Clinical Concept-Based Radiology Reports Classification Pipeline for Lung Carcinoma

Area of Science:

• Medical Informatics
• Natural Language Processing
• Oncology

Background:

• The rising incidence of cancer necessitates efficient data extraction from unstructured medical reports.
• Radiology reports contain crucial information on disease characteristics, treatment, and outcomes, but manual extraction is time-consuming.
• Natural Language Processing (NLP) offers a solution for automating information extraction from free-text clinical data.

Purpose of the Study:

• To develop and compare NLP models for classifying lung carcinoma radiology reports based on clinical concepts.
• To evaluate the performance of rule-based and machine learning models for this classification task.
• To assess the utility of the developed pipeline for automated annotation and data analysis of lung cancer reports.

Main Methods:

• A clinical concept-based classification pipeline was created for lung carcinoma radiology reports.
• Rule-based, XGBoost, and Bidirectional Long Short-Term Memory (Bi-LSTM) deep learning models were developed and compared.
• The models were trained and tested on 1700 radiology reports (CT and PET/CT) and validated on 501 reports from the MIMIC-III database.

Main Results:

• The rule-based algorithm with expert input achieved the highest performance, with an F1 score of 0.94 internally and 0.74 on external validation.
• Among machine learning models, Bi-LSTM_dropout outperformed XGBoost and Bi-LSTM_simple on the internal dataset.
• On external validation, Bi-LSTM_simple showed relatively better performance compared to the other two machine learning models.

Conclusions:

• The developed NLP pipeline demonstrates effectiveness in classifying lung carcinoma radiology reports.
• Rule-based approaches, leveraging expert knowledge, show strong performance in this specific application.
• The pipeline can facilitate automated annotation and efficient analysis of large volumes of lung cancer radiology reports.