Identifying Factors Contributing to Intractable Epilepsy in an Adult Population Using MIMIC IV Data
Abstract number :
1.237
Submission category :
4. Clinical Epilepsy / 4D. Prognosis
Year :
2022
Submission ID :
2204314
Source :
www.aesnet.org
Presentation date :
12/3/2022 12:00:00 PM
Published date :
Nov 22, 2022, 05:24 AM
Authors :
Monika Baker, BS – University of Utah; Samir Abdelrahman, MS, PhD – Assistant Professor, Biomedical Informatics, University of Utah; Joshua Bonkowsky, MD, PhD – Division Chief, Pediatric Neurology, University of Utah
Rationale: Thirty percent of persons with epilepsy become treatment-resistant (intractable) with life-long seizures, and require multiple medications or surgery. There are major gaps in the clinical understanding of intractable epilepsy, including an inability to identify patients at risk for treatment resistance. Our goal was to build a machine learning model that could assist in epilepsy intractability prediction using MIMIC IV, a dataset of de-identified adult visits from the Boston Beth Israel Deaconess Medical Center.
Methods: The model was developed using data only prior to and including the first visit coded for epilepsy. Data included demographics, insurance payer, ICD diagnoses, ICD epilepsy characteristics of a patient’s first epilepsy visit, and length of hospital stay. Intractability was then determined using data after the initial epilepsy visit (Figure 1). For model development, 4 different classification algorithms were tested: xgboost, random forest (RF), support vector machine (SVM), and decision tree. Data-preprocessing and model development was performed in Python.
Results: All models performed well on the testing set, with the highest F1 scores being 0.93 for SVM and 0.92 for RF. To determine which variables contributed most to intractability categorizations, we performed Shapley analysis on the RF model using TreeExplainer (Figure 2). According to our analysis, the Unspecified and Other variables were most significant, possibly due to the large proportion of patients that had this label at their initial seizure presentation. Length of Stay and External Causes were the other two variables positively associated with intractability.
Conclusions: Our best models performed well on categorizing and predicting patients at risk for intractability, with F1 scores over 0.90. In terms of variable analysis, the association of an increased Length of Stay with increased intractability was particularly interesting. This finding suggests that a more severe initial epilepsy presentation is associated with risk for treatment resistance.
Funding: Not applicable
Clinical Epilepsy