AI-assisted interpretation of laboratory results

01 / Research focus

Research questions

• Can analysis across multiple biomarkers help prioritise results for pathologist review?
• Can interpretable flags identify cases where additional markers or parameters merit consideration?
• How can machine-learning models analyse relationships between biomarkers while making their outputs understandable to pathologists?
• How can these methods be evaluated for clinical relevance, missed findings, and unnecessary follow-up suggestions?

02 / Methodology

Research approach

The research approach involves defining clinically meaningful review criteria with pathologists, followed by preparing appropriately governed laboratory datasets, developing computational methods, and evaluating their outputs against expert assessment.

The programme focuses on analysing existing laboratory result data. Development of new blood-testing equipment or assays is outside this research scope.

Intended workflow

1. Laboratory results

   Prepare biomarker results across batches of samples for research analysis.

2. Data-quality checks

   Check completeness, measurement units, and laboratory reference ranges before interpreting patterns.

3. Biomarker-pattern analysis

   Investigate statistical and machine-learning methods for identifying combinations of results that may warrant closer review.

4. Flags with supporting rationale

   Present the results underlying each flag and any suggested additional markers or parameters for consideration.

5. Pathologist review

   Have a pathologist assess the flag in the available clinical context and determine whether further investigation is appropriate.

Research governance

Use of patient-derived laboratory data requires appropriate data-access permissions, privacy safeguards, and any required ethics or institutional approvals. Study protocols must define data handling, access controls, and the use of de-identified data where appropriate.

03 / Computer science & AI

From laboratory data to interpretable healthcare AI

The computer science component focuses on how laboratory results are represented, processed, and analysed at scale. The AI component examines whether models can identify useful relationships across biomarkers and present findings in a form that pathologists can assess. The following areas define the research agenda.

Intended research outputs

The work aims to produce reproducible analysis pipelines, comparative model evaluations, and a prototype interface for pathologist review. These are research objectives; no deployed clinical AI service or validated diagnostic capability is claimed.

04 / Evaluation

Evaluation priorities

Evaluation will examine agreement with expert review, missed findings, false alerts, and the relevance of suggested follow-up investigations. Further evaluation is intended to assess performance across patient groups and laboratory settings.

Model evaluation will use separate development and test datasets, with records from the same patient kept within a single partition to reduce data leakage. Planned comparisons include rule-based baselines, sensitivity to missed review-worthy findings, false-alert burden, and performance across relevant patient groups. Evaluation of the review workflow will consider how pathologists use the outputs, alongside model performance.

Methodological reference: Good Machine Learning Practice guiding principles ↗

The research programme is in progress. These evaluation priorities guide the work; completed validation results and demonstrated clinical benefits are not reported on this page.

05 / Research leadership

Principal investigator

06 / Collaboration

Research collaboration

We welcome conversations with pathologists, laboratory medicine researchers, computer scientists, machine-learning researchers, and health-data scientists interested in shaping clinically relevant research questions and evaluation methods.