The Question
Sleep is one of the most-discussed health behaviors — poor sleep is linked to cardiovascular disease, diabetes, depression, and mortality. But sleep quality is also deeply entangled with the conditions it predicts. If you already know someone has chronic pain, depression, and three chronic diseases, does knowing they sleep poorly add any new information? This investigation tests whether sleep measures improve health decline prediction beyond what standard clinical features already capture.
ADM Prediction (Made Before Running Models)
Predicted winner: ML with sleep, but modest gain. Sleep quality is a published mortality and morbidity predictor (Cappuccio et al., 2010 meta-analysis: short sleep RR 1.12 for all-cause mortality). But sleep quality correlates strongly with depression, pain, and chronic conditions already in the model. The marginal information content of sleep — beyond what depression and disease burden capture — may be small.
Results
ROC Curves
Feature Importance (Top 8)
Sleep Quality vs Health Decline Rate
Multi-Model Comparison
Sleep’s predictive value may differ by age and sex. Do sleep measures help more for younger adults (where baseline risk is lower) or older adults (where comorbidities dominate)?
The ADM Insight
Sleep matters for health — but your doctor already knows most of what sleep would tell them. Adding sleep data improves prediction modestly, confirming sleep’s role as a health indicator. But most of sleep’s predictive value is already captured by depression scores, chronic pain, and functional limitations. Sleep’s independent signal is real but small — consistent with the epidemiological finding that sleep quality is more often a symptom than a cause.
Cohort: HRS RAND respondents aged 50+ with sleep data in waves 10–15 (2010–2020). Outcome: significant health decline (self-rated health worsens by 1+ points) or mortality within 6 years.
Domain baseline: Standard health risk score + published sleep-health relative risks. Poor sleep RR 1.5 for health decline, short sleep (<6 hours) RR 1.3, long sleep (>9 hours) RR 1.2.
ML Model 1 (No Sleep): GradientBoostingClassifier on standard health features — demographics, chronic conditions, functional status, depression, BMI, smoking, exercise.
ML Model 2 (+Sleep): Same GBM architecture with two additional features: sleep quality composite (count of poor sleep indicators: restless sleep, trouble falling asleep, waking early, waking at night) and self-reported sleep hours.
Evaluation: 5-fold stratified cross-validation. Bootstrap 95% CIs from 1,000 resamples. Three-way comparison: domain baseline, ML without sleep, ML with sleep.
Limited follow-up: Sleep data only available from wave 10 (2010) onward, limiting the follow-up period compared to earlier investigations.
Self-reported sleep: All sleep measures are self-reported survey items. No polysomnography or actigraphy data available to validate sleep duration or quality objectively.
No sleep apnea data: Obstructive sleep apnea — a major independent risk factor — is not captured in earlier HRS waves and may be underreported even when asked.
Reverse causation: Poor health may cause poor sleep, not just the other way around. This cross-sectional-to-longitudinal design mitigates but does not eliminate this concern.
Medication effects: Many medications (beta-blockers, SSRIs, corticosteroids) affect sleep quality. Medication use is not controlled for in these models.