AI for Rare Disease Diagnosis: Current Tools

Last updated: October 2, 2026

Diagnosing a rare disease can be one of the most difficult challenges in modern medicine. Symptoms may overlap with common conditions, patients may see several specialists, and genetic or clinical findings can be hard to interpret. Artificial intelligence (AI) is increasingly being studied as a way to support this process.

The reality is more modest than the idea of an «AI doctor.» Rather than replacing physicians, current systems can help organize clinical information, identify relevant phenotypes, prioritize genes or variants, and generate possible diagnoses for further investigation. This article reviews the main types of AI and computational tools currently used or studied for rare disease diagnosis.

How We Researched This Article

This article is a desk-research narrative review, not a systematic one. We did not test any of these tools ourselves. We:

• Used peer-reviewed studies, institutional reports (NIH/NCATS, EURORDIS) and preprints published mainly between 2024 and 2026.
• Noted whether each source is a patient survey, a clinical cohort, a computational benchmark or a research system, and labeled preprints as such.
• Treated figures from developers or research teams as reported results, not independent verification, and reported numbers as the original authors presented them.

Why Are Rare Diseases So Difficult to Diagnose?

Rare diseases are individually uncommon, but collectively they affect a large population. More than 10,000 rare diseases have been identified, and many involve genetic or multisystem abnormalities.

The diagnostic process can take years. A Rare Barometer survey run by EURORDIS, with respondents from 104 countries living with more than 1,900 rare diseases, estimated an average total diagnosis time of 4.7 years in Europe. The delay was not equal for everyone: 5.4 years for women and 10.4 years for adolescents aged 10 to 20.

AI is being investigated partly because clinicians must connect many pieces of information: symptoms, laboratory results, medical history, imaging, family history and genetic data.

How AI Can Help With Rare Disease Diagnosis

AI does not usually diagnose a rare disease from a single symptom. Different systems focus on specific parts of the diagnostic process.

1. Phenotype-Based Tools

One of the most established approaches is phenotype matching. Doctors describe a patient’s clinical characteristics using standardized Human Phenotype Ontology (HPO) terms, and software compares them with known disease profiles to prioritize possible diagnoses.

For example, Phen2Gene is an open-source tool that takes HPO identifiers and prioritizes candidate disease genes. These systems can be useful when symptoms are unusual or appear across several organ systems. However, the output is a list of candidates rather than a confirmed diagnosis, and the quality of the result depends on how accurately the phenotype has been documented.

2. AI and Genetic Variant Interpretation

Genetic sequencing can identify thousands of variants in a single patient. The challenge is working out which ones could actually explain the person’s condition. AI and machine-learning approaches can help specialists prioritize variants according to gene-disease relationships, inheritance patterns and known phenotypes.

Sequencing does not automatically produce a diagnosis. Research from the NIH Undiagnosed Diseases Network (UDN) found that many diagnoses required additional phenotyping, genomic analysis or functional investigations beyond initial exome sequencing.

3. AI Analysis of Electronic Health Records

Another emerging approach analyzes the large amount of information in electronic health records (EHRs). In April 2026, NCATS (part of the NIH) highlighted WEST (WEakly Supervised Transformer), developed by researchers at Harvard Medical School and Boston Children’s Hospital. It uses diagnosis codes, laboratory results and clinical notes, and is designed to cope with incomplete or noisy records, which are common in real-world care.

Research has also explored fully automated pipelines. A 2025 study described PhenoBrain, which extracts phenotypes from clinical records and uses them to generate differential diagnoses. These systems remain an area of active research rather than a replacement for specialist assessment.

4. Facial Phenotype Analysis

Some genetic disorders produce characteristic facial features. Computer vision can compare facial photographs with patterns associated with genetic syndromes, giving an additional clue when combined with medical and genetic information. It should be treated as a supporting tool, not definitive evidence: many syndromes present variably, and similar features can appear across different conditions.

5. Generative AI and Differential Diagnosis

Large language models (LLMs) are the fastest-moving area. Researchers are testing whether they can take clinical features and generate differential diagnoses for physicians to investigate. Results so far are modest, and they vary by dataset:

In that evaluation, ChatGPT-4o had the highest accuracy of the three individual systems, but still identified the correct diagnosis in only about one case in five. Combining it with Exomiser, a widely used genomic tool, performed better than either alone.

A separate benchmark of 5,213 computational case reports (a preprint) found that the best LLM ranked the correct diagnosis first in 23.6% of cases, compared with 35.5% for Exomiser in phenotype-only mode. In other words, LLMs had not yet matched established rare-disease tools. A JAMA Network Open research letter from the UDN reported that ChatGPT-4o included the exact final diagnosis in 13.3% of cases, against 5.6% for the historical clinical review in that cohort, which shows why results depend heavily on how a study defines success.

Newer research systems such as DeepRare combine clinical information, phenotype terms and genetic testing results. Reported results are promising, but they come from research evaluations, not from evidence that such a system can diagnose patients independently in routine care.

What the Studies Show at a Glance

AI Is One Part of a Larger Diagnostic Process

The most realistic role for AI is not replacing the specialist but helping specialists handle complex information more efficiently. The NIH Undiagnosed Diseases Network illustrates this: it combines multidisciplinary clinical evaluation with genomic sequencing, laboratory investigations and research methods, and has helped facilitate difficult diagnoses while identifying new disease-associated genes and syndromes.

Main Limitations

• Limited data: Rare diseases provide fewer training examples than common conditions.
• Incomplete medical records: Symptoms may be missing, miscoded or described differently by different clinicians.
• Bias: Models can perform differently across populations if training data are not representative.
• False positives: A long list of possible diseases does not mean they are equally likely.
• False negatives: A system can miss a disease that is poorly represented in its training data.
• Validation: Strong results on research datasets do not automatically translate into reliable clinical performance.
• Privacy: Genomic and medical data are highly sensitive and need appropriate safeguards.

The Future of AI for Rare Disease Diagnosis

Research is moving toward systems that combine several types of evidence instead of a single data source: clinical notes, HPO phenotypes, genomic sequencing, imaging, laboratory results, family history and scientific literature in one workflow. This could be particularly valuable for patients whose initial genetic testing gives no answer, since UDN research has shown that some difficult diagnoses require investigations beyond standard genomic testing.

The goal is not an AI that simply says «this is the disease.» A more useful system helps clinicians spot overlooked possibilities, prioritize additional testing and connect evidence from different parts of a patient’s history.

Conclusion

Frequently Asked Questions

Related Reading

Sources

1. Time to diagnosis and determinants of diagnostic delays of people living with a rare disease: results of a Rare Barometer retrospective patient survey. European Journal of Human Genetics, 2024. PMC11369105
2. Illumina. Strength in numbers: How the Rare Barometer supports the rare disease community (2025). illumina.com
3. NCATS/NIH. WEST AI Algorithm May Help Speed Diagnosis of Rare Diseases (April 14, 2026). ncats.nih.gov
4. Greco K. et al. A Weakly Supervised Transformer for Rare Disease Diagnosis and Subphenotyping from EHRs with Pulmonary Case Studies (preprint). arXiv:2507.02998
5. Evaluation of ChatGPT-4o, Llama 3.1-8B-Instruct and Exomiser in 424 rare disease patients at the Undiagnosed Diseases Network. IOS Press, 2025. doi:10.3233/SHTI251000
6. Shyr C. et al. Large Language Models for Rare Disease Diagnosis at the Undiagnosed Diseases Network. JAMA Network Open, 2025. PMC12374213
7. Systematic benchmarking demonstrates large language models have not reached the diagnostic accuracy of traditional rare-disease decision support tools (preprint). PMC11302616