Artificial Intelligence Tool May Help in Early Diagnosis of Fragile X

Pre-screening tool may identify potential cases 5 years sooner

Marisa Wexler, MS avatar

by Marisa Wexler, MS |

Share this article:

Share article via email
AI tool may help in early fragile X diagnosis | Fragile X News Today | illustration of two doctors looking at tablet

A new tool that uses artificial intelligence (AI) to analyze healthcare records may aid in the early diagnosis of fragile X syndrome, a new study reports.

“By incorporating a combination of co-occurring conditions, an AI-assisted pre-screening tool was developed and validated to identify potential cases at least 5 years earlier than the time of clinical diagnosis,” the researchers wrote.

The scientists said their AI tool was used successfully to analyze a healthcare database in the U.S. state of Wisconsin. Moving forward, this artificial intelligence-based program could be used across other healthcare databases to better — and potentially much sooner — identify individuals who may be affected by fragile X.

“Our AI-assisted pre-screening approach can facilitate and accelerate the clinical diagnosis of [fragile X syndrome] and decrease the duration of the diagnostic odyssey and degree of stress experienced by patients and their families,” the team wrote.

Recommended Reading
AI tool

Artificial Intelligence Tool May Help Speed Diagnosis

One concern in the move forward is that the AI tool thus far was only used in healthcare systems that were predominately comprised of white patients. More testing and validation is needed in other racial and ethnic patient populations, the researchers said.

The study, “Advancing artificial intelligence-assisted pre-screening for fragile X syndrome,” was published in BMC Medical Informatics and Decision Making.

Diagnosing fragile X

Fragile X syndrome can manifest very differently from person to person, which makes diagnosing the genetic disorder a challenge. Studies have shown a marked gap between the estimated prevalence of fragile X and the actual number of people diagnosed — suggesting that as many as 70% of people affected by fragile X syndrome have not been properly diagnosed.

To bridge that gap, a team led by scientists at the University of Wisconsin-Madison created an artificial intelligence tool aimed at better diagnosing fragile X syndrome. This tool is applied to data that is routinely collected in electronic healthcare records (EHRs).

Their aim was to identify data in such EHRs that could predict the diagnosis of fragile X, even before the disorder itself is formally diagnosed.

“The pre-screening model is not intended to be a replacement for genetic testing, but it can serve as a tool to automatically alert physicians about the presence of multiple [fragile X syndrome]-related phenotypes in the patient’s medical records,” the scientists wrote.

“By prompting the physician to further evaluate such individuals and refer them for genetic testing and counseling, our approach could accelerate the diagnostic process and be instrumental in identifying un-diagnosed individuals in the population and addressing their health conditions,” the team wrote.

To create the tool, the team used EHR data collected from 1979 to 2018 served by the Marshfield Clinic Health System in Wisconsin.

From the data, the team identified 55 people who had been diagnosed with fragile X syndrome at an age of 10 or older. The scientists also used data from 5,500 people without a fragile X diagnosis, who were similar to the patients in terms of age and sex.

For all of these patients, the researchers extracted data from five years before the formal diagnosis of fragile X syndrome, or the equivalent ages for controls.

“All data used in this study are directly collected in a medical setting and are in fact real world data from actual patients, providing further proof of [the AI tool’s] potential utility in real world clinical applications,” the team noted.

With these data in hand, the researchers then “trained” their artificial intelligence algorithm using a mathematical strategy called random forest. Conceptually, the AI tool uses a set of mathematical rules to look for patterns in the diagnostic codes that could differentiate between people with or without fragile X syndrome.

Recommended Reading
orphan drug status | Fragile X News Today | NN1-351 | illustration of woman using megaphone

FMR1 Gene Renamed to Increase Accuracy and Remove Stigmatizing Language

Testing the AI tool

To test the utility of the trained algorithm, the scientists tested it on data collected from UW Health, a separate healthcare system in Wisconsin.

“Our next step, reported here for the first time, was to evaluate the performance of this model in a new unseen dataset, i.e., an external validation study,” they wrote.

In this dataset, the team identified data for 52 fragile X cases and 5,200 people without the disorder, matched for sex and age.

To test the tool’s accuracy, the researchers calculated a statistical measurement called the area under the receiver operating characteristic curve, or AUROC. This is basically a measurement of how well a test can tell the difference between two groups — i.e., fragile X or not. AUROC values can range from 0.5 to 1, with higher values suggesting better ability to discriminate.

In the original Marshfield dataset, the AUROC for the AI tool was 0.798. In the UW Health analysis, it was 0.795.

“The AUROCs of the predictive models created and evaluated using the Marshfield cases and the UW Health cases were almost identical (0.798 vs. 0.795), representing the high level of reproducibility of results in different health care systems,” the scientists wrote.

“Our AI-assisted pre-screening tool could significantly improve the diagnostic process and could provide substantial benefits for patients, families and the health care system,” they concluded.

A noted limitation of this analysis was that nearly 90% of patients in both healthcare systems were white. The researchers highlighted a need to further validate this model in other populations, especially those of non-European ancestry.

“Additional studies on larger populations will provide more precise information on the performance of the model,” they wrote.