Algorithm Predicts Mortality in Carbapenem-Resistant ICU Infections

The Lethal Threat of Resistant Gram-Negative Bacteremia in the ICU

Bloodstream infections caused by antibiotic-resistant bacteria represent a profound clinical challenge in the intensive care unit, driving substantial excess mortality (odds ratio 1.58), prolonged hospital stays, and increased direct medical costs averaging $12,442 per patient [1]. For patients who develop sepsis or septic shock from these pathogens, early recognition and the rapid initiation of optimal antimicrobial therapy are the cornerstones of survival, with recent data showing that combining rapid diagnostic tests with antimicrobial stewardship programs significantly reduces mortality (odds ratio 0.72) and accelerates time to optimal therapy by 29 hours [2, 3, 4]. However, infections driven by carbapenem-resistant gram-negative bacilli, particularly Klebsiella pneumoniae, Pseudomonas aeruginosa, and Acinetobacter baumannii, often defy standard empiric treatments and carry an exceptionally poor prognosis [1, 5]. Specifically, mortality is significantly higher when patients with carbapenem-resistant Klebsiella pneumoniae receive antimicrobial monotherapy rather than combination therapy (odds ratio 1.75), complicating initial treatment decisions [5]. While clinical guidelines emphasize aggressive acute management, physicians frequently struggle to accurately predict which patients will deteriorate despite maximal therapy [3]. A newly developed predictive model now offers clinicians a data-driven tool to forecast short-term mortality in this highly vulnerable patient population.

Tracking 15-Day Mortality in High-Risk ICU Patients

Multidrug and carbapenem-resistant gram-negative bacilli bloodstream infections cause exceptionally high mortality in intensive care units. Because accurately predicting mortality can improve treatment and support complex end-of-life decisions, researchers aimed to develop a machine learning model to predict outcomes in patients suffering from these specific infections. To build this prognostic tool, the investigators conducted a retrospective cohort study at a tertiary care medical center between 2017 and 2023. The study included 197 adult intensive care unit patients with confirmed bloodstream infections caused by multidrug and carbapenem-resistant Klebsiella pneumoniae, Pseudomonas aeruginosa, and Acinetobacter baumannii. The research team collected routine demographic, clinical, and laboratory data to train and validate the predictive algorithms. The clinical severity of the cohort was exceptionally high, as the 15-day mortality rate among included patients was 48 percent. This rapid and severe disease trajectory highlights the necessity for prognostic tools that utilize readily available hospital data to stratify patient risk and guide clinical interventions.

Training the Predictive Classifiers

To process the demographic, clinical, and laboratory data, the researchers first established baseline comparisons between patient groups. They utilized Mann-Whitney U and Chi-square tests to identify initial differences between survivors and non-survivors. Following these baseline comparisons, the investigators conducted a multivariable analysis using binary logistic regression to isolate independent mortality risk factors. This statistical foundation ensured that the subsequent computational models were built upon clinically relevant variables rather than statistical noise. Building on this groundwork, the research team evaluated 10 machine learning classifiers to determine which algorithm could most accurately forecast patient outcomes. To ensure the robustness of these models and prevent overfitting to the specific patient cohort, they employed stratified 5-fold cross-validation (a computational technique that repeatedly divides the dataset into five subsets for training and testing). Because physicians require transparent decision-making tools rather than opaque algorithms, the investigators interpreted the model predictions using SHapley Additive exPlanations (SHAP). This analytical method assigns a specific importance value to each clinical variable, allowing clinicians to see exactly which patient factors are driving the algorithm's mortality prediction.

LightGBM Accuracy and Key Prognostic Drivers

Among the tested algorithms, the Light Gradient Boosting Machine (LightGBM) classifier showed the highest accuracy in forecasting 15-day mortality. This tree-based machine learning model demonstrated high discriminative ability, achieving an Area Under the Receiver Operating Characteristic (AUROC) of 0.94 and an Area Under the Precision-Recall Curve (AUPRC) of 0.952. For clinical application, these metrics indicate the algorithm reliably distinguishes between patients who will survive and those who will succumb to their infections. The LightGBM model achieved an overall accuracy of 0.868, alongside a precision of 0.906 and a recall of 0.822, meaning it correctly identified a high proportion of true mortality cases while minimizing false alarms. Further statistical validation confirmed the model's robustness, as the LightGBM model achieved an F1 score of 0.855 (a harmonic mean of precision and recall) and a Matthews Correlation Coefficient (MCC) of 0.744 (a measure of the quality of binary classifications). Additionally, the algorithm's probability estimates were highly calibrated, reflected by the fact that the LightGBM model achieved a Brier score of 0.131, a metric where lower values indicate a smaller difference between predicted probabilities and actual patient outcomes. To make the algorithm's decision-making transparent for physicians, the researchers utilized SHapley Additive exPlanations (SHAP) analysis to identify the specific clinical variables driving the mortality predictions. The SHAP analysis revealed coagulopathy, rapid access to antibiotics, and septic shock as important predictive features for 15-day mortality. Beyond these primary drivers, the SHAP analysis also revealed Sequential Organ Failure Assessment (SOFA) score, platelet count, C-reactive protein (CRP) level, and time-related parameters as important predictive features. By translating routine physiological and treatment data into an accurate short-term prognosis, this model may support early intervention and assist in complex end-of-life decisions. For intensivists managing highly resistant gram-negative bacteremia, this provides a validated, data-driven tool to justify escalating aggressive therapies immediately or to facilitate timely, evidence-based discussions with families regarding care goals when the predicted likelihood of mortality is overwhelmingly high.

References

1. Allel K, Stone J, Undurraga EA, et al. The impact of inpatient bloodstream infections caused by antibiotic-resistant bacteria in low- and middle-income countries: A systematic review and meta-analysis.. PLoS medicine. 2023. doi:10.1371/journal.pmed.1004199

2. Singer M, Deutschman CS, Seymour CW, et al. The Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 2016. doi:10.1001/jama.2016.0287

3. Rhodes A, Evans L, Alhazzani W, et al. Surviving Sepsis Campaign: International Guidelines for Management of Sepsis and Septic Shock: 2016. Critical Care Medicine. 2017. doi:10.1097/ccm.0000000000002255

4. Peri AM, Chatfield MD, Ling W, Furuya-Kanamori L, Harris PNA, Paterson DL. Rapid Diagnostic Tests and Antimicrobial Stewardship Programs for the Management of Bloodstream Infection: What Is Their Relative Contribution to Improving Clinical Outcomes? A Systematic Review and Network Meta-analysis.. Clinical infectious diseases : an official publication of the Infectious Diseases Society of America. 2024. doi:10.1093/cid/ciae234

5. Li D, Rao H, Xu Y, Zhang M, Zhang J, Luo J. Monotherapy vs combination therapy in patients with Klebsiella pneumoniae bloodstream infection: A systematic review and meta-analysis.. Journal of infection and chemotherapy : official journal of the Japan Society of Chemotherapy. 2024. doi:10.1016/j.jiac.2024.02.007