Identifying patients at high risk for antibiotic treatment following hospital admission: a predictive score to improve antimicrobial stewardship measures

Abstract

Purpose: Identifying patients for clinical studies evaluating strategies to reduce unnecessary antibiotic usage in hospitals is challenging. This study aimed to develop a predictive score to identify newly hospitalized patients with high likelihood of receiving antibiotics, thus improving patient inclusion in future studies focusing on antimicrobial stewardship (AMS) programs.

Methods: This retrospective analysis used data from the PILGRIM study (NCT03765528), which included 1,600 patients across ten international sites. Predictive variables for antibiotic treatment during hospitalization were computed, and an additive score model was developed using logistic regression and 10-fold cross-validation. The PILGRIM score was validated in an independent cohort (validation cohort), with performance metrics assessed.

Results: Data from 1,258 patients was included. In the development cohort 52.8% (n = 445) and in the validation cohort 42.4% (n = 134) of patients received antibiotics. Key predictors included hematologic malignancies, immunosuppressive medication, and past hospitalization. The logistic regression model demonstrated an area under the curve of 0.74 in the validation. The final additive score incorporated these predictors plus "planned elective surgery" achieving a specificity of 92%, a positive predictive value of 78%, a sensitivity of 41%, and a negative predictive value (NPV) of 69%in validation set.

Conclusion: The PILGRIM score effectively identifies newly hospitalized patients likely to receive antibiotics, demonstrating high specificity and PPV. Its application can improve future AMS programs and trial recruitment by facilitating targeted inclusion of patients, especially in the hematological and oncological setting. Further -external and prospective- validation is needed to broaden the model's applicability.

Keywords: Antibiotic treatment; Antimicrobial stewardship; Clinical trial; Prediction score.

Fig 1

Workflow of the explorative data analysis: Starting with the 3:1 non-random split of the available PILGRIM dataset, the score-development was continued with the development dataset by (1) identification and ranking of risk factors, (2) calculation of multiple regression models with identified risk factors in an iterative manner, (3) development of an additive score model based on the regression model with the best performance and (4) final score validation based on the formerly separated validation dataset, Created in https://BioRender.com

Fig 2

Illustration of the 10-fold cross-validation used for the logistic regression model described in step 2 of the methods section: based on the development dataset 10 filters were programmed, each separating the dataset in the ratio 9:1. The dataset size always remained the same. In each iteration of the regression models, the models were calculated 10-fold, each fold with different patients for the 90%-development and 10%-validation part, Created in https://BioRender.com

Fig 3

Process of iterative logistic regression model calculations incl. 10-folds cross validation: in the first iteration, all available variables were included and using 10-folds cross validation an average model performance was calculated. In the following iteration the least significant variable was removed (n- 1) from the model and the model was calculated again and so forth, Created in https://BioRender.com

Fig 4

Forest-plot of variables highly correlated with the endpoint (based on univariate regression models). depts. departments, cardiothor. surg. and card. cardiothoracic surgery and cardiology

Fig 5

Development of key performance metrics over the course of 41 logistic regression models starting with all available variables (iteration 1)

Fig 6

Score performance per threshold using integer weightings based on the Werfel method