NLP ML

NLP ML performs binary classification of medical text data using natural language processing and machine learning to extract diagnostic features from chest x-ray radiology reports.

Key Features:

• Data Handling: Accepts CSV input requiring mandatory 'text' and 'class' columns and requires commas within text to be replaced with spaces prior to processing.
• Vectorization Options: Supports TF-IDF and bag-of-words vectorization as candidate feature representations.
• Model Selection: Employs a random search strategy to select among vectorization techniques and machine learning models.
• Validation: Evaluates model performance using 10-fold cross-validation on the training dataset.
• Imbalance Handling: Applies SMOTE (Synthetic Minority Over-sampling Technique) to address class imbalance.
• Result Storage: Saves experimental results and model outputs in a SQLite database.
• Development Dataset: Was developed using the General Medicine Inpatient Initiative (GEMINI) dataset comprising clinical data from seven hospitals.
• Diagnostic Targets: Demonstrated detection of chest x-ray report features including presence or absence of opacities and endotracheal intubation.
• Performance Metrics: Reported performance for detecting absence of opacities includes accuracy 0.84, precision 0.94, recall 0.81, and ROC AUC 0.86, with reduced performance on highly imbalanced datasets.

Scientific Applications:

• Radiology Report Classification: Automates extraction of diagnostic features such as opacities and endotracheal tubes from chest x-ray reports to support clinical workflows on internal medicine wards.
• Clinical Text Model Development: Provides a framework for developing and evaluating binary classifiers for medical informatics research using imbalanced clinical text datasets.

Methodology:

Accepts CSV input with mandatory 'text' and 'class' columns (commas in text replaced with spaces); performs random search over TF-IDF or bag-of-words vectorizers and multiple ML models; applies SMOTE for class imbalance; evaluates with 10-fold cross-validation; saves results to a SQLite database; developed and evaluated on the GEMINI dataset.