The rapid growth of machine learning (ML) applications in healthcare promises transformative advancements in diagnostics, prognosis, and treatment decisions. However, translating these models into clinical practice faces significant hurdles. This research addresses the critical need for reliable uncertainty quantification (UQ) in clinical ML, emphasizing the importance of conveying uncertainty to clinicians. Current approaches often provide only the "most likely" recommendation and prioritize population-level accuracy metrics, neglecting the unique nature of each patient and overlooking uncertainty related to model predictions. This research aims to solve these issues by achieving three main objectives. Objective 1 focuses on designing conditionally valid and efficient UQ approaches, addressing challenges related to integrating UQ into clinical ML models, and establishing reliable confidence estimates. Objective 2 explores the fusion of UQs within Hybrid ML models, aiming for more adaptive and efficient UQ in a low data regime. Objective 3 tackles the disentanglement of aleatoric and epistemic uncertainties, acknowledging their distinct sources and characteristics. By addressing these objectives, the research aims to contribute to developing trustworthy ML systems for personalized diagnosis, prognosis, and treatment in healthcare. Ultimately, this work seeks to enhance safety and reliability, accelerating the integration of ML research into clinical practice.