Ad hoc classification of radiology reports

Area of Science:

• Medical Informatics
• Natural Language Processing
• Information Retrieval

Background:

• Ad hoc classification involves automatically categorizing text documents into user-defined, nonstandard classes.
• Statistical information retrieval techniques offer a potential solution for classifying large document sets, such as dictated mammography reports.

Purpose of the Study:

• To evaluate the use of statistical information retrieval techniques for the ad hoc classification of dictated mammography reports.
• To develop and assess an automated classification algorithm for medical documents.

Main Methods:

• An automated classification algorithm was generated using positive and negative evidence from relevance-judged documents.
• Documents were sorted into "membership," "exclusion," and "uncertain" categories.
• Negation and conjunction phrases were expanded and tokenized to manage absent findings.
• Classifier performance was evaluated using the F measure, combining recall and precision.

Main Results:

• Single terms were the most effective text features for classification.
• Adding pairs of unordered terms provided minor improvements.
• Excessive classifier enhancement iterations led to overtraining and degraded performance.
• Optimal performance was achieved with balanced proportions of relevant and irrelevant training documents.
• Special handling of negation phrases improved performance, especially with limited terms.

Conclusions:

• The ad hoc classifier system shows promise for classifying large medical document collections.
• The NegExpander tool effectively distinguishes positive from negative evidence, crucial for accurate medical document classification.