Blog February 14, 2022

AI Analysis of Real-World Data to Improve Rare Disease Diagnosis

Scientists are hoping to harness artificial intelligence (AI) to help diagnosis rare diseases. Bionest experts weigh in

One of the main challenges that physicians encounter with rare diseases, which are by definition individually uncommon, is often establishing a definitive diagnosis. Early diagnosis opens the door to more effective disease treatment and management, potentially leading to better outcomes and lower costs to healthcare systems, as well as opportunities for patients to participate in clinical trials. However, diagnosis of rare diseases can be lengthy and expensive. According to research conducted by the rare disease patient advocacy organization Global Genes, achieving an accurate diagnosis for a rare disease can take five or more years from the initial onset of symptoms and involve an average of six to seven physicians. Moreover, initial diagnoses of rare diseases are frequently incorrect, leading to further delays for patients to receive appropriate treatment and higher health care costs. 

The ability to obtain a timely, accurate diagnosis is particularly challenging for patients with conditions that affect multiple organ systems. In such diseases, symptoms can be highly diverse, equally associated with more common conditions, and can manifest differently over a period of years as the disease progresses.

Fabry disease is a multi-systemic and inherited X-linked rare disease that is complex and difficult to diagnose. In this lysosomal storage disorder, mutations in the GLA gene cause an enzyme deficiency that results in the build-up in the kidneys of globotriaosylceramide, a glycolipid. The accumulation of the metabolite leads to progressive, cumulative damage to multiple organ systems. The age of onset and symptomatology can vary widely in Fabry disease, and not all GLA mutations will result in clinically significant disease. Moreover, while early onset Fabry affects primarily males, with first symptoms appearing in childhood or adolescence, later-onset Fabry affects both sexes. Patients, even those with early onset Fabry, are rarely diagnosed with the disease until they are adults and are typically experiencing symptoms related to cardiac disease or progressive kidney damage. The actual incidence of Fabry is unknown, and estimates vary from 1 in 40,000 to 1 in 117,000 births, annually. However, due to the disease’s complexity and variable manifestations, experts believe that it remains substantially under-diagnosed.

Researchers at the University of Tennessee Health Science Center hope to improve the clinical diagnosis of Fabry Disease by using an artificial intelligence (AI) approach. They developed and calibrated a machine learning algorithm with the goal of pinpointing the small segment of patients in the general population at highest risk of having the condition, based on analysis of longitudinal real-world data such as symptoms, lab results, prescribed treatments, and outcomes. Patients identified by the AI tool as being at highest risk could then be further subjected to diagnostic testing, including gene sequencing, to achieve a definitive diagnosis earlier and at lower cost than what is currently possible.

To develop this AI tool, researchers first trained an algorithm and demonstrated the feasibility of their approach using a large set of de-identified health data covering the period from January 2013 to July 2020 from approximately 5,000 patients across the United States with confirmed Fabry’s disease. The AI tool extracted hundreds of phenotypic signals, including both familiar and less-studied disease characteristics, corresponding to different organ systems and types of pathologies. Each signal was assigned a statistical weight for predicting an individual’s likelihood of undiagnosed Fabry disease depending on if, how, and when that signal manifested in the patient’s medical history.

The researchers then further trained the AI tool by searching for a potential link between a patient’s phenotypic disease signatures and a diagnosis of Fabry disease using both the confirmed Fabry patients and a second testing cohort of a million ‘control’ patients with no evidence of Fabry disease. With this analysis, the researchers were able to construct a statistical ‘phenotypic biomarker’ for Fabry disease. As a final step, the AI tool predicted for each patient their likelihood of having Fabry disease by testing the individual’s phenotypic disease signature and its relationship to the Fabry biomarker. The researchers found that for the 1% of the test cohort patients identified by the tool as being at highest risk, Fabry Disease was 24 times more prevalent than in the overall population. A sensitivity analysis for the tool showed strong predictive results in both men and women, who were analyzed separately.

The researchers are now planning to further validate their predictive algorithm by using it in a prospective analysis with patients within the Tennessee health systems over the next year, prior to introducing the tool into clinical practice. By using existing longitudinal medical data to screen patients, identify those at highest risk of Fabry disease, and recommend them for further diagnostic testing, the AI tool may substantially improve the efficiency and lower the cost of Fabry disease diagnosis while continuing to generate new insights into fundamental characteristics of this disease.

Other groups are also exploring uses of AI to improve the diagnostic process for rare diseases that are difficult to diagnose. In 2019, Microsoft teamed with Takeda and German barcode tracking company EURODIS GmbH to better diagnose rare diseases using symptom patterns. The group’s algorithms have been employed in several pilot trials in a group of Spanish hospitals to link subtle symptoms together and detect possible signs of rare diseases in children. Health tech nonprofit startup Foundation 29 is also developing an AI tool called Dx29, with support from the Global Commission to End the Diagnostic Odyssey for Children with a Rare Disease, a multi-sector initiative using technology to speed the time to diagnose a rare disease.