Clinical Text Classification with Bio_ClinicalBERT

A deep learning based project for automated classification of clinical assessment text to identify whether "Diagnosis affecting recovery" is present in initial evaluations.

📋 Overview

This project implements a fine-tuned BERT-based model (Bio_ClinicalBERT) to analyze clinical assessment text and determine if the documentation contains evidence of diagnosis affecting patient recovery. The model processes clinical impressions and evaluations to support healthcare documentation quality assurance by identifying whether recovery-impacting diagnoses are properly documented.

🎯 Key Features

• Pre-trained Clinical Model: Utilizes emilyalsentzer/Bio_ClinicalBERT optimized for clinical text
• Text Preprocessing: Advanced NLP preprocessing with lemmatization and stopword removal
• Binary Classification: Identifies presence/absence of diagnosis affecting recovery
• High Accuracy Threshold: 95% confidence threshold for predictions
• Sentence-Level Analysis: Tokenizes and analyzes text at sentence granularity
• GPU Acceleration: CUDA-enabled training and inference

📁 Project Structure

.
├── fd_train.csv           # Training dataset
├── fd_test.csv            # Testing dataset
├── requirements.txt       # Python dependencies
├── Untitled.ipynb         # Main notebook with training & inference
├── save_tokenizer_bcb/    # Saved tokenizer (generated)
├── save_model_bcb/        # Saved base model (generated)
└── fine_tuned_model_bcb/  # Fine-tuned model checkpoints (generated)

🛠️ Technologies Used

• Deep Learning: PyTorch, Transformers (Hugging Face)
• NLP: NLTK, WordNetLemmatizer
• Data Processing: Pandas, NumPy
• Model Training: Hugging Face Trainer API
• Evaluation: scikit-learn, Hugging Face Evaluate
• Visualization: Seaborn
• GPU Management: CUDA, Numba, pynvml

📦 Installation

1. Clone the repository

   git clone https://github.com/bilalhameed248/Diagnosis-Effecting-Patients-Recovery-Detection.git
   cd Diagnosis-Effecting-Patients-Recovery-Detection

2. Install dependencies

   pip install -r requirements.txt

3. Download NLTK data

   import nltk
   nltk.download('stopwords')
   nltk.download('popular')

🚀 Usage

Training the Model

1. Prepare your data: Ensure fd_train.csv and fd_test.csv contain columns:

   • sentence: Clinical text
   • label: Binary labels (0/1)

2. Run training cells in Untitled.ipynb:

   # Load and preprocess data
   dataset = load_dataset('csv', data_files={
       'train': './fd_train.csv', 
       'test': './fd_test.csv'
   })
   
   # Train model
   trainer.train()

3. Training configuration:

   • Epochs: 10
   • Batch size: 4
   • Max sequence length: 512 tokens
   • Evaluation strategy: Per epoch

Inference

Use the inference_why_skill_care() function for predictions:

content = """
Patient evaluation text here...
"""

inference_why_skill_care(content)

Output: Returns whether diagnosis affecting recovery is found with 95%+ confidence.

📊 Model Performance

The model includes:

• Accuracy metric evaluation
• Confusion matrix analysis
• Classification reports
• Checkpoint saving (top 2 models)

🔧 Configuration

Training Arguments

TrainingArguments(
    per_device_train_batch_size=4,
    per_device_eval_batch_size=4,
    num_train_epochs=10,
    save_total_limit=2,
    evaluation_strategy="epoch"
)

Text Preprocessing

• Lemmatization using WordNet
• Stopword removal (English)
• Lowercasing and special character removal
• Whitespace normalization

💻 Hardware Requirements

• GPU: CUDA-compatible GPU recommended
• VRAM: Minimum 8GB for batch size 4
• Storage: ~2GB for model checkpoints

Monitor GPU usage:

from pynvml import *
nvmlInit()
h = nvmlDeviceGetHandleByIndex(0)
info = nvmlDeviceGetMemoryInfo(h)

📈 Results

The model evaluates clinical documentation to determine if it contains evidence of diagnosis affecting patient recovery. Key metrics tracked:

• Training/validation accuracy
• Confusion matrix
• Per-epoch performance

📧 Contact

For questions or support, please open an issue in the repository.

---