In a recent study, electronic health record-based detection showed increased accuracy in early autism screening.
Autism detection based on electronic health records (EHR) can lead to increased accuracy for early autism screening, according to a recent study.
Children and families are more likely to receive the resources they need for autism management when diagnosed early. While the American Academy of Pediatrics recommends universal screening in children aged 18 to 24 months, the current median age of first diagnosis is 50 months. This makes most children with autism diagnosed too late to receive benefits from early support.
The Modified Checklist for Autism in Toddlers with Follow-up (M-CHAT-F) and its revision are the most common tools for early autism diagnosis, used in children aged 16 to 30 months. The sensitivity of the M-CHAT-F is 39%, while its positive predictive value (PPV) is 15%.
Other tools show improved sensitivity and PPV, but the need for improved accuracy has been observed in these tools. New approaches are needed to reduce biases leading to diagnostic disparities present in current methods.
EHR data has provided promising results for early, accurate diagnosis. To evaluate how the inclusion of EHR data affects early autism screening models, investigators conducted a diagnostic study from August 1, 2020, to April 2, 2022. Data was gathered from the Duke University Health System (DUHS) EHR.
Inclusion criteria included being born from October 1, 2006, to December 1, 2019, having at least 1 recorded encounter with DUHS before 30 days of age, and having at least 2 recorded encounters with DUHS before 1 year of age. Demographic information of participants was also gathered from EHR fields.
Computable phenotypes were used to identify autism spectrum disorder (ASD) and other neurodevelopmental disorders in participants. To be defined as a case, patients needed to have codes for the condition documented on 2 or more separate days and have a code for the condition linked to an encounter with a DUHS clinic specializing in neurodevelopmental disorders.
EHRs for patients when aged 30, 60, 90, 180, 270, and 360 days were used for prediction models. A developmental set was constructed to train models and tune hyperparameters, while a test set analyzed the model’s performance. Data was randomly divided into 1 of these 2 sets.
Performance was determined through the receiver operating characteristic curve, the mean PPV, and the concordance index.
There were 45,080 children in the study, 924 of which fit the criteria for ASD.Additionally, 175 patients with 1 diagnosis code related to autism but not fitting the criteria were included in a secondary analysis. Another 10,782 participants fit the criteria for a separate neurodevelopmental disorder, while 33,374 were used as control participants.
At 30 days, the EHR detection model showed 45.5% sensitivity and 23% PPV at 90% sensitivity. This changed to 59.8% sensitivity and 17.6% PPV at 81.5% sensitivity by day 360. These were meaningful levels of accuracy, making EHR a potentially valuable tool for accurate early autism screening.
Engelhard MM, Henao R, Berchuck SI, et al. Predictive value of early autism detection models based on electronic health record data collected before age 1 year. JAMA Netw Open. 2023;6(2):e2254303. doi:10.1001/jamanetworkopen.2022.54303