capstone ai-internship neurology subspecialties

Neurology Subspecialties Map

All 31 neurology subspecialty fellowships per AAN/ACGME/UCNS, with prevalence, AI maturity, opportunity assessment, and recommended wedge.

This is the first decision you make as a team: which subspecialty to focus on for 8 weeks. Use this map to narrow from “neurology” to 1-2 specific conditions.


How to read this map

Each subspecialty is scored on 5 dimensions (1-5 scale, 5 = best):

DimensionWhat it measures
US prevalenceHow many patients (1=rare <100K, 5=>10M)
AI maturityHow saturated the AI market is (1=greenfield, 5=fully saturated)
Pain severityPatient-reported pain intensity per Reddit research
Doctor shortageHow underserved the subspecialty is (5=most undersupplied)
Build difficultyHow hard for us to ship in 8 weeks (1=trivial, 5=needs imaging+signals+EHR)

Total opportunity = Prevalence × Pain × Shortage ÷ (AI maturity × Build difficulty)

A high total = the most leverage.


The 31 subspecialties (ACGME + UCNS accredited + non-accredited)

SubspecialtySitesUS prevalenceAI maturityPainShortageBuildTotalWedge
Headache / Migraine41 UCNS47M (2nd highest DALY globally)1 (greenfield)5 (massive dismissal)5 (564 accredited specialists vs. 3,700 needed — mdedge 2024)1 (transcript-only)25Scribe + pre-visit synthesis, no imaging layer
Movement disorders (PD, ET, dystonia)481.1M Parkinson’s, 7M ET2 (StrivePD, Cala Health)5 (UPDRS drift, dyskinesia)4 (4,000+ wait lists per Reddit)3 (chart + wearable)15Trajectory tracking + wearable merge
Epilepsy77 ACGME3.4M US, 50M global2 (Ceribell, encevis)5 (EEG dismissal = gaslighting)33 (chart + EEG)15EEG-visit synthesis, seizure calendar
Multiple sclerosis201.1M US3 (NeuroQuant, multiple DMTs)5 (13-year diagnostic journeys)32 (chart + MRI)10Trajectory + lesion tracking
Cognitive / Behavioral / Dementia35 (BNNP), 4 (geriatric)7.4M Alzheimer’s (2026, Alzheimer’s Association Facts & Figures), 100K FTD2 (Cortechs)5 (caregiver pain, 4-year diagnostic)5 (only 4 geriatric neuro sites — re-verify against UCNS directory)2 (chart + caregiver)25Caregiver-in-the-loop, longitudinal

Tier 2 — Solid opportunities

SubspecialtySitesUS prevalenceAI maturityPainShortageBuildTotalWedge
Vascular neurology / Stroke99 ACGME800K strokes/yr, 7M survivors5 (Viz, RapidAI, Brainomix, Aidoc all 510k’d)435 (imaging + workflow)3Saturated — skip unless you have a fresh angle
Neuromuscular49200K ALS, 200K+ others1443 (chart + EMG)8Niche, but underserved
Neurocritical care70 (UCNS+ACGME)(ICU subset)2 (Ceribell, Piramidal)434 (real-time signals)6Real-time ICU EEG, hard to ship
Sleep medicine8470M sleep disorder sufferers1 (mostly dental devices)322 (chart + sleep study)6Big market but not “neurology” in the AI sense
Pain medicine10450M chronic pain1431 (chart only)12Generalist, less neuro-specific
Clinical neurophysiology89(EMG/EEG labs)1333 (chart + signals)3Technical, narrow
Neuro-oncology3425K primary brain tumors/yr, 200K+ metastatic3 (Cercare, NeuroQuant)5 (high mortality)34 (chart + MRI + path)8High-stakes, hard wedge
Autonomic disorders5 UCNSrare145 (only 5 sites)2 (chart + vitals)8Very niche but underserved
Neuroimmunology(under MS)(under MS)3533 (chart + MRI + labs)8Overlap with MS

Tier 3 — Lower opportunity for our 8 weeks

SubspecialtySitesWhy lower
Endovascular Surgical Neuroradiology2 ACGMEPure stroke, saturated
Neuroimaging / Neuroradiology5 UCNSImaging AI is mature, hard to differentiate
Brain Injury Medicine1 ACGMEToo small
Clinical Neuromuscular Pathology5 UCNSPathology, not clinical
Neural Repair and Rehabilitation0 UCNSDoesn’t exist as fellowship
Neurohospitalist2Acute care, hard wedge
Balance Disorders, Neuropharmacology, Neurogenetics, etc.0Too small

Three conditions worth zooming into

1. Headache / Migraine (Total: 25)

Why this is the strongest starting point:

  • 47M US patients — biggest prevalence of any neuro condition
  • 3,700 headache specialists needed vs. 564 currently → severe shortage
  • 70% of migraine patients are undiagnosed or under-treated
  • Reddit research shows massive patient pain: “brushed off,” “yelled at,” “told to drink water”
  • Most chief complaint in general neurology visits (general neuro is mostly headache)
  • AI market is greenfield (no major incumbent)
  • Easiest to build: transcript-driven, no imaging needed
  • Reimbursement: high (Headache is one of the most common reasons for neuro referral)

Patient pain (verbatim from Reddit research):

  • “10-minute appointment, told to take vitamins, drink more water”
  • “Overheard my neurologist laugh at my appointment when I asked for a headache specialist”
  • “She literally told you to look at your scans then got annoyed that you had looked at your scans”
  • 15+ data points per headache visit (onset, location, quality, severity, frequency, aura, triggers, alleviating factors, etc.)

The wedge:

  • Pre-visit synthesis: pull prior visit data, MIDAS scores, abortive/preventive response, red flag screening
  • Generate the “structured headache note” (the 15+ data points)
  • Surface “this patient is on tier 3 preventives and still has 12 headache days/month — what next?”
  • Trigger specialist referral automatically when criteria met

Competitors to study:

  • DeepCura ($129/mo, claims neuro templates) — closest direct competitor
  • Abridge, DeepScribe — generic, no headache templates
  • Mayo Clinic, Cleveland Clinic — academic headache programs, no AI

Risk: No device data, no imaging, no signals — wedge is mostly software. Lower defensibility than movement disorders or epilepsy.

2. Movement disorders (Total: 15)

Why this is strong:

  • 1.1M Parkinson’s patients (Parkinson’s Foundation), 7M essential tremor
  • UPDRS scoring is required at every visit (manual, time-consuming, error-prone)
  • Apple Watch Movement Disorders API gives tremor/dyskinesia stream
  • Rune Labs StrivePD owns the device side but not the in-clinic synthesis
  • Patient pain: caregiver burden, off-time, dyskinesia side effects, no real-time medication adjustment

Patient pain (verbatim from Reddit research):

  • “Mom, 82, new neurologist thinks she doesn’t have Parkinson’s. This is after 13 years of treatment.”
  • “From living independently, driving, shopping to 24/7 care, can’t walk, can’t use the bathroom, can’t feed herself”
  • “The diagnosis of parkinson’s is still very much a subjective process of ruling out very bad brain conditions first, then finding the medicine that relieves parkinson’s symptoms. There is no definitive test that proves a person has parkinson’s disease.”

The wedge:

  • Pre-visit synthesis: pull UPDRS, medication response, wearable tremor/dyskinesia data
  • Generate the “movement disorder visit note” with quantitative trajectory
  • Surface “UPDRS has drifted +0.5 over 18 months, medication wearing off increasing — consider DBS eval”
  • In-clinic: capture the UPDRS exam as discrete data (tremor, rigidity, bradykinesia, postural stability)

Competitors to study:

  • Rune Labs StrivePD — owns wearable
  • Apple Movement Disorders API — owns tremor/dyskinesia capture
  • Medtronic, Abbott, Boston Scientific — own DBS
  • Cala Health — owns non-invasive stim for ET
  • NeuroQuant — owns MRI volumetrics

Risk: Need to integrate with wearable APIs (StrivePD has, Apple Watch does). Higher build difficulty than headache.

3. Cognitive / Dementia (Total: 25)

Why this is the highest social-impact play:

  • 7.4M Alzheimer’s patients in US (2026, Alzheimer’s Association Facts & Figures), projected to ~13M by 2050
  • Only 4 geriatric neurology fellowship sites in the whole US
  • Caregiver is the real information source (per Reddit research)
  • 4-year average diagnostic journey, often misdiagnosed
  • 75% of dementia caregivers report high stress
  • HIPAA makes caregiver information flow awkward
  • AAN workforce report explicitly calls this the most undersupplied subspecialty

Patient pain (verbatim from Reddit research):

  • “I type up two notes prior to all appointments. The first one is for the front desk staff… The second is for the doctor.”
  • “Sit yourself behind her so the doctor can see you. That way you can nod to confirm or shake your head if it’s not true and she will never know.”
  • “Dementia falls into geriatric specialties! THESE are the doctors that have the knowledge and tools to deal with dementia.” (her internist was useless)
  • “My mom proceeded to wreak complete havoc on the drive home. Hitting my dad, trying to pull the steering wheel…”

The wedge:

  • Caregiver-in-the-loop: structured intake for the caregiver (the one who knows the patient)
  • Pre-visit synthesis: prior MMSE/MoCA scores, behavioral changes, ADL/IADL drift
  • Generate the “dementia visit note” with cognitive trajectory
  • Surface “MMSE dropped 4 points in 18 months, behavioral changes in last 3 months, consider medication adjustment”
  • Side benefit: HIPAA-compliant caregiver communication channel

Competitors to study:

  • Cortechs NeuroQuant (volumetric MRI)
  • Cognito Therapeutics (sensory stimulation)
  • Eli Lilly / Eisai (Lecanemab, Donanemab — but these are drugs, not software)
  • Most memory clinics use paper-based caregiver intake

Risk: Long sales cycle (geriatric clinics, memory care centers), FDA considerations if you touch any clinical decision support.


If you want maximum leverage in 8 weeks, focus on:

Subset 1 (primary): Headache / Migraine

  • Easiest to build
  • Largest patient pool
  • Greenfield AI market
  • Direct overlap with what your friend is researching (general neuro chief complaint is mostly headache)

Subset 2 (cross-cutting feature): Caregiver-in-the-loop

  • Works for cognitive/dementia, stroke recovery, ALS, severe MS
  • Differentiates from every scribe in the market (none of them include the caregiver)
  • Maps to the highest pain point in the Reddit research
  • HIPAA-compliant structured communication channel is a defensible product feature

Why not just movement disorders or just MS? They’re solid but narrower, and your friend is interviewing general neurologists who see mostly headache. Headache gets you the most doctor-interview data and the most patient-side data (47M patients).


Quick-reference scoring table

SubspecialtyPrevalenceAI maturity (1=greenfield, 5=saturated)PainShortageBuildTotal
Headache/Migraine5155125
Cognitive/Dementia4255225
Movement disorders4254315
Epilepsy3253315
Multiple sclerosis3353210
Pain medicine4143112
Neuromuscular214438
Autonomic114528
Neuro-oncology235348
Neurocritical224346
Sleep413226
Stroke / Vascular454353
Clinical neurophysiology213333

Top 5 to consider: Headache, Cognitive/Dementia, Movement disorders, Epilepsy, MS.

Bottom line: skip stroke, skip clinical neurophysiology, skip anything where AI maturity is 4 or 5 (saturated). The highest-leverage moves are in conditions with greenfield AI markets and severe doctor shortages.


How this maps to your 8-week plan

WeekHeadache trackDementia trackMovement disorders track
1Map 20-30 competitors (mostly scribe)Map 20-30 competitors (Cortechs, Cognito)Map 20-30 competitors (Rune, Apple, Medtronic)
2Doctor interviews (general neurologists, headache specialists)Doctor interviews (geriatric neuro, memory clinics)Doctor interviews (movement disorder specialists)
3Label Studio on headache transcript dataLabel Studio on caregiver intake dataLabel Studio on UPDRS scoring data
4Pain point survey (Migraine Impact, MIDAS)Caregiver burden survey (Zarit Burden Interview)UPDRS / MDS-UPDRS scoring review
5-6Pre-visit synthesis architectureCaregiver-in-the-loop architectureWearable + chart synthesis architecture
7Deliverables checkDeliverables checkDeliverables check
8Demo: headache note generation + red flag screeningDemo: caregiver note → clinical summaryDemo: wearable data + chart → UPDRS drift alert

Sources