AI and Machine Learning in NHS Pathology and Diagnostics
Artificial intelligence (AI) is moving from conference slides into the working laboratory. Across NHS pathology, machine learning (ML) tools are being trialled and, in a growing number of services, deployed to help detect features on digital slides, prioritise workloads and flag abnormal results. For biomedical scientists (BMS), this is not an abstract future. Whether you work in cellular pathology, blood sciences, microbiology or quality management, you will increasingly be expected to understand how these tools work, where they fit in a regulated diagnostic pathway, and crucially what your professional responsibilities are when AI is involved.
This guide takes a measured, evidence-based view. It avoids the hype, sets out the current 2026 UK regulatory position, and focuses on what you genuinely need to understand to engage confidently and safely with AI in diagnostics.
What We Mean by AI and Machine Learning
The terms are used loosely, so it helps to be precise. Artificial intelligence is a broad label for computer systems that perform tasks normally requiring human judgement. Machine learning is a subset of AI in which a model learns patterns from data rather than following explicitly programmed rules. Most diagnostic AI in pathology today is built on ML, often using a type of model called a deep neural network for image analysis.
A few distinctions matter in practice:
- Locked versus adaptive models. A locked model behaves identically every time once trained. An adaptive (continuously learning) model can change its behaviour after deployment. Adaptive models raise harder regulatory questions because their performance can drift.
- Assistive versus autonomous. Almost all clinically deployed pathology AI is assistive: it supports a qualified professional who remains responsible for the result. Autonomous diagnosis without human sign-off is not the current NHS model.
- Narrow tools. These systems are narrow specialists. A tool validated to detect prostate cancer on biopsy slides tells you nothing reliable about breast tissue or about a different scanner or stain.
Where AI Is Being Applied in NHS Pathology
AI is not a single product but a set of tools addressing specific tasks. Current and emerging NHS applications include:
| Domain | Example application | Typical role | |--------|---------------------|--------------| | Cellular pathology | Tumour detection and grading on whole slide images | Assistive screening, second read, workload triage | | Cellular pathology | Mitotic counting, biomarker quantification (for example Ki-67) | Standardising measurement | | Blood sciences | Flagging and pattern recognition in results and morphology | Decision support, result prioritisation | | Haematology | Digital blood film pre-classification of cells | First-pass triage before BMS review | | Microbiology | Plate reading and growth detection | Screening negative plates, prioritising positives | | Workflow | Case prioritisation and worklist ordering | Operational efficiency, not diagnosis |
The most visible progress is in digital cellular pathology, which depends on whole slide imaging (WSI) — scanning glass slides into high-resolution digital files that software can analyse. Digital pathology is the enabling foundation: without digitised slides, image-analysis AI has nothing to work on. NHS adoption is expanding through coordinated networks such as the National Pathology Imaging Co-operative (NPIC), led by Leeds Teaching Hospitals NHS Trust, although uptake across the whole service remains uneven and is still maturing.
It is important to be realistic. Many tools are at pilot, evaluation or limited-deployment stage rather than in routine front-line use everywhere. Adoption depends heavily on local digital infrastructure, funding and validation capacity.
How an AI Tool Is Built and Where It Can Go Wrong
Understanding the lifecycle helps you ask the right questions. In simplified terms:
1. Data collection. The model is trained on a large set of labelled examples, for example slides annotated by pathologists. 2. Training. The model learns statistical patterns linking inputs to the labelled outputs. 3. Internal testing. Performance is measured on data the model has not seen. 4. External validation. Performance is tested on data from different sites, scanners, populations and laboratory methods. 5. Clinical evaluation. The tool is assessed in a realistic clinical pathway against defined safety and performance criteria. 6. Deployment and monitoring. Real-world performance is tracked over time.
The failure points BMS should appreciate:
- Bias in training data. If the data under-represents certain populations, demographics or disease subtypes, performance for those groups may be poorer.
- Distribution shift. A model trained on one laboratory's stains, scanners and protocols may underperform on yours. This is a central reason local validation matters.
- Overfitting. A model can look excellent on its own test data but generalise poorly.
- Performance drift. Reagents, instruments, populations and protocols change over time; performance can degrade silently.
- Automation bias. Humans tend to over-trust confident-looking machine output. This is a human-factors risk, not a software fault, and it sits squarely within professional practice.
The 2026 UK Regulatory Position
This is the area where memory is unreliable, because the framework is actively changing. The current UK position can be summarised as follows.
AI used for diagnosis or to inform clinical decisions is generally regulated as a medical device. Software that has a medical purpose — including in vitro diagnostic (IVD) software and AI as a Medical Device (AIaMD) — falls under medical device regulation overseen by the Medicines and Healthcare products Regulatory Agency (MHRA).
Conformity marking. To be placed on the Great Britain market, a device must carry an appropriate mark of conformity. The UKCA (UK Conformity Assessed) mark is the Great Britain route. Under transitional arrangements, CE-marked devices meeting EU requirements have continued to be accepted in Great Britain, and in early 2026 the MHRA consulted on longer-term recognition of CE-marked devices. Higher-risk devices require assessment by an independent UK Approved Body; lower-risk devices may use self-assessment with a quality management system (commonly ISO 13485).
The framework is being reformed. The MHRA has published a phased roadmap. Strengthened post-market surveillance requirements came into force in 2025, and further changes covering classification, conformity assessment and international reliance pathways are expected to take effect during 2026. A dedicated regulatory framework specifically for AI as a medical device is anticipated in 2026, informed by the MHRA's wider Software and AI as a Medical Device change programme.
The AI Airlock. The MHRA runs the AI Airlock, described as the first regulatory sandbox for AI as a medical device. Its pilot closed in 2025 and a further phase is running into 2026. It lets manufacturers and regulators work through novel challenges — such as explainability, model change and post-market monitoring — in a controlled setting, with findings feeding the emerging framework.
Because the detail is in flux, the safe professional habit is to check the current MHRA guidance and your trust's procurement and governance position rather than rely on a remembered status.
What Governance, Validation and Oversight Look Like in Practice
Regulatory approval of a product is necessary but not sufficient. A tool that is lawfully marketed still has to be safely adopted in your laboratory. The professional consensus, reflected in Royal College of Pathologists (RCPath) guidance on digital pathology and in the joint RCPath and Royal College of Radiologists position on AI in the NHS, emphasises careful, locally validated, human-supervised adoption rather than autonomous deployment.
Key elements of responsible local adoption:
- Local validation. Confirm the tool performs acceptably on your own samples, scanners, stains and patient population before relying on it. Vendor performance figures from elsewhere do not transfer automatically.
- Defined intended use. Use the tool only for the task and sample type it was validated for. Off-label use shifts risk onto the laboratory.
- Human oversight. A qualified professional remains accountable for the diagnostic result. Assistive AI informs that decision; it does not replace the decision-maker.
- Standard operating procedures. Document how the tool is used, what to do when it disagrees with the human, and how failures are escalated.
- Quality management and audit. Bring AI into existing ISO 15189:2022 quality systems — verification, ongoing performance monitoring, internal quality control and audit.
- Incident reporting. Treat AI-related errors and near-misses like any other device issue, including MHRA reporting where appropriate.
- Transparency and explainability. Tools that can show why they reached a conclusion support clinician confidence and safer oversight.
- Information governance. Patient images and data are personal data; data protection and clinical governance still apply.
What This Means for Your Role as a Biomedical Scientist
AI changes the nature of laboratory work rather than removing the need for it. The Health and Care Professions Council (HCPC) standards and Institute of Biomedical Science (IBMS) professional expectations of accountability, competence and patient safety apply just as fully when an algorithm is in the loop. Practical implications for BMS:
- You may become a validator and monitor. BMS are well placed to run and document local verification, track ongoing performance and spot drift.
- Sample and pre-analytical quality matters more, not less. Poor scanning, staining or specimen quality degrades AI performance. Your technical standards directly affect algorithm reliability.
- You must guard against automation bias. Maintain critical judgement; a confident output is not a correct output.
- New specialist roles are emerging. Digital pathology coordination, AI validation and laboratory informatics are growing career directions.
- Keep learning. Engage with IBMS and RCPath resources, and follow current MHRA guidance as the framework develops.
Frequently Asked Questions
Is AI replacing biomedical scientists and pathologists?
No. The deployed model across NHS pathology is assistive, not autonomous. AI supports detection, triage and measurement, but a qualified professional remains accountable for the diagnostic result. The clearer effect is changing roles and creating new specialist responsibilities, particularly around digital pathology and validation.
Is diagnostic AI regulated in the UK?
Yes. Software with a medical purpose, including AI as a medical device, is regulated by the MHRA as a medical device and must carry an appropriate conformity mark such as UKCA to be placed on the Great Britain market. The framework is being reformed through 2025 and 2026, with a dedicated AI-as-a-medical-device framework anticipated. Always check current MHRA guidance.
Why does a laboratory need to validate a tool that is already approved?
Regulatory marking shows a product met defined requirements in its assessed context. It does not guarantee equivalent performance on your specific scanners, stains, protocols and patient population. Local validation confirms the tool performs acceptably in your service before you rely on it, which is consistent with ISO 15189 quality principles.
What is automation bias and why should BMS care?
Automation bias is the human tendency to over-trust confident machine output and under-apply independent judgement. It is a human-factors risk rather than a software defect, and it can lead to errors being accepted uncritically. Maintaining professional scepticism and following oversight procedures are core safeguards.
Can AI tools change or degrade after deployment?
Yes. Adaptive models can change behaviour over time, and even locked models can appear to degrade as reagents, instruments, populations and protocols shift. This is why ongoing performance monitoring, audit and clearly defined revalidation triggers are essential parts of governance.
Who is accountable if an AI-assisted result is wrong?
Accountability remains with the qualified professionals and the organisation operating within their regulated quality system, not the algorithm. Assistive AI informs a decision; it does not transfer professional responsibility. This is why human oversight, documented procedures and incident reporting are non-negotiable.
Further training
Build your knowledge with related modules on the platform: