STRAIN - Stress Tracking & Response Analysis Network

Homepage of the strain website
EEG scans being processed
Results being generated after processing
Statistics reported
Report generated
Strain Logo
Output of strain agent by strain agent on Prompt Opinion platform
Graph + Image output
Markdown Report Output

Inspiration

Mental health care has a data problem. Clinicians rely almost entirely on
self-reported symptoms - but by the time a patient articulates that they're
overwhelmed, burned out, or unable to focus, they've often been struggling for
months. We asked: what if a 2-second brainwave snapshot could quantify what a person can't put into words?

STRAIN (Stress and Affect Recognition via Neural Inference) was built to answer
that question.

What it does

STRAIN is a real-time EEG analysis platform that processes brainwave signals from a 14-channel EMOTIV EPOC+ headset to generate:

Affective state predictions - Valence, Arousal, and Dominance scores (1–5) load (β/α ratio), mapped to a 0–100 clinical scale
Explainability - top EEG signal contributions (band power, channel variance) driving each prediction
FHIR R4 export - structured clinical data bundles ready for EHR integration

The platform includes a full-stack web dashboard (React + Three.js 3D brain visualization), a FastAPI backend, and an MCP-connected AI agent on Prompt Opinion that can converse with patients, triage their symptoms, redirect them to a scan, and generate a full clinical report - complete with a downloadable Markdown file formatted for doctor consultation.

The demo narrative

Sam Rivera is a 26-year-old PhD student. She opens the STRAIN agent on Prompt Opinion and says she's overwhelmed - racing thoughts, can't focus, three deadlines she can't face. The agent empathises, explains that STRAIN can measure cognitive load directly from her brainwaves, and drops a link to the dashboard.

Sam runs a scan. Her results: Valence 1.0 / 5, Arousal 4.0 / 5, β/α ratio 57.6 (normal range: 1–3), cognitive load 100% - an extreme ADHD-stress overlap signature identified in the DREAMER dataset. She returns to the agent, asks for her analysis, and receives a full neural screening report with EEG electrode topography, VAD charts, risk indicators, and personalised next steps - including a recommendation for formal ADHD screening (Conners ACRS).

How we built it

Data pipeline: DREAMER.mat (Zenodo) → sliding-window export (256 samples, 50%
overlap, 128 Hz) → (14, 256) epoch tensors → Welch PSD feature extraction → subject-holdout Ridge regression (VAD).

Backend: FastAPI orchestrator with agent pipeline (DataCurator → AnalysisEngine → Explainer), FHIR R4 bundle generator, and a matplotlib EEG electrode topography renderer.

Frontend: React 19 + TypeScript, Three.js/react-three-fiber 3D brain with
animated dendrite lines and electrode markers, 9-second animated scan sequence.

AI integration: FastMCP server exposing strain-tools over SSE/Streamable
HTTP. Connected to Prompt Opinion BYO agent - the agent calls analyze_named_patient_tool, which runs the full pipeline and returns a rich
Markdown report with inline charts (QuickChart.io), brain scan image, and a downloadable report URL.

Interoperability: FHIR R4 Observation bundles for all screening outputs -
designed for real EHR integration.

## Challenges

EEG feature leakage: Getting subject-holdout cross-validation right so the
model generalises across people, not just memorises one subject's baseline.
MCP + LLM hallucination: The Po agent repeatedly fabricated URLs and patient names until we locked the tool docstrings, removed example URLs, and added hard
REQUIRED: directives in the system prompt.
Real-time feel from batch data: The 9-second scan animation needed to feel
live while actually processing a pre-recorded DREAMER epoch - deterministic channel power seeding made it reproducible and convincing.
Clinical framing without overclaiming: Every output carries a disclaimer. The language is clinical but the system is explicitly not a medical device - threading that needle in the UI and agent responses took deliberate iteration.

What we learned

Brain-computer interfaces don't need to be invasive or expensive to be clinically
informative. A consumer-grade 14-channel headset, open research data, and a well-structured ML pipeline can produce explainable, FHIR-ready outputs that a
clinician could actually use as a conversation starter - not a diagnosis, but a signal worth discussing.