High efficiency classification of children with autism spectrum disorder

Abstract

Autism spectrum disorder (ASD) is a wide-ranging collection of developmental diseases with varying symptoms and degrees of disability. Currently, ASD is diagnosed mainly with psychometric tools, often unable to provide an early and reliable diagnosis. Recently, biochemical methods are being explored as a means to meet the latter need. For example, an increased predisposition to ASD has been associated with abnormalities of metabolites in folate-dependent one carbon metabolism (FOCM) and transsulfuration (TS). Multiple metabolites in the FOCM/TS pathways have been measured, and statistical analysis tools employed to identify certain metabolites that are closely related to ASD. The prime difficulty in such biochemical studies comes from (i) inefficient determination of which metabolites are most important and (ii) understanding how these metabolites are collectively related to ASD. This paper presents a new method based on scores produced in Support Vector Machine (SVM) modeling combined with High Dimensional Model Representation (HDMR) sensitivity analysis. The new method effectively and efficiently identifies the key causative metabolites in FOCM/TS pathways, ranks their importance, and discovers their independent and correlative action patterns upon ASD. Such information is valuable not only for providing a foundation for a pathological interpretation but also for potentially providing an early, reliable diagnosis ideally leading to a subsequent comprehensive treatment of ASD. With only tens of SVM model runs, the new method can identify the combinations of the most important metabolites in the FOCM/TS pathways that lead to ASD. Previous efforts to find these metabolites required hundreds of thousands of model runs with the same data.

Multivariate techniques enable a biochemical classification of children with autism spectrum disorder versus typically‐developing peers: A comparison and validation study

“…The FDA‐sub model had the best validation performance with a false positive rate of 19/154 at an expected false negative rate of 5%. As expected, these misclassification rates are higher than the false negative rate of 3.6% at a false positive rate of 2.6% found previously due in part to the absence of “% DNA methylation” and “8‐OHG” in the validation data, two of the most important variables for separating ASD and TD cohorts . CART and LR of a subset of metabolites produced results that were slightly worse than PCA with false positive rates of 38/154 and 32/154, respectively; both of these results are at an expected false negative rate of 5%.…”

“… Including the same variables in the training and validation data. Previous analyses found that “% DNA methylation” and “8‐OHG” were two of the most important variables for separation, but these data were not present in the validation set. Future studies should include these variables to allow for the highest possible classification accuracy as these classifiers are considered for clinical translation into a diagnostic test. Including both ASD and TD populations.…”

“…In addition to this variable the subsets chosen for FDA‐sub, LR‐sub, and CART also have three other variables in common: : “SAM/SAH,” : “Glu‐Cys,” and : “fCystine/fCysteine.” Furthermore, “% oxidized glutathione” highlights the importance of oxidative stress for classification while the “SAM/SAH” ratio is directly linked to DNA methylation and epigenetic components. As suggested previously, it is important to include variables that account for both FOCM (DNA methylation) and TS (oxidative stress) pathways in separating ASD from TD cohorts and this is what the performed analysis returned regardless of which technique was used.…”

“…While this presents a large number of variable combinations, this task was completed on a standard laptop computer and runtime was no more than an hour for a fixed number of variables. As the number of variables and/or complexity of the classifier increases, filter methods, such as the report by Li et al that used this training data, can be used to separate the variable selection and classification problems for improved computational efficiency. Furthermore, validation strategies such as testing on separate validation data sets and cross‐validation are needed to mitigate overfitting and estimate the model's predictive power.…”

“…As expected, these misclassification rates are higher than the false negative rate of 3.6% at a false positive rate of 2.6% found previously 29 due in part to the absence of "% DNA methylation" and "8-OHG" in the validation data, two of the most important variables for separating ASD and TD cohorts. 29,44 CART and LR of a subset of metabolites produced results therefore, these methods are likely not ideal for early-stage clinical data. Additionally, at least for the data sets under study in this work, we did not find advantages of using nonlinear classification techniques.…”