AI-assisted haematology triage for faster, safer clinical decision-making

The Triage Workflow That Made Clinical Decisions Harder

Haematology triage is the process of reviewing incoming referrals and deciding how urgently each patient needs specialist care. In the current workflow, nurses and clinicians manually review referral letters, lab values, scanned reports, and clinical notes before assigning a triage pathway.

The problem was not a lack of clinical data.

• The problem was that the data was scattered, unstructured, and cognitively heavy to process under time pressure.
• Nurses had to read long referral letters, cross-check abnormal lab results, judge urgency, and document decisions without a consistent interface designed for haematology triage. This created risk around missed patterns, inconsistent decisions, and delayed prioritisation.
• Triagen was designed as an AI-assisted clinical decision-support interface that helps clinicians quickly prioritise referrals, understand AI reasoning, and retain full control over the final decision.
• The system does not automate triage.
• It structures clinical information, suggests a triage pathway, explains the reasoning, and lets the clinician accept or modify the decision with documentation.

This was a high-stakes workflow design for clinical decision-making. When haematology referrals are difficult to scan, clinicians lose time, cognitive load increases, and urgent cases can be harder to identify. For conditions such as suspected leukaemia or malignant haematology cases, triage delays can directly affect patient care.

A Structured 12-Week Process, Sequenced To Build Evidence

The project began with an experience audit of the current site, followed by a heuristic evaluation to surface usability problems before any users were involved. I then ran a survey with 23 respondents to validate and expand on those issues, used the findings to build personas and a journey map, and moved into moderated usability testing to observe real planning behavior. Design decisions followed directly from that evidence.

Clinicians Did Not Need More Information. They Needed Clearer Decision Support.

The research showed that haematology triage was not failing because clinicians lacked data. It was failing because the available data was hard to interpret quickly. Referral information was often long, inconsistent, and unstructured. Lab values and scanned reports existed, but clinicians still had to manually connect the evidence to the final triage decision.

The strongest finding was simple:
AI support would only be useful if clinicians could verify it.

The biggest problems clustered around four areas.

Fixing The Workflow Before Refining The UI

The early wireframes focused on workflow structure, information hierarchy, and decision logic. The goal was to validate whether clinicians could move through the triage process without unnecessary context switching.

Clinical Feedback Made The Design Sharper

Think-aloud testing with three practising doctors helped evaluate whether the interface matched real clinical reasoning. The testing was especially useful because doctors could identify where the prototype looked clear from a UX perspective but still failed clinical expectations.

What experts told us :

Testing Signals That Changed The Design

Explainable AI. Clinician-Led Decisions.

Based on the research and testing, the design strategy focused on reducing cognitive load while preserving clinician authority. The interface needed to help clinicians understand urgency quickly without making the system feel like a black-box decision-maker.

Making AI-Assisted Triage Easier To Review, Trust, And Control.

The final desktop interface focused on helping triage nurses review referrals faster, understand AI recommendations, and make accountable clinical decisions. The design was not created to let AI decide independently. It was designed to make urgency, evidence, documentation, and final nurse-led decisions easier to manage in one connected workflow.

Case Dashboard: Clearer Referral Prioritisation

The dashboard gives triage nurses a structured view of all incoming cases. Cases are organised by triage decision, status, referral reason, assigned clinician, and completion level. This helps nurses identify urgent referrals faster instead of manually scanning disconnected case information.

Workspace Controls: Better Shift And Case Management

The workspace panel supports real clinical triage work beyond the case list. Nurses can control visible columns, manage alerts, pin doctors, set refresh preferences, and write shift handover notes. This makes the dashboard more useful during active clinical shifts.

Case Review: Evidence And AI Reasoning In One View

The case review screen connects patient details, referral notes, lab values, AI assessment, and triage reasoning in one place. The "Why This Decision?" section explains the clinical factors behind the recommendation instead of only showing a confidence score.

Accepted Decision State: Clear Nurse Accountability

After the nurse accepts the AI triage decision, the interface shows the decision status and who accepted it. This makes the final clinical action visible, documented, and accountable.

Modify Decision Flow: Controlled Override With Reason

When the nurse disagrees with the AI recommendation, the modify flow requires an alternative triage decision and a reason. This keeps clinician authority clear while making changes traceable.

Modified Decision State: Updated Triage Outcome

After modification, the case updates to show the final nurse-selected decision and documentation. The interface keeps the original AI reasoning visible while clearly showing the nurse's final decision.

This project taught me that AI in healthcare should not be designed to replace clinical judgment. It should reduce the effort required to review complex information while keeping the clinician fully in control of the final decision. The most important design challenge was creating trust through visibility. A confidence score alone was not enough. Clinicians needed to see the evidence, understand the reasoning, review the source information, and know exactly how to act on the recommendation. I also learned that safety-critical interfaces need restraint. More features do not automatically make the product better. The value came from simplifying the workflow, making urgent information easier to scan, and turning AI output into something clinically reviewable, explainable, and accountable.