This study, presented by Ahmed et al., aims to analyze the performance of multiple machine learning (ML) algorithms at detecting the presence of and severity level of depression using wearable-actigraphy data. Data was taken from the “Depresjon” database and Adaptive Synthetic Sampling (ADASYN) was applied to fix class imbalance. Feature engineering was then used to extract relevant features, and two types of models, binary models that identify a patient as either depressed or nondepressed and multiclass models that identify depression severity, were designed and implemented. Models were evaluated using accuracy, recall, specificity, precision, F1-score and MCC. The XGBoost model achieved the highest accuracies of 84.94% for the binary classification task and 85.91% for the multiclass classification task and was also first in every other metric across both classification tasks. The Logistic Regression (LR) model performed worst across all metrics in the binary classification task and was second worst in multiclass classification, where Support Vector Machines (SVM) performed worst. Explainable artificial intelligence (XAI) techniques were used on both the XGBoost model and the artificial neural network model to determine the features that had the most impact on classification. Power spectral density mean was identified as a key feature for differentiating depressed and nondepressed patients across both models. Other important features included demographic information such as age and gender and markers related to activity level and circadian rhythm. Classifications between mild and moderate depression also involved heightened irregularities in daily activities. These results highlight the connection between daily behavioral patterns, circadian rhythm irregularities and depression.