# Development and validation of a predictive nomogram for severe adverse drug reactions: a dual-center pharmacovigilance study

## TL;DR

This study developed a predictive tool to identify patients at high risk of severe adverse drug reactions using machine learning and real-world data from two hospitals.

## Contribution

A novel nomogram for predicting severe adverse drug reactions was developed and validated using logistic regression and machine learning techniques.

## Key findings

- The logistic regression model showed the highest predictability with an AUC of 0.707 in the test set.
- The nomogram demonstrated good calibration with a p-value of 0.369 from the Hosmer-Lemeshow test.
- Twenty predictors, including age and specific medications, were identified as significant risk factors for severe adverse drug reactions.

## Abstract

Severe adverse drug reactions (SADRs) pose significant challenges to pharmacotherapy. Machine learning (ML) models hold promise in providing reliable solutions for predicting SADRs. This study is designed to pinpoint the independent risk factors contributing to SADRs through the application of ML techniques, thus constructing a predictive model for SADRs applicable in real-world clinical settings.

This retrospective dual-center cohort study analyzed adverse drug reaction (ADR) cases reported in two Chinese tertiary medical centers from 2014 to 2022. Per the World Health Organization - Uppsala Monitoring Centre severity criteria, cases were classified as SADRs or common ADRs. Independent predictors were identified via univariate and multivariate logistic regression (LR). A random partitioning of the data set resulted in a 75% training set and a 25% test set. The performance of three ML algorithms, including LR, Random Forest and Gradient Boosting Machine, was compared. A nomogram was constructed, model performance was measured by the area under the receiver operating characteristic curve (AUC), concordance index (C index), Hosmer-Lemeshow test (H-L test), Decision Curve Analysis (DCA), and Clinical Impact Curve (CIC).

A total of 508 SADRs were identified. The AUC values of LR model demonstrates the highest predictability among the three ML models. The AUC was 0.707 in the test set and the AUC in the training set was 0.689. A nomogram was established based on the LR model and evaluated. The C-index was 0.714 in the test set and the AUC in the training set was 0.713; The H-L test produced a chi-square value of 9.769 (p = 0.369), indicating good calibration. The DCA and CIC verify that the LR model possesses significant predictive value. According to the LR model, there were 20 predictors, including age ≥54 years, concurrent diseases ≥3, cardiac insufficiency, hemorrhagic disorders, active malignancies, cerebral infarction, bone fractures, anti-infectives, cytotoxic antineoplastics, proton pump inhibitors, antiepileptics, anticoagulants, diagnostic agents, arterial administration.

This study established a predictive nomogram for SADRs based on LR through comparative analysis of three ML approaches. The developed nomogram enables clinically meaningful risk stratification for SADRs, facilitating prophylactic surveillance of high-risk populations.