Dissertation committee:
Γεώργιος Φιλντίσης, Ομότιμος Καθηγητής, Τμήμα Νοσηλευτικής, ΕΚΠΑ (ΕΠΙΒΛΕΠΩΝ)
Παύλος Μυριανθεύς, Καθηγητής, Τμήμα Νοσηλευτικής, ΕΚΠΑ
Δημήτριος Καλλές, Καθηγητής, Σχολή Θετικών Επιστημών και Τεχνολογίας, ΕΑΠ
Κωνσταντίνος Τσουμάκας, Ομότιμος Καθηγητής, Τμήμα Νοσηλευτικής, ΕΚΠΑ
Βασίλειος Βερύκιος, Καθηγητής, Σχολή Θετικών Επιστημών και Τεχνολογίας, ΕΑΠ
Θεόδωρος Κατσούλας, Αναπληρωτής Καθηγητής, Τμήμα Νοσηλευτικής, ΕΚΠΑ
Παντελής Στεργιάννης, Επίκουρος Καθηγητής, Τμήμα Νοσηλευτικής, ΕΚΠΑ
Summary:
Introduction: Hospital-acquired infections, particularly in the critical care setting, have become increasingly common during the last decade, with Gram-negative bacterial infections presenting the highest incidence among them. Multi-drug-resistant Gram-negative infections are associated with high morbidity and mortality with significant direct and indirect costs resulting from long hospitalization due to antibiotic failure. Time is critical to identifying bacteria and their resistance to antibiotics due to the critical health status of patients in the intensive care unit. As traditional susceptibility tests require more than 24 hours after sample collection to determine susceptibility to specific antibiotics, we propose to apply machine learning techniques to help the clinician assess whether bacteria are resistant to individual antimicrobials before antimicrobial susceptibility testing is completed.
Aim: To apply and compare Machine Learning methods using data from the hospital information system and to develop antimicrobial resistance prediction models to support decisions about antimicrobial therapy.
Methods: Various artificial intelligence (Machine Learning) methods were applied and compared to hospital information system data on demographic data, culture results and antimicrobial susceptibility data of patients hospitalized in the ICU and other departments of a Greek hospital over three years. The five individual studies use different machine learning classifiers and techniques such as ClassBalancer and the Synthetic Minority Oversampling Technique (SMOTE) to deal with data imbalance. In addition, various data analysis techniques are used, including association rule mining with the Apriori algorithm and 10-fold cross-validation. Software tools such as WEKA and the R programming language are used to analyze and visualize the results.
Results: From the combined results of the five individual studies, an extensive use of machine learning methods for the assessment and prediction of antimicrobial resistance emerges. Various algorithms and techniques were used and evaluated based on indicators such as TP rate, FP rate, Precision, Recall, F-measure, MMC, area under ROC curve and PRC. Techniques such as kNN, polynomial logistic regression, Multilayer perceptron, JRip and regression classification models were highlighted for their strong performances on different performance measures. The results of the AutoML application confirmed the value of automated machine learning in finding robust predictive models (in particular StackEnsemble), with high performance on weighted metrics such as AUCW, APSW, F1W and ACC. The importance of various characteristics, such as type of antibiotic, sex, age and type of sample, was highlighted as a critical element in the prediction of antimicrobial resistance. Finally, by analyzing association rules based on minimum support and confidence thresholds, rules with particularly high confidence were extracted, revealing strong associations between data features and antibiotic susceptibility.
Conclusion: Applying machine learning algorithms to patient antimicrobial susceptibility data, readily available, from the hospital information system, even in resource-limited hospital settings, can provide informative antibiotic susceptibility predictions to aid clinicians in selecting appropriate empirical antibiotic therapy. These strategies, when used as a decision support tool, have the potential to improve empiric therapy selection and reduce the antimicrobial resistance burden.
