Neuropsychological and clinical variables associated with cognitive trajectories in patients with Alzheimer's disease

Abstract

Background: The NeuroArtP3 (NET-2018-12366666) is a multicenter study funded by the Italian Ministry of Health. The aim of the project is to identify the prognostic trajectories of Alzheimer's disease (AD) through the application of artificial intelligence (AI). Only a few AI studies investigated the clinical variables associated with cognitive worsening in AD. We used Mini Mental State Examination (MMSE) scores as outcome to identify the factors associated with cognitive decline at follow up.

Methods: A sample of N = 126 patients diagnosed with AD (MMSE >19) were followed during 3 years in 4 time-points: T0 for the baseline and T1, T2 and T3 for the years of follow-ups. Variables of interest included demographics: age, gender, education, occupation; measures of functional ability: Activities of Daily Living (ADLs) and Instrumental (IADLs); clinical variables: presence or absence of comorbidity with other pathologies, severity of dementia (Clinical Dementia Rating Scale), behavioral symptoms; and the equivalent scores (ES) of cognitive tests. Logistic regression, random forest and gradient boosting were applied on the baseline data to estimate the MMSE scores (decline of at least >3 points) measured at T3. Patients were divided into multiple splits using different model derivation (training) and validation (test) proportions, and the optimization of the models was carried out through cross validation on the derivation subset only. The models predictive capabilities (balanced accuracy, AUC, AUPCR, F1 score and MCC) were computed on the validation set only. To ensure the robustness of the results, the optimization was repeated 10 times. A SHAP-type analysis was carried out to identify the predictive power of individual variables.

Results: The model predicted MMSE outcome at T3 with a mean AUC of 0.643. Model interpretability analysis revealed that the global cognitive state progression in AD patients is associated with: low spatial memory (Corsi block-tapping), verbal episodic long-term memory (Babcock's story recall) and working memory (Stroop Color) performances, the presence of hypertension, the absence of hypercholesterolemia, and functional skills inabilities at the IADL scores at baseline.

Conclusion: This is the first AI study to predict cognitive trajectories of AD patients using routinely collected clinical data, while at the same time providing explainability of factors contributing to these trajectories. Also, our study used the results of single cognitive tests as a measure of specific cognitive functions allowing for a finer-grained analysis of risk factors with respect to the other studies that have principally used aggregated scores obtained by short neuropsychological batteries. The outcomes of this work can aid prognostic interpretation of the clinical and cognitive variables associated with the initial phase of the disease towards personalized therapies.

Keywords: Alzheimer dementia; MMSE; SHAP analysis; machine learning; mild cognitive impairment; random forest.

Figure 1

Shapley Additive exPlanations (SHAP) analysis results for the best-performing model (Random Forest, using a 70-30% train-test split). Features that consistently ranked within the top ten across at least 60% of runs are reported. For categorical variables (e.g. hypertension and hypercholesterolemia), beeswarm plots are shown, where each point represents a SHAP value for a feature and an individual observation. Blue points indicate low variable values, while red points indicate high values. For continuous variables (e.g., IADL score, Corsi block tapping test, Babcock story recall, and Stroop color-word test), dependence plots are presented, with each point representing a feature score for an individual participant. Higher SHAP values suggest a positive contribution to the model's prediction of MMSE decline at T3.

Figure 2

ROC curves showing the performance of the Random Forest (RF) classifier across selected test set sizes (10%, 20%, 30%, and 40%). The True Positive Rate is plotted against the False Positive Rate for each test size, with the Area Under the Curve (AUC) illustrating classifier effectiveness. Blue line represents the mean ROC, the shaded gray area indicates variability across iterations, and the red dashed line corresponds to the line of no-discrimination (AUC = 0.5).