Automated verbal autopsy classification: using one-against-all ensemble method and Naïve Bayes classifier

Area of Science:

• Public Health
• Biostatistics
• Epidemiology

Background:

• Verbal autopsy (VA) is crucial for determining causes of death (COD) in regions with limited vital registration.
• Existing automated VA COD assignment methods show suboptimal performance compared to physician diagnoses.
• There is a need for improved algorithms to enhance the accuracy of VA-based health system interventions.

Purpose of the Study:

• To enhance the Naïve Bayes Classifier (NBC) algorithm for more accurate verbal autopsy (VA) cause of death (COD) assignment.
• To compare the performance of the enhanced NBC, termed One-Against-All NBC (OAA-NBC), against leading VA classification methods.
• To evaluate the impact of OAA-NBC on both population-level and individual-level COD assignments.

Main Methods:

• Trained multiple NBC algorithms using the one-against-all (OAA) approach on 26,766 verbal autopsy (VA) records from diverse global datasets.
• Compared OAA-NBC performance against Tariff, InterVA-4, InSilicoVA, and standard NBC.
• Assessed performance using cumulative cause-specific mortality fraction (CSMF) accuracy for population-level agreement and cumulative partially-chance corrected concordance (PCCC) and sensitivity for individual-level assignments.

Main Results:

• The One-Against-All Naïve Bayes Classifier (OAA-NBC) demonstrated superior performance in assigning causes of death (COD) that closely align with physician and clinical diagnoses.
• OAA-NBC improved classification sensitivity by 6-8% compared to other leading verbal autopsy (VA) algorithms.
• Population-level agreement metrics (CSMF accuracy) for OAA-NBC were comparable or higher than other tested algorithms, though individual-level assignment accuracy requires further refinement.

Conclusions:

• The One-Against-All approach significantly enhances the Naïve Bayes Classifier (NBC) for verbal autopsy (VA) cause of death (COD) assignment.
• OAA-NBC offers improved accuracy in COD classification, particularly at the population level, supporting evidence-based health strategies.
• While further optimization for individual-level assignments is needed, OAA-NBC represents a promising advancement over existing VA classification methods.