Objectives <p>Emergency department (ED) overcrowding is a global challenge, emphasizing the need for early identification of high-risk patients. This study aims to develop a time-agnostic, interpretable predictive process monitoring approach to forecast the improvement status of emergency patients based on vital signs.</p> Background <p>Predicting patient pathways in the ED has been explored in previous research, often using predictive process monitoring (PPM) to predict future events or process durations. However, interpretable PPM approaches that predict patient outcomes using vital signs, independent of absolute timing, remain underexplored.</p> Methods <p>ED event logs from an academic medical center were encoded using three indexing strategies based on segments of the event sequence rather than absolute time. Features extracted from these sequences were used to train machine learning models, including XGBoost and Random Forest (RF), to predict patient improvement status.</p> Results <p>Among the indexing methods, the last three events yielded the highest predictive performance, while XGBoost achieved the best results with an AUC-ROC of 0.796. The first three events also provided valuable predictive information, demonstrating that early or late segments of a patient’s event sequence can be informative, even without considering absolute time.</p> Conclusion <p>This study demonstrates that a time-agnostic, interpretable PPM approach using vital signs can effectively predict emergency patient outcomes. By leveraging sequential event information, the proposed framework provides insights into both early and late stages of patient care, supporting improved risk stratification and decision-making in the ED.</p>