Clinical Document Classification Using Labeled and Unlabeled Data Across Hospitals.

Area of Science:

• Artificial Intelligence in Healthcare
• Medical Informatics
• Radiology

Background:

• Reviewing emergency department radiology reports is crucial but time-consuming.
• Timely follow-up of abnormal findings significantly impacts patient outcomes.
• Current machine learning methods require extensive labeled data, which is costly and slow to acquire.

Purpose of the Study:

• To investigate a semi-supervised transfer learning framework for radiology report classification.
• To leverage unlabeled clinical data and pre-existing knowledge to improve model performance with limited labeled data.
• To enhance the efficiency and accuracy of identifying critical radiology findings.

Main Methods:

• Developed and evaluated a semi-supervised transfer learning framework using convolutional neural networks (CNNs).
• Trained models across three hospitals, utilizing both labeled and unlabeled radiology report data.
• Compared CNN performance against conventional supervised learning approaches.

Main Results:

• CNNs outperformed conventional supervised learning methods without requiring problem-specific feature engineering.
• Utilizing unlabeled data reduced the need for labeled data by over 50% while maintaining performance.
• Knowledge transfer from a source hospital significantly boosted the performance of semi-supervised CNNs in a target hospital.

Conclusions:

• Semi-supervised transfer learning with CNNs is effective for radiology report classification.
• This framework substantially decreases reliance on large labeled datasets.
• Cross-hospital knowledge transfer enhances model generalizability and performance in clinical settings.