Published 1 day ago

USC Launches Open-Source Medical AI Framework to Trace Data Errors and Improve Reliability

Medical AI systems fail. The harder problem is figuring out why—and doing it without spending weeks on manual investigation that produces results no one else can reproduce.

That is the gap a new NSF-funded project from USC’s Viterbi School of Engineering is designed to close. Led by Ruishan Liu, a WiSE Gabilan assistant professor with joint appointments across computer science, biomedical data science, and radiation oncology, the three-year initiative received nearly $900,000 in funding and launched in August 2026. The result will be an open-source framework that traces AI model failures back to their root causes in the data pipeline—automatically, systematically, and with a human expert in the loop.

4 mins read

6 sections

Key Highlights

USC debuts an open-source framework to trace medical AI failures to specific data errors
Provenance tracking and LLM reasoning turn AI failure analysis into a reproducible pipeline
Human-in-the-loop curation keeps clinical judgment central in high-stakes AI remediation

The Core Problem: Bad Data, No Clear Trail

AI models in clinical settings are only as reliable as the data used to train them. When a model underperforms, the current standard response is slow, manual, and largely non-reproducible investigation. There is no established automated method to connect a model’s poor output to a specific upstream data error—whether that error originated in a scanner protocol, a preprocessing step, or a labeling inconsistency.

Medical data compounds this challenge. Unlike general-purpose datasets, clinical data is expensive and ethically constrained to collect. It requires patient consent and significant clinical expertise. Simply discarding problematic samples is rarely viable. The framework needs to fix data problems, not just flag them.

Three Components, One Closed-Loop System

The framework is structured around three interconnected components.

Algorithmic foundations handle attribution, diagnosis, and intervention. This layer identifies which data caused a model failure, explains why, and recommends the most cost-effective repair. It uses reasoning-based large language models to generate root-cause hypotheses—replacing ad hoc manual troubleshooting with a structured, automated pipeline.

Provenance infrastructure acts as a traceability layer. Every data sample carries a logged history of its transformations, queryable through efficient probabilistic structures such as Bloom filters. This converts the AI “black box” into a transparent system where failures can be traced to specific origins—a particular scanner protocol, a processing step, or a human annotation error.

The Integrated Curation Agent orchestrates both layers into a closed-loop workflow. Critically, it includes a human-in-the-loop feature that routes high-stakes remediation decisions to medical experts for review and approval. This is not a fully automated override system—it is designed to keep clinical judgment in the decision chain where it matters most.

Why Radiation Oncology as the Test Bed

The primary application domain is radiation oncology, a field that depends on complex, multimodal data: CT and MRI imaging, radiation treatment plans, and longitudinal electronic health records. It is a demanding test environment precisely because the data is heterogeneous, the stakes are high, and errors are often context-dependent in ways that generic tools cannot detect.

Liu and her collaborators—Virginia Tech researchers Ruoxi Jia and Wenjie Xiong, whose expertise spans algorithms and computer systems—have designed the framework to generalize beyond oncology. The methodology is intended to apply across medical specialties and, potentially, to other high-stakes domains where data integrity directly affects outcomes in areas such as medical imaging.

What Open Source Changes Here

The decision to release the framework as open source is not incidental. It means clinical data scientists and medical researchers can adopt a standardized, replicable methodology for data curation rather than building bespoke troubleshooting processes from scratch. Standardization matters in healthcare AI: reproducibility and accountability are prerequisites for institutional trust, not optional features.

The framework also addresses a structural inefficiency. Manual data troubleshooting is not just slow—it is costly in a domain where clinician and data scientist time is a constrained resource. Automating root-cause analysis with expert review checkpoints could meaningfully reduce the overhead of maintaining AI systems in production clinical environments.

The Practical Takeaway

For anyone building, deploying, or evaluating medical AI tools, this framework points toward a more defensible architecture: one where model failures are traceable, data histories are queryable, and human experts retain meaningful oversight over high-stakes decisions.

The open-source release will make it possible to assess the framework against real-world clinical workflows rather than controlled benchmarks alone. That is where reliability claims in medical AI either hold up—or don’t.