Developing an AI-based prediction model for anaphylactic shock from injection drugs using Japanese real-world data and chemical structure-based analysis

Abstract

Background

This study aims to develop an AI-based prediction model for injection drugs that cause anaphylactic shock using Japanese Real-World Data (JADER database) and chemical structure-based analysis.

Methods

Data sourced from the JADER database included adverse drug reaction reports from April 2004 to December 2020. Only drugs with an adverse reaction named "anaphylactic shock" were selected for analysis. For model building, various models were constructed to predict anaphylactic shock-inducing drugs, such as logistic regression, LASSO, XGBoost, RF, SVM, and NNW. These models used chemical properties and structural similarities as feature variables. Dimension reduction was applied using principal component analysis. The dataset was split into training (80%) and validation (20%) sets. Six different models were trained and optimized through fivefold cross-validation.

Results

From April 2004 to December 2020, 947 drugs with the adverse reaction name "anaphylactic shock" were extracted from the JADER database. 320 drugs were excluded due to analytical challenges, and another 400 were removed due to their administration route. 227 drugs were finalized as target medicines. For model validation, the performance of each model was evaluated based on metrics like AUCs of ROC curve, sensitivity, and specificity. Additionally, two ensemble models, constructed from the six models were assessed using bootstrap sampling. Interestingly, it was identified that mepivacaine structural similarity had the highest importance in the final model.

Conclusions

The study successfully developed an AI-based prediction model for anaphylactic shock inducing-injection drugs. The model would offer potential for drug safety evaluation and anaphylactic shock risk assessment.

Graphical abstract