# Prediction of infliximab and anti-drug antibody concentrations in patients with inflammatory bowel disease using machine learning models with real-world data from a prospective cohort study

## TL;DR

This study uses machine learning to predict drug and antibody levels in IBD patients, aiming to improve treatment personalization and reduce monitoring needs.

## Contribution

The novel contribution is the development of machine learning models for predicting infliximab and anti-drug antibody concentrations using real-world data.

## Key findings

- Random Forest and XGBoost models achieved high accuracy in predicting infliximab and anti-drug antibody concentrations.
- The anti-drug antibody model showed stable performance across multiple prediction steps.
- Machine learning models may support individualized dosing and reduce the need for frequent drug monitoring.

## Abstract

Although population pharmacokinetic models are the standard approach for identifying inter-individual variability and optimizing infliximab concentration, their development and validation are complex and time-consuming. Therefore, this study aimed to develop and evaluate machine learning (ML) models to predict infliximab and anti-drug antibody (ADA) concentrations in patients with inflammatory bowel disease (IBD) receiving maintenance infliximab therapy.

A total of 1,806 infliximab and ADA concentration measurements were prospectively collected from 149 IBD patients. Recurrent neural networks (RNN)-based models, including long short-term memory (LSTM) and gated recurrent unit (GRU) architectures, as well as regression-based models such as Elastic Net, Support Vector Regression, Random Forest (RF), and extreme gradient boosting (XGBoost), were developed. Recursive multi-step prediction was applied to evaluate short-term forecasting performance.

RF outperformed in predicting infliximab concentrations, and XGBoost yielded the best performance in predicting ADA levels (2-fold accuracy, 86.67% and 96.67%, respectively). The infliximab prediction model maintained acceptable accuracy up to two recursive predictions steps but exhibited a notable performance decline at the third step. In contrast, the ADA model showed robust performance across all three recursive steps, maintaining 2-fold accuracy exceeding 96%.

ML models were developed to predict infliximab and ADA concentrations, with RF and XGBoost showing the best performance for infliximab and ADA prediction, respectively. The ADA model demonstrated stable multi-step forecasting capability. These models may support individualized dosing strategies and reduce the need for frequent therapeutic drug monitoring in clinical practice.