Every day, emergency department providers make critical decisions—like whether to admit a patient—based on limited information. Researchers analyzing these decisions from electronic health records (EHR) face a similar challenge: crucial factors affecting decisions and outcomes are often unmeasured. Many standard statistical methods assume all relevant confounders are measured, leading to potentially misleading results. To solve this problem, we developed a mathematical modeling approach that uses proxy variables—data readily available in the EHR, such as vital signs—to indirectly estimate these otherwise latent confounders. Across multiple studies, our method more thoroughly accounts for confounding compared to standard analyses, frequently reversing their conclusions (e.g., demonstrating that admission is beneficial rather than harmful). Motivated by these findings, we created a broadly applicable, proxy-based framework that is both practical and robust to model misspecification. Our method offers researchers a novel and accessible tool for addressing unmeasured confounding in healthcare settings beyond emergency medicine.