SAFE-Gate is a safety-critical clinical decision support framework for emergency department triage of patients presenting with dizziness and vertigo. The system routes each patient case through six parallel expert gates -- each implementing a distinct clinical or analytical perspective -- and fuses their outputs through an Adaptive Confidence-Weighted Conservative Merging (ACWCM) mechanism that provably preserves patient safety.
Unlike conventional ensemble methods that optimize for accuracy, SAFE-Gate is designed around a conservative-by-construction principle: whenever expert gates disagree, the system defaults to the most cautious risk tier. This eliminates dangerous under-triage (false negatives) at the cost of controlled over-triage, a clinically acceptable trade-off in emergency medicine.
The architecture is grounded in a formal risk lattice
R* ≤ R1 ≤ R2 ≤ R3 ≤ R4 ≤ R5, where R* represents abstention (defer to
human clinician). Three safety properties -- Conservative Preservation, Abstention
Correctness, and Critical Non-Dilution -- are verified at every inference step,
providing mathematical guarantees that no critical patient is misclassified to a
lower-acuity tier.
SAFE-Gate Architecture
========================================================================
Stage 1 Stage 2 Stage 3 Stage 4
INPUT PARALLEL GATES ACWCM FUSION OUTPUT
---------------- ---------------- ------------ ----------
Patient Data -----> [G1] Critical -->|
(52 features) Red Flags |
|
[G2] Cardiovasc. -->| +----------+
Risk | Risk Lattice | | Risk Tier
|--> | ACWCM |-> + Safety
[G3] Data -->| R* ≤ R1 ≤ | Fusion | Certificate
Quality | R2 ≤ R3 ≤ +----------+ + Audit
| R4 ≤ R5 ^ Trail
[G4] TiTrATE -->| |
Patterns | Confidence
| Weights
[G5] Bayesian -->|
Uncertainty |
|
[G6] Temporal -->|
Risk
========================================================================
Safety Properties: Conservative Preservation | Abstention Correctness
Critical Non-Dilution | Zero Safety Violations
| Metric | Value |
|---|---|
| Critical sensitivity (R1--R2) | 100.0% |
| Discharge specificity (R5) | 66.7% |
| False negative rate | 0.0% |
| Overall accuracy | 65.2% |
| Macro F1-score | 68.9% |
| Over-triage rate (ACWCM) | 16.4% |
| Safety violations | 0 / 6,398 |
| Inference latency | 1.70 ms/case |
git clone https://github.com/ChatchaiTritham/SAFE-Gate.git
cd SAFE-Gate
pip install -r requirements.txt
# Optional: install in development mode
pip install -e .from src.safegate import SAFEGate
# Initialize (6 parallel gates with ACWCM fusion)
safegate = SAFEGate(mode="acwcm")
# Define patient presentation
patient = {
'age': 72, 'gender': 'male',
'systolic_bp': 85, 'heart_rate': 125,
'spo2': 88, 'gcs': 13,
'symptom_onset_hours': 1.5,
'vertigo_severity': 'severe',
'dysarthria': True,
'atrial_fibrillation': True,
'hints_head_impulse': 'abnormal',
'hints_nystagmus': 'central',
'hints_test_of_skew': 'positive',
}
result = safegate.classify(patient)
print(f"Risk Tier: {result['final_tier']}")
print(f"Enforcing Gate: {result['enforcing_gate']}")
print(f"Confidence: {result['confidence']:.2f}")
print(f"Latency: {result['latency_ms']:.2f} ms")python tests/test_full_system.pyimport json
with open('data/synthetic/test/synthetic_test_804.json', 'r') as f:
test_data = json.load(f)
results = safegate.batch_classify(test_data)SAFE-Gate/
├── src/
│ ├── safegate.py # Main SAFEGate orchestrator
│ ├── gates/
│ │ ├── gate1_critical_flags.py # G1: Rule-based red flag detection (18 rules)
│ │ ├── gate2_moderate_risk.py # G2: Cardiovascular risk scoring (Eq. 3)
│ │ ├── gate3_data_quality.py # G3: Data completeness (22 fields, Eq. 4)
│ │ ├── gate4_titrate_logic.py # G4: TiTrATE syndrome matching (Hamming)
│ │ ├── gate5_uncertainty.py # G5: BNN MC dropout (52→128→64→5)
│ │ └── gate6_temporal_risk.py # G6: Temporal FSM (5 states)
│ ├── merging/
│ │ ├── conservative_merging.py # ACWCM fusion + conflict resolution
│ │ ├── risk_lattice.py # Risk lattice (R*, R1--R5)
│ │ └── safety_certificate.py # Safety certificate generation
│ ├── baselines/
│ │ ├── esi_guidelines.py # ESI rule-based baseline
│ │ ├── single_xgboost.py # XGBoost baseline
│ │ ├── ensemble_average.py # Ensemble averaging baseline
│ │ ├── confidence_threshold.py # Confidence thresholding baseline
│ │ ├── dempster_shafer.py # Dempster-Shafer combination
│ │ └── bayesian_model_avg.py # Bayesian Model Averaging
│ ├── theorems/
│ │ └── theorem_verification.py # Runtime safety property checking
│ └── utils/
│ ├── audit_trail.py # Clinical audit trail generator
│ └── visualization.py # Plotting utilities
├── data/
│ ├── generation/ # Synthetic data generators
│ └── synthetic/ # Train (4,796) / Val (798) / Test (804)
├── evaluation/ # Evaluation pipeline and figures
├── experiments/ # XAI: SHAP, counterfactual, NMF
├── tests/ # Test suite
├── docs/ # Additional documentation
├── notebooks/ # Jupyter quickstart
├── requirements.txt
├── setup.py
├── CITATION.cff
└── LICENSE
Deterministic screening via 18 atomic Boolean rules across 5 clinical categories (hemodynamic instability, altered mental status, acute focal deficits, severe headache, respiratory compromise). Any single triggered rule immediately escalates to R1 at maximal confidence.
Weighted accumulation model combining demographic, symptom, and clinical history risk factors with XGBoost consistency validation. Captures elevated stroke probability from features that individually fall below critical thresholds but collectively signal cardiovascular concern.
Evaluates completeness ratio across 22 essential clinical fields. When data completeness falls below the safety threshold (ρ < 0.70), the gate outputs R* (abstention), forcing the system to defer to a human clinician rather than risk an unreliable classification.
Implements weighted Hamming distance matching against three characterised benign vestibular syndromes (BPPV, vestibular neuritis, Meniere disease). High similarity to a known benign profile supports safe discharge; low similarity triggers escalation.
Bayesian neural network (52→128→64→5) with Monte Carlo dropout (T=20 forward passes). Computes a composite uncertainty index from predictive entropy and prediction variance. Triggers abstention or escalation when model uncertainty exceeds calibrated thresholds.
Finite-state machine modelling symptom evolution trajectories. Five temporal states (hyperacute, acute stable, acute improving, subacute, chronic) with progression-modified transitions capture dynamic risk that point-in-time assessments miss.
SAFE-Gate enforces three formally verified safety properties at every inference:
Conservative Preservation (CP).
The final merged tier is never more than one tier less conservative than the most
cautious gate output: rank(T_final) ≤ rank(min(T_i)) + 1. Under basic minimum
selection T_final = min(T_i); ACWCM permits bounded one-tier relaxation only
when high-confidence gate consensus supports it.
Abstention Correctness (AC). If any gate outputs R* (abstention), the final system output is R*. The system never overrides a gate's decision to defer to human judgment.
Critical Non-Dilution (CND). A critical assessment (R1 or R2) from any gate cannot be diluted by non-critical assessments from other gates. Critical signals propagate to the final output.
These properties are verified at runtime and logged in the audit trail for every patient classification.
# Run the comprehensive test suite
python tests/test_full_system.py
# Generate evaluation metrics
python evaluation/generate_performance_metrics.py
# Generate manuscript figures
python evaluation/create_manuscript_figures.pyThe evaluation suite includes comparisons against six baseline methods:
| Method | Type |
|---|---|
| ESI Guidelines | Rule-based |
| Single XGBoost | Gradient boosting |
| Ensemble Average | Unweighted fusion |
| Confidence Threshold | Threshold-based |
| Dempster-Shafer | Evidence theory |
| Bayesian Model Averaging | Probabilistic |
python experiments/shap_explainability.py
python experiments/nmf_interpretability.py
python experiments/counterfactual_explanations.pyContributions are welcome. Please follow these guidelines:
-
Fork the repository and create a feature branch from
main. -
Write tests for new functionality in
tests/. -
Follow existing code style (PEP 8, type hints, docstrings).
-
Run the test suite before submitting:
python tests/test_full_system.py -
Submit a pull request with a clear description of changes.
- All gate implementations must include
evaluate(),get_name(), andget_description()methods. - Safety properties (CP, AC, CND) must not be violated by any code change.
- New baselines should follow the interface pattern in
src/baselines/. - Clinical thresholds must be referenced to published guidelines.
Please use GitHub Issues for bug reports and feature requests. Include:
- Python version and OS
- Minimal reproducible example
- Expected vs. actual behaviour
If you use SAFE-Gate in your research, please cite:
@article{tritham2026safegate,
title = {{SAFE-Gate}: An Adaptive Knowledge-Based Expert System for
Emergency Triage Safety with Confidence-Weighted Conservative
Merging and Formal Safety Guarantees},
author = {Tritham, Chatchai and Snae Namahoot, Chakkrit},
journal = {Soft Computing},
publisher = {Springer},
year = {2026},
url = {https://github.com/ChatchaiTritham/SAFE-Gate},
license = {MIT}
}This project is licensed under the MIT License. See LICENSE for details.
- Chatchai Tritham -- Naresuan University, Thailand
- Chakkrit Snae Namahoot -- Naresuan University, Thailand (chakkrits@nu.ac.th)