Actigraphy is a non-invasive method for assessing potential sleep disorders that uses a wrist-worn device to record light exposure and physical activity. Data collection challenges include off-wrist periods when the device is not worn. This study applies machine learning (ML) and deep learning (DL) techniques to address these gaps in actigraphy data from 23 patients with varying activity patterns. A Random Forest classifier detected OW periods and excluded days with significant gaps. We tested nine models to estimate missing values for the activity measure Proportional Integrated Mode (PIM), with each model retrained to fit individual patient activity patterns. The most effective models varied by each patient, with Random Forest Regressor, XGBoost, Decision Tree, and Linear Regression emerging as top performers. Analysis of both original and reconstructed datasets using the nparACT library from R showed minimal or consistent changes in actigraphy metrics post-reconstruction, highlighting the effectiveness of ML techniques in managing missing data in actigraphy studies.