A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients withPseudomonas aeruginosa

A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa LV Xiao-chun , Ren Qi , Zhu Lihong , CHEN Kun , WANG Jian-bing , CHEN Fang , JIN Kai-ling , View ORCID Profile LIN Kai doi: https://doi.org/10.1101/2025.06.19.25329970 LV Xiao-chun 1 Intensive Care Unit, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Ren Qi 1 Intensive Care Unit, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site Zhu Lihong 1 Intensive Care Unit, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site CHEN Kun 2 School of Public Health, Zhejiang University , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site WANG Jian-bing 2 School of Public Health, Zhejiang University , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site CHEN Fang 3 Department of Healthcare-associated Infection Management, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site JIN Kai-ling 3 Department of Healthcare-associated Infection Management, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site LIN Kai 3 Department of Healthcare-associated Infection Management, Zhejiang Hospital , Hangzhou, Zhejiang, China Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for LIN Kai For correspondence: l.kai.cln.epi{at}gmail.com Abstract Full Text Info/History Metrics Data/Code Preview PDF Abstract Background Carbapenem resistance in Pseudomonas aeruginosa is increasing in intensive care units (ICUs). To enhance antimicrobial stewardship and infection control, we aimed to develop and validate a real-time interpretable hybrid Automated Machine Learning (AutoML) ensemble for multi-outcome prediction. Methods We retrospectively analyzed 847 adult ICU admissions with P. aeruginosa isol ates at a tertiary hospital in Hangzhou, China (January 2018 to December 2024). After a three-stage VTF-MI-L1 feature selection pipeline, XGBoost, LightGBM, CatBoost, random forests, and linear/logistic regression were used as base learners and combined via Bagging, Voting, Stacking, and Gradient Boosting. Nested five-fold cross-validation was used to assess model performance (AUC for classification; MSE, RMSE, MAE, and R 2 for regression). Interpretability was provided by SHAP values, and the inference latency was recorded. Results For carbapenem resistance rate (CRR) prediction, the CatBoost regressor (cRMSD = 0.1663; r = 0.8849; R 2 ≈ 0.78) and the Voting Regressor (cRMSD = 0.1675; r = 0.8838) outperformed all other models ( p < 0.05). XGB-R achieved the best accuracy and computational efficiency for the last two tests of the CRR of P. aeruginosa (CRR-PA-Last2) ( p < 0.05). In predicting ICU length of stay, XGB-R led with r = 0.9724, cRMSD = 55.7 d, and σ-ratio = 0.88, significantly surpassing Bagging and CatBoost regressors ( p < 0.05). XGB-R also yielded the lowest composite error for the ICU-to-death interval (cRMSD = 205.1 d; r = 0.7741; σ-ratio = 0.71), again outperforming Bagging and CatBoost (p < 0.05). Across all four regression outcomes, XGB-R obtained the lowest average rank (1.95), whereas CatBoost and Voting regressors showed particular strengths in predicting resistance. SHAP analysis identified age, carbapenem exposure intensity, and duration of mechanical ventilation and catheterization as the key positive contributors. All top-ranked models required < 50 ms per inference, meeting the bedside real-time requirements. Conclusions The proposed hybrid AutoML ensemble delivered highly accurate, interpretable, and millisecond-level predictions of diverse resistance-related outcomes, underscoring its potential for ICU antimicrobial stewardship and infection control. Multicenter prospective studies are warranted to confirm the generalizability of these findings. 1 Introduction Antimicrobial resistance (AMR) is a phenomenon in which target pathogenic microorganisms (e.g., bacteria, fungi, viruses, or parasites) develop resistance to antimicrobial agents because of genetic mutations, selection pressures, or changes in biological adaptation mechanisms, leading to the failure of standard treatment regimens [ 1 , 2 ]. AMR greatly complicates the clinical management of infectious diseases, prolongs hospital stays, increases healthcare costs, and increases mortality rates [ 3 ]. The rapid emergence and global dissemination of multidrug-resistant organisms (MDROs)—often termed “superbugs”—have posed serious challenges to healthcare systems and public health globally [ 4 , 5 ]. Excessive and irrational use of antimicrobial agents is the primary driver of AMR development and dissemination [ 6 ]. Intensive care units (ICUs), which are crucial hospital departments treating critically ill patients, have an antimicrobial usage intensity more than tenfold that of general wards. Over 70% of ICU patients receive antimicrobial therapy during hospitalization, often involving combinations or sequential use of multiple antimicrobial agents [ 7 – 9 ]. Consequently, approximately 35% of ICU patients develop MDRO infections and nearly half of these infections are classified as healthcare-associated infections (HAIs) [ 10 ]. Pseudomonas aeruginosa ( P. aeruginosa ) is one of the most challenging multidrug-resistant pathogens in HAIs, notably posing significant risks within ICUs [ 11 , 12 ]. Data from the Gulf Cooperation Council countries indicate that resistance rates of P. aeruginosa to carbapenems exceed 30%, and the 30-day mortality rate among infected patients is as high as 68.8%. The World Health Organization (WHO) has also listed carbapenem-resistant P. aeruginosa (CRPA) as a “critical priority” pathogen, making its control and treatment among the most urgent issues in the management of global AMR [ 13 ]. The rapid advancement of artificial intelligence (AI) has facilitated the development of clinical prediction tools with high accuracy and good generalizability [ 14 ]. Nevertheless, existing clinical AI research predominantly focuses on traditional clinical endpoints such as mortality, readmission rates, and length of hospital stay. For instance, Olang et al. reviewed 400 ICU studies and identified that only 16 employed AI methodologies, all of which solely utilized mortality as the outcome measure, with a maximum reported area under the curve (AUC) of 0.929; however, AMR was not considered [ 15 ]. Similarly, a large-scale single-center investigation only emphasized length of ICU stay and mortality outcomes without addressing the risk of AMR [ 16 ]. Broader quality assessments also reflect this trend: among 338 clinical AI models evaluated by de Hond et al., 67% targeted mortality, hospitalization duration, or readmission rates, while research addressing drug-related risks, particularly AMR, was described as extremely scarce [ 17 ]. Conversely, two recent systematic reviews targeting AMR prediction models included only 21 to 27 original studies, typically characterized by limitations such as small sample sizes, single-center studies, offline genomic validation only, absence of external validation, and lack of bedside integration [ 18 , 19 ]. Even in antimicrobial stewardship scenarios, among 18 AI-supported antimicrobial stewardship program (AI-ASP) tools evaluated by Pinto-de-Sá et al., which reported AUC values ranging from 0.64 to 0.992, functionalities were primarily restricted to prescription auditing or inappropriate medication detection. Therefore, the capacity for AMR risk prediction and meaningful clinical translation remains limited [ 20 ]. Given these limitations, this study aimed to introduce an innovative multiscale framework to predict AMR risks. The proposed framework integrates Intermediate Feature Fusion with Automated Machine Learning (AutoML) algorithms, to enable precise AMR risk prediction and highly interpretable outputs for ICU patients, ultimately improving the clinical needs of antimicrobial stewardship. 2 Materials and methods 2.1 Data 2.1.1 Data description This retrospective study analyzed the clinical data of patients admitted to our hospital’s ICU from January 1, 2018, to December 31, 2024. Complete medical records of 3,293 patients were collected. The following patients were included: (1) those aged ≥18 years; (2) with an ICU stay of at least 24 h; and (3) at least one specimen (blood, urine, secretions, or excretions) positive for P. aeruginosa during admission. We excluded patients with missing clinical data. Ultimately, 847 patients met the inclusion criteria. Collected variables included the following: Demographic information: Age, sex, and length of ICU stay (ICU-LOS). Comorbidities and medical history: Number of pre-diagnosed comorbidities at admission [ 21 ]. Surgical and invasive procedures: The total number of surgical procedures performed during hospitalization, duration of central venous catheterization (CVC-days), duration of mechanical ventilation (MV-days), and duration of urinary catheterization (UC-days). Antimicrobial prescriptions: Type, dose, administration route, and treatment duration. Microbiological culture results: Microbiological reports from all specimens collected during hospitalization. Antimicrobial susceptibility testing (AST): AMR definitions followed the guidelines of the WHO and the National Health Commission of China [ 1 , 22 ]. AST procedures strictly adhered to the Clinical and Laboratory Standards Institute (CLSI) and European Committee on Antimicrobial Susceptibility Testing protocols, utilizing the minimum inhibitory concentration to determine resistance. Outcome measures: Including antimicrobial susceptibility of the last isolate of P. aeruginosa before discharge (LastPaAST, classified as resistant [R] or susceptible [S]), in-hospital mortality (IHM), ICU length of stay (ICU-LOS), time from ICU admission to death (ICU-Death interval), average carbapenem resistance rate of the last two isolates (CRR-PA-Last2), and carbapenem resistance rate (CRR). 2.1.2 Indicator calculation and data preprocessing (1) Resistance rate calculation CRR serves as the primary indicator for assessing AMR risk. The range of carbapenems was defined according to the CLSI standards (CLSI M100) [ 23 ]. The CRR of P. aeruginosa was calculated using Equation: where n res represents the number of isolates resistant to at least one carbapenem and N total is the total number of isolates tested. (2) Anatomical Therapeutic Chemical/defined daily dose Calculation Antimicrobial usage was standardized using the Anatomical Therapeutic Chemical (ATC) classification system and a defined daily dose (DDD) methodology [ 24 ]. Standardization was quantified using the ATC/DDD per day of therapy (DOT), calculated as follows: Here, ATC/DDD represents the defined daily dose and DOT refers to the number of days of therapy. 2.2 Data preprocessing and feature selection ICU datasets often exhibit high dimensionality, noise, and multicollinearity, which complicate the stable identification of key features [ 25 ]. As shown in Fig 1 , we propose a three-stage feature selection method called VTF-MI-L1. Initially, near-zero variance features were removed using a Variance Threshold Filter (VTF). Next, Mutual Information (MI) identified nonlinear relationships, preserving the candidate features with maximal contributions. Finally, L1 regularization generated sparse solutions, addressing multicollinearity and enhancing interpretability [ 26 , 27 ]. This process reduced the number of feature dimensions from 93 to 48 without compromising the predictive performance. Download figure Open in new tab Fig 1. Feature selection flowchart. First, duplicate records were removed and variables were categorized into numerical and categorical types. Missing numerical data were imputed using median values, whereas mode imputation was used for categorical variables. Numerical variables underwent Z-score normalization [ 28 ], and categorical variables were encoded using one-hot encoding [ 29 ]. These preprocessing steps were implemented using ColumnTransformer to produce dense feature matrices for modeling. Second, a VTF eliminates variables with near-zero variance, thereby eliminating features that lack substantial predictive information. Third, a mutual-information univariate filter was applied. For the regression tasks, Pearson’s correlation coefficients and MI scores were computed while retaining the top-quartile features. For classification tasks, ANOVA F-values and MI scores were used in a similar manner. Fourth, we employed Embedded L1 Regularization. The regression models utilize Least Absolute Shrinkage and Selection Operator (LASSO) regression with cross-validation, retaining nonzero coefficient features. The classification models used logistic regression with an L1 penalty. Finally, the regression and classification feature sets from these steps were subjected to union and intersection operations, producing robust and comprehensive feature combinations for modeling. 2.3 Design and implementation of AutoML To systematically evaluate the predictive performance, an AutoML framework was developed for regression and classification tasks ( Fig 2 ). The design integrated diverse ensemble strategies and robust cross-validation to enhance prediction accuracy and generalization. For post-standardized preprocessing, the model-building module utilized various base learners as follows: for classification, XGBoost Classifier (XGB-C), LightGBM Classifier (LGB-C), CatBoost Classifier (CAT-C), Random Forest Classifier (RF-C), and logistic regression (LR); for regression, XGBoost Regressor (XGB-R), CatBoost Regressor (CAT-R), LightGBM Regressor (LGB-R), Random Forest Regressor (RF-R), Linear Regression (LR), and Voting Regressor (VR). Download figure Open in new tab Fig. 2 Model architecture framework diagram. Four mainstream ensemble methods were evaluated to enhance model robustness and generalization: Bagging: a parallel ensemble method in which multiple base learners are trained on bootstrap samples drawn with replacements from the original training set. The final prediction is calculated by averaging the outputs of the regression tasks or majority voting for the classification tasks. For the bagging regression model (BR), the ensemble prediction y ^ bag is calculated using (3) : where N represents the number of base models, and y ^ i denotes the prediction of the i -th model. For the bagging classification model (BC), the ensemble prediction is described by (4) : where “ mode ” represents the most frequent predicted class. Voting: a method in which the predictions of multiple base models are aggregated by weighted averaging for regression or weighted probabilities for classification (soft voting). For the voting regression model (VR), the final prediction was calculated using (5) : where w i is the weight assigned to the i -th model. For the voting classification model (VC), the class probabilities are aggregated and the predicted class is described by (6) : where p ( c ) i represents the predicted probability of class c by the i -th model. Stacking: A hierarchical ensemble method in which predictions from multiple base models are used as new features for a meta-learner. The final prediction is described by (7) : where f meta represents the metamodel trained on the out-of-fold predictions of the base models. In this study, stacking was implemented in two forms—the stacking classifier (SC) and the stacking regressor (SR)—corresponding to classification and regression tasks, respectively. Boosting: A sequential ensemble method in which each new base learner is trained to fit the residuals or gradients of the preceding learners. The ensemble prediction after m iterations is represented by Equation (8) : where y ^ ( m ) ( x ) represents the prediction at iteration m , η denotes the learning rate, and h m ( x ) is the newly fitted weak learner. For regression and classification outcomes, the Boosting component employed two symmetrical implementation strategies: Continuous outcomes (regression): Gradient Boosting Regression Trees (GBRT) were utilized, with the Mean Squared Error (MSE) as the optimization objective. This model effectively captures nonlinear relationships by iteratively fitting residuals, demonstrating superior performance in numerical indicators such as AMR rates. Discrete outcomes (Classification): Gradient Boosting Decision Trees (GBDT) were implemented with log-likelihood as the optimization objective. This approach generates posterior probabilities, facilitating the evaluation of threshold-independent metrics such as the ROC-AUC. These methods were integrated into an end-to-end AutoML pipeline, featuring unified preprocessing and automated base learner mapping to ensemble meta-learners, and managed via the scikit-learn pipeline framework. The models were trained under consistent computing conditions with predefined hyperparameters to minimize variability and ensure reproducibility. 2.4 Evaluation methods and statistical analysis A nested cross-validation framework was employed to systematically evaluate the predictive performance and generalization capability of both the regression and classification models. For regression tasks, 5-fold cross-validation was utilized, with model performance assessed through metrics, including the Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Coefficient of Determination (R 2 ). The statistical significance of the R 2 values was assessed using an F-test [ 30 , 31 ]. Stratified 5-fold cross-validation was adopted for the classification tasks. The evaluated metrics included accuracy, precision, sensitivity, specificity, F1-score, and the Area Under the Receiver Operating Characteristic Curve (ROC-AUC). ROC curves were constructed using out-of-fold predicted probabilities to estimate the AUC. In the regression tasks, four continuous outcome variables—CRR, CRR-PA-Last2, ICU-Death interval, and ICU-LOS—were analyzed using a unified 5-fold cross-validation procedure, with MSE, RMSE, MAE, and R 2 calculated for each. Each outcome-metric combination was treated as an independent dataset and eight regression models (XGB-R, LGB-R, CAT-R, RF-R, BR, GBRT, VR, and SR) were ranked. The Friedman test was employed to evaluate the overall differences among these models, with pairwise comparisons subsequently performed using the Nemenyi post-hoc test [ 32 ]. In addition, SHapley Additive exPlanations (SHAP) were introduced to interpret models and elucidate the relative importance and directional effects of features on model predictions [ 33 ]. To further assess consistency across multiple models and predictive tasks, Taylor diagrams were used to visualize correlations, standard deviations, and Centered Root Mean Square Differences (CRMSD) between predictions and observations. 2.5 Statistical methods and modeling tools All statistical analyses and model development were performed in Python 3.10, utilizing open-source libraries, including Pandas, NumPy, SciPy, scikit-learn, LightGBM, CatBoost, XGBoost, and TensorFlow. Continuous variables are summarized as mean ± standard deviation (SD) if normally distributed, while non-normally distributed data underwent logarithmic transformation [ln(x+1)] and were summarized using geometric means (GM) with geometric standard deviations (GSD). Categorical variables are reported as frequency and percentage (n [%]). The feature engineering workflow encompasses variance threshold filtering, univariate feature selection (using F-statistics and MI), and embedded feature selection through L1 regularization models (LASSO for regression tasks and logistic regression for classification tasks). Ensemble modeling strategies included Bagging, Voting, Stacking, and Boosting methods, with ridge regression or logistic regression consistently serving as meta-learners for these ensembles. The entire modeling pipeline was automated and standardized using fixed parameter configurations. To statistically compare model performance, regression models were subjected to non-parametric comparisons using the Friedman test, followed by Nemenyi post-hoc tests. Classification models employed bootstrap methods (1,000 repetitions) to compute confidence intervals and two-tailed p-values for pairwise differences in AUC values among models, with false-discovery rate (FDR) correction applied for multiple hypothesis testing. All statistical tests were two-sided, and the significance threshold was set at p < 0.05. 3 Results 3.1 Selected features A total of 48 feature variables were ultimately included in the analysis based on the feature selection methodology. These features included demographic information, comorbidities, medical history, surgical and invasive procedures, antimicrobial exposure, and microbiological characteristics. A summary of the key feature distributions is presented in Table 1 . Among the 847 ICU patients included, 597 were male (70.5%) and 250 were female (29.5%). The cohort was predominantly comprised of older adults, with a median age of 79 years; 77.1% of patients were aged 65 years or older. Regarding surgical history, 427 patients (50.4%) had not undergone surgery, 239 (28.2%) underwent one surgical procedure, and 181 (21.4%) underwent two or more. All non-normally distributed numerical variables were log-transformed prior to statistical analysis, with the results expressed as GM and GSD. The GM durations were 20.28 days (GSD 2.94) for central venous catheterization (CVC-days), 17.86 days (GSD 3.51) for mechanical ventilation (MV-days), and 21.30 days (GSD 3.49) for urinary catheterization (UC-days). The time to the first carbapenem-resistant P. aeruginosa (TFC-CRPA) had a GM of 17.76 days (GSD, 2.14). The proportion of initial carbapenem-resistant isolates (IPA-CARBSUSC) was 62.9%. Carbapenems (CARB) showed the highest usage (ATC-DDD GM 6.70, GSD 2.54), followed by antifungals (AF, GM 4.05, GSD 4.10), and β-lactam/β-lactamase inhibitors (BL/BLI, GM 2.35, GSD 2.96). Other agents such as rifampin (RIF), sulfonamides (SUL), penicillin (PEN), and nitrofurantoin (NIT) were used at low levels (GM ≤ 0.01). View this table: View inline View popup Table 1. Descriptive statistics for baseline characteristics and outcome variables. 3.2 Outcome variables As shown in Table 1 , the primary and secondary outcomes of the study were as follows. The overall CRR was 0.63 (SD 0.35). For the CRR-PA-Last2, 223 patients (26.3%) had a value of 0, 21 patients (2.5%) had a value of 0.5, and 603 patients (71.2%) had a value of 1. Regarding the LastPaAST, 615 patients (72.6%) were resistant (R) and 232 patients (27.4%) were susceptible (S). IHM occurred in 312 patients (36.8%), and 535 patients (63.2%) were discharged alive. ICU-LOS was 28.95 days (SD, 2.69), and the ICU-Death interval was 56.01 days (SD, 3.70). 3.3 Performance evaluation 3.3.1 Classification tasks We first generated raw ROC curves for all the classification models based on their probability outputs and used the AUC to evaluate the overall discriminative performance. The sensitivity thresholds were aligned to approximately 0.800 for all the models, after which the threshold-dependent metrics were calculated. We further applied 1,000-fold bootstrapping and Benjamini– Hochberg correction to obtain FDR-adjusted p-values ( q -values), allowing for a systematic comparison of model performance for two binary outcomes: IHM and LastPaAST. For the IHM outcome, ROC curves ( Fig. 3a ) showed that Voting Classifier (VC) achieved the highest AUC (0.842), indicating a superior ranking ability. When evaluating the threshold-dependent performance at a fixed sensitivity, VC outperformed the other models with the highest specificity (0.751), accuracy (0.770), positive predictive value (PPV: 0.653), and negative predictive value (NPV: 0.866), as shown in Table 2 . Pairwise AUC comparisons ( Table 3 ) revealed significant improvements of VC over LGB-C (ΔAUC = 0.014, 95% CI: 0.007–0.022, q < 0.05) and XGB-C (ΔAUC = 0.011, 95% CI: 0.005–0.018, q < 0.05), while differences with other models were not statistically significant (q = 0.05). Download figure Open in new tab Fig. 3 ROC curves of nine classifiers for two binary outcomes. The panels show (a) IHM and (b) LastPaAST. The curves represent the mean five-fold cross-validation, and the AUC values are listed in the legend. View this table: View inline View popup Download powerpoint Table 2. Discrimination metrics of nine classification models for two binary outcomes. View this table: View inline View popup Download powerpoint Table 3. Pairwise DeLong comparisons of AUC between classification models with 95% confidence intervals and FDR-adjusted q-values. For the LastPaAST outcome ( Fig. 3b ), the ROC curve of LGB-C was closest to the top-left corner, yielding an AUC of 0.981. VC and XGB-C closely followed (AUC = 0.978), whereas CAT-C and RF-C had AUCs values of 0.976 and 0.975, respectively ( Table 2 ). At a fixed sensitivity of 0.800, LGB-C achieved the best threshold-dependent metrics, with a specificity of 0.974, an accuracy of 0.848, and a PPV of 0.988. Pairwise AUC comparisons indicated significant advantages of LGB-C over BC, GBDT, LR-C, RF-C, SC, and XGB-C, with AUCs ranging from +0.003 to +0.036, all of which were statistically significant (q < 0.05). These results highlighted LGB-C as the optimal model for LastPaAST prediction, with both strong overall discrimination and calibrated threshold performance. VC and CAT-C also showed competitive results and could serve as alternatives for deployment. 3.3.2 Regression tasks In CRR prediction ( Fig. 4a ), CAT-R and VR were closest to the observed values on the Taylor diagram: CAT-R achieved a centered root mean square difference (cRMSD) of 0.166, a correlation coefficient (r) of 0.885 (R 2 ≈ 0.78), and a σ-ratio of approximately 0.880. VR followed with cRMSD = 0.168, r = 0.884 (R 2 ≈ 0.78), and σ-ratio ≈ 0.850. XGB-R ranked third with cRMSD = 0.175, r = 0.873, and σ-ratio ≈ 0.900. The other models exhibited progressively higher cRMSD values. The Nemenyi post-hoc test showed significant differences between CAT-R and SR ( p = 0.02) and between VR and SR ( p = 0.01), confirming CAT-R and VR as the top-performing models for ARR prediction ( Fig. 5a ). Download figure Open in new tab Fig. 4 Taylor diagrams summarizing regression performance across four outcomes. (a) CRR; (b) CRR-PA–Last2; (c) ICU-LOS; (d) ICU-Death interval. The radial distance represents the modeled standard deviation, the azimuthal angle shows the Pearson correlation with observations, and the gray semicircles depict the centered root mean square difference (cRMSD). The orange dashed arc represents the standard deviation of the observed values. Download figure Open in new tab Fig. 5 Pairwise Nemenyi post-hoc p-values for eight regressors. Heat maps display false-discovery rate-adjusted p-values comparing model ranks for (a) CRR, (b) CRR-PA–Last2, (c) ICU-LOS, and (d) ICU-Death interval. Darker blue tiles indicate smaller p-values, with cells showing p < 0.05 highlighted in deep blue and annotated in white. In the CRR-PA-Last2 evaluation ( Fig. 4b ), the predicted standard deviations of all models converge on the observed value, forming a compact, nearly superimposed cluster of points on the Taylor diagram. The Friedman test revealed no overall differences (p > 0.05). The Nemenyi test revealed a significant difference only between BR and XGB-R ( p = 0.03); all other comparisons were non-significant ( p > 0.05), suggesting that XGB-R offered better-than-average computational efficiency while maintaining predictive accuracy, and the other models performed similarly ( Fig. 5b ). For ICU-LOS prediction ( Fig. 4c ), XGB-R achieved the best overall performance with r = 0.972, cRMSD = 55.70 days, and σ-ratio ≈ 0.877 (194.03/221.27). GBRT followed with r = 0.959, cRMSD = 62.93 days, and σ-ratio ≈ 0.930. VR and RF-R showed systematic underestimation (σ-ratio ≈ 0.778 and ≈ 0.759), and CAT-R notably underestimated variability (σ-ratio ≈ 0.596). The Nemenyi test indicated that XGB-R significantly outperformed BR ( p = 0.03) and CAT-R ( p 0.05), highlighting the balance between accuracy and amplitude alignment ( Fig. 5c ). In predicting ICU-Death interval ( Fig. 4d ), SR performed robustly with cRMSD ≈ 221.16 days, r = 0.740, and σ-ratio ≈ 0.874 (189.28/322.87). However, XGB-R maintained the best overall balance with a lower cRMSD = 205.07 days, σ-ratio ≈ 0.712, and r = 0.774. The Nemenyi test showed significant advantages of XGB-R over BR ( p = 0.03) and CAT-R ( p 0.05), further confirming the superiority of XGB-R for tasks with heavy-tailed distributions ( Fig. 5d ). Table 4 summarizes the mean ± SD and average ranks for each model across the four outcomes. XGB-R ranked first overall (Avg Rank = 1.95), demonstrating the lowest overall error (MSE 9.76 ×10 3 ± 18.23, RMSE 64.11 ± 84.06, MAE 25.24 ± 32.20) and highest average R 2 (0.77 ± 0.16). View this table: View inline View popup Download powerpoint Table 4. Multi-outcome regression performance and Friedman average ranks of the eight candidate models (mean ± SD) VR (2.95) and GBRT (3.40) followed in second and third places, although VR and GBRT systematically under and overestimated the amplitude, respectively. CAT-R (4.60) was ranked fourth for its stability in ARR and Combined-ARR predictions. RF-R (5.15) and SR (5.25) ranked fifth and sixth, respectively, whereas LGB-R (5.75) and BR (6.95) ranked the lowest. Overall, XGB-R, VR, and GBRT demonstrated the best balance between the performance and computational efficiency across multiple outcomes. 3.4 Feature interpretation 3.4.1 Classification tasks (1) IHM As shown in Fig. 6 , the SHAP analysis of the LGB-C and RF-C models revealed that patient age (Age [years]), TAMT, PMX (ATC/DDD per DOT), UC-days, CARB (ATC/DDD per DOT), CARB (ATC-DDD), and CVC-days were the top contributors to the prediction of IHM. Age was the highest factor in both models, indicating a strong association with poor outcomes in older adult patients. Indicators of antimicrobial exposure and invasive procedures (e.g., combination therapies, urinary catheterization, and central venous catheterization) significantly influenced the model predictions. In particular, the high SHAP values for polymyxin and carbapenem use (regardless of whether measured by ATC/DDD per DOT or total ATC-DDD) highlight their predictive value for mortality risk. These features may partially reflect the severity of the underlying disease and treatment intensity rather than direct causality. Download figure Open in new tab Fig. 6 Global SHAP beeswarm plots for two binary outcomes. (a) LGB-C model for IHM; (b) RF-C model for IHM; (c) LGB-C model for LastPaAST; (d) RF-C model for LastPaAST. Each dot represents an individual SHAP value, with the color indicating a normalized feature value (blue = low; red = high). The features were ranked by the mean absolute SHAP value (|SHAP|), reflecting their relative contributions to the prediction. Only the top 10 features are shown; the SHAP values for all the remaining features are close to zero. Download figure Open in new tab Fig. 7 Global SHAP beeswarm plots for four regression outcomes. (a) CAT-R model for CRR; (b) VR model for CRR; (c) XGB-R model for CRR-PA–Last2; (d) XGB-R model for ICU-LOS; (e) XGB-R model for ICU-Death interval. Each plot shows the top 10 most influential predictors ranked by the mean absolute SHAP value (|SHAP|). Each dot represents an individual SHAP value, with the color encoding the normalized feature value (blue = low, red = high). Positive SHAP values indicate a higher predicted risk, whereas negative values indicate protective effects. The SHAP values for all remaining features are close to zero. (2) LastPaAST For the LastPaAST outcome, the SHAP results showed that variables closely related to TFC-CRPA and IPA-CARBSUSC dominated the feature rankings in both the LGB-C and RF-C models. In addition, exposure to carbapenems and third-generation cephalosporin/β-lactamase inhibitors (CARB [ATC/DDD and ATC/DDD per DOT], and 3GC-BLI [ATC/DDD and ATC/DDD per DOT]) exhibited high explanatory power. Other highly ranked factors included MV-days, NSP, and combination therapy duration, suggesting that invasive procedures and antimicrobial exposure are the key drivers of the development of carbapenem resistance in P. aeruginosa . Across both the Death and LastPaAST outcomes, the feature importance rankings for LGB-C and RF-C were highly consistent. Age, antimicrobial exposure (especially to carbapenems and polymyxins), combination therapy, and invasive procedures have been consistently emphasized. 3.4.2 Regression tasks (1) CRR In SHAP value rankings for CRR prediction, both CAT-R and VR models identified IPA-CARBSUSC_S and IPA-CARBSUSC_R as the two most important features, with average absolute SHAP values of 0.24 and 0.21 for CAT-R, and 0.28 and 0.25 for VR, respectively. These values are substantially higher than those of all other variables. TFC-CRPA (days) ranked third in CAT-R (SHAP = 0.18) and highest in VR (SHAP = 0.16). CARB (ATC/DDD per DOT) ranked third and fourth for CAT-R and VR, respectively, with SHAP values of 0.15 and 0.14. MV-days ranked fourth in CAT-R (SHAP ≈ 0.13) and fifth in VR (SHAP ≈ 0.12). All other candidate features had SHAP values below 0.10, with consistent rankings across both models. Overall, the SHAP analysis demonstrated high consistency between models, with IPA–CARBSUSC_S/R, TFC-CRPA, CARB exposure, and MV-days identified as the top four contributors to CRR prediction. (2) CRR-PA–Last2 In the CRR-PA–Last2 prediction, SHAP rankings from the CAT-R and XGB-R models similarly emphasized the dominant role of resistance status and combination therapy duration. For CAT-R, IPA–CARBSUSC_S had the highest mean SHAP value (0.1716), followed by IPA– CARBSUSC_R (0.1232), TFC-CRPA (0.0916), CARB (ATC/DDD per DOT, 0.0079), and MV-days (0.0066). For XGB-R, IPA–CARBSUSC_R ranked first (0.2053), followed by TFC-CRPA (0.1461), IPA–CARBSUSC_S (0.0544), MV-days (0.0145), and CARB (ATC/DDD, 0.0084). Both models confirmed that early resistance and prolonged multidrug exposure were decisive predictors, whereas MV-days and carbapenem dose made relatively modest contributions. (3) ICU-LOS SHAP analysis from the XGB-R model indicated that ICU length of stay (log-transformed) was primarily driven by five features. MV-days had the highest average SHAP value (0.384) with a wide and positively skewed distribution, indicating a strong marginal impact. UC-days ranked second (SHAP = 0.262), followed by CVC-days (SHAP = 0.183), both of which had a positively skewed SHAP distribution. The DAMT DOT ranked fourth (SHAP = 0.157), with contributions on both sides of the axis but a predominance of positive values. CARB (ATC/DDD) ranked fifth (SHAP = 0.073) with a modest but consistent positive influence. All other features had SHAP values < 0.06. These contributions were visualized using beeswarm plots, which clearly illustrated the marginal effects of each feature on the log-transformed ICU-LOS prediction. A note in the figure caption indicates that log transformation was applied throughout the modeling and interpretation processes owing to the strong skewness in the outcome distribution. (4) ICU-Death intervals In the SHAP analysis of the XGB-R model for the log-transformed ICU-Death interval, CVC-days had the highest average SHAP value (0.703) with a wide and positively skewed distribution, indicating its dominant marginal effect. UC-days followed (SHAP = 0.268), indicating a potential association with increased predicted mortality risk, which may reflect the impact of prolonged urinary catheterization. TFC-CRPA ranked third (SHAP = 0.167), suggesting a positive correlation between prolonged exposure to MDR pathogens and longer time-to-death. MV-days and OXA (ATC/DDD per DOT) ranked fourth (SHAP = 0.150) and fifth (SHAP = 0.054), respectively. While MV-days showed mostly positive contributions, the influence of OXA was modest but concentrated in high-value regions. All other features had SHAP values of < 0.047. These results were visualized in beeswarm plots showing the marginal contributions of each feature, and the captions indicate that log transformation was applied consistently across model training, prediction, and interpretation. Due to the strong skewness in the ICU-LOS and ICU-Death interval distributions, both outcomes were log-transformed, consistently modeled, and interpreted on a log scale. Other continuous variables were imputed using medians and standardized via Z-score transformation. Categorical variables were imputed using modes and one-hot encoding. All transformations were implemented within a unified pipeline and transformed target regressor structure to ensure reproducibility. The final beeswarm plots display the five features with the highest marginal contributions to the prediction of (log-transformed) ICU outcomes. 4 Discussion We innovatively developed and validated a hybrid AutoML ensemble framework based on a retrospective single-center ICU cohort. The pipeline integrates conventional learners such as logistic regression and Random Forest with gradient-boosting algorithms including XGBoost, LightGBM, and CatBoost. Within a unified preprocessing, training, cross-validation, and ensemble pipeline, the framework simultaneously achieved quantitative regression predictions of continuous outcomes, such as CRR, ICU length of stay, and time-to-death intervals, as well as binary predictions for categorical outcomes, such as in-hospital all-cause mortality and antimicrobial susceptibility of the last isolated strains. Compared to previous research focusing solely on individual tasks or requiring separate maintenance of multiple models, this integrated approach significantly reduced clinical deployment and maintenance complexity and surpassed prior ICU AMR prediction outcomes across several metrics [ 34 – 36 ] In binary classification tasks, fixing the sensitivity at 0.80 facilitated a more intuitive comparison of the trade-off between false-positive rates and the overall discriminative capacity. The VC achieved the highest AUC for mortality risk prediction, making it particularly suitable for scenarios requiring maximal detection coverage. Similar to the findings of Chhillar and Singh, who reported an AUC of 0.9985, accuracy of 99.42%, and sensitivity of 99.20% by integrating SVM, XGBoost, and MLP models on the WDBC breast cancer dataset, our study confirmed the superior predictive capability of soft voting ensembles for mortality risk [ 37 ]. Additionally, the Random Forest Classifier (RF-C) demonstrated optimal performance in predicting sepsis severity, mortality, and ICU length of stay (ICU-LOS) [ 38 ]. Our analysis further validated RF-C as the preferred model when minimal false-positive rates were clinically essential. To predict the susceptibility of the last isolated strain, LGB-C combined the highest AUC with the optimal threshold performance, making it the recommended first choice. Moreover, VC and CAT-C exhibited strong performance in multi-model ensembles, providing diversity and weighted optimization. In the regression tasks, the XGB Regressor (XGB-R) consistently demonstrated the lowest RMSE and best average rank performance, aligning with the systematic review by Almeida et al. on hospital length of stay prediction, where XGBoost outperformed traditional methods, particularly in resource-sensitive contexts requiring minimal overall prediction errors [ 39 ]. The SR and gradient-boosted regression trees with Huber loss (GBRT-Huber) showed robust performance at extreme quantiles, consistent with Wang’s robust boosting theory, underscoring the suitability for predicting heavy-tailed distributions such as time-to-death [ 40 ]. CatBoost Regressor (Cat-R), with balanced accuracy and interpretability, is ideal to handle mixed-feature scenarios. Collectively, gradient-boosting methods provided the best compromise across correlation, amplitude matching, and error scale, whereas the traditional Bagging and Random Forest approaches were less suitable owing to systematic amplitude biases. The medical imaging findings of Srinivasu et al. support this conclusion, highlighting CatBoost combined with SHAP as highly discriminative yet transparent [ 41 ]. SHAP analysis was integrated into our model interpretability workflow to achieve a comprehensive explanatory loop in line with the Explainable AI (XAI) framework proposed by Bilal et al. [ 42 ]. By visualizing the global feature importance alongside individual decision pathways, the analysis clarified the clinical logic inherent to model predictions and provided actionable quantitative evidence supporting antimicrobial stewardship and life-support adjustments. Global SHAP explanations validated the model’s sensitivity to well-established clinical risk factors: increased age, prolonged mechanical ventilation, and central venous catheterization were positively correlated with mortality risk ( Fig. 6 ). This finding aligns closely with the established practice of incorporating age and intensity of organ support into the traditional APACHE II scoring system [ 43 ]. IPA-CARBSUSC_R demonstrated significant predictive contributions to mortality risk in both LGB-C and RF-C, with mean absolute SHAP values of 0.45 and 0.42, respectively, highlighting the importance of timely AST and precise medication administration [ 44 ]. Additionally, CARB ATC/DDD and combined drug-use days ranked prominently in the global SHAP analyses of CAT-R and VR, reinforcing that widespread carbapenem use is related to increased resistance in P. aeruginosa [ 45 – 47 ]. XGB-R further identified interactions between “time from admission to event,” dynamic support indicators (MV-days, CVC-days), and cumulative medication intensity, providing quantitative evidence to inform personalized interventions. Furthermore, combination antibiotic regimens involving DAMT and PMX administration positively influenced mortality and last antibiotic-susceptibility outcomes, reflecting the intensity of critical infection therapy but also possibly indicating negative impacts, such as antibiotic-associated dysbiosis or toxicity. PMX exposure influenced both IHM and susceptibility outcomes. Dosage analysis revealed that polymyxin B accounted for a cumulative consumption of 1,270.78 ATC/DDD (217.09 ATC/DDD per DOT), representing 55% of total polymyxin consumption at the study site—significantly higher than polymyxin E and colistin sulfate formulations. Ecological studies from Brazilian ICUs similarly reported significant correlations between rising annual polymyxin B consumption and the incidence of carbapenem/polymyxin co-resistant P. aeruginosa (β = 0.71, p < 0.01) [ 48 ]. Research on the molecular mechanisms further indicated that PMX-induced lipid A remodeling and mutations in the pmrAB and phoPQ two-component regulatory systems reduce outer membrane permeability, restrict carbapenem entry, and promote cross-resistance. These findings suggest that although polymyxin B remains a critical “last-line” treatment against MDR P. aeruginosa , the intense use thereof also exacerbates selection pressure for resistance. Meanwhile, the membrane-disruptive combination therapy of polymyxin B and meropenem has demonstrated potential in vitro and animal studies to inhibit heterogeneous resistance, providing a foundation for future precision dosing and combination therapy strategies [ 49 ]. In addition to pharmaceutical factors, advanced age, extended duration of mechanical ventilation, and central venous catheterization correlated positively with mortality risk, indicating declining physiological reserves and heightened complication risks, consistent with the established APACHE II and SOFA prognostic scoring systems. To assess real-time applicability further, models were trained and evaluated on a standardized hardware platform (Intel® 13th-Gen Core™ i5-13400F processor with 10 cores/16 threads, base 2.5 GHz, turbo up to 4.6 GHz, and 32 GB DDR4 RAM) under a consistent software environment (Python 3.8; scikit-learn 1.1.2; XGBoost 1.5.0; LightGBM 3.3.2; CatBoost 1.0.5), with single-prediction latency measurements. The results demonstrated that classification tasks using a VC and SC averaged inference times below 50 ms ( Table 5 ). Regression tasks employing the VR and SR averaged between 40 and 50 ms, comparable to single-model baselines such as GBRT, XGB, and LGB ( Table 6 ). Despite longer inference times for the Bagging models, voting and stacking ensembles maintained millisecond-level responsiveness, thus satisfying real-time ICU clinical decision-support requirements. View this table: View inline View popup Download powerpoint Table 5. Average prediction time of classification models. View this table: View inline View popup Download powerpoint Table 6. Average prediction time of regression models. Tables 5 and 6 summarize the average single-prediction inference times of the classification and regression models, respectively. Despite the significant findings, this study has limitations. First, its retrospective single-center design necessitates multicenter prospective validation for better generalization. Second, the threshold choices in VTF-MI-L1 methods involve subjective empirical decisions, potentially requiring recalibration for specific scenarios. Finally, the computational complexities of AutoML and cross-validation pose challenges to resource management and real-time deployment, suggesting that future efforts should balance model complexity with practical responsiveness. 5 Conclusion This retrospective single-center cohort study presented a novel hybrid AutoML ensemble framework combining traditional learners (e.g., logistic regression, Random Forest) with gradient-boosting algorithms (XGBoost, LightGBM, and CatBoost). The framework achieved precise prediction and highly interpretable outputs for various ICU outcomes, including CRR, susceptibility to the last isolated strains, IHM, ICU stay duration, and time-to-death. Future studies should emphasize multicenter prospective validation and model optimization to improve generalization and real-time clinical decision support. 6 Funding This study was supported by the Zhejiang Provincial Center for Disease Control and Prevention Science and Technology Program (2025JK198, “Development of an ICU Healthcare-Associated Infection (HAI) Risk Prediction System Based on Environmental Multimodal Feature Fusion”), and the Clinical Medical Research Fund of Zhejiang Medical Association (2024ZYC-A257, “Construction and Evaluation of a Polymyxin Antimicrobial Resistance (AMR) Risk Assessment Model Based on Deep Feature Fusion and Ensemble Learning Algorithms”). 7 Declaration of competing interests The authors declare that they have no competing financial interests or personal relationships that may have influenced the work reported in this study. Data Availability All relevant data are within the manuscript and its Supporting Information files. Supporting information S1 File. Patient-Level Filtered Clinical and Antimicrobial Exposure Features Dataset. It includes detailed patient characteristics such as Patient ID, Age(year), sex, and various metrics related to antimicrobial use (e.g., DAMT(DOT), CAMT(DOT), TAMT(DOT), AG(ATC/DDD per DOT), FQ(ATC/DDD)), as well as clinical exposure indicators (e.g., CVC-days, MV-days, UC-days). The dataset also includes information on antimicrobial susceptibility (e.g., IPA– CARBSUSC) and other relevant clinical parameters. S2 File. This is the Clinical Outcome Indicators Dataset. This file contains the primary and secondary clinical outcome variables for each patient, aligned with the Patient ID in the ‘Patient-Level Filtered Clinical and Antimicrobial Exposure Features Dataset’ (S1 File). Key outcome indicators include IHM, LastPaAST, CRR, ICU-LOS(days), and ICU-Death Interval (days). References 1. ↵ Antimicrobial resistance : World Health Organization ; 2024 . Available from: https://www.who.int/health-topics/antimicrobial-resistance . 2. ↵ Prevention USCFDCa . About Antimicrobial Resistance 2024 . Available from: https://www.cdc.gov/antimicrobial-resistance/about/index.html . 3. ↵ O’Neill J. Tackling drug-resistant infections globally: final report and recommendations . 2016 . 4. ↵ Magiorakos AP , Srinivasan A , Carey RB , Carmeli Y , Falagas ME , Giske CG , et al. Multidrug-resistant, extensively drug-resistant and pandrug-resistant bacteria: an international expert proposal for interim standard definitions for acquired resistance . Clinical microbiology and infection: the official publication of the European Society of Clinical Microbiology and Infectious Diseases . 2012 ; 18 ( 3 ): 268 – 81 . Epub 20110727. doi: 10.1111/j.1469-0691.2011.03570.x . PubMed PMID: 21793988 . OpenUrl CrossRef PubMed 5. ↵ Collaborators AR . Global burden of bacterial antimicrobial resistance in 2019: a systematic analysis. Lancet (London , England ). 2022 ; 399 ( 10325 ): 629 – 55 . Epub 20220119. doi: 10.1016/s0140-6736(21)02724-0 . PubMed PMID: 35065702 ; PubMed Central PMCID: PMCPMC8841637 . OpenUrl CrossRef PubMed 6. ↵ CDC . Antimicrobial Resistance: Causes and How It Spreads 2024 . Available from: https://www.cdc.gov/antimicrobial-resistance/causes/index.html . 7. ↵ Pace MC , Corrente A , Passavanti MB , Sansone P , Petrou S , Leone S , Fiore M . Burden of severe infections due to carbapenem-resistant pathogens in intensive care unit . World J Clin Cases . 2023 ; 11 ( 13 ): 2874 – 89 . doi: 10.12998/wjcc.v11.i13.2874 . PubMed PMID: 37215420 ; PubMed Central PMCID: PMCPMC10198073 . OpenUrl CrossRef PubMed 8. Cortegiani A , Antonelli M , Falcone M , Giarratano A , Girardis M , Leone M , et al. Rationale and clinical application of antimicrobial stewardship principles in the intensive care unit: a multidisciplinary statement. Journal of Anesthesia , Analgesia and Critical Care . 2023 ; 3 ( 1 ): 11 . doi: 10.1186/s44158-023-00095-6 . OpenUrl CrossRef 9. ↵ Abushaheen MA , Muzaheed Fatani AJ , Alosaimi M , Mansy W , George M , et al. Antimicrobial resistance, mechanisms and its clinical significance . Disease-a-Month . 2020 ; 66 ( 6 ): 100971 . doi: 10.1016/j.disamonth.2020.100971 . OpenUrl CrossRef 10. ↵ Li ZJ , Wang KW , Liu B , Zang F , Zhang Y , Zhang WH , et al. The Distribution and Source of MRDOs Infection: A Retrospective Study in 8 ICUs, 2013-2019 . Infect Drug Resist . 2021 ; 14 : 4983 – 91 . Epub 20211127. doi: 10.2147/idr.S332196 . PubMed PMID: 34866918 ; PubMed Central PMCID: PMCPMC8636692 . OpenUrl CrossRef PubMed 11. ↵ Pipitò L , Bono E , Mazzola CV , Rubino R , Anastasia A , Distefano SA , et al. Outcome of Bloodstream Infections Caused by Antibiotic-Resistant Bacteria: A 7-Year Retrospective Study at the University Hospital of Palermo , Italy. Antibiotics (Basel ). 2025 ; 14 ( 5 ). Epub 20250501. doi: 10.3390/antibiotics14050464 . PubMed PMID: 40426531 ; PubMed Central PMCID: PMCPMC12108205 . OpenUrl CrossRef PubMed 12. ↵ Colín-Castro CA , López-Jácome LE , Rodríguez-García MJ , Garibaldi-Rojas M , Rojas-Larios F , Vázquez-Larios MDR , et al. The ongoing antibiotic resistance and carbapenemase encoding genotypes surveillance . The first quarter report of the INVIFAR network for 2024 . PloS one. 2025;20(4):e0319441. Epub 20250416. doi: 10.1371/journal.pone.0319441 . PubMed PMID: 40238844 ; PubMed Central PMCID: PMCPMC12002462 . OpenUrl CrossRef PubMed 13. ↵ Al-Musawi T , Al-Agha R , Al-Khiami S , Al-Shamari H , Baghdadi M , Bosaeed M , et al. Bacteremia in the Gulf Cooperation Council Region: A Review of the Literature 2013-2023 . Infect Drug Resist . 2025 ; 18 : 2329 – 55 . Epub 20250507. doi: 10.2147/idr.S497241 . PubMed PMID: 40357416 ; PubMed Central PMCID: PMCPMC12067724 . OpenUrl CrossRef PubMed 14. ↵ Lapp L , Roper M , Kavanagh K , Bouamrane MM , Schraag S . Dynamic Prediction of Patient Outcomes in the Intensive Care Unit: A Scoping Review of the State-of-the-Art . J Intensive Care Med . 2023 ; 38 ( 7 ): 575 – 91 . Epub 20230405. doi: 10.1177/08850666231166349 . PubMed PMID: 37016893 ; PubMed Central PMCID: PMCPMC10302367 . OpenUrl CrossRef PubMed 15. ↵ Olang O , Mohseni S , Shahabinezhad A , Hamidianshirazi Y , Goli A , Abolghasemian M , et al. Artificial Intelligence-Based Models for Prediction of Mortality in ICU Patients: A Scoping Review . J Intensive Care Med . 2024 : 8850666241277134 . Epub 20240816. doi: 10.1177/08850666241277134 . PubMed PMID: 39150821 . OpenUrl CrossRef PubMed 16. ↵ Iwase S , Nakada TA , Shimada T , Oami T , Shimazui T , Takahashi N , et al. Prediction algorithm for ICU mortality and length of stay using machine learning . Sci Rep . 2022 ; 12 ( 1 ): 12912 . Epub 20220728. doi: 10.1038/s41598-022-17091-5 . PubMed PMID: 35902633 ; PubMed Central PMCID: PMCPMC9334583 . OpenUrl CrossRef PubMed 17. ↵ de Hond AAH , Leeuwenberg AM , Hooft L , Kant IMJ , Nijman SWJ , van Os HJA , et al. Guidelines and quality criteria for artificial intelligence-based prediction models in healthcare: a scoping review . NPJ digital medicine . 2022 ; 5 ( 1 ): 2 . Epub 20220110. doi: 10.1038/s41746-021-00549-7 . PubMed PMID: 35013569 ; PubMed Central PMCID: PMCPMC8748878 . OpenUrl CrossRef PubMed 18. ↵ Ardila CM , González-Arroyave D , Tobón S . Machine learning for predicting antimicrobial resistance in critical and high-priority pathogens: A systematic review considering antimicrobial susceptibility tests in real-world healthcare settings . PloS one . 2025 ; 20 ( 2 ): e0319460 . Epub 20250225. doi: 10.1371/journal.pone.0319460 . PubMed PMID: 39999193 ; PubMed Central PMCID: PMCPMC11856330 . OpenUrl CrossRef PubMed 19. ↵ Ardila CM , Yadalam PK , González-Arroyave D . Integrating whole genome sequencing and machine learning for predicting antimicrobial resistance in critical pathogens: a systematic review of antimicrobial susceptibility tests . PeerJ . 2024 ; 12 : e18213 . doi: 10.7717/peerj.18213 . OpenUrl CrossRef PubMed 20. ↵ Pinto-de-Sá R , Sousa-Pinto B , Costa-de-Oliveira S . Brave New World of Artificial Intelligence: Its Use in Antimicrobial Stewardship—A Systematic Review . Antibiotics . 2024 ; 13 ( 4 ): 307 . PubMed PMID : doi: 10.3390/antibiotics13040307 . OpenUrl CrossRef PubMed 21. ↵ Rahman S , Jiang LY , Gabriel S , Aphinyanagphongs Y , Oermann EK , Chunara R . Generalization in Healthcare AI: Evaluation of a Clinical Large Language Model . ArXiv . 2024 ;abs/2402.10965. 22. ↵ Xun H , Zi-de D , Yu-xing N , Min D , Bi-jie H , Liu-yi L , et al. Chinese experts’ consensus on prevention and control of multidrug resistance organism healthcare associated infection . Chin J Infect Control . 2015 ; 14 ( 01 ): 1 – 9 . OpenUrl 23. ↵ Performance Standards for Antimicrobial Susceptibility Testing . Clinical and Laboratory Standards Institute ., 2024 . 24. ↵ Guidelines for ATC classification and DDD assignment . Oslo: WHO Collaborating Centre for Drug Statistics Methodology , 2024 . 25. ↵ Johnson AEW , Pollard TJ , Shen L , Lehman L-wH , Feng M , Ghassemi M , et al. MIMIC-III, a freely accessible critical care database . Scientific Data . 2016 ; 3 ( 1 ): 160035 . doi: 10.1038/sdata.2016.35 . OpenUrl CrossRef PubMed 26. ↵ Mohtasham F , Pourhoseingholi M , Hashemi Nazari SS , Kavousi K , Zali MR. Comparative analysis of feature selection techniques for COVID-19 dataset . Scientific Reports . 2024 ; 14 ( 1 ): 18627 . doi: 10.1038/s41598-024-69209-6 . OpenUrl CrossRef PubMed 27. ↵ Rihan SDA , Anbar M , Alabsi BA . Approach for Detecting Attacks on IoT Networks Based on Ensemble Feature Selection and Deep Learning Models. Sensors (Basel , Switzerland ). 2023 ; 23 ( 17 ). Epub 20230823. doi: 10.3390/s23177342 . PubMed PMID: 37687798 ; PubMed Central PMCID: PMCPMC10489985 . OpenUrl CrossRef PubMed 28. ↵ Demircioğlu A . The effect of feature normalization methods in radiomics . Insights into Imaging . 2024 ; 15 ( 1 ): 2 . doi: 10.1186/s13244-023-01575-7 . OpenUrl CrossRef PubMed 29. ↵ Abualigah L , Alomari SA , Almomani MH , Zitar RA , Saleem K , Migdady H , et al. Artificial intelligence-driven translational medicine: a machine learning framework for predicting disease outcomes and optimizing patient-centric care . Journal of translational medicine . 2025 ; 23 ( 1 ): 302 . doi: 10.1186/s12967-025-06308-6 . OpenUrl CrossRef PubMed 30. ↵ Chicco D , Warrens MJ , Jurman G . The coefficient of determination R-squared is more informative than SMAPE, MAE, MAPE, MSE and RMSE in regression analysis evaluation . PeerJ Comput Sci . 2021 ; 7 : e623 . Epub 20210705. doi: 10.7717/peerj-cs.623 . PubMed PMID: 34307865 ; PubMed Central PMCID: PMCPMC8279135 . OpenUrl CrossRef PubMed 31. ↵ Trevor Hastie , Robert Tibshirani , Friedman J. Model Assessment and Selection. The Elements of Statistical Learning . Springer Series in Statistics (SSS): Springer ; 2009 . p. 219 – 59 . 32. ↵ Demšar J . Statistical Comparisons of Classifiers over Multiple Data Sets . Journal of Machine Learning Research . 2006 ; 7 ( 1 ): 1 – 30 . OpenUrl 33. ↵ Chen H , Lundberg SM , Lee SI . Explaining a series of models by propagating Shapley values . Nat Commun . 2022 ; 13 ( 1 ): 4512 . Epub 20220803. doi: 10.1038/s41467-022-31384-3 . PubMed PMID: 35922410 ; PubMed Central PMCID: PMCPMC9349278 . OpenUrl CrossRef PubMed 34. ↵ Feretzakis G , Sakagianni A , Loupelis E , Kalles D , Martsoukou M , Skarmoutsou N , et al. Using Machine Learning to Predict Antimicrobial Resistance of Acinetobacter Baumannii, Klebsiella Pneumoniae and Pseudomonas Aeruginosa Strains . Studies in health technology and informatics . 2021 ; 281 : 43 – 7 . doi: 10.3233/shti210117 . PubMed PMID: 34042702 . OpenUrl CrossRef PubMed 35. Martínez-Agüero S , Mora-Jiménez I , Lérida-García J , Álvarez-Rodríguez J , Soguero-Ruiz C . Machine Learning Techniques to Identify Antimicrobial Resistance in the Intensive Care Unit . Entropy (Basel ). 2019 ; 21 ( 6 ). Epub 20190618. doi: 10.3390/e21060603 . PubMed PMID: 33267317 ; PubMed Central PMCID: PMCPMC7515087 . OpenUrl CrossRef PubMed 36. ↵ Lewin-Epstein O , Baruch S , Hadany L , Stein GY , Obolski U . Predicting Antibiotic Resistance in Hospitalized Patients by Applying Machine Learning to Electronic Medical Records . Clinical infectious diseases: an official publication of the Infectious Diseases Society of America . 2021 ; 72 ( 11 ): e848 – e55 . doi: 10.1093/cid/ciaa1576 . PubMed PMID: 33070171 . OpenUrl CrossRef PubMed 37. ↵ Chhillar I , Singh A . An improved soft voting-based machine learning technique to detect breast cancer utilizing effective feature selection and SMOTE-ENN class balancing . Discover Artificial Intelligence . 2025 ; 5 ( 1 ): 4 . doi: 10.1007/s44163-025-00224-w . OpenUrl CrossRef 38. ↵ Su L , Xu Z , Chang F , Ma Y , Liu S , Jiang H , et al. Early Prediction of Mortality, Severity, and Length of Stay in the Intensive Care Unit of Sepsis Patients Based on Sepsis 3.0 by Machine Learning Models . Front Med (Lausanne) . 2021 ; 8 : 664966 . Epub 20210628. doi: 10.3389/fmed.2021.664966 . PubMed PMID: 34291058 ; PubMed Central PMCID: PMCPMC8288021 . OpenUrl CrossRef PubMed 39. ↵ Almeida G , Brito Correia F , Borges AR , Bernardino J . Hospital Length-of-Stay Prediction Using Machine Learning Algorithms—A Literature Review . Applied Sciences . 2024 ; 14 ( 22 ): 10523 . PubMed PMID : doi: 10.3390/app142210523 . OpenUrl CrossRef 40. ↵ Wang Z . Unified Robust Boosting . Journal of Data Science . 2025 ; 23 ( 1 ): 90 – 108 . doi: 10.6339/24-JDS1138 . OpenUrl CrossRef 41. ↵ Srinivasu PN , Jaya Lakshmi G , Gudipalli A , Narahari SC , Shafi J , Woźniak M , Ijaz MF . XAI-driven CatBoost multi-layer perceptron neural network for analyzing breast cancer . Scientific Reports . 2024 ; 14 ( 1 ): 28674 . doi: 10.1038/s41598-024-79620-8 . OpenUrl CrossRef PubMed 42. ↵ Bilal H , Khan MN , Khan S , Shafiq M , Fang W , Khan RU , et al. The role of artificial intelligence and machine learning in predicting and combating antimicrobial resistance . Computational and structural biotechnology journal . 2025 ; 27 : 423 – 39 . doi: 10.1016/j.csbj.2025.01.006 . OpenUrl CrossRef 43. ↵ Yelamanchi R. The APACHE II Scoring Systems and the ICU. In: Rajendram R, Preedy VR, Patel VB, editors. Biomarkers in Trauma, Injury and Critical Care. Cham: Springer International Publishing ; 2022 . p. 1 – 14 . 44. ↵ Yuan Q , Guo L , Li B , Zhang S , Feng H , Zhang Y , et al. Risk factors and outcomes of inpatients with carbapenem-resistant Pseudomonas aeruginosa bloodstream infections in China: a 9-year trend and multicenter cohort study . Frontiers in microbiology . 2023 ; 14 : 1137811 . Epub 20230518. doi: 10.3389/fmicb.2023.1137811 . PubMed PMID: 37260693 ; PubMed Central PMCID: PMCPMC10227572 . OpenUrl CrossRef PubMed 45. ↵ Yuan F , Li M , Wang X , Fu Y . Risk factors and mortality of carbapenem-resistant Pseudomonas aeruginosa bloodstream infection in haematology department: A 10-year retrospective study . Journal of Global Antimicrobial Resistance . 2024 ; 37 : 150 – 6 . doi: 10.1016/j.jgar.2024.03.018 . OpenUrl CrossRef PubMed 46. Tsao L-H , Hsin C-Y , Liu H-Y , Chuang H-C , Chen L-Y , Lee Y-J . Risk factors for healthcare-associated infection caused by carbapenem-resistant Pseudomonas aeruginosa. Journal of Microbiology , Immunology and Infection . 2018 ; 51 ( 3 ): 359 – 66 . doi: 10.1016/j.jmii.2017.08.015 . OpenUrl CrossRef 47. ↵ Tamma PD , Cosgrove SE , Maragakis LL . Combination therapy for treatment of infections with gram-negative bacteria . Clin Microbiol Rev . 2012 ; 25 ( 3 ): 450 – 70 . doi: 10.1128/cmr.05041-11 . PubMed PMID: 22763634 ; PubMed Central PMCID: PMCPMC3416487 . OpenUrl Abstract / FREE Full Text 48. ↵ Ribeiro Á CDS , Crozatti MTL , Silva AAD , Macedo RS , Machado AMO , Silva ATA . Pseudomonas aeruginosa in the ICU: prevalence, resistance profile, and antimicrobial consumption . Rev Soc Bras Med Trop . 2019 ; 53 : e20180498 . Epub 20191220. doi: 10.1590/0037-8682-0498-2018 . PubMed PMID: 31859938 ; PubMed Central PMCID: PMCPMC7083346 . OpenUrl CrossRef PubMed 49. ↵ Wickremasinghe H , Yu HH , Azad MAK , Zhao J , Bergen PJ , Velkov T , et al. Clinically Relevant Concentrations of Polymyxin B and Meropenem Synergistically Kill Multidrug-Resistant Pseudomonas aeruginosa and Minimize Biofilm Formation . Antibiotics . 2021 ; 10 ( 4 ): 405 . PubMed PMID : doi: 10.3390/antibiotics10040405 . OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted June 20, 2025. Download PDF Data/Code Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa LV Xiao-chun , Ren Qi , Zhu Lihong , CHEN Kun , WANG Jian-bing , CHEN Fang , JIN Kai-ling , LIN Kai medRxiv 2025.06.19.25329970; doi: https://doi.org/10.1101/2025.06.19.25329970 Share This Article: Copy Citation Tools A Hybrid AutoML Ensemble Integrating Conventional Learners and Gradient-Boosting Models for Multi-Outcome Prediction in ICU Patients with Pseudomonas aeruginosa LV Xiao-chun , Ren Qi , Zhu Lihong , CHEN Kun , WANG Jian-bing , CHEN Fang , JIN Kai-ling , LIN Kai medRxiv 2025.06.19.25329970; doi: https://doi.org/10.1101/2025.06.19.25329970 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Intensive Care and Critical Care Medicine Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (300) Cardiovascular Medicine (4435) Dentistry and Oral Medicine (444) Dermatology (382) Emergency Medicine (608) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1509) Epidemiology (15228) Forensic Medicine (30) Gastroenterology (1124) Genetic and Genomic Medicine (6598) Geriatric Medicine (668) Health Economics (997) Health Informatics (4536) Health Policy (1368) Health Systems and Quality Improvement (1613) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15916) Intensive Care and Critical Care Medicine (1103) Medical Education (623) Medical Ethics (146) Nephrology (667) Neurology (6599) Nursing (346) Nutrition (998) Obstetrics and Gynecology (1144) Occupational and Environmental Health (957) Oncology (3332) Ophthalmology (974) Orthopedics (369) Otolaryngology (420) Pain Medicine (436) Palliative Medicine (130) Pathology (663) Pediatrics (1693) Pharmacology and Therapeutics (691) Primary Care Research (711) Psychiatry and Clinical Psychology (5447) Public and Global Health (9230) Radiology and Imaging (2198) Rehabilitation Medicine and Physical Therapy (1370) Respiratory Medicine (1196) Rheumatology (593) Sexual and Reproductive Health (712) Sports Medicine (530) Surgery (712) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'a0050b468a521d20',t:'MTc3OTU0OTYyMA=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();