🏥 Context-Aware ICU Alerting System

A Disease-Conditioned Early Warning Framework using MIMIC-IV

---

📌 Overview

Traditional ICU monitoring systems rely on fixed physiological thresholds (e.g., HR > 100 bpm → alert).
However, these thresholds ignore clinical context such as:

• Underlying diagnosis
• Admission type
• Treatment interventions
• Laboratory abnormalities

This project builds a context-aware ICU alerting system that dynamically adapts alert thresholds based on patient-specific clinical context.

The system is developed using the MIMIC-IV (v3.1) critical care dataset.

---

🎯 Motivation

A heart rate of 110 bpm may be:

• Expected in sepsis
• Dangerous in post-operative neurosurgery
• Acceptable in chronic atrial fibrillation

Generic alerting leads to:

• 🚨 Alarm fatigue
• ❌ High false positive rates
• ⚠ Missed deterioration

This project aims to design a disease-conditioned alert framework that improves clinical relevance and personalization.

---

🗂 Dataset

Data source: MIMIC-IV v3.1 (PhysioNet)

ICU Module

• chartevents → Time-series vital signs
• icustays → ICU stay linkage
• inputevents → Treatment administration

Hospital Module

• admissions → Admission context & outcomes
• diagnoses_icd → Diagnosis codes
• d_icd_diagnoses → ICD code descriptions
• labevents → Laboratory results

---

🏗 System Architecture

flowchart TD
  A["Vital Signs: chartevents"] --> B["ICU Stay Linking: icustays"]
  B --> C["Admission Context: admissions"]
  C --> D["Diagnosis Mapping: diagnoses_icd"]
  D --> E["Lab Context: labevents"]
  E --> F["Treatment Context: inputevents"]
  F --> G["Context-Aware Alert Engine"]

Loading

---

🧠 Methodology

1️⃣ Data Processing

• Extract ICU time-series vitals
• Align vitals with ICU stay windows
• Map ICD codes to disease categories
• Aggregate labs into temporal bins
• Encode treatment exposures

2️⃣ Context Feature Engineering

Each time window includes:

• Multivariate vital signs (HR, MAP, RR, SpO₂, Temp)
• Disease category indicator
• Lab abnormality flags
• Treatment flags (vasopressors, sedation, fluids)

3️⃣ Alert Strategies

🔹 Rule-Based Baseline

Disease-conditioned thresholds:

If Sepsis:
  Alert if HR > 120 and MAP < 65

If COPD:
  Alert if SpO₂ < 88

🔹 Statistical Modeling

Compute disease-specific baseline distributions:

$z = (x - \mu_{disease}) / \sigma_{disease}$

Alerts trigger when deviation exceeds learned thresholds.

🔹 Deep Learning Extension (Planned)

• LSTM / Transformer architecture
• Input: Multivariate time-series + diagnosis embedding
• Output: Deterioration risk in next 6 hours

---

🚀 Goals

• Reduce alarm fatigue
• Personalize ICU monitoring
• Improve early detection of deterioration
• Create a reproducible research framework

---

📊 Potential Applications

• Personalized Early Warning Scores
• Contextual anomaly detection
• ICU deterioration prediction
• Clinical decision support research

---

📁 Project Structure

condition-aware_risk_CDSS/
├── data_processing/
│   └── README.md
├── feature_engineering/
│   └── README.md
├── modeling/
│   └── README.md
├── evaluation/
│   └── README.md
├── deployment/
│   └── README.md
├── notebooks/
│   └── README.md
└── README.md

Quick Links

• data_processing/
• feature_engineering/
• modeling/
• evaluation/
• deployment/
• notebooks/
• README.md

---

🔒 Data Access

Access to MIMIC-IV requires:

• PhysioNet credentialing
• CITI training certification
• Data use agreement

No raw patient data is stored in this repository.

---

👤 Author

Samuel Adeniji

---

📜 License

This project is for research and educational purposes only.