Health equity by design: The U-smile method prioritizes underrepresented populations

Abstract Underrepresented populations are often overlooked by prediction models biased toward the majority class and focused on overall accuracy. In public health, this bias can lead to poor targeting of high-risk groups and widened health disparities. Traditional variable selection methods rarely assess whether new predictors improve outcomes for specific subgroups. The U-smile method offers a solution: a class-stratified evaluation framework that quantifies and visualizes the impact of candidate variables on prediction quality in both event and non-event classes. This study examines U-smile's ability to prioritize predictors that enhance model performance for underrepresented populations. We applied U-smile to a binary classification task using the Heart Disease dataset (N = 661) and generated random variables. Logistic regression models were built in a nested setting, simulating class imbalance from 1% to 99% event prevalence. U-smile computed absolute (BA), relative (RB), and prevalence-weighted (I) coefficients to quantify prediction changes by class. U-smile results were compared with traditional evaluation metrics: ΔAUC, Brier score, F1-score, and net reclassification index (NRI), focusing on class-specific insights. U-smile identified variables that improved prediction for the minority class (e.g., RB₁ = +0.22, BA₁ = +0.15) despite negligible ΔAUC ( < +0.02). It also flagged non-informative variables that worsened minority-class predictions (e.g., BA₁ = –0.08). U-smile plots enabled rapid visual interpretation and guided selection of variables that traditional global metrics failed to detect as meaningful for vulnerable populations. U-smile shifts the paradigm in variable selection by evaluating predictors through a class-specific lens. It supports the development of fairer, more inclusive AI models and improves equity in digital health. By revealing overlooked improvements for underrepresented groups, U-smile strengthens decision-making in public health prediction tools. Key messages • The U-smile method helps select predictors that improve risk prediction for underrepresented populations. • Class-stratified evaluation reveals inequities hidden by traditional variable selection metrics.