<![CDATA[Sleep quality is an important indicator of health, shaped by the intricate interactions of various multidimensional elements, including eating habits, physical activity levels, and psychological well-being. The purpose of this study is to assess the efficacy of linear (Logistic Regression) and non-linear machine learning algorithms (Decision Tree and Gradient Boosting) in predicting sleep disorders. The dataset used came from a survey that included 195 participants and underwent several stages of preprocessing, including data cleansing, feature engineering (especially regarding duration), and standardization through Z-score normalization. Model validation is performed using the Stratified 5-Fold Cross-Validation technique to reduce bias and ensure the stability of performance metrics. The findings show that the Logistics Regression model, optimized through threshold adjustment (with an adjustment threshold of 0.55), shows the most favorable and consistent performance, achieving an accuracy rate of 79.5% and a drawback rate of 96.7%, thus surpassing the efficacy of non-linear models such as Gradient Boosting. These results reinforce the principle of simplicity, which states that, in limited datasets, simpler linear models often show greater resilience than their more complex counterparts, which are prone to overfitting. An interpretability assessment conducted through the Analysis of the Importance of Permutation Features showed that bedtime dietary practices and stress levels emerged as the dominant factors affecting sleep quality, surpassing the contribution of physical activity and anthropometric variables (such as height and weight). In light of these findings, this study advocates health interventions that emphasize time-oriented nutrition management (chrononutrition) and recommends the use of linear models as a basic strategy in the screening process for populations characterized by limited sample sizes]]>
