# Auditing and Monitoring Artificial Intelligence Systems in Healthcare: A Multilayer Framework for Bias Detection, Explainability, and Regulatory Compliance

## TL;DR

This paper introduces a comprehensive framework for auditing and monitoring AI systems in healthcare to ensure fairness, transparency, and compliance throughout their lifecycle.

## Contribution

The paper proposes a multilayer framework that integrates bias detection, explainability, performance monitoring, and regulatory compliance into a continuous governance process.

## Key findings

- The framework combines four operational layers to address bias, transparency, performance, and compliance in healthcare AI.
- It emphasizes continuous monitoring and institutional decision-making pathways for accountability.
- The model supports adaptation across regulatory environments and prioritizes fairness metrics based on clinical risk.

## Abstract

Artificial intelligence (AI) is increasingly embedded in clinical decision-making, yet most oversight approaches remain limited to pre-deployment validation or isolated technical evaluation. This gap creates risks related to bias, safety, accountability, and regulatory compliance once systems operate in real clinical environments. This article presents a normative, lifecycle-oriented auditing and monitoring framework for healthcare AI derived from a structured synthesis of literature on trustworthy AI, clinical risk management, and governance practice. The framework integrates four operational layers: (1) bias detection and fairness assessment; (2) explainability and model transparency; (3) performance, safety, and drift monitoring; and (4) regulatory and ethical compliance. Unlike prior models that treat technical validation and governance oversight separately, the proposed approach links continuous monitoring outputs to institutional decision authorities through predefined escalation pathways and role-based responsibilities across developers, clinicians, and governance bodies. The framework is designed for practical use by healthcare institutions, regulators, and AI developers. It provides guidance on monitoring frequency, the prioritization of fairness metrics based on clinical risk, the evaluation of clinically meaningful explanations, and adaptation across regulatory environments. By operationalizing auditing as an ongoing governance process rather than a one-time certification activity, the model supports the accountable and trustworthy deployment of AI systems throughout their real-world lifecycle. This work offers a structured foundation for aligning technical monitoring with clinical governance and regulatory expectations in healthcare AI implementation.