Machine learning performance in a microbial molecular autopsy context: A cross-sectional postmortem human population...

Area of Science:

• Forensic microbiology
• Computational biology
• Bioinformatics

Background:

• The postmortem microbiome offers crucial insights for death investigations and understanding host health.
• Analyzing large, complex microbiome datasets presents analytical challenges.
• Machine learning (ML) offers potential solutions for interpreting postmortem microbiome data.

Purpose of the Study:

• To compare the effectiveness of different ML methods in analyzing postmortem microbiomes.
• To assess the ability of ML models to predict key attributes in death investigations.
• To determine optimal sampling strategies for postmortem microbiome analysis.

Main Methods:

• Collected postmortem microbiome samples from five anatomical areas in 188 death cases.
• Sequenced and analyzed microbiome data.
• Compared three ML methods: boosted algorithms (Xgboost), random forests, and neural networks.
• Evaluated prediction accuracy for postmortem interval (PMI), location, and manner of death.

Main Results:

• All ML algorithms performed well, with distinct predictive strengths.
• Xgboost showed high accuracy but potential for overfitting.
• Random forests demonstrated stability, especially with increased anatomical areas.
• Analyzing more than three anatomical areas yielded diminishing returns in accuracy.

Conclusions:

• The eyes and rectum provided the most informative microbiome data for predicting death circumstances.
• ML methods are valuable tools for analyzing postmortem microbiome data in forensic investigations.
• Specific anatomical sites yield more predictive power than others.