Abstracts

Rationale:
West Syndrome (WS) is an epileptic encephalopathy characterized by infantile spasms, hypsarrhythmia, and developmental regression with a peak incidence at 6 months. Adrenocorticotropic hormone (ACTH) is one of two FDA approved therapies for this potentially devastating condition, and is the drug of choice for some. However, predictors as to who will respond to ACTH remain elusive. With rising cost of ACTHar gel ($47,000/vial), it became imperative to identify those infants who are most likely to respond to ACTH.  We implemented machine learning to better predict which patients would be ACTH responders.
Method:
Retrospective chart review was performed on 74 infants younger than 12 months who presented with WS and treated with ACTH as first-line therapy at Children’s Healthcare of Atlanta from 2015-2019. Clinical data (demographics, development, seizure control at 2 weeks and 6 months from the time of diagnosis) were collected. Abnormal MRI was defined as etiologically relevant abnormalities. All patients had pre-treatment and post treatment EEG. Statistical analysis included Student t test, chi-square, Fischer’s exact test, and logistic regression where appropriate. Bonferroni correction set the significant p-value at < 0.0035.   For machine learning, 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 was used to identify which features were most helpful in improving the area under the curve (AUC) of the receiver operating characteristic (ROC) curve of the LASSO-regularized logistic regression model, with a good result at > 0.7.
Results:
Of 74 patients, 43 were males and mean age of onset of spasms was 5.4 months with mean age of treatment at 6.7 months. 34 (46%) patients had clinical cessation of spasms at 2-week follow up after ACTH treatment and were classified as responders. Responders were more likely to have normal development and less likely to have persistent seizures (both p < 0.0035) at 6 months from diagnosis.   Machine learning methods identified the three most critical historic factors for predicting treatment response at two weeks: corrected gestational age at onset (older age), hypoxic ischemic encephalopathy (HIE) and Trisomy 21, with an AUC of 0.71. Other patient characteristics, including etiology and time to treatment, did not differ between responders versus non-responders (Table 1), thus unhelpful in predicting treatment responses.
Conclusion:
Age of onset, HIE, and Trisomy 21 was identified by machine learning techniques as predictors of ACTH response with an AUC of 0.71. Responders were more likely to have normal development and less likely to have seizures at 6 months. Further prospective investigation is required to determine whether these predictive factors can indeed predict response to ACTH in WS.
Funding:
:NA