Application of an Antimicrobial Stewardship Program in Brazilian ICUs Using Machine Learning Techniques and an Educational Model

Application of an Antimicrobial Stewardship Program in Brazilian ICUs Using Machine Learning Techniques and an Educational Model

Application of an Antimicrobial Stewardship Program in Brazilian ICUs Using Machine Learning Techniques and an Educational Model

Antimicrobial agents are frequently used empirically and include therapy for both Gram-positive and Gram-negative bacteria. In Brazil, multidrug-resistant Gram-negative pathogens are the cause of most nosocomial infections in ICUs. Therefore, the excessive use of antimicrobials to treat Gram-positive bacteria represents an opportunity to reduce unnecessary antibiotic use in critically ill patients. Besides, the success of a program aimed at reducing the use of antibiotics to treat gram-positive bacteria could also evolve to include other microorganisms, such as gram-negative bacteria and fungi. Analyzing data from the ICUs of the associated hospital network, high use of broad-spectrum antibiotics and vancomycin were observed, although MRSA infections rarely occur. Thus, if physicians could identify patients at high risk of infection by gram-positive bacteriaa reduction in antibiotic consumption could occur.. The more accurate treatments could result in better patient outcomes, reduce the antibiotics' adverse effects, and decrease the prevalence of multidrug-resistant bacteria. Therefore, our main goal is to reduce antibiotic use by applying an intervention with three main objectives: (i) to educate the medical team, (ii) to provide a tool that can help physicians prescribing antibiotics, and (iii) to find and reduce differences in antibiotic prescription between hospitals with low- and high-resources. To achieve these objectives, he same intervention will be applied in ICUs of two hospitals with different access to resources. Both are part of a network of hospitals associated with our group. First, baseline data corresponding to patient characteristics, antibiotic use, microbiological outcomes and current administration programs in practice at selected hospitals will be analyzed. TThen, a predictive model to detect patients at high risk of Gram-positive infection will be developed. After that, t will be applied for three months as an educational tool to improve medical decisions regarding antibiotic prescription. After obtaining feedback and suggestions from physicians and other hospital and infection control members, the model will be adjusted and applied in the two selected hospitals for use in real time. For one year, we will monitor the intervention and analyze the data monthly.

This proposal is a five-step quality improvement project. Analysis of baseline data [3 months]: Retrospective data will be collected from ten hospitals of Rede D'Or São Luiz. Patient characteristics, microbiological results and the use of antimicrobial agents will be analyzed. Stewardship programs currently in place will also be recorded. Development of the predictive model [3 months]: Collected data and machine learning techniques will be used to develop a predictive model to identify patients at risk of Gram-positive infection. This model will be evaluated using standard methods (e.g., accuracy and confusion matrix) and through clinical decision curves. This model will be embedded in an app and a web page to provide real-time guidance on the predicted probability of infection due to Gram-positive agents. Educational and calibration phase [3 months]: Firstly it will be used use the predictive model as a simulation tool to educate physicians. For three months, physicians will use the model to understand the main factors associated with Gram-positive infection. They will test the model using real-case data previously collected at the hospitals. The model will provide them information such as the probability of that patient having a Gram-positive infection and the proportion of infected patients in that ICU and hospital. After that, a meeting with all ICU and infection control members from participating hospitals will be held. A specific probability cutoff will be defined for starting gram-positive coverage. For example, the members can define that they feel comfortable not treating empirically gram-positive bacteria if the predicted probability is below a given threshold (say 5%). Quality improvement protocol will also involve other traditional methods to decrease antibiotic use, including audit feedback and daily remembrances to withdraw gram-positive antibiotic coverage. Educational material will be developed and provided for all sites, as well as in-site training. This phase will motivate the involvement of the hospital members, especially physicians, which can improve engagement to the intervention to be implemented afterward. Hopefully, it will also generate insights and feedback from the medical team to improve the tool to be implemented.

A predictive model to identify patients at risk of Gram-positive infection. This model will be embedded in an app and a web page to provide real-time guidance on the predicted probability of infection due to Gram-positive agents. The intervention will be implemented in two selected hospitals, aiming at monthly decreasing the use of broad-spectrum antibiotics while maintaining or reducing the ICU standardized mortality ratio and the standardized resource use.

Application of an antimicrobial stewardship program in Brazilian ICUs using machine learning techniques and an educational model

Firstly it will be used the predictive model as a simulation tool to educate physicians. For three months, physicians will use the model to understand the main factors associated with Gram-positive infection. They will test the model using real-case data previously collected at the hospitals. The model will provide them information such as the probability of that patient having a Gram-positive infection and the proportion of infected patients in that ICU and hospital. This model will be embedded in an app and a web page to provide real-time guidance on the predicted probability of infection due to Gram-positive agents. The intervention will be implemented in two selected hospitals, aiming at monthly decreasing the use of broad-spectrum antibiotics while maintaining or reducing the ICU standardized mortality ratio and the standardized resource use.

It was evaluated through the Defined Daily Dose (DDD): The assumed average maintenance dose per day for a drug used for its main indication in adults; and Duration of Treatment (DOT):Duration of Treatment with antibiotics.

It was evaluated through the Defined Daily Dose (DDD): The assumed average maintenance dose per day for a drug used for its main indication in adults; and Duration of Treatment (DOT): Duration of Treatment with antibiotics

Inclusion Criteria: prescribers from the hospital units participating in the study. Exclusion Criteria: prescribers who do not work in intensive care units. refusal to participate