capstone ai-internship neurology research-papers

Week 1–2 Research Papers — Same Problem Space

Companion to “Week 1 - Neurology Research.md”. Curated list of academic papers that frame the exact same wedge your friend is researching.

All papers below address one or more of:

Ambient AI scribes / clinical documentation burden
LLM-driven clinical reasoning on neurology cases
Multimodal chart + imaging + EEG synthesis
AI–clinician disagreement / missed findings
EHR summarization with retrieval-augmented generation

Tier 1 — Read these first (directly on your wedge)

1. Ambient AI Scribes Reduce Burnout — But Lower Note Quality

Olson KD et al. (Yale + 5 health systems). JAMA Network Open. Oct 2, 2025. doi:10.1001/jamanetworkopen.2025.34976

263 ambulatory clinicians, Abridge for 30 days
Burnout: 51.9% → 38.8% (74% lower adjusted odds)
Equivalent to ~10.8 min saved per workday
Neurology/Psychiatry was 5.3% of sample (small, but directionally same)
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2839542

2. AI Scribe Notes Are Lower Quality Than Human Notes

Reddy A et al. (VA / Univ of Washington). Annals of Internal Medicine. Apr 17, 2026.

5 standardized primary care visits, 11 AI scribes vs. 18 human clinicians, 30 blinded raters using PDQI-9
Human notes scored higher in every case, especially on thoroughness, organization, usefulness
Authors conclude AI scribes are draft tools, not substitutes
https://mednews.uw.edu/news/AI-scribes-lower-quality

3. Abridge at KUMC — Detailed Operational Data

JAMIA Open. Feb 21, 2025. doi:10.1093/jamiaopen/ooaf013

99 post-implementation survey responses
81% said documentation workflow easier
77% said it improved patient care through decreased documentation burden
Median note generation: 76 sec (Jul 2023) → 38 sec (Apr 2024)
https://pmc.ncbi.nlm.nih.gov/articles/PMC11843214/

4. AI Scribes Save Under a Minute Per Note (real-world review)

STAT review. Apr 1, 2026.

Aggregate of published studies: scribes saved clinicians under 1 minute per clinical note
Despite marketing, modest real-world time savings
https://www.statnews.com/2026/04/01/ai-ambient-scribes-modest-time-savings-clinical-documentation/

5. Hype vs Reality in AI Clinical Integration

PMC12700513. 2025.

Overreliance causes clinicians to accept outputs uncritically
Neglecting clinical cues and diminishing independent judgment
https://pmc.ncbi.nlm.nih.gov/articles/PMC12700513/

Tier 2 — Clinical reasoning LLMs on neurology cases

6. LLM Outperforms Physicians on Clinical Reasoning

Brodeur PG, Buckley TA, Kanjee Z, Goh E et al. Science. Vol 392, Issue 6797, pp 524–527. Apr 30, 2026. doi:10.1126/science.adz4433

OpenAI o1 series tested against hundreds of physicians
On landmark diagnostic cases: physicians with GPT-4 = 76%, physicians with conventional resources = 74%, GPT-4 alone = 92%
Real-world ER second-opinion study at Beth Israel Deaconess
Best paper for the “LLM as clinical reasoning layer” argument
https://www.science.org/doi/10.1126/science.adz4433

7. Multi-Agent Framework for Neurological Clinical Reasoning

Sorka M, Gorenshtein A, Aran D, Shelly S (Technion + Mayo). PLOS Digital Health. Dec 4, 2025. doi:10.1371/journal.pdig.0001106

10 LLMs on 305 Israeli Board Certification Neurology exam questions
Five-agent framework (complexity classifier, interpreter, retrieval, synthesis, validator)
LLaMA 3.3-70B: 69.5% (base) → 73.4% (RAG) → 89.2% (agentic)
Domain-specialized models (Meditron-70B, MedLLaMA3) performed worse than general models
Most directly relevant paper for “reasoning layer over multimodal neuro data”
https://pmc.ncbi.nlm.nih.gov/articles/PMC12677565/

8. Neural-MedBench — Reasoning Benchmark for Multimodal Clinical AI

arXiv:2509.22258. Sep 2025.

Compact reasoning-intensive benchmark probing limits of multimodal clinical models
https://arxiv.org/html/2509.22258v1

Tier 3 — Neurology-specific documentation gaps

9. ChatGPT Neurology Referral Letters — Completeness Study

Rattananan W. Medicina (Kaunas). Oct 28, 2025. doi:10.3390/medicina61111931

5 standardized neurology scenarios × 10 letters = 50 letters using ChatGPT-4o
Mean total score 25.76/30 (87% completeness, 84% quality)
72% of letters (36/50) had content gaps in HPI and physical exam sections
Management was the lowest-scoring domain (72.7%)
Direct evidence of the gap your product would fill
https://pmc.ncbi.nlm.nih.gov/articles/PMC12653999/

10. Standardized Note Templates Improve Quality Metrics in Resident Clinics

Breithaupt A et al. (UCSF + ZSFG). Neurology Education. Vol 4 No 1. Mar 5, 2025. doi:10.1212/NE9.0000000000200200

23 neurology residents, cluster-randomized trial
Condition-specific templates for epilepsy, headache, Parkinson’s
Significant quality improvements:
- Epilepsy: driving documentation 92% vs 53% (p=0.002)
- PD: medication-related motor symptoms 95% vs 50% (p=0.01)
- Headache: lifestyle counseling 77% vs 21% (p=0.005)
No efficiency gain — templates made notes longer, not faster
Critical: shows that just having templates improves “missed finding” rate dramatically — same wedge as your product
https://www.neurology.org/doi/10.1212/NE9.0000000000200200

11. Completeness of Neurology Referral Letters (general)

medRxiv 2025.06.12.25329503. 2025.

Letters scored on demographics, chief complaint, HPI, physical exam, management, consultation questions
Wide variability, “clear documentation of attempted therapies to minimal or absent”
https://www.medrxiv.org/content/10.1101/2025.06.12.25329503v1.full.pdf

Tier 4 — Missed findings, AI-clinician disagreement

12. AI–Clinician Disagreement Prediction Pipeline

Sanchez M et al. (Harvard / Stanford). Cell Reports Medicine. Oct 17, 2023. doi:10.1016/j.xcrm.2023.101207

Deployed chest X-ray AI — disagreement rate with radiologists = 6.5%
Built pipeline modeling disagreement + significance + confidence
Expected burden reduction: 4.8%
Key insight: diagnostic model alone insufficient, need ecosystem of models around it
https://www.sciencedirect.com/science/article/pii/S2666379123003749

13. NLP Limits AI–Radiologist Report Discrepancies

Radiology Business. Northwestern Medicine build.

NLP software detected missed diagnoses on high-acuity CT
https://radiologybusiness.com/topics/health-it/enterprise-imaging/natural-language-processing-can-limit-report-discrepancies-between-ai-and-radiologists

14. Patient Reactions to AI–Clinician Discrepancies

PMC12141964. 2025.

How discrepancies between AI-derived and radiologists’ recommendations affect patients
Critical for trust model design
https://pmc.ncbi.nlm.nih.gov/articles/PMC12141964/

15. Liability When AI Finds What Radiologist Missed

AuntMinnie. 2025.

Radiologists more legally culpable if AI catches an abnormality they missed
https://www.auntminnie.com/imaging-informatics/artificial-intelligence/article/15747540/if-ai-finds-an-abnormality-that-a-radiologist-misses-whos-at-fault

16. Missed Radiology Follow-Ups — Northwestern Program

PRNewswire / Northwestern Medicine. 2022.

Built internal AI to trigger follow-up on incidental imaging findings
https://www.prnewswire.com/news-releases/using-artificial-intelligence-to-solve-one-of-health-care-s-most-enduring-problems-301504387.html

Tier 5 — Multimodal foundation models (the AI architecture wedge)

17. Medical Multimodal Foundation Models Review

Sun K, Xue S et al. (Tsinghua). arXiv:2412.02621. Dec 3, 2024.

Comprehensive review across datasets, architectures, clinical applications
Two categories: MMVFMs (vision-only) and MMVLFMs (vision-language)
Covers MRI, CT, X-ray fusion with clinical reports
https://arxiv.org/html/2412.02621v1

18. Foundation Models for Radiology

dirjournal.org. 2025. doi:dir.2025.253445

Applications, opportunities, challenges, risks
https://www.dirjournal.org/articles/foundation-models-for-radiology-fundamentals-applications-opportunities-challenges-risks-and-prospects/doi/dir.2025.253445

19. Vision-Language Foundation Model for 3D Medical Imaging

Nature s44387-025-00015-9. 2025.

3D medical imaging with VL foundation models
https://www.nature.com/articles/s44387-025-00015-9

20. Multimodal Dataset and Benchmarks for Vietnamese PET/CT

NeurIPS 2025.

Adapting general VLMs to medical domain
https://neurips.cc/virtual/2025/poster/121676

Tier 6 — RAG for clinical text

21. RAG on Clinical Notes for Diagnostic Reasoning

arXiv:2401.10733v2. 2024.

Dynamic QA of clinical documents using retrieval-augmented generation
https://arxiv.org/html/2401.10733v2

22. Enhancing Medical AI with RAG (Review)

PMC12059965. 2025.

Combines information retrieval + generative models for medical tasks
https://pmc.ncbi.nlm.nih.gov/articles/PMC12059965/

23. Clinical Entity Augmented Retrieval

Nature npj Digital Medicine. 2024. doi:10.1038/s41746-024-01377-1

LLMs + RAG improved information extraction vs. prior methods
https://www.nature.com/articles/s41746-024-01377-1

24. Clinical Text Summarization with LLMs — Evidence Review

JMIR. 2025. doi:10.2196/68998

State of the art on clinical text summarization using LLMs
https://www.jmir.org/2025/1/e68998

25. Generative LLMs in EHRs — Systematic Review

Du X et al. (Brigham and Women’s / Harvard). JAMIA. 2024. doi:10.1093/jamia/ocaf233

76 studies since ChatGPT release, real EHR data
88.2% used zero-shot prompting
Only 2 studies used multimodal data
6 studies identified hallucinations (e.g., fabricated patient names)
Critical caveats: bias, hallucinations, impersonal tone
https://www.medrxiv.org/content/10.1101/2024.08.11.24311828v2.full-text

Tier 7 — Implementation / operational

26. LLM Verification of Patient Care Documents

Chung P et al. NEJM AI. 2025. doi:10.1056/AIdbp2500418

LLMs effectively summarize EHRs but require verification
https://ai.nejm.org/doi/full/10.1056/AIdbp2500418

27. LLM Performance on Clinical Reasoning Tasks

JAMA Network Open. 2025. doi:10.1001/jamanetworkopen.2024.47679

Differential dx less accurate than diagnostic testing
Final dx, management, miscellaneous reasoning more accurate
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2847679

28. Adapted LLMs Outperform Medical Experts in Clinical Text Summarization

arXiv:2309.07430v4. 2023.

LLMs outperform medical experts on multiple summarization tasks
https://arxiv.org/html/2309.07430v4

29. Neurologic Examination Adds ~6.7 Minutes Per Visit

PMC11620545. 2024.

Subgroup analysis: pain visits +6.7 min
https://pmc.ncbi.nlm.nih.gov/articles/PMC11620545/

30. Documentation Burden Measurement

PMC11534919. 2024.

Total EHR time → decreased patient satisfaction, lower communication ratings
https://pmc.ncbi.nlm.nih.gov/articles/PMC11534919/

Tier 8 — EHR / workflow integration

31. EHR Documentation Burden Crowds Out Health Information Exchange

Health Affairs. 2024. doi:10.1377/hlthaff.2024.00398

Each additional hour spent documenting → 7.1% reduction in HIE use
https://www.healthaffairs.org/doi/abs/10.1377/hlthaff.2024.00398

32. Hard-to-Use EHRs Linked to Physician Burnout

AMA News. 2024.

Every 1-point boost in EHR usability → 3% lower odds of burnout
https://www.ama-assn.org/practice-management/digital-health/new-research-links-hard-use-ehrs-and-physician-burnout

33. Ambient Documentation Technology — Clinician Experience

JAMA Network Open. 2025. doi:10.1001/jamanetworkopen.2024.37847

Survey study pre/post ambient documentation
https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2837847

34. Documentation Philosophy — Shorter Yet Information-Rich Notes

Rodríguez-Fernández JM. Frontiers in Digital Health. 2022. doi:10.3389/fdgth.2022.1063141

Argument for redesigning clinical notes
https://www.frontiersin.org/journals/digital-health/articles/10.3389/fdgth.2022.1063141/full

What this corpus tells us about your wedge

The gap is real and quantified:

Templates improve “missed findings” rate from 53% to 92% on driving status (Breithaupt)
72% of LLM-generated neurology referral letters had content gaps (Rattananan)
AI scribes reduce burnout 13 percentage points but produce lower-quality notes (Olson vs. Reddy tension)
Clinicians spend 2x more time on EHR/admin than patient care (KUMC/AHA data)

The architecture is feasible:

Multi-agent LLM framework hits 89.2% on neurology board questions (Sorka)
LLaMA 3.3-70B + RAG + agents outperforms GPT-4o on neurology reasoning
Multimodal foundation models for medical imaging are now production-grade (Tsinghua review)

The market is validated:

$150M Series E for Aidoc
Abridge = 2025/2026 Best in KLAS
Multiple health systems publishing positive pilot data

The unsolved wedge:

Nobody is merging ambient scribe + prior chart + imaging + EEG into one synthesis
6.5% AI-radiologist disagreement rate (Sanchez) is real and unaddressed by current scribes
Missed follow-ups on incidental findings remain a top-3 safety issue (Northwestern, Inflo Health)

Sources by topic

Ambient scribes:

#1, #2, #3, #4, #5, #33, #43 (JAMIA Open KUMC)

Neurology-specific:

#7, #9, #10, #11, #29

Multimodal / RAG / reasoning:

#6, #7, #8, #17, #18, #19, #20, #21, #22, #23, #24, #25, #26, #27, #28

Missed findings / disagreement:

#12, #13, #14, #15, #16

Documentation burden:

#30, #31, #32, #34

EHR workflow / integration:

#31, #32, #33, #34

Quartz 5

Explorer

Week 1-2 - Research Papers