Soft-Label Machine Learning for Verbal Autopsy: Pediatric Mortality Data from Uganda

356 pediatric post-discharge mortality verbal autopsy records from a multisite cohort study in Uganda. The dataset was created by Manav Doshi and last updated in May 2026. It was collected using encrypted tablets and a REDCap database, with verbal autopsies independently reviewed by two physicians.

Use Cases

• Compare soft-label versus hard-label supervised learning approaches based on physician-assigned cause-of-death probabilities.
• Train models for population-level cause-specific mortality fraction (CSMF) estimation based on verbal autopsy data.
• Investigate the impact of diagnostic uncertainty in medical data modeling based on physician confidence levels.
• Benchmark XGBoost performance on probabilistic classification tasks using clinical and demographic features.

Strengths

• 356 records provide a foundation for modeling pediatric post-discharge mortality.
• Data includes physician-assigned causes with confidence levels, enabling soft-label modeling.
• Population-level model performance reported as 0.88 ± 0.04 CSMF accuracy.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

Borealis Harvested Dataverse

Collection Method

Collected at point of care using encrypted tablets and a REDCap database, with verbal autopsies reviewed by two physicians.

Freshness

Last updated 2026-05-09 04:10:47; freshness should be verified.

Geography

Uganda