Case Studies

Centre for Doctoral Training in AI for Medical Diagnosis and Care

University of Leeds | Dr Anna-Grace Linton & Dr Zuzanna Wójcik

Amplifying Patient Voices in AI & Healthcare
Despite researching distinct projects, both Anna-Grace Linton and Zuzanna Wójcik’s PhDs highlighted the importance of amplifying patient voices in healthcare, and the transformative ways in which artificial intelligence (AI) could assist with this. Both looked at Patient-Reported Outcome Measures (PROMs), which are questionnaires that aim to capture the patient’s perspective of their current treatment plan and, in turn, their quality-of-life. While Anna-Grace used AI to analyse cancer patients’ descriptions of their health and quality-of-life, Zuzanna explored how machine learning models might better understand and, ultimately, predict how cancer patients respond to chemotherapy.

As the summaries of their projects below will make clear, both Anna-Grace and Zuzanna emphasise that patient perspectives, as gathered through PROMs, are more than just “data” for AI to analyse. Rather, they contain rich and often complex representations of real people’s lives and daily struggles. Their individual yet complimentary PhD projects reveal that whether AI is interpreting people’s healthcare stories or symptoms, ethical, patient-centred AI in healthcare begins with ensuring that patient experiences are accurately captured and understood.

Dr. Anna-Grace Linton
When patients fill in PROMs, they not only answer tick-box questions, but can also respond with short comments about how they feel. These responses hold value because they often detail the patient’s personal insights about their symptoms, day-to-day concerns and quality-of-life. However, because thousands of patients across the country complete PROMs, these written comments are often too time-consuming for clinicians to read and analyse manually. This means that crucial, patient-reported information can go ignored, leaving our health services less aware of patients’ lived experiences.

Anna-Grace’s PhD project set out to rectify this by exploring how AI could make these hidden patient voices more visible, looking in particular at patient comments from cancer PROMs datasets. Using natural language processing (NLP) – a field of AI that allows computers to understand human language – Anna-Grace’s research developed two frameworks to automate the analysis of PROMs from patients with prostate and colorectal cancer. The first framework used NLP to organise patients’ responses into meaningful themes such as emotional and physical well-being, daily challenges, and social functioning. The second framework used large language models – AI systems that understand and generate human-like language – to condense patient responses into short, meaningful summaries that capture their experience.

Using various NLP models, Anna-Grace discovered interesting patterns within patient responses. For instance, patients experiencing day-to-day pain often also described difficulties caring for their family members. This link highlighted how a cancer patient’s physical symptoms can impact the emotional, social and practical aspects of their life. By uncovering these complex implications, Anna-Grace’s work highlighted the value of analysing PROMs, which could one day shape patients’ support programmes or treatment plans. Her project not only involved training AI to recognise, interpret and understand patient narratives, but drew attention to the multifaceted ways in which living with chronic illness impacts a patient’s life.

Working with sensitive data provided by patients and collected by the NHS, Anna-Grace relied on the CDT’s high-performance computing facilities, which were critical for running her NLP models efficiently and securely. Guided by CDT supervisors with expertise in computing, healthcare and behaviour science, her project also – and perhaps most importantly – discussed the benefits and limitations of AI-based analysis of PROMs from the patient and clinical stakeholder’s perspective. These discussions raised issues regarding the value of patient involvement in AI research, as well as healthcare inequalities. In fact, Anna-Grace’s thesis rightfully dedicates an entire chapter to this issue, exploring how missing voices in PROMs arise from social and demographic factors such as language barriers, computer access, and digital literacy. This is where Anna-Grace’s research made one of its most compelling observations: PROMs analysis could risk excluding already underrepresented groups in healthcare data. By making this issue explicit, Anna-Grace’s project set the stage for future research to ensure that AI models are developed to incorporate underrepresented patient voices. If AI can learn to capture and elevate these voices, then it serves as a powerful opportunity to humanise healthcare data, producing more representative datasets and culturally aware analyses of PROMs going forward.

Today, in her postdoctoral work as a data scientist at Queen Mary University of London, Anna‑Grace continues to explore healthcare challenges, now focusing on rare cardiovascular conditions. The interdisciplinary skills, networks and technical training she developed through the CDT have laid the foundation for a career dedicated to responsible, patient‑centred AI in healthcare.

Dr. Zuzanna Wójcik
Chemotherapy remains a cornerstone of cancer treatment, offering an effective way to target and destroy rapidly dividing cancer cells. However, it can also cause significant side effects, including chemotherapy toxicity (which refers to the harmful effects of chemotherapy drugs when they damage healthy cells alongside cancer cells). In some cases, these complications become serious enough to require hospital admission or additional clinical support, placing further physical and emotional strain on patients. Being able to identify which patients are most at risk of experiencing toxicity could greatly improve treatment planning and overall quality of care. Traditional clinical tools often miss patients’ subjective reports of symptoms and cannot always capture early signs of chemotherapy toxicity. Zuzanna’s project asked a simple but powerful question: What if artificial intelligence could learn directly from patients’ own experiences? Using PROMs reported by patients with breast, gynaecological and colorectal cancer, Zuzanna investigated how AI (specifically machine learning models) could analyse patient perspectives to predict changes in their quality of life during chemotherapy as well as their risk of unplanned hospital admissions.

While PROMs are increasingly used in AI to predict patient outcomes, the absence of clear guidance has led to inconsistent data handling, model development, and evaluation. These issues limit how effectively PROMs and AI can be used in practice. Zuzanna’s project was unique in that it proposed a new, patient-centred framework for applying AI to PROMs. Unlike previous machine learning models, which rely solely on clinical or biomedical data, Zuzanna’s project used models that centralised patient-reported information, as gathered through PROMs and symptom-severity reports. The results found that AI models work better when they include the patients’ own perspectives. In fact, they outperformed those models that use clinical data alone, highlighting just how essential patient centred data is in AI research. These findings, published in the National Library of Medicine in 2025, also concluded that routinely collecting patient-reported symptoms can identify emerging patterns of chemotherapy toxicity earlier. This is significant because it could provide both clinicians and patients an opportunity to anticipate potential complications and intervene sooner.

As part of the CDT’s diverse research community, Zuzanna was able to discuss her project with both a clinical oncologist, a group of patient representatives from the Leeds Institute of Medical Research at St James’s Hospital, and the independent, patient-led UK organisation Use My Data. This collaborative process shaped the design of Zuzanna’s AI models, not only improving the model’s relevance to patients, but building public trust in medical AI research more generally. Put simply, Zuzanna did not simply decide what her AI models should predict using previous datasets or medical studies. Instead, she asked patients and clinicians what outcomes mattered most to them. For instance, feeling well enough to perform usual activities was most important to patients, whereas avoiding emergency hospital visits was integral for clinicians. What truly set Zuzanna’s project apart, then, was its inherently inclusive design that was guided by listening to others. Ultimately, her project found that AI models which take patient stories into consideration have the potential to predict more accurate outcomes for those undergoing chemotherapy and, more significantly, show patients that their lived experiences matter in shaping their healthcare journeys.

Her project continues to align with ongoing, contemporary trends in digital healthcare around the globe, shedding further light on one of the many ways in which AI can revolutionise the early detection of illnesses, the personalisation of patient treatments, and the overall reduction of pressure on our healthcare systems. In her current role as a Postdoctoral Research Associate in Digital Health Research at the University of Bristol, Zuzanna remains involved in patient-centred research, developing and evaluating novel AI-driven sensing technologies for the continuous monitoring of symptoms of Parkinson’s Disease.

Centre for Doctoral Training in AI for Medical Diagnosis and Care
University of Leeds | Case Study: National Pathology Imaging Co-operative (NPIC)

Since its conception in October 2019, the UKRI Centre for Doctoral Training (CDT) in AI for Medical Diagnosis and Care has built and nurtured a cohort of highly talented, interdisciplinary researchers with a range of expertise from STEM and health backgrounds. Dr. Lucy Godson, Dr. Andrew Broad, and Dr. Jason Keighley began their PhDs as part of the CDT’s first cohort and have since graduated to work as Digital Pathology AI Scientists at the National Pathology Imaging Co-operative (NPIC) in Leeds. Having gained world-class skills from the CDT, Lucy, Andrew and Jason now work – through their individual projects – towards developing artificial intelligence in digital pathology to ultimately improve diagnosis and care for NHS patients.

But what exactly is digital pathology, and how does it differ from traditional histopathology? Put simply, both fields examine tissues to diagnose disease. But digital pathology marks a decided transition from histopathology by harnessing computer technologies, high-resolution scanners, and AI to analyse and interpret pathology slides digitally as opposed to using a traditional microscope. This not only allows pathologists to examine pathology slides more easily but creates a huge storehouse of histopathology images which can be used to develop and innovate AI methods for diagnosis and treatment. NPIC is a national leader within this sphere of digital healthcare, helping NHS hospitals store pathology slides in a secure, readily accessible form. The NPIC system covers more than twenty per cent of NHS labs in England, with the hope to one day create the largest national pathology database in the world. Below there are summaries of Lucy, Andrew and Jason’s PhD research within the CDT before they began their exciting career pathways at NPIC.

Dr. Lucy Godson
Before Lucy began her career at NPIC, she completed her PhD within the CDT at Leeds. Her research bridged digital pathology and artificial intelligence to address the various challenges involved in cancer diagnosis, specifically melanoma cancer. A core issue of Lucy’s research had to do with identifying which patients responded to certain treatments for melanoma. With the development of digital slide scanners, tissue slides can now be digitized to produce whole slide images (WSIs). This digitised process was central to Lucy’s development of artificial neural networks (ANNs) which she used to classify melanoma patients into “immune subgroups”. These subgroups indicated the varying levels of immune cells within a patient’s tumour: if a patient had higher levels of immune cells, for instance, this could be used to determine potential treatments options and survival outcomes. With only a small percentage of pathology departments in the UK fully staffed, delays in cancer diagnosis remain an ongoing challenge.

But Lucy’s research was foundational in highlighting how AI innovation in healthcare could address these workforce shortages, diagnosing cancer faster and more cost-effectively. The CDT provided powerful, state-of-the-art computing technologies at the University of Leeds. These resources became vital for Lucy’s development of AI models. As part of a CDT that aims to train its researchers to deploy AI as responsibly as possible, Lucy frequently worked alongside a melanoma pathologist. This meant that her research was not purely technical but geared towards building AI that ensures compassionate patient care and clinical value. Further to this, the CDT allowed Lucy to access real patient datasets, namely the Leeds Melanoma Cohort. As Lucy explains: “Thanks to the CDT’s leadership, there was strong collaboration between experts in AI and clinicians working in healthcare. This teamwork helped shape a supportive research environment and gave students access to resources that would otherwise be hard to reach”. Just as important as collaborative resources, however, is the ability to communicate complicated scientific research to the public. This is precisely what the CDT equips their researchers with: transferable skills in public engagement and outreach. Lucy developed these by attending various conferences and, on one occasion, presenting her findings to a panel of patient advocates in a Patient Dragon’s Den organised by the CDT. These opportunities allowed Lucy to keep her research patient-focused and clinically meaningful, ensuring that her findings were always rooted in patient needs.

In her current role at NPIC, Lucy continues to develop advanced AI algorithms for melanoma diagnosis. While pathologists will always have oversight with patient diagnosis, Lucy’s pioneering research works to expand future possibilities where artificial intelligence can operate as an assistive technology to support pathologists with cancer detection, diagnosis and care.

Dr. Andrew Broad
Andrew’s PhD research looked at applications of AI in digital pathology for colorectal cancer diagnosis. Currently, pathologists examining WSIs need to examine many tiny sections of the image (often called tiles), making the whole practice of extracting useful diagnostic information a labour-intensive, time-consuming and inconsistent one across pathologists. While there have been a variety of proposed AI solutions to this, they often involve examining thousands of small image tiles across the WSI which is a slow process requiring significant computing resources. Andrew’s research directly intervened in this challenge by setting out to find more efficient ways of extracting information from WSIs, starting with “attention-inspired artificial intelligence”. This refers to a category of artificial intelligence techniques that are designed to mimic the biological principles of attention in the human brain. If a human looks for a particular object for instance, the brain focuses on specific stimuli while ignoring others. Attention-inspired AI aims to do a similar thing: to prioritise certain stimuli over others when making decisions. In the context of colorectal cancer imaging data, Andrew was examining whether AI could analyse such images and locate regions of cancerous tumour without overlooking diagnostically important features. The CDT provided supervisory support and guidance, with access to the high-powered computing infrastructure Andrew required to investigate this hypothesis.

His results were significant, showing that AI could locate regions of cancer significantly faster than traditional methods while still accurately identifying diagnostically important features like tumour regions, and metrics such as “tumour stroma ratio” (which helps predict how aggressive the cancer could be). The “ground truth” (the human pathologists’ original evaluation) versus the “prediction” (the AI analysis) achieved over eighty percent overlap. In other words, Andrew’s research proved that AI attention methods could speed up pathology AI processes without sacrificing accuracy, potentially paving the way for quicker diagnoses and improved patient care. Andrew’s research was rooted in interdisciplinary interests, investigating the extent to which AI might mimic the complexities of the human brain (in terms of pattern recognition, visual attention etc) to extract diagnostic information from digitised images of colorectal cancer. Andrew published his findings in the Journal of Pathology Informatics (2022), and Scientific Reports (2024), marking a fundamental contribution to scientific knowledge.

In line with the CDT’s ethics on AI in digital healthcare, Andrew’s research is an important step towards ensuring the accuracy and responsibility of AI tools in real-world clinical practice. The end-goal is not to replace medical professionals but to elevate AI as an assistant in clinical settings. While Andrew’s findings have not yet been used in real-world scenarios, his research and publications demonstrate the ways in which AI can improve the accuracy of analysing cancer tissue. Now at NPIC, Andrew develops software to support various research projects, furthering ways to implement AI algorithms into digital pathology.

Dr. Jason Keighley
With expertise in computer science and software development, Jason begun his PhD with the aim to explore how AI could make digital pathology more efficient through three core areas: compression, classification and generation. The first area of his research looked at whether AI algorithms could compress (reduce the file size of) WSIs more effectively than traditional formats like JPEG. WSIs are scanned at a very high resolution to preserve microscopic visual details. Because of this, they are extremely large, gigapixel-scale images. Every year, the NHS spends a huge amount of money to store these digital histopathology slides. But if AI could act as a possible alternative to these traditional formats, this would save the NHS immense storage costs. Jason found that while AI-based compression was less efficient than traditional methods, the AI systems were multipurpose, able to support various tasks like file compression and the automation of manual tasks.

The second area of Jason’s research looked at “pathology feature classification using AI”: whether AI could examine a dataset and identify patterns (such as cellular structures or stroma) on its own without relying on a clinician’s manual annotations. Can AI detect a tumour without human intervention, for instance? Jason found that AI algorithms could indeed do this, informing a clinician exactly what type of tissue (benign/malignant, tumour/normal) was present in a pathology slide. The final area of Jason’s research had to do with “generation”: whether AI models could automatically create or generate dataset annotations. With access to high-performance computing clusters through the CDT, Jason found that AI could automate the kinds of labelling work that clinicians spend time on manually, saving time and space to focus on higher-value tasks. Before publishing aspects of these findings in his poster paper with SPIE Medical Imaging, Jason attended the SPIE Medical Imaging Conference in San Diego which not only allowed him to share ideas with international audiences, but to reflect on issues such as the implementation of future AI policy, data privacy, and patient care. In fact, even though Jason’s PhD was technologically specialist, CDT-organised conferences opened channels of communication between patients and researchers to allow for informal discussion in collaborative settings.

Jason now works in software development, infrastructure management, and AI training and integration at NPIC. His projects are highly technical, yet the long-term implications of these are extraordinary: moving scalable AI solutions from research into real-world scenarios that will not only advance clinical safety and build trust in AI-enabled diagnostics but will transform patient pathways and care.

Conclusion
With healthcare systems facing a global work shortage of pathologists and diagnostic backlogs only increasing, the work that Dr. Godson, Dr. Broad, and Dr. Keighley are involved in at NPIC has never been more important. Their projects innovate and modernise digital pathology, but are currently in their experimental, research phase. After all, it takes rigorous governance and regulation to trial, evaluate and eventually implement the kinds of AI systems mentioned above to NHS hospitals and real-world diagnostics. The balance between technical innovation and clinical feasibility is a challenging one. But what remains clear is that the work being done by NPIC is a future-oriented investment, an example of world-leading research that is not only crucial to successful digital pathology, but an important precedent in building public trust in AI-assisted healthcare.

Images
Image 1 description: One of the main imagining technologies used in digital pathology is Whole Slide Imaging (WSI). The image below is a WSI that shows a section of tissue under high magnification. The pink and purple colouring indicates the use of H&E (haematoxylin and eosin) stain which then allows pathologists to identify cell structures, mutations, or abnormalities.