New AI tool for doctors’ notes may save patients years to be diagnosed
Model improved its accuracy as patients accumulated more health data
Computer-based analysis of doctors’ notes within healthcare records could dramatically speed up the diagnosis of rare diseases, such as AADC deficiency, a new study shows.
By analyzing descriptions of a patient’s symptoms and history, this method could help flag those who face the long and often difficult journey to a correct diagnosis.
The study, “Information content as a health system screening tool for rare diseases,” was published in npj Digital Medicine.Â
Although AADC deficiency and other rare diseases are individually rare, as an aggregate, they are fairly common. By some estimates, more than 1 in 20 people worldwide living with a rare disease.
Overcoming ICD code limitations
With the advent of new tools, such as computer-based healthcare record systems and machine learning, researchers have been exploring whether it is possible to better identify patients with rare diseases through analyses of large-scale healthcare record data. Studies in this area have often utilized ICD codes, which are designations primarily used for insurance billing purposes. However, ICD codes are specifically designed to label the disease a patient has, so their utility as a sole tool for identifying undiagnosed patients with rare diseases is inherently limited.
In the study, researchers explored an alternative approach, where they analyzed SNOMED Clinical Terms, a system that enables clinicians to record descriptions of a patient’s symptoms and history. Essentially, the researchers’ analysis aimed to identify combinations of descriptions among these clinical terms that are uncommon in the general population, which may be indicative of a rare disease.
To test their approach, the researchers analyzed more than 35,000 unique SNOMED Clinical Terms from more than 1.2 million people in the Singapore health system. About 3.5% of these patients had a confirmed diagnosis of a rare disease, with more than 900 distinct rare diseases represented.
The researchers found their terms-based approach could identify patients with rare diseases with some accuracy. The model became more accurate as patients accumulated more data over time, with AUC ranging from 0.662 for patients with at least two encounters to 0.717 for those with five or more encounters. AUC, or Area Under the Curve, is a statistical measure that tests how well two groups can be distinguished (i.e., rare disease or not). Scores range from 0.5 to 1, with higher numbers reflecting better distinction ability.
Although this approach showed promise for identifying rare diseases generally, the researchers noted its accuracy was markedly reduced for rare diseases that are underdiagnosed.
“The takeaway is that [this type of] screening works well as an initial filter for rare diseases in general, but for the subset of rare diseases that tend to be under-recognized, its performance gap highlights a need for more nuanced or intensive screening methods,” they concluded.
About the Author
