Will artificial intelligence live up to its promises for rare diseases?

François Houÿez, Director of Treatment Information and Access at EURORDIS – Rare Diseases Europe, explores opportunities and limitations of AI in the context of rare diseases

The generalised use of AI in healthcare is already a reality. Clinicians use AI to support diagnosis, select personalised treatments in shared decision-making, follow up patients, keep track of medical progress, and process the vast scientific literature required for continuous education.

AI is also embedded throughout the research and development of health technologies, medicines, medical devices and digital tools – from identifying health needs to analysing data, reviewing scientific literature, evaluating products, and repurposing treatments. It is used to monitor drug utilisation, predict supply shortages, and analyse real-world data to compare technologies’ relative effectiveness.

Algorithms capable of diagnosing disease without a healthcare professional’s opinion are already available. Some are developed by commercial companies that operate networks of healthcare centres. The safeguards proposed by ethicists and policymakers – that AI should always be backed up by prohibitively close levels of human oversight – may prove difficult to maintain in practice.

Healthcare systems are also under strain. In the UK, where estimates point to a substantial shortfall of healthcare professionals, the government has suggested that AI could help reduce this gap. (1) In such contexts, AI tools that facilitate and accelerate diagnosis and support treatment selection may be increasingly relied upon. In certain areas, AI systems have even demonstrated performance that rivals or exceeds that of healthcare professionals, including in diagnosing rare diseases.

For AI to fulfil its promise in healthcare for people living with rare diseases, patient involvement is essential

Patients can now access AI-based tools themselves to obtain medical insights. Yet they may still need to wait to consult a specialist before treatment can be initiated. If this delay results in lost opportunities for effective therapy, the requirement for human validation may clash with the need for rapid decision-making. Waiting times for appointments in Centres of Expertise or with specialists can be long, and general practitioners may not feel sufficiently confident to validate an AI-generated diagnosis or initiate treatment independently.

This raises a question: if patients don’t receive timely treatment due to waiting for human validation of AI outputs, could they argue that the requirement for excessive supervision itself caused harm?

Simplifying complex medical information

At the same time, AI – including chatbots based on large language models – offers potential solutions to a longstanding problem: health literacy. Despite decades of discussion about improving access to understandable information, tangible progress has been limited. Standardised template documents are produced, often designed for adults at OECD level 2 literacy – individuals who can match text with information and make basic inferences. Yet many citizens struggle to understand consent forms for clinical trials, information provided before complex surgery or even medicine package leaflets. Information is not always available in all languages spoken within a country. Resources such as Orphanet or Wikipedia may exist for complex rare diseases, but they often remain too technical or incomplete for many users. Helplines attempt to fill the gap, yet we must acknowledge our collective difficulty in reconciling citizens with scientific language.

However, AI also has limitations, particularly in rare diseases. Its effectiveness depends on the quality and size of training datasets. Rare diseases, by definition, involve small populations. While machine learning can support high-level computational research workflows, predictive algorithms may struggle when trained on limited data. Projects such as HTx ‘Next Generation HTA’ have explored methods for deriving predictions from smaller datasets in conditions such as head and neck cancer or myelodysplastic syndrome. These efforts suggest possible ways forward. Nevertheless, significant challenges remain before AI predictions in rare diseases can reliably complement the expertise of specialised healthcare professionals. After all, distinguishing a running zebra from a running horse based solely on sound is not straightforward – even for sophisticated systems.

The importance of patient engagement and trust

For AI to fulfil its promise in healthcare for people living with rare diseases, patient involvement is essential. AI projects intended for direct healthcare or research applications should involve patient representatives from the outset, through testing, certification and implementation. Frameworks to assess AI-based technologies, such as CHEERS-AI (2) or MAS-AI for imaging, (3) have themselves engaged patient advocates in their development.

Trust will ultimately determine whether AI is adopted and accepted. This requires patients to test AI applications, provide feedback and engage in dialogue about perceived barriers and benefits. The 2018 Compare e-cohort study of 1,183 people living with chronic conditions illustrates this tension: although 20% believed the benefits of AI clearly outweighed the risks and only 3% felt the opposite, 35% said they would refuse at least one AI-based intervention in their care. (4)

AI will continue to reshape healthcare. Whether it lives up to its promises will depend not only on technical performance, but on governance, transparency, realistic oversight, and meaningful engagement with the people it is intended to serve.

References

1. https://www.gov.uk/government/speeches/the-2024-budget-and-nhs-productivity
2. Consolidated Health Economic Evaluation Reporting Standards for Interventions That Use Artificial Intelligence: https://www.valueinhealthjournal.com/article/S1098-3015(24)02366-0/fulltext
3. Model for Assessment of Artificial Intelligence: https://cimt.dk/en/assessment/mas-ai
4. Tran, VT., Riveros, C. & Ravaud, P. Patients’ views of wearable devices and AI in healthcare: findings from the ComPaRe e-cohort. npj Digit. Med. 2, 53 (2019). https://doi.org/10.1038/s41746-019-0132-y