical Protocol Context is a section in Meta Analyst’s instruction prompt, showing core assessment indicators and clinical significance. The left column shows the original Chinese version, and the right column shows the English translation.

Panel of Figure 10 — Figure 10: An example of the prompt for generating cognitive profile reports. The left column shows the original Cantonese prompt, and the right column shows the English translation.

Panel of Figure 11 — Figure 11: An example of the model-generated cognitive function summary report. The left column shows the original Chinese output, and the right column shows the English translation.

Panel of Figure 12 — Figure 12: An example of the model-generated cognitive function summary report for a normal case. The left column shows the original Chinese output, and the right column shows the English translation.

Panel of Figure 13 — Figure 13: An example of the LLM-CoT baselineprompt with step-by-step reasoning for Alzheimer’s diagnosis assessment. The left column shows the original Chinese prompt, and the right column shows the English translation.

Panel of Figure 14 — Figure 14: An error case in the MoCA Serial Subtraction task. The examiner generates hallucinated numbers. With the verifier’s feedback, the examiner simply removes hallucinated numbers and gives up on extracting other new numbers. The left column shows the original Cantonese prompt, and the right column shows the English translation.

Panel of Figure 15 — Figure 15: An error case in the MoCA Abstraction task. In this boundary case, it”s hard to judge if ”use wheels to move” is an abstract concept, while the assessor judges it as correct. The left column shows the original Cantonese prompt, and the right column shows the English translation.

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

Explainable and Generalisable LLM-based Cognitive Decline Detection with Spontaneous Speech
Sep 28, 2026Ziyun Cui, Wen Wu, Chuan Shi +10Explainable Artificial Intelligence Speech-Based Dementia Detection
Tsinghua University, Beijing, China. · Shanghai Artificial Intelligence Laboratory, Shanghai, China. · Peking University Sixth Hospital, Beijing, China. +3
AD-CARE: A Guideline-grounded, Modality-agnostic LLM Agent for Real-world Alzheimer's Disease Diagnosis with Multi-cohort Assessment, Fairness Analysis, and Reader Study
Mar 26, 2026Wenlong Hou, Sheng Bi, Guangqian Yang +16Alzheimer's Disease Classification Clinical Decision Support
Department of Biomedical Engineering, The Hong Kong Polytechnic University, Hong Kong SAR, China. · Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, Beijing, China. · Amazon, USA. +10
Tabular LLMs for Interpretable Few-Shot Alzheimer's Disease Prediction with Multimodal Biomedical Data
Mar 17, 2026Sophie Kearney, Shu Yang, Zixuan Wen +8Alzheimer's Disease Classification Tabular Foundation Models
University of Pennsylvania, Philadelphia, PA, USA · Stony Brook University, Stony Brook, NY, USA · University of North Carolina at Chapel Hill, Chapel Hill, NC, USA +2

Sep 28, 2026J/K move · Enter open · S save

Explainable and Generalisable LLM-based Cognitive Decline Detection with Spontaneous Speech

Ziyun Cui, Wen Wu, Chuan Shi, Shuguang Yang+9

Tsinghua University, Beijing, China. · Shanghai Artificial Intelligence Laboratory, Shanghai, China. · Peking University Sixth Hospital, Beijing, China.+3

Explainable Artificial Intelligence Speech-Based Dementia Detection Speech Language Models

Open paper AskPDF arXiv

Alzheimer's disease (AD) and mild cognitive impairment (MCI), which may precede AD, manifest early through subtle linguistic and acoustic alterations. Traditional diagnostics, however, are often resource-intensive and lack scalability for mass screening. To address these challenges, we introduce a novel bilingual speech large language model framework for automated, explainable cognitive screening. Unlike conventional pipelines that rely on error-prone automatic speech recognition, our system directly processes raw speech to learn joint acoustic-semantic representations, preserving critical prosodic cues often lost in transcription. Utilising our newly collected PUTH-AD dataset alongside multiple open-source corpora, we implemented a multi-task learning objective that simultaneously performs cognitive status classification and generates clinician-understandable natural language explanations. Our system achieved the highest average accuracy and AUROC across six dataset/task conditions, comparing three representative baselines. The system demonstrated cross-task transfer to held-out PUTH-AD task subsets, maintaining classification accuracy on an entirely unseen cognitive task without task-specific fine-tuning. Furthermore, clinician evaluation confirms that the generated explanations are both clinically relevant and largely consistent with the underlying speech evidence, supporting their potential utility in clinical interpretation. This study provides a scalable, objective, and explainable framework for speech-based cognitive screening, combining cognitive status classification with natural language explanations that clinicians can assess and verify, bridging the gap between advanced AI and clinical utility.

Abstract

Alzheimer's Disease Classification

Explore similar work

Explainable and Generalisable LLM-based Cognitive Decline Detection with Spontaneous Speech

AD-CARE: A Guideline-grounded, Modality-agnostic LLM Agent for Real-world Alzheimer's Disease Diagnosis with Multi-cohort Assessment, Fairness Analysis, and Reader Study

Tabular LLMs for Interpretable Few-Shot Alzheimer's Disease Prediction with Multimodal Biomedical Data