TLDR: MedLog is a proposed universal protocol for logging every interaction of AI models in healthcare, similar to how syslog works for computer systems. It aims to bring transparency and accountability to medical AI by recording nine core fields for each AI invocation, including model details, user identity, inputs, outputs, and outcomes. This standardized logging will enable continuous surveillance, real-time safety monitoring, bias detection, and iterative improvement of AI models, ultimately transforming how medical AI is evaluated and governed globally.
Artificial intelligence (AI) is rapidly integrating into healthcare globally, from assisting with diagnoses to streamlining administrative tasks. However, this swift adoption has highlighted a critical gap: the lack of a standardized way to record how, when, by whom, and for whom these AI models are used. Without this transparency, it’s challenging to measure real-world performance, detect adverse events, or correct biases and data shifts.
Introducing MedLog: A Universal Protocol for Medical AI
To address this, researchers have introduced MedLog, a new protocol designed for event-level logging of clinical AI interactions. Inspired by syslog, a universal protocol for recording critical events across computer systems, MedLog aims to bring similar transparency and accountability to medical AI. Essentially, every time an AI model is invoked—whether interacting with a human, another algorithm, or acting independently—a MedLog record is created.
The Nine Core Fields of a MedLog Record
Each MedLog record is structured around nine core fields, providing a comprehensive and consistent account of AI activity:
- Header: Contains essential information like timestamps, system identifiers, and a unique event ID.
- Model instance: Identifies the specific AI model and its version, linking to its model card and datasheet.
- User identity: Records who or what invoked the AI model, whether a clinician, patient, or another automated system.
- Target identity: Refers to the entity the model’s output is about, such as a patient ID or a claim ID.
- Inputs: Captures the data provided to the model, including prompts for generative AI or feature vectors for predictive models.
- Internal artifacts: Logs computational details generated during inference, like reasoning traces, retrieved context, or uncertainty estimates, primarily for technical audiences.
- Patient- or clinician-facing outputs: Records the actual outputs presented to human users, such as risk scores, generated text, or recommendations.
- Outcomes: Where possible, links the AI’s recommendation to downstream clinical actions or patient outcomes, even if observed with a delay.
- User feedback: Captures any feedback from users, from structured ratings to free-text comments, crucial for model refinement.
MedLog records are assembled incrementally, meaning information can be added over time as it becomes available, allowing for comprehensive tracking even in complex, multi-stage AI workflows.
Ensuring Privacy and Scalability
Implementing MedLog requires robust measures for patient privacy and data security, similar to those used for Electronic Health Record (EHR) systems. This includes strict access controls, pseudonymization, and compliance with regulations like HIPAA and GDPR. While capturing every AI interaction will generate substantial data, the protocol supports risk-based sampling and lifecycle-aware retention policies to manage storage efficiently. MedLog can be integrated using existing open standards like W3C PROV, OpenTelemetry, and FHIR, making it adaptable across different healthcare systems and vendors. For more details, you can refer to the full research paper here.
Transforming Medical AI with MedLog
The widespread adoption of MedLog promises to transform medical AI in several key ways:
- Human-AI Epidemiology: It will enable a new form of epidemiology that studies how AI influences decision-making, clinical actions, and patient outcomes, providing quantitative evidence for health policy.
- Real-time Safety Surveillance: Regulators and health systems can use MedLog records to detect adverse events, near misses, and model failures in real time, supporting continuous auditing and compliance.
- Detecting Dataset Shifts: By logging inputs and outputs, MedLog can identify when AI model behavior deviates from expectations due to changes in patient demographics, clinical practices, or technologies.
- Monitoring for Bias: It allows for systematic assessment of AI performance across different demographic groups, helping to identify and correct inequities.
- Improving AI Models: Real-world error cases and user feedback captured by MedLog can be used to refine and improve AI models, leading to more accurate and reliable tools.
- International Evaluation: Standardized logs can facilitate global benchmarking of AI models, assessing their generalizability across diverse populations and health systems.
- Advancing Transparency: MedLog creates traceable documentation of AI-generated content, allowing clinicians to review rationales and patients to audit information available to their personal health AI models.
A case study from Clalit Health Services demonstrated MedLog’s potential by detecting a subtle but significant data drift in a hospitalization risk prediction model, caused by a change in a laboratory testing kit. This early detection prevented potential prediction errors that could have impacted patient care.
Also Read:
- Agentic-AI Healthcare: A New Framework for Secure, Multilingual, and Explainable Digital Health
- MedPAO: Streamlining Medical Report Structuring with AI Agents
A Call to Action
The researchers emphasize that systematic monitoring of medical AI interactions is urgently needed to ensure safety, fairness, and effectiveness. MedLog provides the essential framework for this, much like syslog became indispensable for modern computing. Policymakers, regulators, healthcare leaders, and the AI community are called upon to adopt these standards to realize the full promise of AI in medicine while safeguarding patient outcomes and public trust.
