Background: Identifying the neurocognitive profiles that define the preclinical (Subjective Cognitive Decline, SCD) and prodromal (Mild Cognitive Impairment, MCI) stages of Alzheimer´s disease (AD) is crucial to detect individuals at risk and improve early intervention strategies (Jack et al., 2024). We developed an optimized model based on neuropsychological measures to classify cognitively unimpaired (CU), SCD, single-domain MCI (sd-MCI), and multi-domain MCI (md-MCI) along the AD continuum, using machine learning approaches.

Method: We studied 318 participants (138 CU, 73 SCD, 53 sd-MCI, and 54 md-MCI) from the Compostela Aging Study (CompAS) and used 107 baseline neuropsychological and sociodemographic measures (Juncos-Rabadán et al., 2012). We compared three single algorithms (Support Vector Machine [SVM], Random Forest [RF], and Logistic Regression [LR]) and a Soft-Voting Ensemble, evaluating performance using ROC curve analysis, and reporting class-wise AUC under leave-group-out cross-validation. Prioritizing early stages, we selected SVM based on SCD performance and optimized the feature set to 13 features using performance-based selection and a collinearity-reduction procedure. The importance of the included features was assessed using permutation importance analysis.

Results: The final SVM model (13 features) showed high cross-sectional discrimination between the diagnostic groups considered, with class-wise AUC values of 0.82 (CU), 0.83 (SCD), 0.80 (sd-MCI), and 0.86 (md-MCI) (Figure 1). The greatest performance decreases under permutation were observed for short-delay free recall measure (TAVEC_SDFR; 0.190), memory complaints reported by the patient (MFE_Patient; 0.067), and age (0.058) (Figure 2), highlighting episodic memory, cognitive complaints, and age as the most influential variables in differentiating the groups along the AD continuum.

Conclusion: The neuropsychological SVM model demonstrated good discrimination across the preclinical and prodromal stages of AD, highlighting episodic memory and patient-reported memory complaints as key predictors. These findings suggest that this algorithm could support clinical practice by helping to identify individuals at higher risk in the early stages and guiding future clinical decision-making.