CARE (Clinical Assessment & Risk Engine)

A local-first, advice-only system for real-time ED workflow optimization using Gate Engine, CatBoost, and SPC.

🏆 Winner of the Hackathon of Emergency Medicine at ICEM 2026!

A big thank you to the Hackathon team for Presentation Design & Clinical Framing – we had a lot of fun and not much sleep 🚀

Call for Collaborators: I am looking for informaticians and EM physicians to take over maintenance, EHR integration, and further development. If you are interested in stewarding this tool, please contact me via GitHub Issues or my clinic homepage.

---

🔧 System Overview

CARE is designed to support—not replace—clinical judgment. It transforms fragmented emergency department data into unified, actionable insights while maintaining strict human oversight.

Core Components

1. Gate Engine

   • Function: Rule-based system for capacity management, handoff protocols, and equipment tracking.
   • Features: Configurable thresholds (e.g., "page if ICU >90% full") and pilot-stage ICU negotiation rules.
   • Philosophy: Logic is separated from configuration; hospitals can adapt rules to local SOPs without code changes.

2. CatBoost Models

   • Predictions:
     • Length of Stay >6h
     • ICU transfer (3h)
     • Intubation risk (2h)
     • Sepsis bundle activation (2h)
   • Safety: Trained on MIMIC-IV with cross-validation; outputs include SHAP explanations so clinicians understand why a risk was flagged.

3. SPC Monitoring

   • Function: Shewhart/EWMA control charts for admissions, respiratory complaints, and lab turnaround times.
   • Alerting: Triggers when metrics breach statistical control limits, providing an early warning system for outbreaks or resource strain.

4. Local LLM Integration

   • BioBERT: For structured note extraction and semantic analysis.
   • Meditron-7B: For generating natural language explanations (deployed locally via llama.cpp to ensure data sovereignty).

---

🛡️ Clinical Safety & Ethics

"The right information, at the right time, to the right people."

This system operates under a strict Advice-Only Protocol. All outputs require explicit clinician approval before any action is taken. No automated orders are generated.

• Read the Full Clinical Safety Manifesto

Key Safety Principles

• Sustained Ambiguity: The system is designed to hesitate and present uncertainty rather than force a rushed decision.
• Explainability: Every prediction includes SHAP values detailing the contributing factors (e.g., "Flagged due to rising lactate + low SpO₂").
• Fail-Safes:
  • False Positives: Mitigated by requiring CatBoost confidence >0.9 for critical alerts.
  • False Negatives: Fallback to SPC-only alerts if model confidence drops.
  • Data Drift: Automatic weekly retraining triggers and SPC updates.

---

📦 Data Pipeline

Feature Engineering

def create_features(df):
    df["admissions_rolling_6h"] = df["admissions"].rolling(6).mean()
    df["respiratory_pct_change_24h"] = df["respiratory_complaints"].pct_change(24)
    df["cold_snap"] = (df["temp"] < 5).astype(int)
    df["school_holiday"] = df["date"].isin(school_holidays)
    return df

Labeling

df["outbreak_flag"] = 0
df.loc[(df["admissions"] > df["admissions_ucl"]) & (df["vomiting_pct"] > 0.15), "outbreak_flag"] = 1
df.loc[df["date"].between("2023-01-03", "2023-01-05"), "outbreak_type"] = "norovirus"

---

🤖 Model Training

CatBoost

from catboost import CatBoostClassifier

model = CatBoostClassifier(
    depth=6,
    learning_rate=0.05,
    iterations=500,
    l2_leaf_reg=3,
    cat_features=["triage_acuity", "chief_complaint", "allergies"]
)
model.fit(X_train, y_train["outbreak_type"])

SPC

from pyculiarity import detect_cusum
cusum_results = detect_cusum(df["respiratory_complaints"], threshold=1.0, drift=0.1)
df["cusum_alert"] = cusum_results["alert"]

---

📈 Evaluation

Metric	Target
Precision	> 0.9
Recall	> 0.8
F1 Score	> 0.85
Time-to-Detection	< 6 hours

from sklearn.metrics import classification_report
print(classification_report(y_true=df["outbreak_type"], y_pred=model.predict(X_test)))

---

🚀 Deployment

FastAPI Endpoint

from fastapi import FastAPI
app = FastAPI()

@app.post("/predict")
def predict(data: dict):
    features = preprocess(data)
    spc_alert = check_spc(features)
    catboost_probs = model.predict_proba(features)
    return {
        "spc_alert": spc_alert,
        "catboost_probs": catboost_probs.tolist(),
        "recommended_actions": get_actions(spc_alert, catboost_probs)
    }

Logging

import logging
logging.basicConfig(filename="outbreak_logs.csv", level=logging.INFO)

---

💡 Explanations & Transparency

SPC Alerts Example

"Admissions breached UCL (80/hr vs. baseline 60/hr)."

SHAP Visualization

Every high-risk prediction generates a plot showing feature contribution:

import shap
explainer = shap.TreeExplainer(model)
shap.plots.waterfall(explainer.shap_values(real_time_data)[0])

---

⚠️ Failure Modes

Failure	Mitigation
False Positives	Require CatBoost confidence > 0.9
False Negatives	Fallback to SPC-only alerts
Data Drift	Weekly retraining + SPC updates

---

🔄 Continuous Improvement Loop

- Log: Record all predictions and subsequent clinician decisions.
- Review: Monthly review of false positives/negatives with clinical leads.
- Retrain: Weekly retraining cycles to account for seasonal shifts and data drift.
- Test: A/B test protocol changes before full deployment.

📋 Roadmap

- Feature engineering pipeline (Complete)
- CatBoost baseline (Complete)
- FastAPI + Logging (Complete)
- Clinician feedback loop implementation
- EHR integration adapters
- Mobile dashboard for triage nurses

Next Steps for Adopters

- Prototype: Run against historical de-identified data.
- Shadow Deploy: Run alongside manual surveillance without acting on alerts.
- Iterate: Adjust thresholds based on local workflow and feedback.

⚠️ Important Notices

Advice-Only Disclaimer This is an advice-only system. All outputs require explicit clinician approval. The software does not make autonomous medical decisions. AI Training Opt-Out. This repository is explicitly excluded from all AI training datasets under GitHub’s DMCA policy and ethical guidelines. Violations of this opt-out will be reported.

License

Apache License 2.0 - Clinical Safety Manifesto

Free to use, modify, and distribute for non-commercial and commercial purposes, provided attribution is given.

Author: Dr. Katharina Jacoby Research Focus: AI Safety, Systems Engineering, Computational Topology Contact: k.jacoby at posteo.de

---