Healthcare AI has a trust problem – and patients with complex conditions are paying the price

Artificial intelligence could transform care for the millions of people in England living with multiple long-term conditions. Large NHS datasets hold decades of patient records. Sophisticated algorithms can spot patterns no clinician could find alone. But a stubborn gap persists between what AI can do in a research lab and what it actually delivers in a GP surgery.

The problem is not a shortage of clever models. It is a shortage of infrastructure – standardised ways to prepare messy real-world data, rigorously compare competing approaches, and deploy tools that clinicians and patients actually trust. Without that infrastructure, promising AI projects stall as pilots, and the people who most need better care see no benefit.

Our approach and partners

Rather than build another standalone AI tool, this project created an end-to-end ecosystem for clinical AI development – from raw data to implementation-ready system.

The work began with data. A high-performance encoding pipeline transformed 6.8 million patient records from the Clinical Practice Research Datalink (CPRD) into analysis-ready formats, harmonising primary care records with linked NHS England hospital, A&E, and mortality data. Eight machine learning model families – from logistic regression to deep neural networks – were then systematically trained and evaluated on this data. A reinforcement learning framework benchmarked these models against each other, replacing ad hoc selection with a principled, automated approach.

But technical infrastructure alone does not build trust. A four-day Human-Centred AI (HcAI) Design Bootcamp, developed with imec-SMIT (Vrije Universiteit Brussel), Stanford University's Product Realization Lab, and The Alan Turing Institute, brought clinicians, patients, researchers, and designers together to co-create AI solutions grounded in real clinical needs.

Connecting these strands is the Health xAI Implementation Framework – published in iScience – which sets out how to deploy AI across multiple healthcare sites while managing privacy, explainability, and ongoing monitoring.

What we found – and why it matters

1. Simpler models often outperformed complex ones. When tested on 2.8 million patient records, classical machine learning – logistic regression for mortality prediction, gradient boosting (XGBoost) for hospitalisation – frequently matched or beat deep learning on tabular clinical data. This has significant implications for where the NHS invests its AI resources.
2. Automated model selection works. The reinforcement learning framework reached 75% accuracy in identifying the best-performing model within just 2,000 evaluation rounds. All three bandit algorithms converged on the same model rankings, validating the robustness of the approach.
3. Co-design with diverse stakeholders produces usable tools. The HcAI Bootcamp attracted over 260 registrants. Twenty-eight selected participants from 15 different professional backgrounds developed three working prototypes: an LLM-powered health literacy tool providing personalised plain-language explanations, PILLAR – an AI system for precision medication dosing based on kidney function, and an automated patient visit summary tool using speech recognition. Teams that included patient representatives showed greater attention to practical usability.
4. Open access lowers the barrier to entry. All resources – the CPRD data encoder, benchmarking code, trained model configurations, and bootcamp design templates – are freely available through the Open Science Framework, enabling other research teams to build on this work without starting from scratch.

What this means

This project provides a reproducible blueprint for moving the NHS from isolated AI experiments to systematic digital transformation. For patients with multiple long-term conditions, it means AI tools designed around their actual needs rather than technical convenience. For the health system, it means an evidence-based approach to selecting, deploying, and governing clinical AI that prioritises safety and equity alongside performance.

The key predictors identified across models were clinically coherent – age, haemoglobin, cancer incidence, inflammation markers, and multimorbidity burden – supporting integration into existing care pathways and building the clinical trust essential for adoption.

What needs to happen next

Three things need to change. First, the tools must move from offline simulation to real-time validation within NHS trusts – the gap between retrospective accuracy and prospective utility remains significant. Second, lighter-weight versions of the pipeline and benchmarking tools are needed for research teams without access to high-performance computing. Third, regulators – particularly MHRA and NICE – need to work with the research community to establish reinforcement learning and human-centred design methods as recognised standards for clinical AI governance.

Without regulatory engagement, even well-validated tools will struggle to move from research output to approved clinical system.