Using machine learning of neuroimaging and mitochondrial haplogroups to predict maternal family history of Alzheimer’s disease

Maternal family history (FHm) increases the risk for Alzheimer’s disease (AD) compared to paternal. However, family history information is often unknown, complicating identification of higher-risk patients. Parent-of-origin effects on brain atrophy and matrilineally inherited mitochondrial genomic variation have been reported, but their utility for risk stratification is unclear. Machine learning approaches show promise for identifying traits associated with FHm, potentially improving FHm prediction in patients with unknown family history. We assessed whether structural Magnetic Resonance Imaging variables and/or mitochondrial haplogroups could accurately predict self-reported FHm as a proxy for individuals with high AD risk. Datasets were divided into 80% training and 20% testing using stratified sampling, with 4-fold cross-validation for hyperparameter tuning and model optimization. Models were compared based on accuracy and AUC. Feature importance was evaluated to identify the most informative variables. Neuroimaging data were available for 1,129 participants with known family history (mean age 69 years; 61% female; 44% FHm-positive; 71% normal cognition; 20% without normal cognition). For neuroimaging only models, XGBoost (Accuracy = 0.59, AUC = 0.59) and naïve Bayes (Accuracy = 0.58, AUC = 0.58) performed best, with precuneus volume being the most important feature correlated with reported FHm. Inclusion of mitochondrial haplogroups reduced the sample size (n = 717) but had minimal impact on model performance using XGBoost (Accuracy = 0.62, AUC = 0.62). The R0 intermediate clade was the most important mitochondrial haplogroup, with the precuneus remaining the most important overall feature. This study is among the first to integrate neuroimaging and mitochondrial data to identify patients with self-reported FHm of AD. Our findings reinforce the precuneus as an AD-vulnerable region linked to FHm. The addition of mitochondrial haplogroups maintained accuracy similar to imaging alone, despite a reduced sample size. This highlights the potential of combining neuroimaging and mitochondrial genetics for identifying traits associated with reported FHm of AD when family history is unknown.