Bridging AI Scalability and Clinical Accountability: A Human-Centered Decision Support Architecture

This research proposes a reliable AI clinical decision support system that incorporates human-in-the-top-loop verification to guarantee dependability, safety, and accountability in high-stakes medical situations. The system uses modality-specific AI models, such as CNN-Transformer hybrids for physiological signals like ECGs and BioClinicalBERT for language, to interpret a variety of clinical reports, including radiography (X-ray, CT), electrocardiograms (ECG), pathology findings, and discharge summaries. A confidence score, highlights of the explainable information, and a first classification (such as "normal," "suspicious," or "critical") are all included in each AI-generated interpretation. To ensure that only validated insights impact patient care, the system carefully refers high-risk or doubtful instances to trained physicians, such as radiologists, cardiologists, or pathologists, for final assessment and approval rather than working on its own. A secure, interoperable tech stack with FastAPI-based microservices, a React clinician dashboard with FHIR integration for smooth EHR communication, and strong audit capabilities using PostgreSQL and the ELK stack forms the foundation of the system. In addition to improving diagnostic precision and workflow effectiveness, this design incorporates the fundamental tenets of reliable AI: accountability through thorough audit trails, openness through explainability, and fairness through bias monitoring. The approach strikes a compromise between the scalability of AI and the invaluable judgment of medical professionals by putting the human expert at the front of the decision loop—not as a passive overseer, but as an active validator—and ultimately promoting safer, more dependable, and morally good therapeutic outcomes.