Abstracts

Rationale: Adrenocorticotropic hormone (ACTH) and vigabatrin (VGB) are first line treatments for infantile spasms (IS) but predictors for treatment response remain elusive. Rising costs of ACTH make finding the identifying features an imperative though. We used machine learning to predict responses to ACTH, VGB, and effects on 6-month outcomes.

Methods: Retrospective chart review of clinical data (2015-2019) of infants < 12 months with IS and treated with first line ACTH or VGB at Children’s Healthcare of Atlanta. VGB patients treated with non-ACTH or non-steroid as first line prior to VGB were also included to have enough data for machine learning. Responders were defined complete cessation of clinical spasms at 2 weeks for ACTH and up to 8 weeks for VGB. Tuberous sclerosis (TSC) was only treated with VGB and classified as structural and genetic. Statistical analysis performed using student’s t-test, Chi-square test, Fischer’s exact test, and logistic regression (STATA 16). Bonferroni correction set the significant p-value < 0.0038 for patient characteristics, and p < 0.017 for 6-month outcomes.

For machine learning (ML), a logistic regression model with Least Absolute Shrinkage and Selection Operator (LASSO) regularization was trained and evaluated through 10-fold cross validation (CV) with an inner, 10-fold CV loop to select the shrinkage parameter. Performing Monte Carlo simulations over random seeds, a feature selection algorithm identified the most helpful features to improve the area under the curve (AUC) of the receiver operating characteristic (ROC) curve of the LASSO-regularized logistic regression model, with a good result for an AUC ≥ 0.7.

Results: N=120. ACTH was first line in 74 and VGB in 46. At 2 weeks, 46% responded to ACTH and 54% to VGB. VGB responders were 63% at 8 weeks and fewer needed other antiseizure medications at 6 months than VGB non-responders (p = 0.003), but VGB response did not affect development or epilepsy at 6 months. ACTH group (57%) had more normal development than the VGB group (20%) at 6 months (p < 0.0001). Multivariate analysis associated ACTH first line with sex, corrected gestational age at onset, and normal development at onset with development at 6 months.

ML methods identified three critical characteristics to predict treatment response): corrected gestational age at onset, hypoxic ischemic encephalopathy (HIE) and trisomy 21 (T21), with an AUC of 0.71. Older age of spasm onset, HIE and T21 were associated with better response to ACTH. For predicting 6-month development, prognostic features of poor outcomes included severe HIE and abnormal development at time of diagnosis. Patients with other genetic or structural etiology had better development on ACTH than on VGB (AUC = 0.77). VGB patients had worse development overall compared to ACTH.

Conclusions: Machine learning identified predictors of ACTH versus VGB response. First-line ACTH showed more normal development compared to VGB, especially when HIE, T21, structural, and genetic etiology received first-line ACTH, except in TSC. Further prospective work is needed to corroborate these predictors of treatment response in IS.

Funding: Please list any funding that was received in support of this abstract.: Not funded.