In an observational cohort study evaluating 1,243 hospitalization episodes among 969 adults with acute leukemia, computational association rule mining successfully identified co-occurring clinical baseline patterns linked to incident pulmonary infection, yet converting those patterns into binary predictors failed to outperform conventional multivariable logistic regression. Conducted by medical researchers at Sichuan Provincial People’s Hospital and published in Frontiers in Oncology, the investigation sought to determine whether uncovering multidimensional combinations of clinical characteristics through an algorithmic approach could enhance baseline risk prediction for hospital-acquired pulmonary complications compared to traditional regression-based methods. While the mined patterns revealed clinically interpretable profiles, the standard regression model demonstrated superior discrimination across both development and temporal validation cohorts.
Hospital-acquired respiratory infections represent a major clinical hazard for patients undergoing treatment for acute leukemia, a category of hematologic malignancies marked by rapid clonal proliferation of abnormal hematopoietic stem or progenitor cells. The aggressive clinical course of the disease necessitates intensive chemotherapy regimens that induce profound myelosuppression and systemic immunosuppression. In tandem with leukemia-related disruption of normal leukocyte development, this therapeutic marrow suppression leaves patients highly vulnerable to severe opportunistic complications. Previous clinical reports indicate pulmonary infection rates ranging from 42 percent to 44 percent in adult acute myeloid leukemia cohorts, with invasive fungal infections contributing substantially to treatment-related morbidity and mortality during intensive induction therapy.
Evaluating baseline vulnerability remains challenging because respiratory infection risk arises from complex, simultaneous interactions between patient-specific characteristics, hematologic impairment, inflammatory status, and treatment phases. Traditional multivariable regression models typically estimate the independent statistical contribution of individual candidate variables, which may overlook how specific clinical characteristics cluster together in clinical reality. To address this limitation, the researchers deployed association rule mining via the Apriori algorithm. Unlike traditional regression, association rule mining functions as an exploratory, data-driven technique that systematically identifies frequently co-occurring sets of features within a dataset, quantifying the relationships through standardized metrics of support, confidence, and lift.
The study analyzed adult patients admitted between January 2019 and July 2024 with confirmed diagnoses of acute myeloid leukemia, acute lymphoblastic leukemia, or mixed-phenotype acute leukemia. Hospitalization episodes, treated as the primary analytical unit, were split chronologically to enable robust temporal validation. The development cohort comprised 994 hospitalization episodes occurring between January 2019 and September 2023, while the temporal validation cohort included 249 consecutive episodes recorded between October 2023 and July 2024. Patients presenting with active pulmonary infection at the time of hospital admission were excluded, ensuring that the primary outcome focused specifically on new-onset pulmonary infections occurring during hospitalization.
Candidate predictors were strictly limited to demographic factors, disease classifications, clinical treatment phases, and initial laboratory measurements obtained within the first 24 hours of hospital admission. Laboratory thresholds were prespecified based on clinical standards, categorizing white blood cell counts into groups below 4, between 4 and 10, or above 10 times 10 to the ninth power per liter, and defining neutropenia as an absolute neutrophil count below 0.5 times 10 to the ninth power per liter. Platelet counts were categorized as below 20, 20 to 50, or above 50 times 10 to the ninth power per liter, serum albumin was stratified at 35 grams per liter, and procalcitonin was evaluated at a threshold of 0.05 nanograms per milliliter. Missing data in baseline variables were addressed within the development cohort using multiple imputation by chained equations across 20 imputed datasets.
Incident pulmonary infections occurred in 620 of the 994 episodes in the development cohort, representing an incidence of 62.4 percent, and in 163 of the 249 episodes in the temporal validation cohort, representing 65.5 percent. Applying the Apriori algorithm with prespecified thresholds—requiring a support of at least 0.10, confidence of at least 0.70, lift exceeding 1.0, and stability across at least 95 percent of imputed datasets—isolated five non-redundant association rules. These recurrent combinations primarily paired induction chemotherapy with distinct laboratory abnormalities, including absolute neutrophil reduction, elevated procalcitonin, elevated white blood cell counts, and moderate thrombocytopenia, alongside a specific combination linking elevated white blood cell counts with elevated procalcitonin.
Among the identified patterns, the combination of induction chemotherapy and absolute neutrophil reduction demonstrated the strongest statistical association with pulmonary infection, exhibiting a support of 0.104, a confidence of 0.904, and a lift of 1.45. This indicates that 90.4 percent of hospitalization episodes exhibiting this specific baseline combination subsequently developed a pulmonary infection, reflecting a 1.45-fold increase over the baseline likelihood. Other stable combinations included induction chemotherapy paired with elevated procalcitonin, which achieved a support of 0.178, confidence of 0.855, and lift of 1.37, as well as induction chemotherapy paired with leukocytosis, which showed a support of 0.176, confidence of 0.849, and lift of 1.36.
Parallel multivariable logistic regression modeling in the development cohort confirmed several independent risk factors. Induction chemotherapy carried the highest odds of incident infection compared with episodes without documented chemotherapy, showing an adjusted odds ratio of 4.92 with a 95 percent confidence interval of 3.12 to 7.76. Pre-chemotherapy status was also associated with increased risk, yielding an odds ratio of 3.67. Independent laboratory predictors included leukocytosis with an odds ratio of 2.61, baseline neutropenia with an odds ratio of 2.30, severe thrombocytopenia below 20 times 10 to the ninth power per liter with an odds ratio of 2.24, and hypoalbuminemia with an odds ratio of 1.72. Notably, elevated procalcitonin alone was not independently associated with infection risk, showing an odds ratio of 1.11, despite its frequent co-occurrence within the mined association rules.
To evaluate predictive utility, the five stable association rules were converted into binary indicators to establish a rule-derived logistic regression model, which was compared directly against the conventional multivariable model. The conventional model demonstrated significantly higher discrimination in the development cohort, achieving an area under the receiver operating characteristic curve of 0.808 with a 95 percent confidence interval of 0.780 to 0.836, compared to 0.760 with an interval of 0.730 to 0.790 for the rule-derived model. In the temporal validation cohort, discrimination remained higher for the conventional model at an area under the curve of 0.781 versus 0.751 for the rule-derived model, though this temporal difference did not reach statistical significance. Both models exhibited acceptable calibration and positive net benefit on decision curve analysis, with the conventional model maintaining higher net clinical benefit.
The study authors highlighted several methodological considerations and limitations inherent to the analysis. Because this investigation was conducted retrospectively at a single medical center, external validation across broader multi-institutional cohorts remains necessary before applying these findings across diverse clinical settings. Furthermore, association rules describe statistical co-occurrence rather than causal pathways, and evaluating infections across entire hospital stays means longer hospitalizations could increase detection opportunities. Ultimately, while rule-derived indicators did not improve quantitative predictive accuracy over established multivariable regression, association rule mining demonstrated clear utility as an exploratory tool for uncovering meaningful clinical combinations and generating hypotheses regarding multidimensional infection vulnerability in acute leukemia.
To ensure diagnostic consistency across the study cohort, acute myeloid leukemia cases were classified according to the fifth edition of the World Health Organization guidelines, while acute lymphoblastic leukemia followed the 2024 Chinese adult diagnostic consensus. The protocol excluded acute promyelocytic leukemia, chronic myeloid leukemia in blast crisis, boundary lymphoma diagnoses, and patients receiving hematopoietic stem cell transplantation. Diagnostic adjudication of pulmonary infections required independent review by two hematologists referencing Chinese hospital-acquired pneumonia guidelines, integrating respiratory symptoms, imaging, and microbiological tests such as sputum cultures, bronchoalveolar lavage, and galactomannan assays. Furthermore, post-admission interventions, including mechanical ventilation, parenteral nutrition, and subsequent glucocorticoid therapy duration, were omitted from baseline modeling to prevent post-baseline contamination.
Because individual patients frequently contributed multiple hospitalization episodes over the study period, the researchers conducted sensitivity analyses with generalized estimating equations using an independence working correlation structure. This clustered modeling confirmed the primary multivariable regression findings. Acute lymphoblastic leukemia exhibited significantly lower odds of infection compared with acute myeloid or mixed-phenotype leukemia, showing an adjusted odds ratio of 0.71 with a 95 percent confidence interval of 0.52 to 0.98. In addition, baseline hypertension independently increased infection likelihood, yielding an odds ratio of 1.60 with a confidence interval of 1.03 to 2.46, while moderate thrombocytopenia between 20 and 50 times 10 to the ninth power per liter produced an odds ratio of 1.77 with a confidence interval of 1.15 to 2.71.
Methodological safeguards also governed biomarker handling and model multicollinearity. Procalcitonin was not numerically imputed due to selective clinical testing patterns; instead, unmeasured tests were captured with an explicit indicator variable alongside measured elevation status. Sensitivity analyses testing higher cutoffs of 0.25 and 0.50 nanograms per milliliter, as well as models excluding procalcitonin entirely, demonstrated consistent lack of independent association. Imputation checks that excluded the infection outcome confirmed the stability of all five Apriori rules. Variance inflation factors among rule indicators ranged from 1.04 to 1.28, verifying minimal collinearity. Although demographic parameters remained comparable, temporal validation episodes featured more induction chemotherapy (57.8 percent versus 44.6 percent) and fewer pulmonary nodules (15.3 percent versus 23.3 percent), demonstrating model generalizability across evolving clinical case mixes.
Subject of Research: Comparison of association rule mining and logistic regression for predicting pulmonary infection in adults with acute leukemia
Article Title: Association rule mining identifies baseline clinical patterns associated with pulmonary infection in adults with acute leukemia: development and temporal validation of rule-derived prediction models
Article References: Xiao, N., Xiao, T., Huang, B., Yang, J., Huang, X., Miao, Y., Wu, H., & Wang, W. (2026). Association rule mining identifies baseline clinical patterns associated with pulmonary infection in adults with acute leukemia: development and temporal validation of rule-derived prediction models. Frontiers in Oncology, 16, Article 1902892. https://doi.org/10.3389/fonc.2026.1902892
Image Credits: AI Generated
DOI: 10.3389/fonc.2026.1902892
Keywords: acute leukemia, pulmonary infection, association rule mining, Apriori algorithm, logistic regression, temporal validation, induction chemotherapy, neutropenia, leukocytosis, thrombocytopenia, procalcitonin, hematologic malignancies, risk stratification, clinical prediction models, ROC curve, calibration slope, decision curve analysis, hospital acquired infection
News Source: Nathaniel Bowman. (November 21, 2026). Data Mining Finds Leukemia Lung Infection Patterns but Misses Predictive Edge. Scienmag.



