Night Signals
Sleep × Doomscrolling
A data-audit and leakage-aware modelling case study on 1,000 public Kaggle records described by the uploader as synthetic.
System Architecture
Data quality and model evaluation documented in Night Signals.
Public synthetic dataset
A Kaggle file with 1,000 records and uncertain label construction is audited.
Data quality review
Response patterns, outliers, and missing values are audited across 29 analytic variables.
Full EDA
Statistical summaries, distributions, correlations, and cross-tabulations reveal behavioral patterns.
Internal label evaluation
Outcome columns are excluded; nested CV, calibration, and holdout diagnostics evaluate recovery of the Poor label.
Research integrity
NHANES and PhysioNet are separate context layers, not external model validation.
An auditable synthetic-data case study with explicit validity limits.
Project Overview
Night Signals audits a public synthetic Kaggle dataset, including label inconsistencies, missingness, and suspicious structure.
Its binary classifier excludes measured sleep outcomes, uses nested development cross-validation, calibration, and an internal holdout. The current holdout balanced accuracy is 75.96%; it measures recovery of this file's label, not real-world risk.
NHANES and PhysioNet provide unrelated background evidence. Neither validates the Kaggle classifier or its doomscrolling findings.
The project documents the limits of cross-sectional synthetic data and reports exploratory associations without causal or clinical claims.
Problem
The relationship between doomscrolling habits, news consumption patterns, and sleep outcomes is poorly quantified, with most discussion relying on anecdote rather than structured data.
Solution
Audited 1,000 public synthetic records and built an internally evaluated, leakage-aware label classifier; examined NHANES and PhysioNet separately as background context.