Prediction of subjective cognitive functioning based on psychosocial variables using regressions models and machine learning in Mexican adults

Background In Mexico, cognitive decline is a growing problem where mild cognitive impairment has an estimated prevalence of between 20% and 25%.Aim Identify the predictive capacity of psychosocial variables for assessing the level of subjective cognitive function.Methods A nonexperimental and cross-sectional design was used, as well as a non-probabilistic convenience sampling method. Psychosocial variables were assessed to identify the association with subjective cognitive functions in 596 Mexican participants. Descriptive analysis, correlation and machine learning algorithms (Lineal regression, Logistic regression, Naïve Bayes, Decision Trees, Random Forest & Support Vector Machines) were used. For continuous model prediction, the original numerical score from the FACT-Cog was used, whereas for predictive classification, that numerical score was dichotomized using k-means clustering.Results For predictive classification, the Naïve Bayes model achieved the highest numerical performance (AUC = 0.830, 95% CI [0.765, 0.894], accuracy = 0.783), followed closely by Logistic Regression (AUC = 0.817, 95% CI [0.751, 0.883]) and Support Vector Machines (AUC = 0.789, 95% CI [0.713, 0.865], accuracy = 0.758). For continuous prediction of FACT-Cog scores, Linear Regression yielded the most optimal explanatory power (R2 = 0.340, cross-validated R2 = 0.375, RMSE = 0.786) compared to non-linear tree-based algorithms. Sensitivity analyses confirmed the stability of these estimates.Conclusions A comparison of various types of models (linear, nonlinear, and conditional) suggests that subjective cognitive functioning is associated with subjective well-being and neuroceptive abilities based mainly on the variables analyzed in this study.