As artificial intelligence (AI) increasingly supports screening, diagnosis, and surgical planning, the gap between high predictive performance and clinical utility remains a critical challenge. This dissertation proposes a Responsible AI framework for clinical decision support systems, arguing that reliability, transparency, and actionability are not secondary features but core design requirements. The research is structured across three primary aims, beginning with an investigation into reliability in single-modality systems. Using audio-based respiratory screening, this work demonstrates how integrating confidence calibration and predictive uncertainty into AI pipelines ensures models communicate the limits of their own knowledge. The second phase extends this framework to multi-modal clinical data by investigating how representation choice and data fusion affect model robustness in sleep disorder assessment. By utilizing concept bottleneck models and knowledge-informed representations, this work aligns AI reasoning with human-meaningful clinical concepts to increase transparency. Finally, the research applies these responsible AI methods to surgical planning via an image-guided pipeline for 3D airway analysis in an orthognathic surgery population. This automated segmentation tool provides standardized, patient-specific metrics to evaluate postoperative remodeling at scale. Collectively, this research demonstrates that for AI to be safely integrated into healthcare, systems must move beyond "black-box" predictions toward architectures that express uncertainty and provide interpretable evidence. The overarching contribution is a cross-modal methodology that bridges the gap between AI development and clinical trust.