Robust sleep prediction under missing wearable sensor data

Background

Consumer wearables record heart rate, movement and skin temperature through the night, and sleep quality can be estimated from them. In practice the data is incomplete: a strap loosens, a battery dies, a sensor is covered. Most published models are evaluated on clean data, so their behaviour under realistic gaps is poorly understood.

Research question

When part of a wearable device's data is missing, how does the sleep-prediction performance of lightweight machine learning models change, and at what cost in compute?

Method

Standard track. Using a public dataset with a research license, compare logistic regression, random forest and gradient-boosted trees on the full data with a fixed train/test split and a single agreed metric. Produce a results table and an error analysis.

Advanced track. Simulate missing sensor data at increasing proportions and across different sensors. Re-evaluate each model, add a simple imputation baseline, and report accuracy, robustness (the slope of performance against missingness) and training and inference cost. Ablate the imputation method and the feature set. Compare against a published result reproduced in stage nine.

Findings

Not applicable for a sample project. A real project would report the results table, the robustness curves with confidence intervals, and the compute measurements.

What the student would learn

Train/test discipline, baselines, controlled experiments, ablation, error analysis, and writing a methods section that someone else could run.