Clinical assessment and prediction model construction for older patients with upper gastrointestinal bleeding

Objective To construct a clinical assessment prediction model for the safe discharge of older patients with upper gastrointestinal bleeding via machine learning algorithms. Methods A retrospective analysis was conducted on 1,032 older patients with upper gastrointestinal bleeding who were admitted through the emergency departments of Zhongda Hospital, Shanghai Changhai Hospital, Lishui Central Hospital, and Zhengzhou First Hospital from January 2018 to December 2020 to obtain relevant epidemiological, treatment, and prognostic data on Chinese older patients with upper gastrointestinal bleeding. This study aimed to clarify the comorbidities and medication history, onset conditions, causes, treatment interventions, and prognostic status of older patients with upper gastrointestinal bleeding, as well as the high-risk factors associated with death. A total of 218 older patients with upper gastrointestinal bleeding admitted to Zhongda Hospital from January 2021 to December 2022 who met the same criteria were subsequently selected as the validation group. A clinical assessment prediction model suitable for the older population was constructed via machine learning to predict safe discharge from upper gastrointestinal bleeding. Results The results revealed that the nine diagnostic variables of HB, SBP, INR, BUN, Alb, CR, HR, age, and CCI were used for algorithm modeling. The importance of the elements was consistent with the ranking in the machine learning algorithm. Among the machine learning algorithms, the random forest algorithm was the best clinical prediction model. For the validation group, the AUC of the RF model for the prediction of safe discharge was 0.889. The model prediction accuracy was 0.830 (0.781–0.880), the sensitivity was 0.868 (0.806–0.928), the specificity was 0.786 (0.705–0.867), the positive predictive value was 0.832 (0.765–0.898), and the negative predictive value was 0.828 (0.751–0.905), all of which were better than those of traditional scoring systems (CANUKA, AIMS65, MAP, ABC, and GBS). Conclusion Machine learning algorithms can form more accurate prediction models than traditional scoring systems.