Depression is characterized by an insidious onset and marked symptom heterogeneity. Current diagnostic practice still mainly relies predominantly on clinical interviews and rating scales, which are constrained by substantial subjectivity, limited capacity for early identification, and insufficient performance in subtype differentiation and prognostic evaluation. With the growing availability of objective data, including neuroimaging, electroencephalography, speech and digital behavioral features, clinical scales, and multi-omics profiles, machine learning has opened new avenues for the objective identification of depression. Focusing on this central issue, the present review summarizes recent advances in the application of machine learning to neuroimaging and other unimodal objective data, further discusses the integrative value of multimodal data in the diagnosis, subtyping, disease-course assessment, and treatment-response prediction of depression, and provides a systematic overview of common fusion strategies, model interpretability, and issues related to clinical translation.