Interpretable machine learning models for predicting incestuous relationships from short tandem repeat profiles in sexual assault-related pregnancies for forensic intelligence.

Pregnancies resulting from sexual assault present complex ethical and investigative challenges, especially in cases that involve minors or individuals with intellectual disabilities. When suspect samples are unavailable, identifying whether the biological father is a close relative can substantially narrow the scope of the investigation. This study presents interpretable machine learning models to evaluate the likelihood of a consanguineous relationship between an alleged perpetrator and a victim in cases of pregnancies resulting from suspected sexual assault, using short tandem repeat (STR) profiles of the fetus or newborn and the mother. We simulated 100,000 STR profiles across four incestuous relationship scenarios: parent (father-daughter), sibling (brother-sister), grandparent (grandfather-granddaughter), and uncle (uncle-niece) scenarios. Logistic regression and decision tree models were trained using two STR-derived features: the number of homozygous loci and the number of loci matching the maternal genotype. Both models demonstrated excellent predictive performance, particularly in the parent and sibling scenarios (area under the curve >0.9). The models' interpretability ensures the transparency essential for forensic casework. This approach enables informed investigative decisions and may prevent unnecessary DNA testing, thereby improving forensic efficiency while protecting human rights.