Classification of cancer-related death certificates using machine learning

Area of Science:

• Computational linguistics
• Medical informatics
• Public health surveillance

Background:

• Timely notification of cancer cases is essential for monitoring and prevention.
• Automating cancer case abstraction from free-text documents like death certificates is challenging and time-consuming.

Purpose of the Study:

• Investigate automated methods for detecting notifiable cancer cases from free-text death certificates.
• Improve the efficiency and accuracy of cancer registry data collection.

Main Methods:

• Employed machine learning classifiers, including Support Vector Machines (SVM).
• Utilized natural language processing techniques and the Medtex toolkit for feature extraction.
• Incorporated features such as stemmed words, bi-grams, and SNOMED CT medical concepts.

Main Results:

• Achieved high performance in identifying notifiable cancer cases, with an SVM classifier reaching an F-measure of 0.9866.
• The combination of SNOMED CT concepts and stemmed tokens yielded the lowest variance (0.0032) and false negative rate (0.0297).
• SVM demonstrated robust performance with minimal variance (0.001141) across evaluated runs.

Conclusions:

• Feature selection significantly impacts classifier performance; stemmed tokens with SNOMED CT concepts are highly effective.
• The choice of classifier, such as SVM, also influences detection accuracy.
• Feature weighting schemas had a negligible effect on overall performance.