NBA Sleep Tracking Data Imputation

Abstract

This thesis investigates imputation methods for nights of missing sleep wearable data from NBA Academy athletes. Sparsity in sleep tracking data arises as a result of behavioral non-compliance or device malfunction, hindering the NBA Academy's ability to provide actionable insights that improve player sleep, a crucial component for player development. Motivated by existing work on time series data imputation, four main techniques are evaluated: K-Nearest Neighbors Regression, Linear Interpolation, Linear Regression, and Quadratic Regression. Each technique is applied and evaluated on key sleep metrics such as sleep duration, rMSSD (Root Mean Square of the Successive Differences between Heartbeats), and average heart rate. Results indicate K-Nearest Neighbors Regression and Linear Interpolation, with access to data in the past and future (offline imputation), as the best-performing sleep imputation methods. Furthermore, this thesis utilizes the NBA Academy's shooting and jumping datasets in conjunction with the sleep dataset to explore a relationship between sleep and athletic performance, finding a generally weak correlation between sleep and athletic performance data, regardless of the time lag. This research has applications in all areas of sport and performance as well as in domains where data sparsity is problematic.