Evaluasi Prediksi Stres Berdasarkan Tidur dan Kognitif Menggunakan Machine Learning

Abstract

Sleep and cognitive indicators are theoretically relevant to stress, but this relevance does not guarantee individual prediction in small datasets. This exploratory study evaluates the learnability of Stress_Level in the public Sleep Deprivation and Cognitive Performance dataset containing 60 observations and 12 predictors. Descriptive analysis is complemented by nested cross-validation with five outer folds, three repetitions, and three inner folds. Encoding, scaling, and feature selection are fitted exclusively on training data. Six regression families are compared with a training-mean baseline, least absolute shrinkage and selection operator (LASSO), and recursive feature elimination (RFE) with Ridge. All marginal 95% correlation intervals include zero; the largest correlation concerns sleep quality (0.190). All feature-based models yield negative mean coefficients of determination, ranging from −0.470 to −0.235, indicating failure to generalize in this internal evaluation. The baseline achieves the lowest mean absolute error of 9.65 ± 1.33, compared with 9.69 ± 1.35 for RFE–Ridge and 9.76 ± 1.41 for Ridge; values after ± represent standard deviations across outer folds. Averaged cross-validated Ridge predictions concentrate within 14.67–19.72 despite the target range of 0–40. Tuning and feature selection do not demonstrate added predictive value over the baseline. These negative findings diagnose the information limits of this dataset and do not support stress screening or diagnosis. Small sample size, limited instrument documentation, and absent external validation restrict inference. Future studies require validated stress measurements, longitudinal observations, and temporally aligned physiological and psychosocial features.