Sepsis is a leading cause of mortality and high healthcare costs in intensive care unit patients. Its pathophysiology is complex and clinical manifestations are highly diverse, so early identification and targeted interventions are essential to improve outcomes. With the proliferation of electronic health records and critical care data, the development of sepsis prediction models using machine learning and deep learning has become an active area of research. The review summarizes advances from traditional scoring systems such as SOFA and qSOFA to machine learning algorithms such as gradient boosting trees and random forests to deep learning time series models such as LSTM and Transformers. It includes data sources, feature engineering, methodologies, evaluation standards such as TRIPOD AI, and early warning, risk stratification, and organ dysfunction prediction models. It discusses challenges such as generalizability, fairness, model shift, workflow integration, alarm fatigue, and practical utility. It highlights opportunities in multimodal data fusion, causal inference, federated learning, and digital twins for a new generation of clinically applicable models. It recommends moving from algorithmic accuracy to clinical value through external validation, calibration, prospective evaluation, and clinical evaluation.