Multilingual information retrieval in thoracic radiology: feasibility study.

Area of Science:

• Medical Informatics
• Computational Linguistics

Background:

• Electronic medical records primarily store essential patient information as unstructured text, posing challenges for automated data extraction.
• Natural language processing (NLP) offers a solution to unlock and interpret clinical information embedded within free-text medical documents.

Purpose of the Study:

• To evaluate the efficacy of coupling machine translation (MT) with NLP for extracting clinical findings from Portuguese chest X-ray reports.
• To assess the performance of this methodology in identifying normal reports and reports containing device descriptions.

Main Methods:

• Utilized the MEDLEE NLP processor, originally designed for English, by first machine translating Portuguese chest X-ray reports into English.
• Applied a serial coupling of machine translation and NLP to process 165 chest X-ray reports.
• Established a reference standard for report classification based on the consensus of two experienced radiologists.

Main Results:

• Achieved a sensitivity of 1 and specificity of 0.71 for identifying normal chest X-ray reports.
• Attained a sensitivity of 0.97 and specificity of 0.97 for identifying chest X-ray reports with device descriptions.
• Demonstrated high accuracy in classifying report types using the combined MT-NLP approach.

Conclusions:

• The coupling of machine translation and natural language processing is a viable method for extracting clinical findings from Portuguese chest X-ray reports.
• This integrated approach shows significant potential for improving automated data retrieval and analysis in multilingual clinical settings.
• Further research can explore expanding this methodology to other medical text types and languages.