Using machine learning of neuroimaging patterns combined with mitochondrial genomics to predict maternal inheritance of Alzheimer’s disease
Background Maternal family history (FHm) contributes increased risk for developing Alzheimer’s disease (AD) compared to paternal. Family history is often unknown, making it difficult to identify these higher‐risk patients. Atrophy patterns indicative of AD relate to parent‐of‐origin effects. Risk fa...
Saved in:
Published in | Alzheimer's & dementia Vol. 19; no. S24 |
---|---|
Main Authors | , , , , , |
Format | Journal Article |
Language | English |
Published |
01.12.2023
|
Online Access | Get full text |
Cover
Loading…
Summary: | Background
Maternal family history (FHm) contributes increased risk for developing Alzheimer’s disease (AD) compared to paternal. Family history is often unknown, making it difficult to identify these higher‐risk patients. Atrophy patterns indicative of AD relate to parent‐of‐origin effects. Risk factors may also include mitochondrial variation (mtDNA). Machine learning (ML) approaches could identify predictive traits for FHm, helping detect higher‐risk patients with unknown family history. We aimed to determine if structural MRI variables and mitochondrial haplogroups could classify FHm.
Method
Using data from the National Alzheimer’s Coordinating Center and NIA Genetics of Disease Storage Site, we determined if a feature set, comprised of structural MRI variables and mitochondrial haplogroups, could accurately predict FHm. Haplogroups were called with HaploGrep v2.4.0 based on the PT17‐FU1 reference. Two‐thirds of the dataset was used for training and one‐third for testing. For hyper‐parameter tuning and optimization of model performance, we used 4‐fold cross‐validation. To identify traits that aid in predicting FHm, feature importance was calculated for the best performing model (selected based on accuracy and AUC).
Result
ML models were initially trained on imaging variables for 961 individuals (60% female) with confirmed positive (69%) or negative (31%) FHm. Average age at MRI was 68±11 years old. XGboost emerged as the best method (Accuracy = 0.62, AUC = 0.56). Feature importance calculations indicated precuneus was the most important. Including haplogroups significantly reduced the sample size (n = 602) and model performance. Most of the dataset (87%) were assigned to mitochondrial haplogroups within the R branch. Notably, although model performance was reduced, precuneus remained one of the most important variables.
Conclusion
This analysis was among the first to explore if structural or genomic traits predicted FH. Although performance for the initial models showed moderate accuracy, we are currently testing additional ML approaches to enhance predictivity. We identified the precuneus as a key‐AD‐vulnerable region associated with FHm. Our study contributes to a growing literature showing structural, neurodegenerative, cerebral metabolic and increased amyloid‐β burden in the precuneus of individuals with FHm. Analyses aimed at more comprehensive evaluation of mitochondrial genomic variation that expands upon haplogroup classification are ongoing to clarify the role mtDNA plays in maternally inherited AD. |
---|---|
ISSN: | 1552-5260 1552-5279 |
DOI: | 10.1002/alz.083076 |