Medical Student Wayne State University School of Medicine
Rationale: Identifying epilepsy patients who are at the highest risk for near-future hospitalization or emergency medical care would be invaluable for reducing healthcare costs and protecting patients. Machine learning applications for predictive modeling have been used in epilepsy to predict outcome of surgical and medical managements. We aimed to utilize this technology on an established epilepsy clinic database to make predictions about future hospitalization risk. Methods: Our dataset consisted of 976 single-center epilepsy clinic visits from 258 patients from 2014 to 2016. We included only patients with two or more visits. We loaded patient visit data into a SQL Server relational database and aggregated it into approximately 2500 datapoints per visit. Patient data consisted of information such as demographics, medication history, seizure etiology, seizure frequency, duration since last seizure, EEG findings, surgery history, family history, social history, comorbidities (aggregated with Charlson comorbidity index) etc. We included calculated datapoints for duration between visits. We omitted data related to each patient's most recent visit, except for epilepsy-related hospitalization, which was the dependent variable used to train our model. We then trained the model with an apriori algorithm in Microsoft SQL Server Analysis Services (an enterprise-grade multidimensional datamining platform). For validation, we generated ten random-seed models with 70/30 holdout partitions. For each model, we calculated confusion matrices and generated receiver operating characteristic (ROC) curves in R with pROC. Results: Two hundred fifty-eight patients had a mean age of 44.17 years (SD=15.14) and 63% were female. Blacks constituted 78% of patients. Most common seizure types were GTCs (41%), focal with impaired awareness seizures (31%) and focal to bilateral tonic-clonic seizure (31%). Most common etiology was unknown (33%) followed by TBI (21%). Patients were on an average of 1.23 active AEDs (SD=0.90). The model successfully classified patients needing hospitalization between their most recent and second-most recent visits with 83% accuracy (95% CI 0.78 to 0.88). Positive and negative predictive values were 75% (95% CI 0.68 to 0.83) and 90% (95% CI 0.86 to 0.94) respectively. Mean area under the ROC curves (Figure 1) was 75% (95% CI 0.74 to 0.76). The mean duration between the most recent and second-most recent visit was 162 days (95% CI 148 to 176). Conclusions: We developed a promising model to predict epilepsy hospitalization based solely on patient demographic and clinical variables. These results suggest that a model such as this may assist clinicians in decision-making and identification of at-risk patients for counselling and providing rescue medications. A highly valid and accurate model could also allow preemptive increase in dose of AEDs. The results also suggest there may be important features found in patient histories that could be utilized in a hospitalization risk stratification index. Further development of this model to identify most-weighted variables and utilizing additional longitudinal data is planned. Funding: Please list any funding that was received in support of this abstract.: None Click here to view image/table
