Support Vector Machine versus Multiple Logistic Regression for Prediction of Postherpetic Neuralgia in Outpatients with Herpes Zoster.

BACKGROUND

Postherpetic neuralgia (PHN), as the most common complication of herpes zoster (HZ), is very refractory to current therapies. Studies of HZ have indicated that early aggressive pain interventions can effectively prevent PHN; therefore, accurately predicting PHN in outpatients with HZ and treating HZ promptly, would be of great benefit to patients. Multiple logistic regression (MLR) has often been used to predict PHN. However, support vector machine (SVM) has been poorly studied in predicting PHN in outpatients with HZ.

OBJECTIVE

The aim of our retrospective study was to analyze the data of outpatients with HZ to evaluate the use of SVM for predicting PHN by comparing it with MLR.

STUDY

DESIGN

A retrospective study.

SETTING

Department of Anesthesiology in China.

METHODS

The data of 732 outpatients with HZ from January 1, 2015 to May 31, 2020 were reviewed. Risk factors for having PHN in outpatients with HZ were screened using least absolute shrinkage and selection operator (LASSO) algorithm. Then, SVM and MLR were used to predict PHN in outpatients with HZ based on screened risk factors. The data from 600 patients were used for training set and another 132 patients for test set. The receiver operating characteristic (ROC) curve was drawn from the 132 test set of patients. The prediction accuracy of the models was assessed using the area under curve (AUC).

RESULTS

The incidence of having PHN in outpatients with HZ was 19.4%. The risk factors selected by LASSO algorithm were gender, age, VAS scores, skin lesion area, initial treatment time, anxiety, sites of HZ (multiple skin lesions), types of HZ (bullous) and types of pain (knife cutting). The AUC for the SVM and MLR in test set were 0.884 versus 0.853. According to the ROC curve, the specificity and the sensitivity were 0.879 and 0.840 for SVM, and 0.780 and 0.840 for MLR, respectively.

LIMITATIONS: Retrospective study and relatively small sample size.

CONCLUSIONS

Both SVM and MLR had good discriminative power, but SVM has better performance in predicting PHN in outpatients with HZ, regarding the prediction accuracy and specificity.