Automated Detection of Radiology Reports that Require Follow-up Imaging Using Natural Language Processing Feature...

Area of Science:

• Medical Informatics
• Radiology
• Machine Learning

Background:

• Radiologists frequently recommend follow-up imaging for potential cancer findings.
• A significant percentage of patients (35-50%) do not adhere to these follow-up recommendations.
• This non-adherence increases the risk of delayed or missed cancer diagnoses and adverse patient outcomes.

Purpose of the Study:

• To develop and evaluate an automated algorithm for detecting follow-up recommendations in free-text radiology reports.
• To leverage natural language processing (NLP) and machine learning (ML) for improved patient management in cancer screening.

Main Methods:

• Utilized a dataset of 6000 free-text abdominopelvic imaging reports.
• Employed NLP techniques for feature engineering, extracting 1500 features including unigrams, bigrams, and trigrams after tokenization and stemming.
• Trained and compared Naive Bayes, Decision Tree, and Maximum Entropy machine learning models.

Main Results:

• The Decision Tree model achieved the highest performance with an F1 score of 0.458 and an accuracy of 0.862.
• Outperformed Naive Bayes (F1: 0.381) and Maximum Entropy (F1: 0.387) models.
• Identified key predictive features, such as "renal neoplasm" and "evalu with enhanc," crucial for recommendation detection.

Conclusions:

• An automated NLP and ML approach can effectively identify radiology reports requiring follow-up.
• Feature engineering and appropriate ML model selection are critical for maximizing algorithm performance.
• This technology has the potential to improve patient adherence to follow-up recommendations and mitigate risks of delayed cancer diagnosis.