Methods: Data were drawn from a representative, epidemiological dataset held by the Marshfield Clinic. The Marshfield Clinic provides integrated, comprehensive care to 97% of the population in northern, central, and western Wisconsin. We retrospectively analyzed diagnostic codes associated with de-identified EHRs for 125 decedents with ASD and 125 sex-matched decedent community controls. We used a machine learning (random forest) algorithm to classify participants into groups (ASD or control) based on their ICD-9 codes, V-codes, and E-codes using a 10-fold cross validation procedure. Random forest, a robust and reliable classification method with low generalization error and high predictive performance, fits multiple decision trees and chooses a class that best aggregates the results of these trees. Diagnoses related to developmental disabilities and mental health conditions (i.e., Chapter 5: Mental Disorders) were excluded from our random forest models to reduce overclassification within the random forest algorithm.
Results: Diagnostic patterns distinguished decedents with ASD from matched decedent community controls with 86% accuracy solely based on their ICD-9 codes, V-codes, and E-codes in the five years before death, indicating that our machine learning algorithm can correctly predict whether participants were decedents with ASD or decedent community controls with greater than 6:1 odds. Patterns of diagnoses that differentiated decedents with ASD from matched decedent community controls included higher rates (in ASD) of epilepsy, long-term medication use, choking, injuries, skin conditions, infections, feeding and swallowing difficulties, cardiovascular screening, respiratory problems, dental problems, and non-specific lab tests and encounters and lower rates (in ASD) of osteoarthritis, peripheral vascular disease, and cancer screening, treatment, and diagnosis.
Conclusions and Implications: We found distinctive profiles of health problems among individuals with ASD in the five years before death. These findings may signify an underlying biological vulnerability in ASD or disparities in receipt of and access to care and effective prevention programs throughout life. Future work should further probe the possibility of underlying biological vulnerability in ASD. In addition, interventions aimed at reducing morbidity and increasing life expectancy should target health and lifestyle issues that may lead to increased morbidity, as well as adequate and effective communication between individuals with ASD and their health care providers.