Talk to Your Heart (PulseForgeAI)

On-Campus Cardiac Rehab Intelligence — Edge AI on NVIDIA DGX Spark

Architecture • Features • Live Demo • Getting Started

Cardiac rehabilitation is a Class 1A recommended therapy that reduces all-cause mortality by 13% and hospitalizations by 31%. Yet only 24% of eligible Medicare beneficiaries ever attend a session, and fewer than 27% of those who start complete the program. The CMS Million Hearts initiative estimates closing this gap would save 25,000 lives and prevent 180,000 hospitalizations every year.

Talk to Your Heart solves the cardiac rehab supervision bottleneck with an AI copilot that runs entirely on-campus. The system ingests live Polar H10 ECG/ACC data via BLE, extracts interpretable clinical features through a dual-window signal processing pipeline, and coordinates two specialized AI agents — a patient-facing Nurse and a clinician-facing Clinical Assistant — grounded in deterministic safety checks, reference cohort matching, and RAG-retrieved AHA/AACVPR guidelines.

Proof of concept is live: pulse-forge-ai.vercel.app Production target: NVIDIA DGX Spark with zero PHI egress.

---

🧠 Architecture

+=============================================================================+
|                     NVIDIA DGX Spark (Production Target)                    |
|              (Current proof-of-concept: Vercel + Ollama tunneling)          |
|                                                                             |
| +-----------+  BLE   +------------------+  MQTT   +----------------------+  |
| | Polar H10 |------->| Dual-Window      |-------->| MQTT Broker          |  |
| | ECG 130Hz |        | Signal Engine    |         | (EMQX / Mosquitto)   |  |
| | ACC 100Hz |        | 5s: SQI,HR,HAR   |         +----------+-----------+  |
| +-----------+        | 30s: HRV,Morph   |              sub   v              |
|                      +------------------+                    |              |
| +-----------+                                                |              |
| | Google Fit|---> Intake JSON       +------------------------+-----------+  |
| | 7-day     |        |              | Lead Orchestrator                  |  |
| | Baseline  |        v              | (Role-Based Router + Safety)       |  |
| +-----------+ +---------------+     +----------+--------------------+    |  |
|               | ChromaDB      |                |          |              |  |
|               | • RAG Lit.    |                v          v              |  |
|               | • Cohorts     |           +--------+ +-----------+       |  |
|               | • Live MQTT   |           | Nurse  | | Clinical  |       |  |
|               +---------------+           | Agent  | | Assistant |       |  |
|                                           +----+---+ +-----+-----+       |  |
|                                                |           |             |  |
|                                           Patient Chat  Doctor Chat      |  |
|                                           + Education   + SOAP Notes     |  |
+=============================================================================+

⚡ Key Features

1. Dual-Window Signal Processing Pipeline

• 5-second window: SQI (template + SNR + motion), instantaneous HR, 4 HAR features
• 30-second window: RMSSD, SDNN, LF/HF (Lomb-Scargle), DWT morphology (QRS, QT/QTc, ST)
• Pan-Tompkins + Hamilton consensus QRS detection via NeuroKit2
• All metrics map directly to published cardiac rehab literature — fully interpretable

2. Two-Agent Clinical Intelligence

• Nurse Agent (Qwen3): Patient-facing wellness companion with Spanish-language support. Strict wellness-framing guardrails — never diagnoses, never recommends medication changes.
• Clinical Assistant (MedGemma-27B): Clinician-facing interactive Q&A and SOAP note generation. RAG-grounded in AHA/AACVPR guidelines with SQI-conditioned confidence reporting.

3. 4-Layer Deterministic Safety Architecture

• Layer 1: EnergySafeWindow — deterministic HR/SQI threshold alerts, pre-LLM
• Layer 2: Emergency keyword classifier — hardcoded response, no LLM
• Layer 3: Output validator — blocks diagnostic language in Nurse responses
• Layer 4: System prompt guardrails — role boundaries enforced via instruction

4. Live MQTT → ChromaDB → LLM Context Pipeline

• Polar H10 vitals published to MQTT topics per patient
• mqtt_subscriber.py upserts live telemetry into ChromaDB live_patients collection
• Agent orchestrator retrieves real-time context for every query

5. Reference Cohort Matching

• PhysioNet Wearable Exercise Frailty Dataset (47 post-cardiac-surgery patients)
• ECG-FM 768-dim embeddings + clinical metadata in ChromaDB
• Interpretable patient comparisons (surgery type, age, frailty, gait velocity)

6. Google Fit Longitudinal Baseline

• 7-day historical HR (15-min buckets), steps, sleep stages, body temperature
• Between-session context for the 4–5 days/week patients aren't in clinic

7. Activity Recognition (HAR Fusion Model)

• ResNet1D (PAMAP2, 88.9%) + HARNet10 SSL (PhysioNet, 73.3%)
• 1152-dim feature fusion, 8 unified activity classes
• Subject-wise train/test splits prevent data leakage

🌐 Live Demo

The proof of concept is deployed at pulse-forge-ai.vercel.app — demonstrating the full pipeline:

• PDF Upload → AHA/AACVPR guidelines chunked into ChromaDB RAG
• Role-Based Chat → toggle Clinician / Patient for different agent personas
• Live Telemetry → real-time Polar H10 HR, HRV, activity via MQTT WebSocket
• SOAP Notes → automated structured clinical documentation
• Deterministic Safety → EnergySafeWindow evaluates before every LLM call

The Vercel deployment uses Ollama tunneling for LLM inference. The production deployment targets NVIDIA DGX Spark for hardware-enforced HIPAA compliance — same codebase, only infrastructure changes.

🚀 Getting Started

Prerequisites

• Python 3.10+
• Polar H10 chest strap (or use --mock mode)
• Ollama with MedGemma-27B or compatible model

1. Signal Processing Application (PyQt5 Dashboard)

cd Application/Polar_Livestream-analysis-Python
pip install -r requirements.txt
python main.py          # Real Polar H10 via BLE
python main.py --mock   # Mock sensor for testing

2. Backend API Server

cd backend
pip install -r requirements.txt
python ingest_literature.py   # Ingest AHA/AACVPR PDFs into ChromaDB
python ingest_cohorts.py      # Ingest PhysioNet reference cohort
python mqtt_subscriber.py &   # Start MQTT → ChromaDB sync (background)
uvicorn main:app --host 0.0.0.0 --port 8000

3. MQTT Live Telemetry Bridge

# From the signal processing app, vitals publish to EMQX broker
# The mqtt_subscriber.py process syncs these into ChromaDB live_patients
# The web dashboard polls /api/live/metrics for real-time display

🔑 Key Files for Evaluator Validation

To verify the components mentioned in our documentation, please review the following key backend implementation files:

• backend/main.py — Contains the FastAPI endpoints, including the /api/session/{patient_id}/soap generation endpoint and the ChromaDB KNN logic (lines 135-155).
• backend/agent_orchestrator.py — Contains the PulseForgeOrchestrator role-based prompt assembly logic with emergency/safety bounds propagation (Layer 2 safety).
• backend/safety_engine.py — Contains EnergySafeWindow implementing deterministic threshold checks (Layer 1 safety) before any LLM invocations.
• backend/ingest_cohorts.py — Contains the script for ingesting physiological datasets into ChromaDB for cohort matching.

📁 Repository Structure

PulseForgeAI/
├── backend/                           ← FastAPI + Vercel web application
│   ├── main.py                        ← FastAPI server (271 lines)
│   ├── agent_orchestrator.py          ← Role-based prompt assembly
│   ├── safety_engine.py               ← EnergySafeWindow deterministic safety
│   ├── mqtt_pipeline.py               ← MQTT pub/sub with QoS differentiation
│   ├── mqtt_subscriber.py             ← EMQX → ChromaDB live state sync
│   ├── ingest_literature.py           ← PubMedBERT RAG ingestion pipeline
│   ├── ingest_cohorts.py              ← PhysioNet cohort → ChromaDB
│   ├── utils.py                       ← Ollama inference with tunnel bypass
│   ├── vercel.json                    ← Vercel serverless deployment
│   └── static/                        ← Web dashboard (HTML/CSS/JS)
│       ├── index.html                 ← Clinical reasoning dashboard
│       ├── app.js                     ← Chat, upload, MQTT WebSocket, Chart.js
│       └── style.css                  ← Glassmorphism design system
│
├── Application/
│   └── Polar_Livestream-analysis-Python/  ← PyQt5 BLE signal processing
│       ├── main.py                    ← Entry point (real or --mock)
│       ├── intake_state.json          ← Clinical + Google Fit schema
│       └── polar_ecg/
│           ├── workers/
│           │   ├── processing_worker.py  ← Dual-window (5s/30s) pipeline
│           │   ├── ble_worker.py      ← Polar H10 BLE connection
│           │   └── mqtt_worker.py     ← QThread MQTT publisher
│           ├── ui/
│           │   ├── dashboard.py       ← Real-time ECG/HRV visualization
│           │   └── intake_form.py     ← Clinical intake form
│           └── utils/
│               ├── google_fit_fetcher.py  ← Google Fit 7-day baseline
│               ├── har_inference.py   ← HAR fusion model inference
│               ├── mock_sensor.py     ← Synthetic ECG/ACC generator
│               └── ring_buffer.py     ← Efficient circular buffer
│
├── Act_Recoginition/                  ← HAR dual-model training pipeline
│   ├── Data_Preparation/              ← PAMAP2 + PhysioNet preprocessing
│   ├── Train_Model/                   ← ResNet1D + HARNet10 SSL fine-tuning
│   ├── Fusion_Model/                  ← 1152-dim feature fusion classifier
│   └── *.pth                          ← Trained model weights
│
├── ECG_Embedding/
│   └── ECG_frailty-db_feature_lookup.py  ← ECG-FM 768-dim + signal metrics
│
├── system_prompts.md                  ← Production agent system prompts (505 lines)
├── master-plan.md                     ← Master plan + execution strategy
├── master-plan_fin.md                 ← Extended master plan with technical detail
├── tests/                             ← Safety engine unit tests
└── archive/                           ← Previous plan versions

🛡️ Safety & Compliance

• HIPAA by architecture — DGX Spark production deployment: zero PHI egress, no BAAs
• Deterministic safety — EnergySafeWindow fires before LLM, catches HR threshold violations in <1s
• 4-layer guardrails — Critical alerts never depend on LLM availability
• General Wellness framing — FDA January 2026 guidance compliant; SOAP notes as administrative documentation
• Emergency keyword bypass — "chest pain", "can't breathe" trigger hardcoded response without LLM

---

Talk to Your Heart — On-Campus Cardiac Rehabilitation Intelligence PoC: pulse-forge-ai.vercel.app | Production: NVIDIA DGX Spark Disclaimer: PulseForgeAI is a research prototype. Not FDA-approved as SaMD.