AI-Augmented Approval Workflows: A Dual-Authority Framework for Clinical Decision Making

Abstract

High-stakes clinical decisions are increasingly influenced by artificial intelligence systems, yet no formal framework exists specifying when AI may approve or recommend medical interventions. Existing advisory-only mechanisms are prone to alert fatigue, with override rates exceeding 90% for warnings of drug-drug interactions in commercial systems. This paper addresses such challenges by proposing a dual-authority framework for establishing explicit approval responsibilities of both human clinicians and AI systems. The framework includes risk-based authority on delegation models, formal disagreement resolution protocols, and liability allocation mechanisms. A lightweight dual-signature protocol enables cryptographic verification of the parties' signoffs over crucial decisions. An integration with established EHR systems and sub-500ms response times allow the system architecture to work conveniently. The proposed evaluation targets sensitivity above 90% for safety-related decisions, reducing override rates below 20%. By codifying shared accountability, the framework meets regulatory requirements such as California’s 2024 Physicians Make Decisions Act while also facilitating meaningful AI inclusion in clinical decision-making.

Ähnliche Arbeiten