LW-LSTM Based Hepatitis Prediction Prognosis using Lightweight Long Short-Term Memory Networks

1. Introduction

      Hepatitis is a serious liver disease that can lead to severe health complications, including liver failure and death if not diagnosed and treated early. With the increasing availability of healthcare data, machine learning and deep learning techniques have become essential tools for disease prediction and prognosis.

         Traditional clinical diagnosis relies heavily on laboratory tests and expert interpretation, which may be time-consuming and prone to human error. Therefore, intelligent automated systems are required to assist medical professionals in making faster and more accurate decisions.

        In recent years, deep learning models—especially Long Short-Term Memory (LSTM) networks—have demonstrated strong capabilities in handling sequential and complex data patterns. However, conventional LSTM models can be computationally expensive.

         To address this, this work proposes a Lightweight Long Short-Term Memory (LW-LSTM) model that efficiently predicts hepatitis patient outcomes (Live or Die) while maintaining high accuracy and reduced computational complexity.

2. Problem Statement

Hepatitis prediction poses several challenges:

1.  Medical datasets often contain missing and noisy values
2.  Data imbalance between survival and death cases
3.  Presence of both categorical and numerical features
4.  Need for efficient and lightweight models
5.  Accurate prognosis is critical for patient survival

This work aims to develop a lightweight deep learning model that can handle these challenges effectively.

3. System Overview

The proposed LW-LSTM system consists of the following stages:

1. 1. Dataset loading
2. 2. Data analysis and visualization
3. 3. Data balancing (oversampling)
4. 4. Data preprocessing
5. 5. Feature scaling
6. 6. Dataset splitting
7. 7. Model training (LW-LSTM)
8. 8. Model evaluation
9. 9. Prediction

4. Input and Output Specifications

Input

 Patient clinical attributes:

1. o Age, Sex
2. o Symptoms (fatigue, malaise, anorexia, etc.)
3. o Medical indicators (bilirubin, albumin, SGOT, etc.)

Output

 Binary classification:

1. o 0 → LIVE
2. o 1 → DIE

 Prediction probability score

5. Methodology

Step 1: Dataset Loading

The dataset is obtained using KaggleHub:

1.  Dataset: Hepatitis dataset
2.  Contains patient records with clinical features and survival status

Step 2: Data Analysis

Exploratory Data Analysis (EDA) is performed using:

1.  Bar charts (class distribution)
2.  Pie charts (percentage distribution)

This helps identify class imbalance and feature characteristics.

Step 3: Data Balancing

To handle imbalance:

1.  Oversampling is applied using resampling technique
2.  Each class is balanced to equal size

This ensures unbiased learning.

Step 4: Data Preprocessing

Data preprocessing includes:

 Encoding categorical variables:

        o Sex → male (1), female (0)

 Converting boolean features to binary (1/0)

 Handling missing values:

1. o Numerical → median
2. o Categorical → mode

This step ensures clean and consistent input data.

Step 5: Feature Scaling

1.  StandardScaler is applied
2.  Ensures all features are on the same scale

Step 6: Dataset Splitting

1.  Training set: 80%
2.  Testing set: 20%
3.  Stratified sampling used for balance

Step 7: LW-LSTM Model Training

Architecture Components

1.  LSTM Layer (32 units) Captures relationships in feature patterns
2.  Dropout Layer (0.2) Prevents overfitting
3.  Dense Layer (16 neurons, ReLU) Learns complex feature interactions
4.  Output Layer (Sigmoid) Produces binary classification

Training Configuration

1.  Loss Function: Binary Crossentropy
2.  Optimizer: Adam
3.  Metrics: Accuracy
4.  Epochs: 10
5.  Batch Size: 32

Step 8: Model Evaluation

The model is evaluated using:

1.  Accuracy
2.  Precision
3.  Recall
4.  F1 Score
5.  Specificity
6.  Error Rate
7.  ROC-AUC Score
8.  Confusion Matrix

These metrics ensure robust performance evaluation.

Step 9: Prediction

Prediction workflow:

1. 1. Load trained model and scaler
2. 2. Input new patient data
3. 3. Apply preprocessing
4. 4. Scale features
5. 5. Reshape for LSTM
6. 6. Generate prediction

Output:

1.  Probability score
2.  Final class:

                   o LIVE

                   o DIE

6. Expected Outcomes

The LW-LSTM model is expected to:

1.  Achieve high prediction accuracy
2.  Handle imbalanced datasets effectively
3.  Reduce false predictions
4.  Provide fast and lightweight computation
5.  Assist doctors in decision-making

7. Conclusion

  This work presents a Lightweight LSTM-based hepatitis prediction system for medical prognosis. By integrating preprocessing, data balancing, and an efficient LSTM architecture, the model successfully predicts patient survival outcomes.

The proposed system can be applied in:

1.  Clinical decision support systems
2.  Hospital management systems
3.  Early disease prognosis tools

It provides a reliable and efficient approach for improving healthcare analytics.

7. References

1.Hochreiter, S., & Schmidhuber, J. (1997). Long Short-Term Memory

2.Chollet, F. (2015). Keras Deep Learning Framework

3.Pedregosa et al. (2011). Scikit-learn: Machine Learning in Python