The Impact of NOACS versus VKAS on Absolute and Relative Cognitive Function Decline Over Time: A Machine Learning Approach

Figures

• 

  Fig. 1. Flowchart of patients admitted to the study and reasons for exclusion.

• 

  Fig. 2. Receiver Operating Characteristic (ROC) curve for the random forest model. The curve illustrates the trade-off between sensitivity (true positive rate) and 1-specificity (false positive rate) across different classification thresholds. The area under the curve (AUC) is 0.87, indicating high discriminative ability of the model in distinguishing patients with and without cognitive decline. The diagonal line represents a no-skill classifier (AUC = 0.5) for reference.

• 

  Fig. 3. Precision–recall (PR) curve for the random forest model. This curve shows the relationship between precision and recall across various classification thresholds. The Area Under the Curve (PR AUC) is 0.76, reflecting the model's ability to identify true positive cases of cognitive decline in the context of class imbalance. A higher PR AUC indicates better performance in correctly identifying patients at risk, particularly when the positive class (cognitive decline) is underrepresented.

• 

  Fig. 4. Confusion matrix of the random forest classifier using a threshold of 0.5. The matrix summarizes model predictions on the test set. The classifier correctly identified 125 patients without cognitive decline (true negatives) and 34 with cognitive decline (true positives) but misclassified 30 patients with cognitive decline as nondecline (false negatives), and 12 patients without decline as decline (false positives). These results correspond to a recall of 0.53 and a precision of 0.74 for the decline class.

• 

  Fig. 5. Kernel density estimation of predicted probabilities of cognitive decline for NOAC- and VKA-treated patients. The plot displays the distribution of predicted probabilities from the RANDOM FOREST model for the two treatment groups. Patients on VKAs exhibit a right-skewed distribution with higher predicted probabilities of cognitive decline, while those on NOACs show a left-skewed distribution with lower predicted risk. This separation is statistically confirmed by a Kolmogorov–Smirnov test (KS statistic = 0.385, p < 0.001), indicating a significant difference in model predictions between groups.

• 

  Fig. 6. Top 15 predictors of cognitive decline ranked by permutation importance. This bar plot displays the mean decrease in ROC AUC associated with randomly permuting each feature in the SMOTE-balanced random forest model, evaluated on the original test set. Error bars represent the standard deviation across 10 permutations. The variable “anticoagulant therapy” emerged as the most influential predictor, followed by glycemia, systolic blood pressure, BMI, albumin, and potassium levels. This analysis confirms the dominant role of anticoagulant type in predicting cognitive outcomes, even within a multivariable and nonlinear framework.