Multi-label machine-learning models identified distinct combinations of clinical history, hematologic disease, hospitalization and chemotherapy exposure associated with bloodstream infections in people receiving chemotherapy. The retrospective study analyzed 1,435 confirmed bloodstream-infection episodes among 1,220 patients treated at the Cancer Institute of the State of Sao Paulo in Brazil from 2017 through 2022. Its strongest test-set discrimination was for infections involving carbapenem-resistant Gram-negative bacilli, while performance for the other microbiological groups was more variable. The investigators described the results as hypothesis-generating and said the models require external validation before they could support clinical decisions.

The research team used a multi-label framework because an infection episode could contain more than one relevant microorganism or susceptibility profile. The five outcomes were bloodstream infections associated with carbapenem-resistant Gram-negative bacilli, methicillin-resistant staphylococci, third-generation cephalosporin-resistant Enterobacterales, fungal pathogens, and non-multidrug-resistant, non-fungal organisms. Episodes involving more than one applicable group were assigned to each group rather than being forced into mutually exclusive categories. This structure was relevant because 166 episodes, or 11.6% of the cohort, were polymicrobial. The final dataset contained 1,640 microorganisms isolated from the 1,435 episodes.

Most infections occurred early in the treatment course: 60.9% arose during a patient’s first chemotherapy cycle. Non-multidrug-resistant or fungal agents accounted for 1,193 episodes, or 83.1% of the total, while carbapenem-resistant Gram-negative bacilli were identified in 109 episodes, or 7.6%. The most frequently isolated species were Escherichia coli, representing 26.3% of organisms; Klebsiella pneumoniae, 15.5%; and Pseudomonas aeruginosa, 9.0%. Among Enterobacterales isolates, 6.6% were resistant to third-generation cephalosporins and 4.7% were resistant to carbapenems. Across Gram-negative bacteria, carbapenem resistance occurred in 10.8%, equivalent to 114 isolates and 109 episodes. Thirty-day mortality ranged from 30.1% to 40.6% among the infection groups and reached 53.2% in the carbapenem-resistant Gram-negative bacilli group.

Each analytical observation represented one chemotherapy cycle rather than one unique patient, so patients undergoing multiple cycles could contribute more than one record. Candidate explanatory variables included age, sex, cancer type, body measurements, laboratory results, chemotherapy drugs and classes, prior infections, antimicrobial exposure and days spent in hospital. Temporal information was encoded through a first-cycle indicator, cumulative counts of previous infections in each pathogen category, and cumulative antimicrobial and hospitalization exposure during the 90 days preceding the cycle. A binary variable also marked cycles initiated during the COVID-19 pandemic period. The initial collection included 1,505 infected cycles from 1,272 patients. After records with missing age or sex were excluded, extreme height, weight and body-mass-index values were removed, and missing blood-test results were flagged, the analysis retained 1,435 cycles from 1,220 patients.

Four algorithms—logistic regression, random forests, multilayer perceptrons and gradient boosting—were implemented in classifier chains. In this arrangement, models predict infection labels sequentially, allowing the output for an earlier label to inform later predictions and thereby representing potential dependencies among infection groups. Hyperparameters were tuned with grid search and empirical refinement within fivefold cross-validation. The researchers reserved one independently partitioned subset for a single final evaluation. To preserve the clinical distribution of uncommon outcomes, no resampling or reweighting was applied. Model selection and overall performance were assessed with a weighted area under the receiver operating characteristic curve, or AUC-ROC, across the five groups. SHapley Additive exPlanations, known as SHAP, were then used on the test set to estimate how individual variables contributed to predictions.

The highest-performing configuration was a classifier chain of 20 random forests, with each forest containing 100 trees and a maximum depth of four. During cross-validation, it produced a weighted mean AUC-ROC of 60.90%, with a standard deviation of 2.52%; on the held-out test set, the weighted average was 66.37%. The model’s highest infection-specific result was for carbapenem-resistant Gram-negative bacilli, for which test-set AUC-ROC reached 83.25%, with a 95% confidence interval of 73.91% to 91.64%. Average precision ranged from 6.40% for third-generation cephalosporin-resistant Enterobacterales to 90.84% for non-multidrug-resistant, non-fungal infections. The neural network and logistic regression each achieved a cross-validation weighted AUC-ROC of 59.96%, with standard deviations of 2.89% and 2.81%, respectively; gradient boosting averaged 56.89%, with a standard deviation of 1.03%. The differences among model families were not large relative to their variability, and the study did not establish the statistical superiority of any algorithm.

SHAP analysis identified days of hospitalization, palliative chemotherapy, hematologic malignancy, monocytopenia and thrombocytopenia as the most influential variables across the infection categories. For carbapenem-resistant Gram-negative bacilli, a previous infection with that group was the leading associated feature, followed by hospitalization duration, hematologic malignancy, thrombocytopenia and low hematocrit. Prominent variables for third-generation cephalosporin-resistant Enterobacterales were previous glycopeptide exposure, hospitalization, acute leukemia, the pandemic-period indicator and antimetabolite treatment. Methicillin-resistant staphylococcal infections were associated with hospitalization duration, hematologic malignancy, lymphopenia, previous fungal infection and male sex. For fungal infections, the leading variables were the pandemic period, male sex, days of glycopeptide use, alkylating-agent treatment and hospitalization duration. These values describe associations encoded by the models rather than independent causal effects.

Chemotherapy classes contributed less overall than hospitalization, prior infection and hematologic disease, but their rankings varied by infection group. Alkylating agents were associated with candidemia, antitumor antibiotics with carbapenem-resistant Gram-negative bacilli, and antimetabolites with third-generation cephalosporin-resistant Enterobacterales. The investigators said these relationships require further study. The pandemic-period indicator was the leading feature associated with fungal infection, which the authors interpreted cautiously: it may have represented changes in healthcare delivery, clinical practice and antimicrobial use rather than a direct effect of SARS-CoV-2 infection. Likewise, prior glycopeptide use may chiefly have marked greater healthcare exposure and susceptibility to infection rather than being a specific risk factor for resistant Enterobacterales or fungi. Palliative chemotherapy was associated with non-multidrug-resistant, non-fungal infection, but that result was relative to the other chemotherapy patients included in the analysis.

The design places important limits on how the findings can be interpreted. Because the dataset was retrospective and restricted to chemotherapy cycles that already had confirmed bloodstream infections, the models estimated the probability of different microbiological profiles among existing infections rather than the risk of developing an infection in all chemotherapy patients. The single-center population may also differ from patients treated elsewhere in their cancer profiles and patterns of bacterial resistance. Relatively few events occurred in uncommon infection groups, especially fungal infections, making their estimates less stable. Although data were partitioned at the chemotherapy-cycle level, partitioning at the patient level would provide a more conservative assessment of performance in previously unseen patients. The wide confidence interval around the strong carbapenem-resistant result also indicates substantial uncertainty, and the retrospective design cannot establish that chemotherapy class, antimicrobial exposure, hospitalization or any other measured factor caused the infections.

Overall, the study shows that multi-label modelling is feasible for exploring several bloodstream-infection profiles within the same chemotherapy cohort while preserving polymicrobial episodes. It also shows that explicitly allowing dependencies among infection labels did not automatically improve predictive performance: the chained random-forest model performed only marginally better during cross-validation than an independent model that fitted separate binary classifiers. The clearest signals involved prior resistant infection, hospitalization duration, hematologic disease, cytopenias and treatment exposures, while chemotherapy classes showed smaller, group-specific associations. Because the study was observational, confined to infected cycles and centered on one Brazilian cancer center, the models are better viewed as a source of hypotheses for larger, externally validated studies than as validated instruments for selecting antimicrobial treatment.

Subject of Research: Multi-label machine learning of bloodstream-infection risk factors in chemotherapy patients

Article Title: Application of machine learning to identify risk factors for bloodstream infections caused by different antimicrobial resistance phenotypes in chemotherapy patients

Article References: Villani Junior, A., Freire, M. P., De Padua Covas Lage, L. A., Lazar Neto, F., Oliveira, M. S., Abdala, E., Nunes, F. L. S., & Levin, A. S. S. (2026). Application of machine learning to identify risk factors for bloodstream infections caused by different antimicrobial resistance phenotypes in chemotherapy patients. Supportive Care in Cancer, 34(11), Article 1127. https://doi.org/10.1007/s00520-026-11379-2

Image Credits: AI Generated

DOI: 10.1007/s00520-026-11379-2

Keywords: machine learning, bloodstream infections, chemotherapy, antimicrobial resistance, cancer care, carbapenem resistance, fungal infections, clinical prediction, infectious diseases, retrospective cohort, Application, machine