AI-Assisted Radiology Decision Support

Status: Research design & prototyping · Focus: diagnostic safety, assistive AI, and clinical workflow

Repo structure includes: agent_design, causal_analysis, experiment_proposal, docs

The problem

Clinicians and radiologists make time-pressured, high-stakes decisions that directly affect patient outcomes. AI support tools promise to “help,” but right now we don’t have strong evidence about how they actually change human decision-making under real workflow conditions.

The core questions are not just “Is the model accurate?” but:

  • Does AI support make the final diagnosis more correct?
  • Does AI support reduce dangerous errors (for example, missing pneumonia)?
  • Does AI support reduce false alarms that lead to unnecessary escalation / treatment?
  • Does AI support change clinician behaviour in a risky way — like over-trusting a suggestion?
  • Who benefits more from AI help: junior readers or experienced readers?

What I designed

I designed an experimental framework to evaluate AI-assist systems in chest X-ray interpretation, focusing on suspected pneumonia. The goal is to measure not only model performance, but clinician + AI team performance.

  • Multi-reader, multi-case design (MRMC): Multiple clinicians read the same curated image set. This lets us compare performance across different readers, not just across different images.
  • Multiple study arms: Each reader is randomly assigned different levels of AI support. For example:
    • Arm A: image only (baseline, no AI).
    • Arm B: AI suggestion (“pneumonia: yes/no”).
    • Arm C: AI suggestion + AI confidence score (“pneumonia likely: 92%”).
  • Behavioural capture: For each case, the clinician not only records the diagnosis but also:
    • their confidence,
    • their read time,
    • whether they would escalate / flag for further review.

This is important because we can measure not just “Did the AI help?” but “Did the AI change behaviour safely?”

Core modules in this project

agent_design/

Designs the “AI assistant” behaviour: how suggestions are surfaced to clinicians, how confidence is communicated, and when intervention happens in the workflow. This matters because poor UX can create blind trust or alert fatigue.

causal_analysis/

Investigates counterfactual questions like “Would this decision have been different without AI?” and fairness questions like “Does AI affect escalation decisions differently across subgroups?” This is where safety, bias, and liability live.

experiment_proposal/

Study protocol drafts, inclusion/exclusion criteria for image sets, reader assignment logic, and instructions to clinicians. This is essentially the “trial design” layer for a human+AI evaluation.

docs/

Written analysis and Markdown reports: methodology notes, interpretation of expected results, governance framing (how do we certify or approve these tools responsibly?), and communication to non-technical stakeholders.

Why this matters

This is more than “AI for healthcare.” It’s about accountability in augmented decision-making. Hospitals, regulators, and vendors all need to answer the same question: Under what conditions does AI support actually improve clinical care, and under what conditions does it just shift responsibility?

This project provides an auditable way to measure that. The work is designed so it can feed both:

  • Clinical leadership (Is this safe to deploy to staff?)
  • Policy / governance bodies (Does this align with responsible AI and human oversight claims?)

My role

I architected the evaluation approach, defined the study arms, scoped the behavioural metrics, and wrote the protocol language to make it usable not only for technical review but also for governance and decision-making. I positioned this project at the intersection of ML performance, clinical workflow reality, and accountability.

Access

The full repository (code, study design, and documentation) is currently private. I’m open to sharing details or collaborating on this work in research, evaluation, or policy contexts.

To discuss this project, reach out via LinkedIn, GitHub, or Telegram using the links in the footer.