Some children diagnosed with autism may fall into distinct subgroups based on their symptoms and other diagnoses, researchers report in the January issue of Pediatrics1.
The three subgroups identified in the study show some overlap in symptoms, but each is characterized by a distinct set of features: seizures, general health problems such as gastrointestinal distress, and psychiatric problems.
“We have identified the ones that are the purest,” says lead researcher Isaac Kohane, professor of pediatrics at Harvard Medical School.
The analysis relied on the largest database yet, and looked at symptoms over time. Even so, it was only able to categorize about 10 percent of the nearly 5,000 cases evaluated, highlighting the complex and diverse nature of autism symptoms.
The lack of a control group and the reliance on diagnostic codes used primarily for insurance purposes both limit the study’s impact, experts say.
Still,the analysis shows potential for using medical databases to help unravel autism’s complexity, says Kohane. The people in each subgroup may share the same underlying causes of autism, for example, and may benefit from the same treatment approaches.
Clinicians enter symptoms and diagnoses into medical databases using codes dictated by the International Classification of Diseases, a standardized manual of symptoms and disorders. Researchers can combine medical data from multiple sites to build large datasets that yield clear patterns.
Kohane and his team took a good first step in this direction, says Nigam Shah, assistant professor of medicine at Stanford University in California, who was not involved in the new study.
“These kinds of studies are the way in which the natural history of diseases should be understood,” Shah says. “I like that they made very few assumptions.”
The researchers scanned medical records from Boston Children’s Hospital and found nearly 14,000 children who had at least one entry for an autism diagnosis. They then narrowed their search to the 4,934 children who were at least 15 years of age at the time of the study.
They identified 45 categories of symptoms, such as intellectual disability or diseases of the esophagus, that cropped up in at least 5 percent of the children. They noted the symptoms’ incidence in the children’s records in every six-month window from birth to 15 years of age.
From these, the researchers used a statistical clustering analysis that automatically identifies patterns — in this case, the groups of children in whom similar clusters of symptoms appear at around the same age. The patterns fell out into three distinct subgroups.
One group includes 120 children, the majority of whom have seizures but not many other symptoms. Another 197 individuals tend to have general health problems, such as gastrointestinal distress, ear infections or asthma.
Many of the children in the third group had been diagnosed with a psychiatric disorder, such as bipolar disorder, between 5 and 15 years of age.This last group also tends to be higher functioning than the other two: About 28 percent of these individuals have intellectual disability, compared with 60 percent and 49 percent in the first two groups, respectively.
There is significant overlap in symptoms between the groups, however. For example, 42 percent of children in the subgroup with general health problems and 33 percent in the psychiatric subgroup also have seizures.
This limits the results from a clinical point of view, says Lisa Croen, director of the Autism Research Program at Kaiser Permanente, an integrated insurance and hospital network. “There’s got to be more work and more refinement of this to identify groups that can be more distinctly identified.”
Croen is using medical health records from the Kaiser Permanente hospital networks in California, the Northwest and Georgia to look for patterns among the symptoms seen in children with autism. Her study has numbers similar to those of Kohane’s study, but includes a control group. That may help clarify that a certain clustering of symptoms is specific to autism, she says.
Another inherent limitation is that the study relies on diagnostic codes, which are primarily used to communicate with insurance companies, and may be biased by non-medical factors.
“As a clinician, I can tell you that some codes are used more than others for reasons that are not just medically based,” says Sylvie Goldman, clinical assistant professor of pediatrics at the Albert Einstein College of Medicine in New York, who was not involved in the study.
The researchers acknowledge these limitations, and point out caveats of their own. For example, the age at which a code appears in the records does not necessarily match the age at which a child first experienced this symptom.
“There’s a ton of noise in the data. [A symptom] gets reported when both the patient reports it and the doctor chooses to record it,” says Finale Doshi-Velez, a postdoctoral associate in Kohane’s laboratory.
Still, using medical codes is one of the only ways to look at large enough numbers of participants for identifying patterns, Doshi-Velez says.“There’s a tradeoff with using perhaps much cleaner data, in which people are surveyed on a regular basis — but then you only have a very small population.”
Using larger numbers of participants would yield cleaner data from medical records, and may identify more subgroups, says Kohane. His team is trying to extract more information from medical records, by categorizing keywords from doctors’ written notes, as they may include observations not captured by the diagnostic codes.
Kohane’s team had initially hoped to analyze records from 496 boys with autism from Wake Forest Baptist Medical Center in North Carolina and compare results with the Boston group. But they concluded that the Wake Forest group is too small for subgroup analysis.
Combining and comparing information across multiple sites is crucial, as clinics may have certain specialties, such as epilepsy, and may attract children with those symptoms.
“Now it’s possible for many other health centers to do these studies, because many of them have electronic health records,” says Kohane. “We’re looking forward to collaborating with many of them to do some finer-grained work.”
1: Doshi-Velez F. et al. Pediatrics 133, e54-63 (2014) PubMed