Robust cross-study body fluid identification using an extreme gradient boosting classifier
Authors
Niklaus Johner
CHUV
Trestan Pillonel
CHUV
Claire Bertelli
CHUV
[sponser-meet-now-chat][/sponser-meet-now-chat]
Identifying unknown biological stains at crime scenes is critical in forensic science, as body fluid origin significantly influences legal outcomes. While microbial profiles are promising biomarkers, existing machine learning models often suffer from study-specific biases. To address this, we developed an extreme gradient boosting classifier (XGBC) using the Microbiome Forensics Database (MFDB). We filtered MFDB to retain sites represented by more than three distinct studies and discarded OTUs (97% clustering) singletons. The training set comprised ~125k profiles across ten body sites: feces, saliva, vagina, skin, blood, urine, breast milk, nostril swabs, amniotic liquid, and semen.
The model utilizes ~8k features based on relative abundances across all taxonomic ranks. Using an 80% training and 20% test split, the model yielded an overall F1-score of 0.97. To ensure robustness and prevent study data leakage, we performed stratified k-fold (k=5) cross-validation using study IDs as groups. While performance decreased for several sites, predictions for feces, saliva, and vagina remained highly accurate (average F1-score 0.93; SD=0.01). This was further confirmed on an independent validation set of 660 samples (feces, saliva, vagina, urine) sequenced internally at CHUV, where the model achieved F1-scores of 0.85 (feces), 0.99 (saliva), and 0.92 (vagina). Furthermore, SHAP value analysis identified body fluid specific species, like Lactobacillus iners and Porphyromonas catoniae, driving classifier decision towards vaginal and salival samples, respectively.
Our results highlight an accurate and interpretable machine learning classifier, capable of predicting saliva, feces, and vaginal sample origins across studies with diverse sequencing technologies. This model provides a robust tool for forensic investigators, though further research is required to validate performance on body site mixtures.
Contact Attendee
Warning: Attempt to read property "user_email" on string in /home/1276969.cloudwaysapps.com/ydbgzhdjeq/public_html/wp-content/plugins/my-conference-now/functions.php on line 2826