Predictive Value of Machine Learning and Nomogram Models Based on Brain Amyloid SUVR in Alzheimer's Disease

This study aimed to develop machine learning and nomogram models based on brain amyloid standardized uptake value ratio (SUVR) for the prediction of Alzheimer's disease (AD). Least absolute shrinkage and selection operator (LASSO) regression was employed to identify key brain regions associated with amyloid SUVR, which were then integrated into a composite SUVR_Score. A nomogram model was subsequently constructed to support clinical decision-making.

A total of 751 subjects diagnosed with either cognitively normal (CN; n=550) or Alzheimer's disease (AD; n=201) were selected from the Alzheimer's Disease Neuroimaging Initiative database (2007-2023). Clinical data and amyloid SUVR values from multiple brain regions were collected. Subjects were randomly assigned to training (n=525) and validation (n=226) cohorts at a 7:3 ratio. LASSO regression was used to identify region-specific SUVR values significantly associated with AD, which were combined into a composite SUVR_Score. Seven machine learning algorithms, comprising decision tree (DT), random forest (RF), eXtreme Gradient Boosting, support vector machine, k-Nearest Neighbors, LightGBM, and Naive Bayes, were trained using seven variables. The model with the best performance was selected, and a nomogram incorporating key predictors was developed to estimate AD risk. Model performance was assessed using receiver operating characteristic curves, calibration plots, and decision curve analysis (DCA).

LASSO regression identified nine brain regions with amyloid SUVR values significantly associated with AD, which were combined into a composite SUVR_Score. Among the machine learning models, RF demonstrated the best performance in the validation cohort (AUC=0.974; accuracy=0.911; sensitivity=0.857; specificity=0.965). A nomogram was then constructed using four predictors: sex, marital status, education, and SUVR_Score, achieving an AUC of 0.965. Calibration curves showed high agreement between predicted and observed outcomes, while DCA confirmed favorable clinical utility.

Machine learning and nomogram models based on brain amyloid SUVR effectively distinguish AD from CN individuals and offer valuable support for clinical diagnosis and risk prediction.