Assistant Professor Wake Forest University School of Medicine Lewisville, North Carolina, United States
Background: Fragile X syndrome (FXS) is the most common inherited cause of intellectual disability and autism. There is no evident physical phenotype at birth, and the manifestation of the condition varies among patients, leading to challenges in diagnosis. Not all patients with FXS meet the current screening criteria (i.e., unexplained developmental delay, intellectual disability, and/or autism), which can result in a high rate of underdiagnosis. Additionally, most past studies on FXS are based on White patients, which could lead to a limited understanding of the clinical manifestation of FXS in patients from diverse groups. Incorporating a diverse patient population in research and developing more inclusive screening strategies is critical to achieving equitable health care for all children. Objective: Developing a machine learning algorithm for early and equitable detection of children with FXS. Design/Methods: This is a retrospective analysis of electronic health records (EHRs) from patients in Atrium Health, the clinical entity of the Southeast Region of Advocate Health Enterprise. Children diagnosed with FXS after age 5 were included in the analysis to ensure sufficient data was available for training the model. We selected a control group matched to the cases based on sex, race, and age. Using a supervised machine learning approach, we created a predictive model to distinguish FXS from the controls based on medical conditions reported 5 years prior to the FXS diagnosis. The area under the receiver operating characteristic curve (AUROC) was used to assess the success of the classification. Results: Thirty-seven children met the inclusion criteria, including 12 females and 25 males. Sixty-five percent of patients were White, 27% were Black/African American, and the rest declined to respond. We matched the cases to 1,850 controls. We were able to predict FXS five years prior to clinical diagnosis, with an AUROC of 0.915. This classifier performs better than our previous models (AUROC = 0.795), which primarily included a White adult population. The model successfully identifies patients who did not meet the key clinical criteria for testing but were later diagnosed with FXS.
Conclusion(s): Machine learning can assist in accelerating the diagnostic process for children with FXS from diverse populations. It identifies non-linear patterns in the data, which enable the detection of patients whose phenotypic profiles do not necessarily match the current screening criteria. Therefore, this approach can enable equitable detection of undiagnosed FXS cases.
