Multimodal Diagnostic Copilot for Rare Neurological Disorders

Track: Genomic Diagnostics

Rare neurological disorders like Rett syndrome, Dravet syndrome, and Angelman syndrome carry an average diagnostic delay of 5–7 years. Children visit an average of 7 specialists before receiving a genetic diagnosis. Whole exome sequencing costs $4,000–6,000 and takes weeks. Targeted gene panels cost $300–500 and take days — but clinicians don't know which to order.

NeuroPhenotype tells clinicians which gene to sequence first — and why.

It fuses two independent data streams into a single actionable recommendation:

The output is not a diagnosis. It is the next best diagnostic step: a prioritized genomic panel recommendation with clinical justification, cost savings estimate, and a structured clinician note.

INPUT LAYER
├── Half 1: Passive Biosignals              Half 2: Clinical Record
│   ├── EEG (23 features) (Live)                   ├── HPO term checklist
│   ├── HRV - Apple Watch (7 features) (Live)    ├── Age of onset / severity
│   ├── GSR - Arduino BLE (4 features)      ├── Family history
│   ├── Movement - MediaPipe (7) (Live)        ├── Prior test history / PDF upload
│   ├── Speech - librosa (6 features) (Live)       └── Inheritance pattern
│   ├── rPPG - webcam (4 features)
│   └── Keyboard/Mouse (4 features)
│
├── FEATURE EXTRACTION → 55-dim vector
│
├── FUSION & CLASSIFICATION
│   └── XGBoost 4-class (Rett / Dravet / Angelman / Control)
│
├── CLAUDE REASONING LAYER
│   ├── Synthesized recommendation
│   ├── Next best diagnostic step
│   ├── Uncertainty statement ("confidence would increase with...")
│   ├── SOAP note (clinician view)
│   └── Plain English summary (patient/family view)
│
└── OUTPUT
    ├── Clinician View — SOAP note, genomic panel, cost, biomarker evidence
    └── Patient View — plain language summary, what happens next

• Python 3.10+
• (Optional) Arduino Nano 33 BLE Sense for IMU/GSR
• (Optional) OpenBCI Cyton or Muse for EEG
• (Optional) Apple Watch for HRV via BLE

# Clone
git clone https://github.com/NSang22/hurd26.git
cd hurd26

# Create virtual environment
python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # macOS/Linux

# Install dependencies
pip install -r neurophenotype/requirements.txt

cd neurophenotype
python dashboard/server.py

Open http://localhost:5050 in your browser. You'll see the sign-in page with two roles:

You can also Sign in with Google or click Skip (Demo) to bypass auth for quick hackathon demos.

• Demo Mode — Select a synthetic profile (Rett / Dravet / Angelman / Control) and run the pipeline with literature-derived mock data. No hardware needed.
• Clinical Mode — Run the live biosignal pipeline with real sensor input. Toggle individual sensors on/off as available.

neurophenotype/
├── modalities/                 # 7 modality implementations
│   ├── base.py                 # Abstract base class
│   ├── eeg.py                  # 23 features — OpenBCI/Muse/DIY circuit
│   ├── hrv.py                  # 7 features — Apple Watch BLE
│   ├── gsr.py                  # 4 features — Arduino BLE
│   ├── movement.py             # 7 features — MediaPipe + Arduino IMU
│   ├── speech.py               # 6 features — librosa, sounddevice
│   ├── rppg.py                 # 4 features — OpenCV webcam
│   └── keyboard_mouse.py       # 4 features — pynput
├── fusion/
│   └── fusion.py               # Config-driven modality loader + concatenation
├── classifier/
│   ├── model.py                # XGBoost wrapper with label encoder
│   ├── train.py                # Synthetic profile generation + training
│   └── inference.py            # Feature vector builder for inference
├── clinical/                   # Half 2 — Clinical Record Integration
│   ├── hpo.py                  # HPO term encoding, condition match scoring
│   ├── intake.py               # Clinical intake data model
│   ├── pdf_parser.py           # Claude API prior test PDF extraction
│   ├── soap.py                 # SOAP note + patient summary generation
│   └── claude_client.py        # Claude API client wrapper
├── dashboard/
│   ├── server.py               # Flask backend (port 5050) + auth
│   ├── landing.html            # Sign-in page (doctor/patient/Google/skip)
│   ├── index.html              # Main diagnostic dashboard
│   └── app.py                  # Streamlit fallback dashboard
├── arduino/
│   └── nano_ble_sense.ino      # IMU + GSR BLE streaming sketch
├── data/
│   ├── public/                 # Public EEG datasets
│   └── private/                # Private datasets (gitignored)
├── config.yaml                 # Toggle modalities on/off
├── main.py                     # CLI entry point
└── requirements.txt

The pipeline is fully modular. If a sensor isn't available, set it to false in config.yaml — the fusion layer concatenates only active modality vectors and the classifier will handle the missing features.

• Left — Input Layer: Sensor acquisition status, sensor toggles, HPO term checklist, clinical intake (onset, family history, prior testing)
• Center — Fusion + Classification: Live biosignal capture modals (movement via webcam, speech via mic), combined condition priorities, live clinical document with inline editing
• Right — Recommendation Layer: Top match, recommended genomic panel, confidence score, estimated savings, turnaround, HPO condition scores, uncertainty notes

After biosignal results return, the center column displays an interactive clinical paper document:

• HPO terms render as clickable styled spans (confirmed / absent / uncertain / unassessed)
• Onset, family history, and prior testing fields are editable via inline popovers
• Changes trigger live probability recalculation without a server round-trip
• Typewriter animation effect for text updates

• Username/password with SQLite backend (auto-created on first run)
• Google Sign-In (OAuth 2.0)
• Pre-seeded demo accounts for both doctor and patient roles
• Skip (Demo) button for hackathon presentations

• Pipeline continues with any subset of sensors — explicit confidence degradation flagged to the user
• HPO-only path available when all hardware is unavailable (low-resource deployment)
• Explicit low-confidence flagging rather than silent unreliable output
• Uncertainty surfaced for atypical presentations (late-onset, male Rett, etc.)
• Patient view designed for families navigating the diagnostic journey — plain language, no jargon

EEG — Rett (MECP2): Roche et al. 2019, J. Neurodevelopmental Disorders; Portnova et al. 2022, J. Personalized Medicine

EEG — Dravet (SCN1A): Hall et al. 2024, Ann. Child Neurology Society; Specchio et al. 2015, ScienceDirect

EEG — Angelman (UBE3A): Sidorov et al. 2017, J. Neurodevelopmental Disorders; Ostrowski et al. 2021, Ann. Clinical and Translational Neurology; Elber et al. 2022, Brain Communications

HRV / Autonomic: Julu et al. 2017, ScienceDirect; Singh et al. 2024, MDPI

Movement / Stereotypy: Temudo et al. 2007, Neurology; Dy et al. 2017, Movement Disorders; STOPme Project 2025, ScienceDirect

Keystroke Dynamics: Alfalahi et al. 2022, Scientific Reports; Gajos et al. 2020, Movement Disorders

This system is a research prototype for diagnostic triage. It does not provide a clinical diagnosis. All outputs are probabilistic recommendations intended to prioritize the next diagnostic step. Results should be interpreted by a qualified clinician in the context of the full clinical picture.