Using machine learning to identify patterns of lifetime health problems in decedents with autism spectrum disorder

Abstract

Very little is known about the health problems experienced by individuals with autism spectrum disorder (ASD) throughout their life course. We retrospectively analyzed diagnostic codes associated with de-identified electronic health records using a machine learning algorithm to characterize diagnostic patterns in decedents with ASD and matched decedent community controls. Participants were 91 decedents with ASD and 6,186 sex and birth year matched decedent community controls who had died since 1979, the majority of whom were middle aged or older adults at the time of their death. We analyzed all ICD-9 codes, V-codes, and E-codes available in the electronic health record and Elixhauser comorbidity categories associated with those codes. Diagnostic patterns distinguished decedents with ASD from decedent community controls with 75% sensitivity and 94% specificity solely based on their lifetime ICD-9 codes, V-codes, and E-codes. Decedents with ASD had higher rates of most conditions, including cardiovascular disease, motor problems, ear problems, urinary problems, digestive problems, side effects from long-term medication use, and nonspecific lab tests and encounters. In contrast, decedents with ASD had lower rates of cancer. Findings suggest distinctive lifetime diagnostic patterns among decedents with ASD and highlight the need for more research on health outcomes across the lifespan as the population of individuals with ASD ages. As a large wave of individuals with ASD diagnosed in the 1990s enters adulthood and middle age, knowledge about lifetime health problems will become increasingly important for care and prevention efforts. Autism Res 2018, 11: 1120-1128. © 2018 International Society for Autism Research, Wiley Periodicals, Inc. LAY SUMMARY: This study looked at patterns of lifetime health problems to find differences between people with autism who had died and community controls who had died. People with autism had higher rates of most health problems, including cardiovascular, urinary, respiratory, digestive, and motor problems, in their electronic health records. They also had lower rates of cancer. More research is needed to understand these potential health risks as a large number of individuals with autism enter adulthood and middle age.

Keywords: aging; health; machine learning; mortality; older adult.

Figure 1. Age of death in decedents with autism spectrum disorder (ASD; N=131) and decedent community controls (N=16,981)

This figure displays the distribution (kernel density estimate) of age of death for the full sample of decedents with ASD and decedent community controls.

Figure 2. Random forest classifier performance based on lifetime ICD-9 codes, V-codes, and E-codes

This receiver operating characteristic (ROC) curve provides a comprehensive visualization of the performance of our predictive model. The area under the ROC curve (AUC) illustrates how well our random forest algorithm can distinguish between decedents with ASD and matched decedent community controls. The ROC curve displays the false-positive rate, or 1 – specificity versus sensitivity. The current classifier has an AUC of 0.88 which is significantly higher than the baseline AUC of 0.5.

Figure 3. Random forest classifier precision-recall curve

This precision-recall curve displays the model-wide relationship between precision and recall. Precision, also called positive predictive value, represents the ratio of correctly predicted positive cases to all cases that have been predicted as positive by the classifier. Recall (sensitivity) represents the ratio of correctly predicted target cases in relation to all cases in the target class.