Artificial Intelligence Models for Risk Analysis of Carbapenem-Resistant Klebsiella pneumoniae: A Comparative Performance Analysis Towards Logistic Regression

Chang Cai

This study investigates the application of Artificial Intelligence (AI) models to improve the prediction and management of Carbapenem-resistant Klebsiella pneumoniae (CRKP), a critical challenge in the context of Antimicrobial Resistance (AMR). The primary objective is to compare the performance of various predictive models—logistic regression, Classification and Regression Trees (CRT), Random Forest (RF), and Artificial Neural Networks (ANN)—in risk stratification and outcome prediction for patients with CRKP infections. Following this comparative evaluation, the study aims to identify key risk factors and prognostic determinants by analysing a dataset comprising genetic and clinical information from collected individuals. To support this objective, the research follows a structured methodological approach, beginning with the collection of clinical and whole-genome sequencing (WGS) data from individuals diagnosed with CRKP infections. To manage the complexity of high-dimensional genetic information, a Two-Step Clustering method was employed to condense the WGS data into interpretable variables suitable for integration with clinical features. This dimensionality reduction enabled a more holistic analysis of potential risk factors. Univariate analysis was then conducted to explore initial associations between variables and patient outcomes. A traditional logistic regression model served as the baseline for comparison, after which advanced AI models—including CRT, RF, and ANN—were developed and evaluated. To address the imbalance in the dependent variable, data balancing techniques, particularly boosting, were implemented. The CRT, RF, and ANN models were subsequently re-applied to the balanced dataset to assess improvements in predictive accuracy and robustness. The results demonstrate that the application of data balancing techniques substantially enhanced model performance by addressing class imbalances, thereby enabling more equitable training and improving the reliability of predictions across outcome categories. AI-based models, particularly RF and ANN, consistently outperformed traditional logistic regression in terms of predictive accuracy and discriminatory power. Among the evaluated models, RF and ANN exhibited the highest effectiveness in classifying CRKP infection risk and predicting patient outcomes. Following the implementation of data balancing, the ANN model showed the most notable improvement in predictive precision, while RF maintained a strong balance between sensitivity and specificity. These findings underscore the ability of AI to capture complex, non-linear relationships within integrated genomic and clinical datasets—analytical capabilities that traditional models often lack. In parallel, the analysis identified key risk factors associated with CRKP infection and prognosis, including carriage of the ST11 strain, intensive care unit (ICU) admission, and infection site, which emerged as significant predictors across multiple models. Together, these insights highlight the potential of AI-driven risk stratification tools to enhance CRKP surveillance, enable early identification of high-risk patients, and inform targeted clinical interventions. The implications for clinical practice are considerable. The integration of AI models into clinical decision support systems (CDSS) could enhance patient management, improve antimicrobial stewardship, and inform more effective infection control strategies. Furthermore, the study advocates for ongoing refinement of AI methodologies, including the development of hybrid modeling approaches and the incorporation of longitudinal patient data, to further enhance the robustness and clinical applicability of predictive models. This study highlights the potential of AI model techniques in AMR outcome predictive modelling, demonstrating improved risk assessment and prognosis for CRKP infections through the integration of genomic and clinical data. The findings support data-driven approaches to AMR management and underscore the importance of interdisciplinary collaboration.

Artificial Intelligence Models for Risk Analysis of Carbapenem-Resistant Klebsiella pneumoniae: A Comparative Performance Analysis Towards Logistic Regression

Files and links (1)

Abstract

Details

Metrics