HepSense (DeepHepaStrat)

An AI-driven Risk Stratification System for Early Liver Failure Prediction

Shifting the clinical paradigm from static "disease staging" to dynamic, longitudinal future prediction.

---

📖 Table of Contents

• Project Overview
• The Clinical Problem
• Methodology
  • 1. Vision Pipeline (HepSense_Vision.py)
  • 2. Clinical Temporal Pipeline (MIMIC-IV)
• Getting Started
  • Prerequisites
  • Installation
• Usage
• Literature & Background
• Development Timeline
• Expected Outcomes

---

🧬 Project Overview

Liver cirrhosis is the 11th leading cause of death globally (over 1.3 million deaths annually). HepSense (DeepHepaStrat) is designed as a multi-modal framework that integrates both computer vision (for imaging analysis) and time-series/NLP data (for clinical health records) to provide a holistic, dynamic risk assessment of liver cirrhosis. It acts as a temporal AI "early warning radar" to reduce preventable ICU mortality by predicting acute decompensation events before they occur.

⚠️ The Clinical Problem

Current clinicians rely on outdated, static scoring equations (e.g., MELD-Na, FIB-4).

• Snapshot Bias: Static scores analyze a single day's blood test, creating a dangerous bias that misses declining trajectories.
• Blind Spots: Doctors are often blind to impending, fatal crises like variceal hemorrhage or encephalopathy.
• Reactive vs Proactive: Most existing models predict mortality after a patient enters critical condition in the ICU. HepSense bridges this gap by forecasting the actual future onset of an emergency 7 to 14 days before it happens, allowing for preventative intervention.

---

🔬 Methodology

1. Vision Pipeline (HepSense_Vision.py)

A deep learning computer vision model built to classify the progression of liver cirrhosis stages (F0 to F4) from medical imaging.

• Architecture: Utilizes a pre-trained MobileNetV2 as the core feature extractor with custom dense classification layers.
• Techniques: Implements data augmentation to handle class imbalances, dropout layers to prevent overfitting, and sparse categorical crossentropy for robust multiclass classification.

2. Clinical Temporal Pipeline (MIMIC-IV)

A quantitative, retrospective data-mining structure utilizing multi-modal clinical histories to capture the velocity of a patient's decline over a 14-day window.

• Data Collection: Executes authorized, secure extraction of de-identified critical care records directly from MIT's PhysioNet repository (mimic-hosp, mimic-icu, mimic-ed, mimic-note).
• Data Analysis & NLP:
  • Missing-value imputation via "Last Observation Carried Forward" (LOCF).
  • NLP (HuggingFace transformers & spaCy) to extract unstructured symptom features (e.g., "vomiting blood") from clinical admission notes.
• Predictive Modeling: Time-Series Gradient Boosted Trees (XGBoost) / LSTMs for forecasting, cross-validated against held-out testing cohorts.
• Explainability (XAI): Employs SHAP (SHapley Additive exPlanations) values to visualize feature importance, ensuring absolute medical transparency and physician trust.

---

🚀 Getting Started

Prerequisites

This project uses modern Python tools and depends heavily on deep learning libraries. It is managed using uv for lightning-fast dependency resolution.

• Python >= 3.12
• uv package manager

Installation

1. Clone the repository:

   git clone <repository_url>
   cd EarlyLiverFailurePrediction

2. Sync the environment: Use uv to install dependencies as defined in the pyproject.toml / uv.lock.

   uv sync

3. Dependencies: The project requires the following core libraries (see requirements.txt):

   • tensorflow (>= 2.13.0)
   • numpy, pandas, scikit-learn, scipy
   • xgboost, catboost, lightgbm
   • matplotlib, seaborn
   • flask

---

💻 Usage

Running the Vision Model Training

The current working implementation includes the Vision Pipeline. To train the MobileNetV2-based model on your image dataset:

1. Ensure your dataset is structured correctly in the Dataset/ directory, with subfolders representing the cirrhosis stages (e.g., F0, F1, F2, F3, F4).
2. Run the vision script:

   python HepSense_Vision.py

3. The script will output the training and validation progress and save the finalized model locally as hepsense_vision_v1.keras.

---

📚 Literature & Background

Our theoretical framework, Explainable Temporal Machine Learning Theory, posits that physiological collapse in chronic end-stage liver disease is not sudden; it is preceded by subtle, long-term degradation across multiple routine biomarkers.

Author	Title & Inference	Research Gap Addressed
Kumar, A., et al.	A Deep Learning Framework for Forecasting Hepatic Decompensation... LSTM on 14-day trailing EHR identified bleeding early vs single-day MELD score.	We use an accessible dataset (MIMIC-IV) for broader replication rather than a private hospital database.
Zhang, L., et al.	Predicting Mortality in Patients with Liver Cirrhosis Using Machine Learning... XGBoost on MIMIC-IV predicted mortality with AUC 0.86 using bedside labs.	We focus on dynamic time-of-onset prediction rather than 24-hour snapshots.
Martinez, R., et al.	Explainable Artificial Intelligence for Clinical Risk Stratification... Black-box ML models are rejected unless SHAP/LIME explains features to doctors.	We integrate dynamic time-series logic seamlessly with XAI.
Chen, S., et al.	Multimodal Fusion of Clinical and Imaging Data for Sarcopenia Detection... Automating SMI calculations from CT scans dramatically improves mortality forecasts.	We fuse imaging capabilities with sequential blood labs and NLP symptom tracking.

Trends in Literature

• Static to Dynamic: Rapid movement away from fixed clinical equations (MELD) toward algorithms that process sequential, time-series data.
• Interpretable AI (XAI): Strict requirement for local explanations (like SHAP) in medical journals to increase physician adoption.
• NLP Integration: Critical predictive signals exist in the unstructured text of doctors' notes, prompting the fusion of NLP with standard numeric lab data.

---

📅 Development Timeline

1. Initiation (Weeks 1-2): Define objectives, secure CITI training, and gain MIMIC-IV access. Configure virtual environments.
2. Planning (Weeks 3-4): Map SQL extraction for MIMIC-IV tables and finalize ML architecture (XGBoost vs. LSTM algorithms).
3. Execution (Weeks 5-10): Execute codebase for missing-value imputation, program NLP modules for unstructured notes, and train baseline models. Integrate SHAP into the Flask web dashboard.
4. Closure (Weeks 11-13): Benchmark against MELD score, finalize the frontend dashboard, and draft the final research paper for presentation.

---

🎯 Expected Outcomes

• Target Performance: Expected AUROC of >0.85, establishing proven, quantitative superiority over the static MELD-Na clinical score.
• Publications: Targeting premier bioinformatics conferences (e.g., IEEE EMBS or BHI) and high-impact hepatology journals (e.g., IEEE JBHI).
• Patents: Formal patent filing for the explicit architectural fusion of the SHAP dashboard and LSTM temporal forecasting.
• Clinical Impact: Empowers clinicians to schedule preemptive, prophylactic interventions safely (e.g., executing an endoscopy before a fatal hemorrhage occurs), optimizing hospital triage resources.