Alina Zalounina

Performance of the TREAT decision support system in an environment with a low prevalence of resistant pathogens

OBJECTIVES To evaluate a decision support system (TREAT) for guidance of empirical antimicrobial therapy in an environment with a low prevalence of resistant pathogens. METHODS A retrospective trial of TREAT has been performed at Copenhagen University, Hvidovre Hospital. The cohort of patients included adults with systemic inflammation and suspicion of community-acquired bacterial infection. The empirical antimicrobial treatment recommended by TREAT was compared with the empirical antimicrobial treatment prescribed by the first attending clinical physician. RESULTS Out of 171 patients recruited, 161 (65 with microbiologically documented infections) fulfilled the inclusion criteria of TREAT. Coverage achieved by TREAT was significantly higher than that by clinical practice (86% versus 66%, P = 0.007). There was no significant difference in the cost of future resistance between treatments chosen by TREAT and those by physicians. The direct expenses for antimicrobials were higher in TREAT when including patients without antimicrobial treatment, while there was no significant difference otherwise. The cost of side effects was significantly lower using TREAT. CONCLUSIONS The results of the study suggest that TREAT can improve the appropriateness of antimicrobial therapy and reduce the cost of side effects in regions with a low prevalence of resistant pathogens, however, at the expense of increased use of antibiotics.

A Probabilistic Network for Fusion of Data and Knowledge in Clinical Microbiology

A problem in clinical microbiology is that of inappropriate antibiotic therapy. Various decision-support systems have been proposed to aid physicians in this domain, and we discuss the a priori advantages of using a probabilistic network over other approaches. The Treat project uses a probabilistic network to combine clinical signs, symptoms and laboratory results, and we discuss the problem of obtaining probabilities for the network. Finally, we consider how such a system can be tested in clinical practice and outline the results of our tests.

A stochastic model of susceptibility to antibiotic therapy-The effects of cross-resistance and treatment history

OBJECTIVE Selection of antibiotic therapy is a complicated process, depending on, among others, the effect of cross-resistance between antibiotics. We propose a model, which incorporates information about treatment history in the form of information on the success or failure of the current treatment and which combines this with data on cross-resistance to predict the susceptibility to future antibiotic treatments, thus providing a systematic basis for revision of antibiotic treatment. METHODS AND MATERIAL The stochastic model was built as a causal probabilistic network (CPN). Data used in the model were based on a bacteriology database including data on patient and episode unique pathogens cultured from a microbiological sample. RESULTS In this paper, we develop a CPN that can exploit knowledge about cross-resistance between two consecutive treatments, explore the properties of this CPN and consider how the CPN can be integrated into a complete decision support system for selection of antibiotic therapy. CONCLUSION The model presented may be useful both as a theoretical tool describing cross-resistance between antibiotics and as a part of complete decision support system for selection of antibiotic therapy.

Learning Susceptibility of a Pathogen to Antibiotics Using Data from Similar Pathogens

OBJECTIVES Selection of empirical antibiotic therapy relies on knowledge of the in vitro susceptibilities of potential pathogens to antibiotics. In this paper the limitations of this knowledge are outlined and a method that can reduce some of the problems is developed. METHODS We propose hierarchical Dirichlet learning for estimation of pathogen susceptibilities to antibiotics, using data from a group of similar pathogens in a bacteremia database. RESULTS A threefold cross-validation showed that maximum likelihood (ML) estimates of susceptibilities based on individual pathogens gave a distance between estimates obtained from the training set and observed frequencies in the validation set of 16.3%. Estimates based on the initial grouping of pathogens gave a distance of 16.7%. Dirichlet learning gave a distance of 15.6%. Inspection of the pathogen groups led to subdivision of three groups, Citrobacter, Other Gram Negatives and Acinetobacter, out of 26 groups. Estimates based on the subdivided groups gave a distance of 15.4% and Dirichlet learning further reduced this to 15.0%. The optimal size of the imaginary sample inherited from the group was 3. CONCLUSION Dirichlet learning improved estimates of susceptibilities relative to ML estimators based on individual pathogens and to classical grouped estimators. The initial pathogen grouping was well founded and improvement by subdivision of the groups was only obtained in three groups. Dirichlet learning was robust to these revisions of the grouping, giving improved estimates in both cases, while the group-based estimates only gave improved estimates after the revision of the groups.

Hierarchical Dirichlét Learning – Filling in the Thin Spots in a Database

Estimation of probabilities by classical maximum likelihood estimators can give unreliable results when the number of cases is small. A Bayesian approach, where prior probabilities with Dirichlet distributions are used to temper the estimates, can reduce the variance enough to make the estimates useful. This is demonstrated by using this approach to estimate mortalities of severe infections from different sites, lungs, skin urinary tract, etc. The prior probabilities are provided in a hierarchical way, i.e. by deriving them from the same database, but without distinguishing between different sites of infection.

Interpretative reading of the antibiogram - a semi-naïve Bayesian approach

BACKGROUND An antibiogram (ABG) gives the results of in vitro susceptibility tests performed on a pathogen isolated from a culture of a sample taken from blood or other tissues. The institutional cross-ABG consists of the conditional probability of susceptibility for pairs of antimicrobials. This paper explores how interpretative reading of the isolate ABG can be used to replace and improve the prior probabilities stored in the institutional ABG. Probabilities were calculated by both a naïve and semi-naïve Bayesian approaches, both using the ABG for the given isolate and institutional ABGs and cross-ABGs. METHODS AND MATERIAL We assessed an isolate database from an Israeli university hospital with ABGs from 3347 clinically significant blood isolates, where on average 19 antimicrobials were tested for susceptibility, out of 31 antimicrobials in regular use for patient treatment. For each of 14 pathogens or groups of pathogens in the database the average (prior) probability of susceptibility (also called the institutional ABG) and the institutional cross-ABG were calculated. For each isolate, the normalized Brier distance was used as a measure of the distance between susceptibility test results from the isolate ABG and respectively prior probabilities and posteriori probabilities of susceptibility. We used a 5-fold cross-validation to evaluate the performance of different approaches to predict posterior susceptibilities. RESULTS The normalized Brier distance between the prior probabilities and the susceptibility test results for all isolates in the database was reduced from 37.7% to 28.2% by the naïve Bayes method. The smallest normalized Brier distance of 25.3% was obtained with the semi-naïve min2max2 method, which uses the two smallest significant odds ratios and the two largest significant odds ratios expressing respectively cross-resistance and cross-susceptibility, calculated from the cross-ABG. CONCLUSION A practical method for predicting probability for antimicrobial susceptibility could be developed based on a semi-naïve Bayesian approach using statistical data on cross-susceptibilities and cross-resistances. The reduction in Brier distance from 37.7% to 25.3%, indicates a significant advantage to the proposed min2max2 method (p<10(99)).

Transferability modelling in the TREAT decision support system

One of the key-components for success of a decision support system is in its flexibility and applicability to different clinical locations. The present study is devoted to a system which is capable of successful transfer to a distant environment. We have developed a decision support system for antibiotic treatment (TREAT), which was adapted to four different hospitals in Europe. The system is based on a causal probabilistic network (CPN). The purpose of this paper is to present the models for transferability used in TREAT. The problem of transferability is addressed in the context of CPNs, emphasising the advantages of use of CPNs for solving the problem. The process of adapting TREAT is relatively easy; that is due to the modularity of the system. The system has been built using a modular architecture that allows rapid transfer of the system to different clinical environments. Such modularity can be archived by simple means which include the universal and modular structure of the CPN, the establishment of a large group of conditional probabilities in the CPN that are assumed to be independent of time and place, and the use of hierarchical Dirichlet methods for learning of data. Due to the universal structure of the CPN, the problem of transferability in TREAT concerns only the medical domain factors, not the topology of the system.

Model-Based Medical Decision Support - A Road to Improved Diagnosis and Treatment?

The hypothesis is advanced that model-based medical decision support is a methodology, which may be appropriate for construction of medical decision support systems. The methodology, which is based on a combination of structural modeling and decision theory is outlined through three medical applications.

Steps on the Road to Clinical Application of Decision Support --- Example TREAT

The decision support system TREAT advices on antibiotic treatment of severe infections. A multicenter randomized clinical trial has demonstrated that Treat reduces inappropriate treatment by 50%. This paper will show that TREAT satisfies several features closely correlated with decision support systemss ability to improve clinical practice. Examples of such criteria are: providing recommendations, not just assessments; transparent line of reasoning; convenience in use. Additional design features, such as transferability and addressing an important clinical problem, will also be discussed.

Design of a Method for Prediction of Susceptibility to Antimicrobials

Abstract The results from the in vitro susceptibility tests are usually available only for a limited set of antibiotics. This paper provides a practical method for prediction of a posteriori probability for coverage for all relevant antibiotics. The method combines susceptibility results for a limited set of antibiotics with data on cross-resistance between antibiotics. The Brier distance was used to measure the accuracy of the predicted coverages. Across all pathogen/antibiotic combinations in the derivation database, the Brier distances for a priori coverages was 39%, reduced to 25% for predicted a posteriori coverages, indicating that there is a significant advantage to the method proposed (p −99 ).