Extracting Cause of Death From Verbal Autopsy With Deep Learning Interpretable Methods

Area of Science:

• Public Health
• Medical Informatics
• Computational Linguistics

Background:

• Medical certification is the standard for determining cause of death (CoD).
• In low and middle-income countries, most deaths occur outside healthcare facilities, necessitating alternative methods like Verbal Autopsy (VA).
• Existing automated VA methods primarily analyze questionnaire data, excluding valuable narrative information.

Purpose of the Study:

• To develop and evaluate models for automatic CoD ascertainment using textual information from Verbal Autopsy (VA).
• To integrate open-ended narrative responses with closed-ended questionnaire data for improved CoD determination.
• To assess the efficacy of Deep Learning models in VA data analysis.

Main Methods:

• Development of a set of computational models for automatic CoD ascertainment.
• Focus on analyzing textual information, including open-ended responses from VA interviews.
• Comparative analysis of models utilizing closed-ended questions, open responses, and combined data.

Main Results:

• Open-ended VA responses provide significant information for CoD ascertainment.
• The combination of closed-ended questions and open-ended responses yields the highest accuracy in CoD determination.
• Deep Learning models demonstrate promising interpretability and performance for VA data.

Conclusions:

• Automated analysis of VA narratives, particularly open-ended responses, enhances cause of death ascertainment.
• Integrating diverse data sources within VA, including free-text, improves diagnostic accuracy.
• Deep Learning approaches offer a robust and interpretable framework for advancing automated VA systems.