Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis

Choonghan Kim; Hyunmin Hwang; Hangeol Chang; Jaemin Kim; Jinse Park; Jae-Sung Lim; Jong Chul Ye

cs.CLJan 6, 2026

Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis

Authors: Choonghan Kim, Hyunmin Hwang, Hangeol Chang, Jaemin Kim, Jinse Park, Jae-Sung Lim, Jong Chul Ye

Organizations: Graduate School of AI, KAIST, Republic of Korea · Haeundae Paik Hospital, Inje University, Republic of Korea · Asan Medical Center, University of Ulsan College of Medicine, Republic of Korea

Abstract

While Large Language Models (LLMs) have shown strong performance on clinical text understanding, they struggle with longitudinal prediction tasks such as dementia prognosis, which require reasoning over complex, non-monotonic symptom trajectories across multiple visits. Standard supervised training lacks explicit annotations for symptom evolution, while direct Reinforcement Learning (RL) is hindered by sparse binary rewards. To address this challenge, we introduce Dementia-R1, an RL-based framework for longitudinal dementia prognosis from unstructured clinical notes. Our approach adopts a Cold-Start RL strategy that pre-trains the model to predict verifiable clinical indices extracted from patient histories, enhancing the capability to reason about disease progression before determining the final clinical status. Extensive experiments show that Dementia-R1 achieves the best overall performance on the AMC real-world unstructured cohort, reaching an AUROC of 84.02% and outperforming models up to 10x larger. The framework also generalizes to Parkinson's disease dementia prediction in an independent hospital cohort, achieving an AUROC of 78.37%. On the ADNI benchmark, our 7B model attains the highest AUROC among all LLM baselines at 83.17%, demonstrating strong longitudinal reasoning over fluctuating cognitive trajectories. Code is available at https://anonymous.4open.science/r/dementiar1-CDB5.

Figures & tables

Figure 1: Multi-dimensional Performance Profile. Dementia-R1 demonstrates a consistent and balanced performance gain across all dimensions, including intermediate clinical reasoning tasks (e.g., MMSE, CDR-SB, ADAS-Cog) and the final dementia prognosis (F1, AUROC, AUPRC).

Figure 2: Overview of the Dementia-R1 Framework. Stage 1 pre-trains longitudinal reasoning via GRPO on clinical index forecasting; Stage 2 fine-tunes for the final dementia prognosis.

Panel of Figure 3 — Figure 3: Experimental setup: sample construction and cohort composition.

Figure 4: Performance across time gaps. AUROC (top) and AUPRC (bottom) by prediction horizon for (a) AMC, (b) HPH, and (c) ADNI. Dementia-R1 achieves the highest AUROC at the longest prediction horizons across all three cohorts.

Figure 5: Clinical Index Density. ADNI (

{\sim}

4.3

\times

) and AMC (

{\sim}

1.6

\times

) have more index visits per patient than HPH.

Figure 6: Neurologist Blind Pairwise Evaluation . Results from two cohorts (AMC and HPH) comparing Dementia-R1 against Qwen2.5-32B-Instruct across five clinical dimensions.

Figure 7: Stage 1 Pre-training Impact. AUROC over training epochs on (a) AMC, (b) HPH, and (c) ADNI. Solid lines show 500-step binned averages; dashed lines show raw checkpoint values.

Figure 8: Pretraining Density Ablation. Best overall AUROC at varying Stage 1 index density levels on (a) AMC and (b) ADNI. Performance degrades monotonically with reduced density.

Figure 9: Note vs. Index Ablation. AUROC by prediction window on (a) AMC and (b) HPH. Three input settings are compared: Note Only (raw clinical text), Index Only (structured MMSE/GDS/CDR), and Note + Index (combined).

Figure 10: Clinical Index Masking Ablation. AUROC change (

\Delta

) when masking clinical indices at inference. Conditions are named by the index removed: M = MMSE, G = GDS, C = CDR in (a) AMC and (b) HPH, and M = MMSE, C = CDRSB, A = ADAS-Cog/memory in (c) ADNI. Full retains all indices and

-

All masks every index. Hatching marks conditions in which MMSE is masked. Exact values are in Appendix A.11 (AMC) and A.10 (ADNI).

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Clinical Index	Range	Tolerance ( $\delta$ )
MMSE	0–30	$\pm 2$
CDRSB	0–18	$\pm 1.0$
ADAS-Cog 11	0–70	$\pm 5$
ADAS-Cog 13	0–85	$\pm 6$
ADAS-Cog Q4	0–10	$\pm 1$
RAVLT (Learning)	-20–20	$\pm 3$

Appendix

Table 8: Tolerance Thresholds for ADNI Indices. Error margins (

\delta

) were scaled proportionally to the range of each metric.

Figure 11: Time Gap Distribution by Cohort (Test Sets). The histograms show the interval between the last available clinical note and the diagnosis date across three cohorts. Intervals $\geq$ 18 months are consolidated into a single $>$ 18m bucket. The AMC and HPH cohorts are concentrated in shorter intervals (6–18m), reflecting relatively dense clinical follow-up prior to diagnosis. In contrast, the ADNI cohort displays a substantially larger proportion of long-term samples ( $>$ 18m), reflecting the longitudinal nature of MCI progression monitoring.

Figure 12: Multi-dimensional Performance Profile (AMC). Radar chart combining Stage 1 clinical index accuracy (MMSE, GDS, CDR) and Stage 2 dementia prognosis (F1, AUROC, AUPRC). Dementia-R1 achieves the most balanced profile overall.

Figure 13: Multi-dimensional Performance Profile (HPH). Radar chart combining Stage 1 clinical index accuracy (MMSE, GDS) and Stage 2 PDD prognosis (F1, AUROC, AUPRC). Dementia-R1 dominates on all prognosis metrics.

Model Method	Size	MMSE	GDS	CDR	Average
External LLMs
HuatuoGPT-o1	8B	44.3 $\pm$ 0.5	36.9 $\pm$ 2.5	43.7 $\pm$ 8.7	41.1 $\pm$ 0.6
MedGemma	27B	52.4 $\pm$ 0.3	44.2 $\pm$ 0.5	42.3 $\pm$ 2.1	47.8 $\pm$ 0.3
Qwen2.5-7B-Inst	7B	53.1 $\pm$ 1.6	49.4 $\pm$ 1.2	58.3 $\pm$ 3.2	52.1 $\pm$ 1.5
Qwen2.5-32B-Inst	32B	55.2 $\pm$ 1.2	48.3 $\pm$ 2.2	69.0 $\pm$ 3.6	53.9 $\pm$ 0.2
Qwen2.5-72B-Inst	72B	56.1 $\pm$ 1.3	52.4 $\pm$ 1.8	71.0 $\pm$ 1.0	56.3 $\pm$ 1.0

Appendix

Table 10: Performance on Clinical Index Prediction for the AMC cohort. We evaluate the accuracy of predicting MMSE, GDS, and CDR scores. Bold / underline = best/second-best. Mean

\pm

std over 3 seeds.

Model Method	Size	MMSE	GDS	Average
External LLMs
HuatuoGPT-o1	8B	40.5 $\pm$ 4.3	26.3 $\pm$ 1.0	33.4 $\pm$ 1.7
MedGemma	27B	43.7 $\pm$ 1.9	32.1 $\pm$ 1.0	37.9 $\pm$ 0.9
Qwen2.5-7B-Inst	7B	48.4 $\pm$ 2.1	35.5 $\pm$ 1.2	41.9 $\pm$ 0.5
Qwen2.5-32B-Inst	32B	49.0 $\pm$ 0.9	36.1 $\pm$ 0.6	42.6 $\pm$ 0.2
Qwen2.5-72B-Inst	72B	49.9 $\pm$ 2.6	40.6 $\pm$ 1.6	45.2 $\pm$ 1.2

Appendix

Table 11: Performance on Clinical Index Prediction for the HPH cohort. MMSE and GDS accuracy. Bold / underline = best/second-best. Mean

\pm

std over 3 seeds.

Clinical Index	Slots	Accuracy (%)
MMSE	200	97.5
GDS	200	98.5
CDR	200	99.5
Overall	600	98.5

Appendix

Table 25: LLM Extraction Audit Results (Initial). Slot-level accuracy of the Qwen2.5-14B-Instruct extractor on 200 randomly sampled clinical notes (600 total slots).

Error type	Count	Description
Fabrication	3	A value is extracted although only the index keyword appears in the note.
False negative	4	The value is missed because it falls outside the extraction boundary window.
Misattribution	4	An overloaded abbreviation is resolved to the wrong scale (e.g., “GDS” as Geriatric Depression Scale rather than Global Deterioration Scale).
Total	11

Appendix

Table 27: Error Type Breakdown (Expanded Audit). The 11 extraction errors are case-level random errors rather than a systematic, directional bias.

Injected noise	Best dementia AUROC	$\Delta$ vs. 0%
0%	0.7945	—
30%	0.7926	$-0.19$ pp
50%	0.7929	$-0.16$ pp
70%	0.7902	$-0.43$ pp

Appendix

Table 28: Downstream Robustness to Stage 1 Extraction Noise. Best dementia AUROC after injecting above-tolerance noise into the Stage 1 clinical indices.

Figure 14: Clinical Index Distribution (AMC Test Set). Number of MMSE, GDS, and CDR measurements per patient. CDR annotations are substantially sparser than MMSE and GDS.

Target	AMC	HPH
All (combined)	79.68	70.76
MMSE	78.82	69.95
GDS	81.03	70.88
CDR	78.88	–

Appendix

Table 30: Pretraining Target Ablation. Overall AUROC (%) when Stage 1 uses individual vs. combined targets. Best per cohort in bold .

Figure 15: Temporal Density Ablation. Overall AUROC at varying measurement density levels across three cohorts. (a) AMC: 0/50/100%; (b) HPH: 0/100%; (c) ADNI: 25/50/75/100%. Performance degrades monotonically with reduced density.

Panel of Figure 16 — Figure 16: Example of longitudinal clinical notes. To protect patient privacy, all personal identifiers (e.g., names) were removed, and dates were randomly shifted.

Panel of Figure 17 — Figure 17: A comparison of the reasoning processes generated by Dementia-R1 and the Qwen2.5-32B for the clinical case shown in Figure 16 . While Dementia-R1 produces a structured and grounded rationale, the baseline generates a dense narrative containing medical hallucinations.

Panel of Figure 18 — Figure 18: Example of longitudinal clinical notes from the HPH cohort for PDD prediction. Unlike the AMC cohort’s SOAP-format notes, HPH records consist of structured fields (S/O/A/P) combined with separate clinical index records, autonomic tests, neuroimaging, gait analysis, and motor symptom descriptions. To protect patient privacy, all personal identifiers were removed, and dates were randomly shifted.

Panel of Figure 19 — Figure 19: A comparison of the reasoning processes generated by Dementia-R1 and Qwen2.5-32B for the HPH clinical case shown in Figure 18 . Both models correctly predict non-progression to PDD, but Dementia-R1 provides a more structured analysis that systematically separates cognitive, functional, and motor assessments, while the baseline produces a less organized narrative.

Panel of Figure 20 — Figure 20: Pre-training prompt template for the AMC dataset. While MMSE is shown as an example, the model is pre-trained to predict various global cognitive scores, including GDS and CDR , based on unstructured clinical notes.

Panel of Figure 21 — Figure 21: Pre-training prompt template for the ADNI dataset. The model predicts future indicators (e.g., MMSE, CDRSB) by analyzing longitudinal structured assessment data. The input includes a variable glossary to assist in interpreting clinical indicators.

Panel of Figure 22 — Figure 22: Fine-tuning prompt template for the AMC dataset. The task requires detecting dementia presence based on unstructured clinical notes and integrated scoring indicators.

Panel of Figure 23 — Figure 23: Fine-tuning prompt template for the HPH dataset. The task requires predicting Parkinson’s disease dementia (PDD) progression based on longitudinal clinical notes including cognitive assessments, motor symptoms, and gait analysis.

Panel of Figure 24 — Figure 24: Fine-tuning prompt template for the ADNI dataset. The model predicts MCI-to-dementia conversion using longitudinal trends of clinical indices and biomarkers.

Panel of Figure 25 — Figure 25: Prompt template for generating diagnostic rationales for Supervised Fine-Tuning.

Clinical Outcome Prediction

Clinical outcome prediction uses patient records, imaging, signals, and biomarkers to estimate risks, disease trajectories, treatment response, and survival.

See all 265 papers

Dementia-Agents: A Multi-Modal Multi-Agent System for Dementia Staging and Phenotyping
Jun 19, 2026Yaling Shen, Maja Christensen, Yiwen Jiang +4Medical Diagnosis Multimodal Agents
Monash University · Eastern Health · Lived Experience Advisor
Cerebra: A Multidisciplinary AI Board for Multimodal Dementia Characterization and Risk Assessment
Mar 23, 2026Sheng Liu, Long Chen, Zeyun Zhao +16Clinical Risk Prediction Clinical Decision-Making
Department of Biomedical Data Science, Stanford University · Department of Radiation Oncology, Stanford University · Center for Data Science, New York University +13
Training Large Language Models to Predict Clinical Events
May 12, 2026Benjamin Turtel, Paul Wilczewski, Kris SkotheimHealthcare Language Model Calibration
Lightning Rod Labs

Jun 19, 2026J/K move · Enter open · S save

Dementia-Agents: A Multi-Modal Multi-Agent System for Dementia Staging and Phenotyping

Yaling Shen, Maja Christensen, Yiwen Jiang, Jenna Dennison+3

Monash University · Eastern Health · Lived Experience Advisor

Medical Diagnosis Multimodal Agents Multi-Agent LLMs

Open paper AskPDF arXiv

Dementia diagnosis requires integrating multi-modal clinical assessments from diverse informants and clinicians under incomplete and heterogeneous data conditions. Yet most AI-driven approaches remain Alzheimer's disease (AD)-centric, framing the problem as binary AD detection or three-stage AD progression modeling within well-curated research settings. This pathology-driven paradigm overlooks the broader, syndrome-level nature of dementia, which spans multiple stages, phenotypes, and etiologies. In this paper, we propose Dementia-Agents, a clinically aligned multi-agent framework for real-world dementia staging and phenotyping. The framework follows a three-step workflow: (1) a data agent translates structured clinical records into semantically faithful textual representations that preserve missing-data signals and routes them to domain-aligned experts; (2) five fine-tuned expert agents generate domain-level predictions; and (3) a coordinator agent performs probabilistic aggregation to produce final staging and phenotyping decisions. We develop and evaluate Dementia-Agents on a real-world clinical cohort of 1,066 patients from two cognitive neurology services. Compared with monolithic multi-modal large language models (MLLMs) and prior medical multi-agent systems, our approach achieves consistent improvements in diagnostic performance for real-world syndrome-level dementia staging and phenotyping, while preserving domain-level interpretability.

Abstract

Clinical Outcome Prediction

Explore similar work

Dementia-Agents: A Multi-Modal Multi-Agent System for Dementia Staging and Phenotyping

Cerebra: A Multidisciplinary AI Board for Multimodal Dementia Characterization and Risk Assessment

Training Large Language Models to Predict Clinical Events