An alternative model for assessing mortality risk in Stevens Johnson syndrome/toxic epidermal necrolysis using a random forests classifier: A pilot study

Abstract

Introduction: Mortality risk prediction is an important part of the clinical assessment in the Stevens-Johnson syndrome and toxic epidermal necrolysis (SJS/TEN) patient. The SCORTEN and ABCD-10 scoring systems have been used as predictive clinical tools for assessing this risk. However, some of the metrics required in calculating these scores, such as the total body surface area (TBSA) involvement, are difficult to calculate. In addition, TBSA involvement is calculated in a variety of ways and is observer dependent and subjective. The goal of this study was to develop an alternative method to predict mortality in patients with SJS/TEN. Methods: Data was split into training and test datasets and preprocessed. Models were trained using five-fold cross validation. Out of several possible candidates, a random forests model was evaluated as being the most robust in predictive power for this dataset. Upon feature selection, a final random forests model was developed which was used for comparison against SCORTEN. Results: The differences in both accuracy (p = 0.324) and area under the receiver operating characteristic curve (AUROC) (p = 0.318) between the final random forests model and the SCORTEN and ABCD-10 models were not statistically significant. As such, this alternative method performs similarly to SCORTEN while only requiring simple laboratory tests from the day of admission. Discussion: This new alternative can make the mortality prediction process more efficient, along with providing a seamless implementation of the patient laboratory tests directly into the model from existing electronic health record (EHR) systems. Once the model was developed, a web application was built to deploy the model which integrates with the Epic EHR system on the Fast Healthcare Interoperability Resources (FHIR) Application Programming Interface (API); this only requires the patient medical record number and a date of the lab tests as parameters. This model ultimately allows clinicians to calculate patient mortality risk with only a few clicks. Further studies are needed for validation of this tool.