Big Data/Machine Learning/AI
Using machine learning to identify novel predictors of Down syndrome Alzheimer dementia Salina Tewolde* Salina Tewolde Anthony Joseph Rosellini Amy Michals Brian Skotko Juan Fortea Jennifer Weuve Marcia Pescador Jimenez Yorghos Tripodis Eric Rubenstein
Background: Down syndrome Alzheimer Dementia (DSAD) is the leading cause of mortality for people with Down syndrome (DS), with a majority experiencing DSAD during their lifetime. In the general population, several conditions are associated with Alzheimer dementia; people with Down syndrome are at high risk for these conditions. Our goal was to use machine learning models to identify predictors of incident DSAD.
Methods: Data were from a cohort of >130,000 Medicaid- and Medicare-enrolled adults with DS in the US from 2011-2019. We identified DSAD through established Alzheimer dementia algorithms. We r We explored 60 co-occurring conditions a priori selected by a team of clinical experts, determined their onset in relation to the index date, ae. We trained an extreme gradient boosting model to predict incident DSAD in 80% of the observations and tested model performance in the 20% hold-out sample.
Results: Our sample included 128,231 adults, of whom 46% developed DSAD. I. Based on the machine learning model, top predictors of incident DSAD included: epilepsy three years before index date; Black race, and ever having a claim for Peripheral vascular disease, or deafness. Model discrimination was good AUC = 0.82, [0.81-0.83]
Conclusion: Unique DSAD predictors include conditions associated with AD dementia and those common in DS. Identifying them can enable earlier diagnosis through DSAD risk scores from claims data.
