Covariate Selection Using Machine Learning for Clinical Models

Management Summary

Previously, covariate selection relied on time-consuming and computationally intensive model runs and could take several days. HMS developed an interactive application that combines statistical analysis and machine learning to systematically analyze, select, and evaluate relevant covariates. Model performance metrics, visualizations, and SHAP values support expert assessment of the results. The application reduces covariate selection from several days to minutes per analysis and generates reproducible analysis notebooks for further use.

Project summary

#lifesciences

Global Pharmaceutical Company

Industry: Pharmaceuticals
Project Start: 10/2025
Focus: Data Science & Machine Learning
Use Case: Predictive Machine Learning

Project Objective

Support covariate selection for pharmacometric models in preparation for NonMEM analyses

Key Features

Development of an interactive application for analyzing and selecting relevant external variables
Combining correlation analyses and machine learning methods for feature selection
Integration of model evaluation and visualization to support expert decision-making

Tech Stack

machine learning snowflake r r shiny shap xgboost catboost

We developed an application for structured covariate selection using machine learning, enabling significantly faster pharmacometric analyses.
Luis Wirth
Senior Data Scientist

The Starting Point

Clinical pharmacometricians analyze study data to identify suitable covariates for modeling pharmacokinetic relationships, such as drug absorption and distribution in the body. Selecting these covariates provides the basis for subsequent modeling and can, among other factors, inform dose selection.

Covariate selection is traditionally performed using nonlinear mixed-effects models. The underlying model runs are computationally intensive and require multiple iterations, meaning that the analysis can take several days. At the same time, evaluating the results requires a high level of domain expertise.

Simple process adjustments are not sufficient because the modeling itself remains computationally intensive and provides no standardized support for recurring analyses. Without suitable tools, the process remains time-consuming and highly dependent on individual experts.

The HMS Solution

ML-Based Analysis and Covariate Selection

HMS developed an application that supports the systematic evaluation and selection of relevant covariates for pharmacometric analyses.

The application analyzes relationships between potential covariates using mutual information as well as Pearson and Spearman correlations and visualizes the results as heatmaps with defined thresholds. Based on these results, subject-matter experts can select relevant covariates and exclude redundant variables, such as BMI and body weight.

ML-Based Feature Selection

The selected variables are fed into machine learning models and evaluated using cross-validation. In each iteration, the Boruta method identifies relevant variables by comparing them with randomized features. Variables that positively contribute to predictive performance are selected.

Model Evaluation and Result

The application uses SHAP values to show variable importance and the direction of each variable’s effect on the target variable. It also provides goodness-of-fit measures and model performance metrics. This information supports the assessment of model quality and the relevance of individual covariates.

Reproducibility for Regulated Environments

At the click of a button, the application generates a Quarto notebook that reproducibly documents the analyses performed and can be executed in an independent programming environment. This supports traceable and reproducible analyses in regulated environments.

Customer Benefits

Faster Covariate Selection with ML-Based Analysis

The ML-based application accelerates and streamlines covariate selection for pharmacometric models.

Faster

From Days to Minutes

The application reduces covariate selection from several days to minutes per analysis.

More Transparent

Traceable Decisions

Standardized evaluation criteria help subject-matter experts assess and select relevant covariates.

More reproducible

Reproducible Analysis Documentation

Automatically generated Quarto notebooks document the analyses performed and enable reproducible execution.

More Efficient

Targeted Preparation for Subsequent Modeling

Relevant covariates are systematically evaluated and selected for subsequent pharmacometric analyses.

Our Strengths, Your Benefits

ML Solutions for Clinical and Pharmacometric Analysis

In this project, HMS combines pharmacometric expertise, machine learning methods, and software development to create an interactive analytical application. The reproducible analysis process also addresses requirements relevant to regulated environments.

What You Can Expect from HMS

For similar projects, we combine domain expertise, machine learning, and software engineering to turn analytical requirements into integrated applications.

Development of analytical applications for clinical and pharmacometric use cases
Integration of machine learning methods into existing data environments
Development of traceable and reproducible analysis processes

Nicolai Henrichs

Principal Sales & Solutions Consultant | Partner

Are you planning a similar project?

Talk to us about your current situation and the requirements for a potential implementation. We’ll help you assess the project from a business and technical perspective and determine the best next steps.

Discuss the project

Related Services