Questions and studies across medical AI.
We investigate questions across specialist education, clinical workflows, medical imaging, forecasting and physiology. For hospitals, research institutes, medical societies and investment partners, these tracks explain the work underway and the evidence needed to guide its next stage.
The current statuses (Completed, Ongoing, and Validation planning) describe research progress. They do not imply product availability, deployment, approval for clinical use, or readiness for routine care.
Explore the question, the projects and opportunities to work together.
For a clinician learning nephrology, a useful explanation needs a clear basis in medical evidence. This track examines how clinical language models draw on training data, retrieved references and external tools to support specialist education.
Project status: Validation planning
NephroGPT explores a language model that uses external tools to support nephrology training and continuing education, with clinical reasoning grounded in evidence. Validation is being planned.
NephroGPT: A Tool-Augmented Large Language Model for Nephrology Training and Continuing Education
Evaluation should establish whether answers are clinically correct and whether the evidence supports them. Specialist review should also examine how errors are handled and whether the model's expressed confidence matches its reliability.
We welcome discussion with nephrology departments and medical societies about the educational setting and the questions worth testing. Together, we could scope the specialist review needed to assess the answers.
Cardiology decisions can depend on information in different forms and on the tools used to interpret it. This track investigates how an AI agent could bring that context together and make its steps visible to clinical reviewers.
Project status: Ongoing
This ongoing project investigates an AI agent that combines clinical information in different forms with external tools to support cardiology decision workflows.
An Evidence-Grounded Multimodal Agentic AI Co-Pilot for Cardiologists
Review should examine the evidence available when the agent makes a plan or uses a tool, then follow the intermediate steps behind its response. It should also establish where the workflow can fail and whether those limits are apparent to a cardiologist.
We invite cardiology departments to discuss a defined decision workflow and the practical constraints on a research evaluation. That conversation could clarify which information and tools the task needs, and how clinical reviewers would inspect the agent's work.
Public health leaders need to understand changes in disease trends in relation to when interventions took place. This track studies prediction and intervention effects over time, including the completed analysis of Polish infectious disease notifications from 2014 to 2022.
Project status: Completed
The completed study examined how pandemic-era interventions affected trends in notifiable infectious diseases in Poland from 2014 to 2022.
The Impact of COVID-19 Non-Pharmaceutical Interventions on Notifiable Infectious Diseases in Poland
Intervention analysis should make the timing of measures explicit and examine whether conclusions hold across observation periods. For forecasting, evaluation should use later data withheld from model development and check whether predicted risks match observed outcomes as patterns change. Both require methods that other researchers can reproduce.
We welcome discussion with public health institutes about the policy question an analysis needs to answer and which surveillance records could support it. Data access, governance and a reproducible analysis plan would be useful starting points.
Chest imaging research connects what appears in an X-ray with how it is described in a report and what it may mean for patient outcomes. The three projects in this track examine automated report labeling, a model for chest X-ray understanding, and predicted age as a possible marker of mortality risk.
Project status: Completed
A completed benchmark of language models for assigning structured CheXpert-style labels to chest X-ray reports, examining reliability in report labeling workflows.
Benchmarking Large Language Models for Automated CheXpert-Style Labeling of Chest X-Ray Reports
Project status: Completed
A completed project on a multimodal language model developed through additional training for chest X-ray understanding and radiology report interpretation.
ChexQwen: A Post-Trained Multimodal LLM as a Foundation CXR Model
Project status: Completed
A completed study of age predicted from chest X-rays as an imaging-derived marker in relation to mortality.
Chest X-Ray-Predicted Age as a Biomarker for Mortality
Evaluation should check report labels against the findings described in the source report and image models against findings visible in the X-ray. Review should examine performance across patient groups and whether confidence matches correctness. For predicted age, the question is whether its relationship to mortality supports a meaningful biomarker.
We invite radiology departments and research institutes to discuss which imaging question matters in their setting and what data could support an evaluation. Useful scoping would establish access to images and reports under appropriate governance, along with the clinical review needed to interpret the results.
Electrical recordings and ultrasound offer different views of the heart. This track investigates whether models can connect those observations within a representation of a patient's changing physiologic state, including generated images whose uncertainty is made explicit.
Project status: Ongoing
EchoWorld is an ongoing study of generating nine echocardiography (heart ultrasound) views from an electrocardiogram (ECG), which records the heart's electrical activity, and clinical context. Its outputs are intended to express uncertainty.
EchoWorld: Generative Echocardiography from ECG and Clinical Context
Evaluation should compare generated views with clinical observations withheld from model development and examine whether they agree across views and over time. It should test whether the model reflects changes in patient state or interventions, and whether its stated uncertainty matches the errors observed.
We welcome a scoping discussion with cardiology and imaging teams about a research setting with paired electrical recordings and ultrasound examinations. The starting points would be data governance and how clinical reviewers could judge whether generated views fit the recorded physiology and communicate uncertainty.