AI Trustworthiness

Overview

Artificial intelligence (AI) holds tremendous potential for transforming healthcare — from automated diagnostics to precision treatment planning. However, as these models move from research settings to clinical environments, the question of trust becomes paramount. Can we trust AI to perform equitably across diverse populations, institutions, and imaging protocols? Can we quantify and mitigate bias without sacrificing clinical accuracy?

The AI Trustworthiness vertical of the BioIntelligence Lab aims to answer these questions through a multi-pronged research program that combines empirical auditing, conceptual analysis, and algorithmic innovation. Our work spans the full pipeline of medical AI — auditing deployed systems, defining and quantifying bias, identifying hidden vulnerabilities, and building generative solutions for fairness. This research directly informs the safe, equitable, and accountable use of AI in medicine.

1. Audits: Evaluating AI in the Wild

We began by asking a fundamental question: Do existing AI systems perform fairly once deployed? To answer this, we conducted some of the first large-scale fairness audits of medical AI systems across multiple modalities — including CT, X-ray, and MRI — and even natural language processing models used in radiology reporting. These studies reveal that while models may achieve high overall accuracy, they often exhibit hidden performance gaps for specific demographic groups, such as age, sex, or insurance status.

By exposing these disparities and establishing benchmark audit protocols, our work laid the foundation for bias assessment in clinical AI and influenced ongoing regulatory discussions around ethical AI deployment.

2. Understanding Bias: Definitions, Context, and Representation

Before we can mitigate bias, we must first understand what it is. In this sub-vertical, we formalize the definitions, measurement strategies, and conceptual underpinnings of bias in medical AI. Our work has demonstrated that many fairness failures arise not from overt discrimination but from representation leakage — where demographic traits become encoded in latent feature spaces without explicit labels.

We also highlight the limitations of current bias reporting practices, showing that coarse demographic labels and incomplete dataset documentation can mask true disparities. By rethinking how we define, measure, and report bias, we aim to create a shared scientific foundation for equitable AI development.

3. Security: Hidden Vulnerabilities in AI Systems

Fairness is not only an ethical concern — it is also a security risk. Our work has revealed that fairness vulnerabilities can be exploited through undetectable adversarial attacks that disproportionately target underrepresented groups. These findings expose a new category of risk: adversarial bias attacks, where imperceptible perturbations can systematically degrade model performance for specific populations while leaving global metrics unchanged.

By combining adversarial machine learning and fairness analysis, this research establishes a new frontier at the intersection of AI security and ethics.

4. Solutions: Generative and Adversarial Bias Mitigation

Finally, our research turns toward solutions — methods that can make AI models more equitable and trustworthy without compromising diagnostic accuracy. We introduce Generative Counterfactual Augmentation (GCA), a novel approach that generates realistic, demographically balanced training examples using generative modeling. Unlike adversarial debiasing, which suppresses demographic information and may harm model utility, counterfactual augmentation acts as a regularizer — improving representation balance while maintaining clinical fidelity.

Building on this, we also explore adversarial representation alignment in 3D CT foundation embeddings, offering a pathway toward fairness-aware foundation models that generalize across populations and institutions.

Impact

The AI Trustworthiness program redefines fairness evaluation and mitigation in medical imaging — from real-world audits to representation-level solutions. Our work has established both theoretical and practical frameworks for the ethical deployment of AI systems in healthcare, influencing radiology practice guidelines and ongoing federal initiatives for safe AI in medicine.