How AI is Transforming Rare Disease Diagnosis

Collectively rare diseases are anything but rare – they impact 30 million people in the US and ten times that amount globally. And 1 in 3 children suffering from a rare disease will not survive beyond the age of five.

One of the biggest challenges facing clinicians is making a quick, accurate diagnosis – on average patients visit eight physicians and receive two to three misdiagnoses before being correctly diagnosed, a process that takes US patients around 7.6 years, and is often referred to as a diagnostic odyssey.

This can in part be explained by the sheer number of rare diseases – there are around 7,000 disorders, which collectively have 12,000 unique characteristics. The complexity surrounding diagnosis is compounded by the overlap of symptoms between diseases – for example, a patient may present with encephalopathy and seizures, which are features of 1,500 rare diseases. In such cases, physicians often turn to their own prior experience to inform their diagnosis, however, statistically, it is highly unlikely that the clinician would have come into contact with the patient’s particular rare disease in the past, or even heard of it. 

Shortening the Diagnostic Odyssey

For those patients presenting with severe disease – of whom 50% are children – shortening the diagnostic odyssey is a pressing challenge. For some sick babies, each minute lost before a diagnosis and precise treatment is started may increase the likelihood of permanent neurological damage or even death.

Diagnosis opens the door to potential interventions that could significantly improve health outcomes and quality of life, as well as reduce the length of stay in hospital and the cost of care. 

Going beyond whole genome sequencing

The ability to sequence the genome is the essential first step for rare disease diagnosis. DNA sequencing has undergone an astonishing revolution in the last decade, as demonstrated by the emergence of numerous consumer-focused genetic mapping tools such as Ancestry.com and 23andMe, as well as companies like Illumina which have pioneered an industrial-scale processing capability that enables a variety of rapid DNA sequencing techniques. 

The gold standard is Whole Genome Sequencing (WGS) – an incredibly powerful and thorough method that once took years to undertake but can now be performed relatively cheaply and in a matter of hours.

Most rare diseases have a genetic component and WGS is the way to detect the genetic abnormality associated with the disorder. But, it’s complicated. The same disease may have slightly different genetic variants in different patients, and a single person may be a carrier of genetic markers associated with multiple rare disorders, but actually only suffer from one. Interpreting this information, therefore, takes highly skilled geneticists, and even they still face a considerable challenge in making a diagnosis due to the potential breadth and ambiguity of the data.

The answer to this problem lies in overlying genomic information with the phenotype – the physical manifestation of the underlying disorder. It requires a painstaking comparison of an individual patient’s characteristics and clinical findings against the thousands of phenotypes associated with rare diseases.

This so-called “deep” phenotyping is laborious and highly technical, and is reliant on experienced physicians with the ability to match patient phenotype with known Human Phenotype Ontology (HPO) rare disease characteristics – all 12,000 of them. This is a manual process that takes hours to complete for each patient, even when undertaken by a few highly trained geneticists with relevant experience. In short, while huge strides have been made to scale and automate genetic analysis, the corresponding and necessary phenotype analysis is still manual and time-consuming and does not scale.

Harnessing the power of AI

This is where AI comes in. 

AI-led technologies are already improving diagnostic speed and accuracy by automating deep phenotyping to complement the broad genetic data generated from WGS. 

The technology is able to quickly analyze the lengthy, unstructured data held within a patient’s Electronic Medical Record, which comprises approximately 80 percent of the meaningful data which previously needed to be analyzed by manual reviewers. In a matter of seconds, AI is able to match patient phenotype data with potential phenotypes known to be associated with a specific rare disease. This in turn helps speed up the overall process of diagnosis from days or weeks to hours, an end-to-end process pioneered by Rady Children’s Institute of Genomic Medicine (RCIGM) and reported recently in the New England Journal of Medicine. 

The vital clinical narrative is unlocked with Clinical Natural Language Processing (CNLP), a highly specialized branch of AI that enables machines to understand human language and hundreds of thousands of detailed clinical concepts. By recognizing and analyzing the clinical and social data within this unstructured clinical narrative, CNLP-based technology can process lengthy, chronologically ordered content and make it computable at scale. This enables detailed patient information to be matched with the entire HPO library in seconds, which when coupled with rapid WGS and additional analysis helps to produce a diagnosis within hours of admission for a critically-ill newborn. 

In the case reported by RCIGM, typical of this kind of scenario, the child was born without any apparent problems but was brought back to the hospital when he was around 6 weeks old and extremely ill. Unbeknownst to his parents and the clinicians looking after him, he had an extremely rare genetic disease. Deteriorating by the hour, despite all attempts to support him, the RCIGM team used the automated AI-supported technique to establish the diagnosis for the child. 

In this particular case, the disease was treatable with vitamin supplements and when these were added to his feed, he recovered rapidly and was discharged a few days later. When reviewed in the clinic 6 months later he was a healthy and thriving baby. The authors noted that tragically his sibling, born 9 years earlier before these new technologies were available, had died at 18 months without a diagnosis – though the likelihood is that it was the same, treatable rare disease but impossible to diagnose without AI.  

AI tools thus empower physicians to make faster diagnoses, increasing patients’ chance of survival and paving the way to improved health outcomes. 

Looking ahead: improving population health

It’s not only critically ill babies and children that stand to benefit from AI-powered technology in the field of rare diseases. There are also significant numbers of undiagnosed – and therefore untreated – adults with milder forms of rare disease. 

In the future, it is likely that AI will be used to look for tell-tale symptoms in otherwise unremarkable medical records to find adults with a partial, and therefore less severe, expression of disease. They are likely to have suffered continual, unexplained health challenges – for example, regular fractures or bone breakages – which would have impacted their lives significantly, but are unlikely to have ever received a definitive diagnosis. By automating the deep phenotyping process, AI will be able to identify patients with combinations of characteristics that suggest rare diseases as the underlying cause, opening the door to medical interventions for those individuals with treatable disorders.

Revolutionizing the outlook for rare disease patients

Patients with rare diseases face a multitude of challenges – a lack of information, the scarcity of specialist physicians and the life-changing health impact of these diseases.

The good news is that considerable progress is happening – we understand more than ever how rare diseases work, and how they then manifest. Pharmaceutical companies are investing more in drug development, thanks in part to government incentives and tax breaks, and in particular the Orphan Drugs Act passed in the US in 1983. These factors have come together to improve treatment options for rare disease sufferers, a vulnerable group of patients who, until recently, it has been very difficult to help.

These new AI-based approaches are now able to help diagnosis at speed and at scale, giving patients access to new treatments as they emerge, and disease management plans. Not only does this improve the survival rate and quality of life for patients with rare diseases, but it also reduces the amount of healthcare they need, reducing the burden on over-stretched health systems. 

About Chris Tackaberry

Chris Tackaberry is the co-founder and CEO of Clinithink, a technology company built around CLiX, the world’s first Healthcare AI capable of truly understanding unstructured medical notes.

Chris is a qualified physician and MSc Computer Science graduate who spent nine years in clinical practice in anaesthesiology and intensive care before embarking on a career in healthcare IT. His combined expertise in medicine, computer science and leadership has been the foundation for his stewardship of Clinithink’s strategic direction and growth.

ai
electronic medical record
healthcare
medicine
hospital
medical
pharmaceutical