We obtained daily patient data from 2019 to 2023 through Cerner Corporation, Kansas City. The data originates from University Health (UH) in Kansas City, Missouri, which encompasses 35 facilities, including two main campuses: University Health Truman Medical Center and University Health Lakewood Medical Center, as well as several primary care clinics. University Health Truman Medical Center is a 314-bed urban hospital with a Level I trauma center designation, offering both acute care and outpatient services. University Health Lakewood Medical Center, a 101-bed facility, provides similar comprehensive care. Both campuses are located in Kansas City, Missouri.

In this study, hospital location and nursing unit data were recoded based on the incidence and prevalence of infections to facilitate analysis. Locations with low incidence rates were categorized as “Other” to streamline the data into more manageable groups. Hospital locations were grouped into four main categories: UH Lakewood Medical Center, UH Truman Medical Center, UH Truman Medical Center BH, and Other. Similarly, nursing units were categorized according to their primary functions, which included Inpatient Units, Mental Health Units, Surgery Units, Rehabilitation and Observation Units, Critical Care and Intensive Care Units, and Others. The Inpatient Units included wards such as Inpatient-3 Blue (indicated third floor of the hospital inpatient unit blue region), Inpatient-3 Gold, Inpatient-3 Red, Inpatient-4 Blue, Inpatient-4 Green, Inpatient-4 Red, and Inpatient-5 North. The Mental Health Units comprised Behavioral Health Services, Center of Mental Health, and Geriatric Psychiatry. Surgery Units included Surgery ICU, Med/Surg North, and Med/Surg South. Rehabilitation and Observation Units comprised Medical Rehab and 4 Gold Observation, while the Critical Care and Intensive Care Units included ICU, MICU II, Progressive Care Unit, and Cardiac Care Unit. All remaining nursing units were classified under the “Other” category.

This recoding process was undertaken to ensure that similar units were grouped together, thereby facilitating more meaningful comparisons across different hospital settings. Table 1 provides a summary of data across three categories: HA-UTI, CA-UTI, and UTI-free (i.e., patients without UTI) from 2019 to 2023.

We compared two primary groups: HA-UTI versus CA-UTI, and HA-UTI versus UTI-free. This approach allowed us to investigate differences in patient characteristics and outcomes between UTI associated with community versus healthcare settings, as well as the impact of UTI presence or absence in patients with prolonged hospital stays. It is widely accepted that a hospital-acquired infection (HAI) is defined by a minimum length of stay (LOS) of 48 hours. In this study, however, we adopted a slightly different approach. We categorized UTI as community-associated if the LOS was less than 96 hours (four days), and healthcare-associated if the LOS was four days or longer. This distinction is supported by research indicating that UTIs typically result in an additional four days (95% CI: 3.1–5.0 days) of hospitalization [37]. As such, hospitalized patients who develop a UTI are expected to remain in the hospital for at least four days on average. Due to the significant imbalance between UTI and non-UTI patients in the CA group, we did not compare UTI versus non-UTI patients within that group to avoid skewed results.

In this study, we included patient demographics, hospital variables, and socioeconomic factors. Specifically, demographic predictors included age, gender, ethnicity, marital status, and employment status. Hospital-related variables included the hospital location and the specific nurse unit. Socioeconomic factors, such as poverty level, the percentage of the population, level of education, median income, and population density, were not directly obtained from individual patients. Instead, these factors were assigned based on the ZIP code of the patient’s residence, with data acquired from [11]. This means that while we did not have individual-level socioeconomic data, we used the median household income and other factors associated with the patient’s ZIP code as proxies.

We employed several other machine-learning techniques for predictive modeling, including Logistic Regression (LR), Decision Tree (DT), Neural Network (NN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). We began with a multivariable logistic regression model [18], employing a stepwise approach to select the most relevant variables. This method iteratively adds variables to the model, ceasing when no further improvement is observed, as determined by the Akaike information criterion. DTs [19] are one of the most commonly used classification methods in machine learning, as they partition datasets into smaller, manageable groups based on specific criteria. This process of splitting is performed in a hierarchical structure, beginning from a root node and moving toward leaf nodes, where final classifications are made based on Boolean decisions at each step [12]. In our study, the DT segmented the patient population using features such as age, nurse unit, and socioeconomic factors, optimizing these splits at each decision point independently. The Gini coefficient was used to measure node impurity, allowing the model to choose the most effective splits. This approach is advantageous due to its simplicity and interpretability, especially in clinical settings where understanding the decision-making process is critical [12]. NNs, on the other hand, consist of multiple layers that transform input data through a combination of linear and nonlinear functions, capturing complex relationships between variables [20]. The NN used in our study was constructed with two hidden layers and employed a sigmoid activation function at the output layer. RF, an ensemble learning method, constructs multiple decision trees during training and aggregates their outputs by taking the mode of the classes (for classification) or the mean prediction (for regression [21]). This method combines decision trees with bootstrapping and aggregation techniques, providing a versatile framework capable of handling both regression and classification tasks. RF is particularly effective in improving prediction accuracy and reducing overfitting by averaging the results of individual trees. Additionally, due to its structure, RF is highly scalable and can be parallelized to handle large datasets, making it suitable for Big Data applications [13]. Additionally, XGBoost, a powerful gradient-boosting technique, improves predictive accuracy by sequentially adding new models to correct errors made by previous models [22]. Known for its scalability and efficiency, XGBoost has been optimized to handle large datasets with sparse or structured data and is renowned for its high computational performance by leveraging parallel and distributed computing. Its innovations, such as handling sparse data and implementing a weighted quantile sketch, make it highly effective for real-world machine-learning tasks [14].

To handle missing values, we applied an imputation technique as follows. We utilized the K-Nearest Neighbor (K-NN) imputation method to estimate and fill in missing values, following the approach described by Malarvizhi and Thanamani [15]. This step ensured that our dataset was complete and that the models could be trained effectively without bias due to missing data.

To address the class imbalance issue present in the data, we used the SMOTE (Synthetic Minority Over-sampling Technique [23]). Using the SMOTE, we enhanced the representation of the minority class, i.e., the incidences of HAIs, and minimized the likelihood of potentially biased model predictions. Briefly, the SMOTE generates synthetic samples by interpolating between randomly selected minority class examples and their nearest neighbors, identified using the K-NN algorithm [16]. This approach contrasts with traditional oversampling methods [24], which simply duplicate existing minority classes to reduce overfitting while improving model generalization.

After applying SMOTE, the dataset was split into training and testing sets, with 80% of the data allocated for training and the remaining 20% reserved for testing. Cross-validation was also conducted to refine the models further and ensure robustness in the evaluation. Model performance was assessed using the area under the curve (AUC) metric, and confusion matrices were generated for each model with a classification threshold set at 0.5. Additionally, the structure of the decision tree model was presented for transparency and interpretability, allowing for a clearer understanding of the decision-making process within the model.

Table 2 provides descriptive statistics of 2019–2023 patients with healthcare-associated infections (HAIs). We can observe distinct differences between those with and without urinary tract infections (UTI) based on various demographic, hospital, and socioeconomic factors. Patients with UTI tend to be older, with an average age of 63.5 years compared to 48.2 years for those without UTI. A higher percentage of UTI patients are female (65.7%) compared to those without UTI (56%). The length of stay (LOS) is significantly longer for UTI-free patients (78.5 days) compared to HA-UTI patients (20.1 days). This difference stems from the study’s selection criteria, which included only patients hospitalized for four or more days in the UTI-free group, thereby excluding short-stay, non-UTI cases. Consequently, this group mainly comprises long-term hospitalized patients with chronic conditions or post-surgical complications. LOS also varied across years, with annual averages ranging from 55.1 to 91.3 days, contributing to the overall median of 78.5 days. UTI patients are associated with specific nurse units (58.7%) and hospital locations (72.9%) compared to those without UTI, who have much lower association with specific nurse units (8.3%) and different healthcare facilities (18.2%). In terms of socioeconomic factors, there are minor differences between the two groups.

Table S1 compares patients with healthcare-associated UTIs (HA-UTI) and community-associated UTIs (CA-UTI). There are notable differences based on demographic, hospital, and socioeconomic factors. HA-UTI patients are older compared to CA-UTI patients (average age of 63.5 years compared to 49.7 years). A higher percentage of CA-UTI patients are female (84.7%) compared to HA-UTI patients (65.7%). The length of stay is significantly longer for HA-UTI patients, averaging 20.1 days, compared to 0.6 days for CA-UTI patients. HA-UTI patients are more frequently located in the inpatient units (58.7%), mainly at UH Truman Medical Center (72.9%), compared to CA-UTI patients (4.0% and 30.1%, respectively). In terms of socioeconomic factors, both groups show similar levels of income, poverty, and education, with minor differences.

Tables 3 and S2 summarize the performance of the DT, NN, LR, RF, and XGBoost models across two predictive tasks: differentiating between HA-UTI and UTI-free cases(Table 3) and distinguishing between HA-UTI and CA-UTI (Table S2). The Logistic Regression (LR) model performs robustly in both tasks, with an overall accuracy of 95.9% (HAI with or without UTI) and 93.2% (HA-UTI vs. CA-UTI). It achieves high specificity (96.4% in Table 3, 92.6% in Table S2) and AUC (94 and 96, respectively), positioning it as a reliable general-purpose model. The Decision Tree (DT) model excels in sensitivity, with 87% in Table 3 and 95% in Table S2, making it effective in identifying positive cases, though it trades off some Positive Predictive Value (PPV) (8% in Table 3, 56% in Table S2). The Neural Network (NN) model, while having lower sensitivity (48%−88%), compensates with strong specificity, particularly in Table S2 (93.4%).

The Random Forest (RF) model strikes a balance between sensitivity and specificity, with the PPV of 13.5% for the HA-UTI prediction (Table 3) and 60.2%−64.5% for HA-UTI vs. CA-UTI (Table S2). It also shows high cross-validation stability, achieving 99.1% (Table 3) and 96.2% (Table S2), demonstrating robustness across different data subsets. XGBoost performs well in both tasks, especially in AUC (92.1 in Table 3 and 95.7 in Table S2), highlighting its effectiveness for structured data and consistency in producing accurate predictions.

We utilized stepwise logistic regression to develop predictive models in two distinct analyses: first, to differentiate between (HAI) and (CAI) among UTI patients, and second, to distinguish between (HAI) in patients with and without UTI. Stepwise logistic regression works by iteratively adding or removing variables based on their statistical significance in predicting the outcome. It begins with the most significant variable and adds one more variable at each step, testing the model’s performance and removing any variable that no longer contributes meaningfully. This process continues until the best-fitting model, with only the most relevant predictors, is constructed.

The results of the stepwise logistic regression analysis of HA-UTI versus UTI-free are shown in Table 4. An Odds Ratio (OR) greater than 1 suggests an increased likelihood of HA-UTI, while an OR less than 1 indicates a decreased likelihood of HA-UTI. For instance, in 2022, patients in UH Truman Medical Center (UHTMC) had 12.66 times higher odds of having an HA-UTI compared to those not in other facilities, highlighting the elevated risk in these settings due to factors such as prolonged stays or invasive procedures. Also, in 2019, the Mental Health nurse unit showed a decreased likelihood of patients acquiring an HA-UTI compared to those without (OR = 0.38). This trend continues in 2022 and 2023, with ORs of 0.38 and 0.27, respectively. Age is also a significant predictor, with increasing odds of HA-UTI in older patients, with ORs of 1.65 in 2020 and 1.51 in 2022. Gender also plays a role, with females showing a lower likelihood of acquiring HA-UTI compared to males without UTI, as seen in 2021 (OR = 0.33) and 2022 (OR = 0.41). Both UH Truman Medical Center (UHTMC) and UH Lakewood Medical Center (UHLMC) show higher odds of HAI with UTI in 2020, with ORs of 12.66 and 7.81, respectively, indicating a greater risk of acquiring HA-UTI in these locations. Ethnicity emerges as a predictor only in 2022, with an OR of 3.18, suggesting that certain ethnic groups may face a higher risk of acquiring UTI among HAI patients.

Similarly, Table S3 identifies significant predictors for HA-UTI versus CA-UTI. An Odds Ratio (OR) greater than 1 suggests an increased likelihood of HA-UTI, while an OR less than 1 indicates a decreased likelihood of HA-UTI. For instance, in 2019, patients in mental health units had 45 times higher odds of having an HA-UTI versus CA-UTI compared to those not in these units (OR = 45.32), highlighting the elevated risk in these settings due to factors such as prolonged stays or invasive procedures. The inpatient nurse unit also shows a positive association with HA-UTI, especially in 2022 (OR = 1.83), suggesting that patients in these units face higher infection risk, potentially due to longer hospital stays or more complex care needs. Both UH Lakewood and UH Truman Medical Centers show varying degrees of positive associations with HA-UTI. For example, UH Truman demonstrated strong associations with HA-UTI in 2020 (OR = 17.63), indicating an increased risk in these hospital locations. Age consistently appears as a predictor, with increasing age associated with a slightly higher likelihood of HA-UTI (OR ~1.4 in 2020 and 2021). Overall, this analysis underscores the importance of specific nurse units and hospital locations in influencing the risk of HA-UTI compared to CA-UTI, and it suggests the need for targeted interventions, particularly in mental health and inpatient units, as well as ongoing monitoring of risk factors like age.

Normalized importance in CRT represents how much each variable contributes to improving the model’s classification accuracy. Table 5 shows the normalized importance of variables for distinguishing between HA-UTI and CA-UTI from 2019 to 2023. The Nurse Unit (Other) consistently ranks as the most significant predictor across all years, with importance scores ranging from 73.3% in 2020 to 97.3% in 2021, underscoring its central role in distinguishing HA-UTI from CA-UTI. Other variables such as Age, Poverty, Level of education, and Median Income show varying degrees of importance over time but are notably less influential compared to the Nurse Unit (Other). For instance, Age reaches its highest importance in 2022 (2.9%), and Poverty is particularly significant in 2020 (11.3%). Facility locations, like UH Truman Medical Center (UHTMC), also feature intermittently, particularly in 2019 (0.9%) and 2020 (5.5%), reflecting its shifting dynamics in predicting HA-UTI vs. CA-UTI over the years. This variation underscores the evolving nature of the contributing factors, especially for socioeconomic variables, which are more prominent in some years than others. Overall, Table 5 clearly indicates that the Nurse Unit (Other) remains the dominant factor, with other demographic and socioeconomic variables playing supporting but less consistent roles over time.

Similarly, Table S4 presents the normalized importance factors for predicting HA-UTI versus UTI-free from 2019 to 2023. Again, the Nurse Unit (Other) consistently ranks as the most influential predictor across all years, with importance scores ranging from 58.2% in 2020 to 93.6% in 2021, underscoring its central role in distinguishing between HA-UTI and UTI-free cases. Other variables, such as Gender (Male), Age, and Income, contribute less consistently. For example, Income shows a strong influence in 2020 (22.7%), while Gender (Male) holds importance in 2019 (3.3%) and 2022 (3.4%). Age also plays a role, particularly in 2022 (6.7%). Hospital locations, such as UH Truman Medical Center (UHTMC) and UH Lakewood Medical Center (UHLMC), are relevant in some years, with UHTMC showing a 7.7% importance in 2020 and UHLMC contributing 3.0% in 2023. Socioeconomic factors, such as Population Density and Poverty, emerge sporadically, with Population Density having a higher significance in 2022 (3.2%) and Poverty playing a role in 2023 (1.5%).

Overall, Table S4 highlights the dominant role of the Nurse Unit (Other), with demographic, socioeconomic, and hospital location factors contributing variably over the years.

The tree diagrams corresponding to each decision tree model are provided in Section S5 of the supplementary document. The size of each compartment reflects the proportion of the population that belongs to the respective group. Additionally, distinct colour codes have been applied to differentiate between HA-UTI, CA-UTI, and UTI-free cases. As illustrated in Figures S3–S12, the tree diagram results align with the normalized variable importance rankings shown in Tables 5 and S4.

The decision tree in figure 1 classifies UTIs as either Healthcare-Associated Infection (HAI) or Community-Acquired Infection (CAI) based on several factors. The root node splits based on Nurse Unit (Other). If the nursing unit is classified as “Other,” the left branch is taken. This branch further splits into “Rehabilitation and Observation Units,” then “Nurse Unit Surgery,” and finally “AGE.” The right branch, where the nursing unit is not “Other,” splits based on “Poverty,” then “Location: UH Lakewood Medical Center,” and “Graduate degree. The tree suggests that the type of nursing unit is a strong initial predictor. Rehabilitation and Observation Units show a higher propensity for HAI. Within the “Other” nursing unit category, surgical units also lean towards HAI, though age plays a role too. For cases not in the “Other” category, poverty, location at UH Lakewood Medical Center, and lower levels of graduate education are associated with CAI. The colours blue and orange represent HAI and CAI, respectively. The Gini index indicates the impurity of each node, with higher values indicating a more even mix of HAI and CAI.