A clinical text classification paradigm using weak supervision and deep representation.

Area of Science:

• Natural Language Processing (NLP)
• Machine Learning in Healthcare
• Clinical Informatics

Background:

• Automatic clinical text classification leverages NLP to extract information from clinical narratives.
• Machine learning models are effective but typically require extensive manual data labeling and feature engineering.
• This study addresses the challenge of reducing human effort in clinical text classification.

Purpose of the Study:

• To propose a novel clinical text classification paradigm using weak supervision and deep representation.
• To reduce the human effort associated with creating labeled training data and conducting feature engineering.
• To evaluate the effectiveness of this paradigm across multiple clinical text classification tasks.

Main Methods:

• Developed a rule-based NLP algorithm for automatic training data labeling (weak supervision).
• Utilized pre-trained word embeddings as deep representation features.
• Evaluated four machine learning models (SVM, RF, MLPNN, CNN) on smoking status and hip fracture classification tasks.

Main Results:

• Convolutional Neural Networks (CNNs) achieved the highest performance (F1 scores of 0.92 for smoking status and 0.97 for hip fracture).
• Word embeddings outperformed traditional features like tf-idf and topic modeling.
• CNNs captured additional patterns beyond rule-based NLP, though they showed sensitivity to training data size and limitations in complex multiclass tasks.

Conclusions:

• The proposed weak supervision and deep representation paradigm effectively reduces human effort in clinical text classification.
• The approach is validated by strong performance on institutional and public datasets.
• This method offers a more efficient pathway for applying machine learning to clinical text data.