Predictive model for initial response to first-line treatment in children with infantile epileptic spasms syndrome

Wenrong Ge, Lin Wan, Wang Zong, Lijun Fu, Guang Yang2025 · DOI

This study, published in the Italian Journal of Pediatrics, presents a machine learning model designed to predict how infants with Infantile Epileptic Spasms Syndrome (IESS) will respond to their first round of treatment. By leveraging an XGBoost algorithm on standard clinical data, the researchers achieved high predictive accuracy (AUC ~0.85), offering a practical tool for clinicians to decide between standard monotherapy and more aggressive combination treatments early in the diagnostic process.

TL;DR

Researchers developed a machine learning model that uses routine clinical data—like age of onset and MRI results—to predict if an infant with IESS will respond to initial treatment. This "data-driven" approach helps clinicians identify high-risk patients early, potentially bypassing ineffective monotherapy in favor of more aggressive combination treatments.

What Problem Does This Paper Address?

Infantile Epileptic Spasms Syndrome (IESS), formerly known as West syndrome, is a devastating form of early-childhood epilepsy. While the goal is rapid seizure control to protect neurodevelopment, the "first-line" treatments (usually hormonal therapy or Vigabatrin) only work about 30–40% of the time.

Clinicians face a dilemma: start with a single drug (monotherapy) to minimize side effects, or jump to a combination of drugs which is more effective but carries higher risks and costs. While previous research identified complex EEG patterns or blood markers as predictors, these require specialized engineering expertise not available in most clinics. This paper addresses the need for a predictive tool that uses "clinically accessible" data—information already sitting in a patient’s medical chart.

Methods

The researchers built their model using a retrospective cohort of 532 children. To handle the "data imbalance" (since most patients do not respond to initial treatment), they employed the Synthetic Minority Over-sampling Technique (SMOTE) to create a balanced training set.

They compared several machine learning architectures, including Random Forest, Support Vector Machines (SVM), and K-Nearest Neighbors (KNN), but found that XGBoost (Extreme Gradient Boosting) provided the best performance. To ensure the model wasn't a "black box," they used SHAP (SHapley Additive exPlanations) values to rank which clinical factors most influenced the predictions.

Datasets

The study utilized a dataset from a single major hospital in Beijing, China, covering patients treated between January 2018 and June 2023.

  • Total Cohort: 532 children with IESS.
  • Outcome Split: 160 responders vs. 372 non-responders.
  • Features: 16 independent variables, including sex, seizure type, age of onset, hypsarrhythmia presence, lead time (delay to treatment), and detailed MRI subtypes (e.g., malformations of cortical development, perinatal brain injury).

Key Insights and Ideas

  • The "Top Six" Predictors: The model’s performance was primarily driven by six factors: age of seizure onset, age of spasm onset, lead time, MRI subtype, treatment choice, and age at treatment.
  • The Importance of "Lead Time": Confirming previous clinical intuition, the model showed that the delay between the first spasm and the start of treatment is a critical determinant of success. Shorter lead times significantly increased the probability of a positive response.
  • Early Onset = Harder to Treat: Children whose seizures began very early (under 3 months of age) were less likely to respond to initial monotherapy, likely reflecting more severe underlying neurological issues.
  • MRI as a Proxy for Etiology: Since genetic testing isn't always available or timely, the model relies heavily on MRI subtypes. Structural abnormalities like Malformations of Cortical Development (MCD) were strong indicators that standard first-line drugs might fail.

Key Results

The XGBoost model outperformed all other tested algorithms across multiple metrics:

  • AUC (Area Under the Curve): 0.8516 ± 0.0165 (A score above 0.80 is generally considered excellent for clinical prediction).
  • Accuracy: 0.7836 ± 0.0229.
  • F1 Score: 0.7833 ± 0.0229.

Qualitatively, the SHAP analysis provided a "map" for clinicians, showing exactly how much each year of age or specific MRI finding shifted the probability of treatment success.

Limitations and Future Work

The authors acknowledge several hurdles:

  1. Retrospective Bias: The data comes from past medical records, which can be subject to inconsistent documentation by different physicians.
  2. Selection Bias: As a major referral center, the hospital saw many patients who had already failed treatment elsewhere, potentially skewing the "non-responder" data.
  3. Small Subgroups: The number of patients treated specifically with Vigabatrin (VGB) was relatively small compared to those on hormonal therapy.

Future Directions: The team aims to develop this model into an installable software tool that integrates directly with Electronic Health Records (EHR). They are also seeking multi-center validation to ensure the model works across different populations and healthcare systems.

Summary updated

Abstract

The study findings suggest that this model may help effectively predict the initial response to first-line treatment, supporting clinical decision-making for children with IESS. Key predictors such as the age of seizure onset and MRI subtype enable early, data-driven intervention strategies in clinical practice.
BibTeX
@inproceedings{ge2025predictive,
  author = {Wenrong Ge and Lin Wan and Wang Zong and Lijun Fu and Guang Yang},
  title = {{Predictive model for initial response to first-line treatment in children with infantile epileptic spasms syndrome}},
  booktitle = {˜The},
  year = {2025},
  doi = {10.1186/s13052-025-01959-z},
}
Figure