Integrating Explainable Machine Learning and Mixed-Integer Optimization for Personalized Sleep Quality Intervention
A new predictive-prescriptive AI model uses SHAP and optimization to generate minimal, high-impact sleep recommendations.
A team of researchers has published a novel AI framework that moves beyond simply predicting poor sleep to prescribing actionable, personalized interventions. The system, detailed in a new arXiv paper, first uses a supervised machine learning classifier trained on survey data to predict an individual's sleep quality with high accuracy (93.66%) and an F1-score of 0.9544. Crucially, it employs SHAP (SHapley Additive exPlanations), an explainable AI (XAI) technique, to quantify the influence of modifiable behavioral and environmental factors on the prediction.
These quantified influences are then fed into a mixed-integer optimization (MIO) model, which acts as a 'prescriptive engine.' This model identifies the minimal set of behavioral adjustments needed for meaningful sleep improvement, explicitly modeling a user's potential resistance to change through a penalty mechanism. The result is a concise, personalized plan—often suggesting just one or two high-impact changes, like adjusting caffeine intake or bedtime—that balances expected improvement against the effort required. The framework demonstrates a clear trade-off between intervention intensity and benefit, showing diminishing returns as more changes are introduced.
- Achieves 93.66% accuracy and a 0.9544 F1-score in predicting sleep quality from survey data.
- Uses SHAP explainability to identify key modifiable factors, then prescribes minimal changes via a mixed-integer optimization model.
- Generates highly personalized plans, often recommending only 1-2 behavioral adjustments to maximize impact while minimizing user burden.
Why It Matters
It bridges the gap between AI prediction and real-world action, offering a blueprint for personalized health coaching that is both effective and practical.