Real-Time ICU Intelligence Ecosystem & Decision-Support Prototype
CANONICAL DISCLAIMER
SepsisGuard v3.0 is a technical demonstration and research prototype. This system has not been clinically validated for real-patient decision making.
- Intended Use: Technical demonstration, machine learning research prototyping, educational evaluation of medical ML pipelines.
- Out-of-Scope Use: Clinical diagnosis of sepsis, issuing medication/treatment orders, autonomous patient triage, or deployment on real ICU patients before prospective clinical trials and regulatory clearance.
- Clinical Judgment Required: All outputs represent statistical model probabilities requiring independent clinical verification by qualified medical professionals.
SepsisGuard AI is an end-to-end machine learning decision-support architecture for early sepsis detection in intensive care unit (ICU) environments:
-
ML Inference Engine: Versioned XGBoost risk classifier (
v2_2026-08-12, operating threshold$0.27$ ) predicting sepsis probability from 20 clinical features. -
PhysioNet 2019 Dataset: Trained and evaluated on leakage-free patient splits from the PhysioNet 2019 Challenge dataset (
$40,336$ patients,$1.4\text{M}$ clinical hours). -
Model-Faithful SHAP Explainability: Genuine
shap.TreeExplainerfeature attributions identifying top risk drivers per prediction (e.g. Temperature, Respiratory Rate, 6h vital sign trends). - Clinical Rule Baselines: Real-time SIRS and partial qSOFA score evaluation displayed alongside ML risk predictions.
- AI Copilot & Narrative Synthesis: Gemini LLM clinical observation synthesis with strict non-clinical safety boundaries (no treatment directives or dosing orders).
-
Simulated ICU Telemetry Stream: Real-time WebSocket streaming (
icu_unit_a) delivering 6 simulated patient beds to interactive ICU dashboard cards. (Note: Telemetry streams represent simulated demo data).
Detailed evaluation reports are available in TEST_REPORT.md, clinical_baseline_comparison.md, and validation/GENERALIZATION_REPORT.md.
Evaluated on untouched held-out test set (data/processed/test.csv, 9,962 longitudinal rows across 244 patients):
| Evaluation Metric | Measured Result | Clinical Significance |
|---|---|---|
| Sensitivity / Recall | 96.93% | Detects |
| Specificity | 32.31% | Identifies non-sepsis ICU hours |
| Positive Predictive Value (PPV) | 2.33% | Low prevalence ( |
| Negative Predictive Value (NPV) | 99.84% |
|
| F1 Score | 0.0455 | Harmonic mean of PPV and Recall |
| ROC-AUC | 0.8250 | Discrimination across risk spectrum |
| PR-AUC | 0.1171 | Precision-Recall curve area |
| Brier Score | 0.1524 | Probabilistic calibration error |
-
SIRS Baseline: ROC-AUC
0.5821 -
SepsisGuard XGBoost: ROC-AUC
0.8250($+0.2429$ improvement over SIRS, DeLong test$p < 0.001$ )
- Status:
BLOCKED — independent dataset unavailable - Audit Result: PhysioNet Set A and Set B were merged during Phase 2 development splits; no external hospital cohort was available in the repository. Detailed in validation/EXTERNAL_DATASET_CARD.md.
# Clone repository
git clone https://github.com/SepsisGuard/SepsisGuard-AI.git
cd SepsisGuard-AI
# Create virtual environment & install dependencies
python -m venv venv
venv\Scripts\activate # On Windows
pip install -r requirements.txt
# Configure environment variables
cp .env.example .env
# Run SepsisGuard Telemetry Server
python backend/app.py# Build & launch container
docker-compose up -d --build
# Verify health status
curl http://localhost:5000/health- ARCHITECTURE.md — Modular architecture, dependency flow, and Socket.IO room scoping.
- DEPLOYMENT.md — Deployment, WSGI configuration, Docker, and health monitoring.
- model/MODEL_CARD.md — Research-grade model card for model
v2_2026-08-12. - data/DATASET_CARD.md — PhysioNet 2019 dataset card & preprocessing specification.
- openapi.yaml — OpenAPI 3.0 specification for REST API endpoints.
Distributed under the MIT License. PhysioNet 2019 data used under CC-BY 4.0.