Algorithmic Detection of Chronological Sleep Decay: An L2-Regularized Logistic Regression Analysis of CDC Surveillance Data (1991–2023)
Background & Objective: Chronic sleep deprivation among adolescents has accelerated over the past three decades, aligning with widespread digital media saturation and reported cognitive focus issues. Traditional public health tracking often evaluates short‑term trends, overlooking macro‑generational shifts. This research presents a supervised machine learning workflow designed to quantify macro‑generational sleep degradation by comparing pre‑digital adolescent sleep baselines with those observed in the modern, smartphone‑saturated era. Methodology & Computational Pipeline: Secondary surveillance microdata were extracted from the Centers for Disease Control and Prevention (CDC) Youth Risk Behavior Surveillance System (YRBSS), aggregating N > 10,400 independent adolescent records across two distinct generational cohorts: 1991 (pre‑digital baseline) and 2023 (digital saturation). Harmonization & Preprocessing: Data cleaning and structural alignment were performed in Python via Google Colab using pandas and NumPy. Incomplete or null records were removed, 30‑year categorical header mismatches were resolved, and qualitative behavioral strings were converted into numerical feature matrices. Standardization: To eliminate magnitude‑driven coefficient bias, all vectorized features were normalized using scikit‑learn’s StandardScaler (μ = 0, σ² = 1), preventing high‑scale variables from disproportionately influencing model gradients during training. Leakage Control: The dataset was partitioned into a strict 80:20 randomized train–test split, deploying 8,320+ profiles for training and isolating 2,080+ unseen holdout records as a pure out‑of‑sample evaluation set to reduce data leakage and overfitting. Model Architecture: A supervised binary logistic regression classifier was engineered to predict an adolescent’s chronological era based solely on rest and behavioral metrics. An L₂ regularization penalty (Ridge configuration) was embedded directly into the optimization loss function to penalize extreme parameter weights and enhance decision‑boundary generalization. Results & Performance Metrics: The regularized model achieved an out‑of‑sample classification accuracy of 98.42% across N = 4,634 validation records. Confusion‑matrix results show that the model cleanly separates the 1991 and 2023 cohorts in feature space, indicating a strong macro‑generational boundary in adolescent sleep architectures. Statistical Significance: A two‑sample t‑test confirmed a large, statistically significant decline in adolescent sleep baselines over the 30‑year interval (t = 89.016, p < 0.001). Conclusions & Public Health Implications: The pronounced feature‑space separation between the 1991 and 2023 cohorts suggests that contemporary digital environments are closely associated with substantial changes in adolescent sleep patterns relative to the pre‑digital baseline. Declining rest hours parallel reported concentration difficulties, supporting the hypothesis that intensified digital media exposure may contribute to impaired daytime executive function and a sharp macro‑generational behavioral boundary. Acknowledgements: This research was completed as part of the Bridge to STEM Academy Summer Session. Special thanks to program mentors and Metea Valley High School for their guidance and institutional support.