Early prognosis of ARDS is crucial in clinical practice, particularly within the first 48 hours following diagnosis, as the mortality rate of patients surpasses 5%, and the persistent prevalence rate exceeds 40%. Therefore, short-term monitoring of patients' ARDS status and the assessment of mortality risk are of utmost importance. The purpose of this study was to develop an algorithm that accurately predicts patient status assessment after 24 hours of ARDS onset by classifying patient assessment status into three categories: cured, ongoing episodes, and death. To accomplish this, we first conducted data preprocessing, screening out ARDS patients based on the Berlin definition, and extracting 59 available variable features, including 5 static variable features and 54 missing value features. Subsequently, we developed an XGBoost classification model based on multi-feature fusion, which was further improved by the Optuna optimizer and ensemble learning framework. The analysis was performed on the MIMIC-IV dataset, which provided data from 26,792 ICU patients. Finally, we used the eICU dataset which include 200859 patients from Physionet for external validation dataset, where the death prediction AUC was as high as 0.96, and the average AUC was 0.82. The proposed algorithm has the potential to aid clinicians in making accurate and timely decisions regarding patient management, including timely treatment intervention and the allocation of resources to reduce mortality rates. The study provides a foundation for future research in early ARDS prognosis prediction and the development of effective treatment strategies.