Early sepsis prediction in elderly patients with urinary tract infections: a machine learning

2.1 Study design and setting

This retrospective observational study was conducted in the ED of a tertiary care hospital. Ethical approval for the study was obtained from Mardin Artuklu University Non-Interventional Clinical Research Ethics Committee (date: 05 November 2024; approval number: 2024/11-12). Due to the retrospective nature of the study, the ethics committee waived the requirement for informed consent. The research was conducted in accordance with the Declaration of Helsinki.

2.2 Study population

Patients aged over 65 years who presented to the ED with a diagnosis of urinary tract infection (UTI) between January 2019 and December 2023 were eligible for inclusion. The diagnosis was based on the International Classification of Diseases (ICD)-10 codes assigned by emergency physicians at the time of admission, in line with institutional protocols, as detailed in Supplementary Table 1. These codes were retrospectively reviewed to identify eligible cases. Inclusion criteria encompassed patients with typical UTI symptoms, such as dysuria, urgency or frequency. These symptoms were supported by urinalysis findings. These included urine microscopy showing ≥10 white blood cells/µL, and positive dipstick tests for leukocyte esterase and nitrites. We also included patients with atypical symptoms often seen in older adults, such as altered mental status, general malaise or functional decline. These cases were eligible only if they did not meet Sepsis-3 criteria at presentation [7, 11]. Exclusion criteria included a diagnosis of sepsis upon ED arrival, defined by Sepsis-3 as an acute increase of two or more points in the SOFA score; the presence of concurrent secondary infections such as pneumonia or cellulitis; and incomplete clinical or laboratory data. All diagnoses were made by the treating physician and verified retrospectively through ICD-10 coding, without modification, by the study team. Urine blood and protein were defined as positive if dipstick testing or microscopy confirmed hematuria or proteinuria, respectively. The presence of urine microorganisms referred to bacteria observed via microscopic urinalysis.

2.3 Data quality and management

Data quality was ensured through independent extraction by two reviewers. Inter-observer reliability was assessed using a random sample of 120 patient records, representing approximately 10 percent of the cohort, which were independently reviewed by both assessors. Every variable included in the study—demographics, comorbidities, vital signs, laboratory values and clinical assessments—was cross-checked between reviewers. Agreement was excellent, with Cohen’s Kappa for categorical variables reaching 0.94 and intraclass correlation coefficients (ICCs) for continuous variables exceeding 0.98. Discrepancies were resolved by a third reviewer blinded to the initial assessments. A prespecified multiple imputation protocol was established to address missing data if any variable exceeded 5% missingness. However, as no variable exceeded the 5% missing data threshold, complete case analysis was used without the need for imputation.

2.4 Data collection and outcome measures

Demographic data, chronic comorbidities, vital signs at presentation, Charlson Comorbidity Index scores, laboratory parameters at presentation, urine analysis parameters, and the development of sepsis within 72 hours were recorded. Data were retrieved from the hospital information management system. All data were anonymized before analysis to maintain patient confidentiality. Data collection was performed by the researchers to ensure consistent and accurate information gathering. Body temperature values were obtained retrospectively from the triage vital signs recorded at ED presentation, but the method of measurement was not consistently specified in the electronic medical records. Sepsis was defined, according to the Sepsis-3 criteria, as life-threatening organ dysfunction caused by infection, indicated by an acute increase of ≥2 points in the SOFA score [3]. The SOFA score was calculated by assigning a score from 0 to 4 for dysfunction in six organ systems—respiratory, coagulation, liver, cardiovascular, kidney and nervous—resulting in a total score ranging from 0 to 24 [12]. The primary outcome of the study was the development of sepsis within 72 hours of the ED visit. SOFA scores were retrospectively calculated using clinical and laboratory parameters recorded in the electronic health record (EHR) at the time of admission. All values used in the SOFA calculation were derived from physician and nurse documented data. Parameters not measured at admission were not included in the SOFA score calculation. If a parameter required for SOFA calculation was not measured, a default score of 0 was assigned for that component, as per standard retrospective sepsis study methodology. This reflects clinical practice, where unmeasured values are often considered normal unless clinically indicated. Sepsis development within 72 hours was determined by reviewing subsequent clinical and laboratory data recorded in the EHR during hospitalization. If a patient was discharged before 72 hours, their last available documented SOFA score was used to assess progression. The secondary outcomes assessed included mortality, intensive care unit (ICU) admission, vasopressor use, antibiotic changes and clinical deterioration during follow-up. Mortality was defined as death occurring at any point during hospitalization or follow-up. ICU admission was recorded for patients who required transfer to intensive care. Vasopressor use was defined as the administration of any vasoactive medication (e.g., norepinephrine, dopamine or vasopressin) due to hypotension. Antibiotic change was recorded when a patient’s antimicrobial regimen was modified during hospitalization due to treatment failure, antimicrobial resistance or clinical deterioration. Clinical deterioration was defined as a worsening of hemodynamic parameters, respiratory distress, or the need for escalation of care despite appropriate antibiotic therapy. These outcomes were extracted from electronic medical records and verified by independent reviewers. All predictor variables, including vital signs, laboratory data and clinical features, were recorded at the time of ED presentation.

2.5 Sample selection

For observational studies involving logistic regression analysis, it is generally recommended that the sample size includes at least 100 patients, plus an additional 50 patients for each independent variable in the final model. Alternatively, a minimum of 500 patients is advised to ensure statistical robustness [13]. Given the study’s aim and available data, the sample size of 1176 patients was deemed appropriate to meet these requirements without compromising the reliability of the results.

2.6 Analysis

All statistical analyses were performed using R (version 4.4.2, R Foundation for Statistical Computing, Vienna, Austria). Descriptive statistics were used to summarize baseline characteristics. Categorical variables were reported as counts (n) and percentages (%), while continuous variables were presented as medians with interquartile ranges (IQR) or means with standard deviations (SD), as appropriate. The Kolmogorov-Smirnov test, accompanied by visual inspections using histograms, was applied to assess the normality of continuous variables. Univariable logistic regression analysis was performed to assess the association between individual variables and sepsis development within 72 hours. Continuous variables were dichotomized at their median values and results were expressed as odds ratios (OR) with 95% confidence intervals (CI). Variables in the univariable analysis with an OR whose 95% CI did not include 1 were selected for inclusion in the multivariable logistic regression model to identify independent predictors while controlling for confounding factors. A stepwise backward elimination approach was applied to retain variables with p < 0.05 in the final model, and results were reported as adjusted ORs (aOR) with 95% CI. All ML models used the same predictors identified as significant in multivariable logistic regression to ensure comparability. This approach ensured that any differences in performance were due to the modeling technique rather than discrepancies in variable selection. GAM, LASSO regression, and decision trees were implemented using age, BUN, CRP, creatinine, systolic blood pressure, respiratory rate, and temperature as predictors. These variables were chosen based on their statistical significance in the logistic regression model, allowing for direct comparison across methods. For logistic regression, continuous variables were dichotomized at their median values to facilitate interpretability and address non-linearity. In contrast, ML models (GAM, LASSO and decision tree) were constructed using continuous predictors without transformation, allowing for non-linear relationships assessment and threshold-based effects to emerge naturally.

GAM was employed to capture potential non-linear relationships among these predictors. LASSO regression was applied to assess whether penalization techniques could improve predictive accuracy while maintaining interpretability. A decision tree model was developed to explore threshold-based risk stratification for bedside utility. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), Brier scores for calibration, and DeLong’s test for statistical comparisons between models. To assess the robustness of the logistic regression model, 5-fold cross-validation was performed. The dataset was randomly divided into five equal parts, and the model was trained on four parts and tested on the remaining part in an iterative manner. The AUC was calculated for both the training and testing datasets in each fold to evaluate the discriminatory ability of the model. The overall AUC values were also calculated by pooling the results across folds, with 95% CIs. Comparative model performance for GAM, LASSO, and decision tree models was evaluated using the test dataset from the 80/20 train-test split. This allowed for direct comparison using identical data inputs. Logistic regression was also evaluated via 5-fold cross-validation for internal validation, with its final performance also reported on the test set for consistency across models. Multicollinearity among predictors was checked using the variance inflation factor (VIF), with a threshold of VIF <5 considered acceptable. Missing data were minimized by careful data extraction. The handling of missing data followed a predefined strategy approved by the ethics committee. If missing values were present in fewer than 5% of cases for a given variable, multiple imputation would have been applied to maintain statistical robustness [14]. For variables with more than 5% missing data, affected cases would have been excluded from the analysis to minimize potential bias. However, in the final dataset, no variable exceeded the 5% threshold for missingness, and therefore, all available data were included in the analysis without the need for imputation or case exclusion. Summary of missingness across predictors and outcome is given in Supplementary Table 2.

3.1 Incidence of sepsis and univariable analysis

In this study, 1176 elderly patients presenting with UTI to the ED were analyzed (Fig. 1), of which 139 patients (11.8%) developed sepsis within the first 72 hours, whereas 1037 patients (88.2%) did not. Among patients who developed sepsis, the median SOFA score at the time of diagnosis was 4 (IQR 3–6), while the median SOFA score in the non-sepsis group remained 1 (IQR 0–1). Univariable analysis of patient characteristics and vital signs (Table 1) showed that older age (OR 1.15, 95% CI 1.12–1.19), diabetes mellitus (OR 3.18, 95% CI 2.18–4.61), and dementia (OR 2.86, 95% CI 1.76–4.54) were associated with a higher risk of sepsis. Patients with higher respiratory rates (OR 1.32, 95% CI 1.27–1.38), altered mental status (OR 2.55, 95% CI 1.78–3.66), and elevated temperature (OR 2.09, 95% CI 1.72–2.55) had increased odds of sepsis. In contrast, higher systolic (OR 0.86, 95% CI 0.84–0.88) and diastolic blood pressure (OR 0.54, 95% CI 0.47–0.60) were negatively associated with sepsis.

Fig. 1.Patient flowchart.

Univariable analysis of laboratory parameters (Table 2) identified elevated CRP (OR 4.44, 95% CI 2.93–6.91), higher glucose (OR 3.00, 95% CI 2.04–4.50), BUN (OR 2.08, 95% CI 1.44–3.05), creatinine (OR 1.92, 95% CI 1.32–2.82), and lactate (OR 1.90, 95% CI 1.32–2.76) as significant risk factors. Higher platelet count (OR 0.52, 95% CI 0.36–0.74) was inversely associated with sepsis. The presence of urinary microorganisms (OR 2.10, 95% CI 1.47–3.01) was a significant predictor, while urinary blood and proteinuria were not significantly associated with sepsis.

3.2 Multivariable modeling

Logistic regression analysis identified age (aOR 1.10, 95% CI 1.05–1.14), BUN (aOR 1.03, 95% CI 1.00–1.05), CRP (aOR 1.04, 95% CI 1.03–1.05), creatinine (aOR 1.79, 95% CI 1.11–3.14), systolic blood pressure (BP) (aOR 0.95, 95% CI 0.92–0.97), respiratory rate (aOR 1.10, 95% CI 1.05–1.16), and temperature (aOR 2.52, 95% CI 1.65–4.38) as statistically significant predictors of sepsis development (Table 3). Age, elevated BUN, CRP, creatinine, increased respiratory rate, and elevated temperature were associated with higher odds of developing sepsis, while lower systolic BP was also independently associated with sepsis risk. Factors such as Diabetes Mellitus, Dementia, white blood cell count, glucose, platelets, lactate, urine microorganism, altered mental status, and Charlson comorbidity index were not statistically significant in the multivariable analysis for predicting sepsis development.

The area under the curve (AUC) for the multivariable logistic regression model was evaluated for both training and testing datasets. The overall training AUC was 0.792 (95% CI 0.742–0.841, p < 0.001), while the overall testing AUC was 0.824 (95% CI 0.733–0.915, p < 0.001) (Table 4). Fold-specific training AUCs ranged from 0.779 to 0.824, with testing AUCs ranging from 0.730 to 0.840 across the five folds (Table 4). To facilitate clinical interpretation, we also analyzed age as a continuous variable in 10-year increments. In this additional analysis, a 10-year increase in age was associated with 4.20 times higher odds of sepsis (95% CI 3.22–5.56, p < 0.001). The goodness-of-fit of the model was evaluated using the Cox & Snell R Square, Nagelkerke R Square, and the Hosmer and Lemeshow test. The Cox & Snell R Square was 0.320, and the Nagelkerke R Square was 0.433. The Hosmer and Lemeshow test indicated an acceptable fit (Chi-square = 6.654, df = 8, p = 0.284).

To assess predictive performance, we compared logistic regression with GAM, LASSO regression and Decision Tree models (Table 5). The area under the curve (AUC) for the multivariable logistic regression model was 0.792 (95% CI 0.742–0.841) in the training set and 0.824 (95% CI 0.733–0.915) in the test set. GAM analysis demonstrated an AUC of 0.9545 (95% CI 0.9317–0.9773, p < 0.001). Unlike traditional regression models, GAM allows for smooth, non-linear relationships between predictors and sepsis risk. As shown in Fig. 2, the risk of sepsis increases non-linearly with age, particularly accelerating beyond 78 years. Similarly, blood urea nitrogen and C-reactive protein exhibit threshold effects, where risk significantly rises beyond 35 mg/dL and 70 mg/L, respectively. The effect of respiratory rate is also non-linear, with a notable increase in risk occurring beyond 22 breaths/min. The deviance explained was 63.5%, with an adjusted R² of 67.8%, supporting the improved predictive performance of GAM over logistic regression.

Fig. 2.Predicted effects of age, blood urea nitrogen, C-reactive protein, and respiratory rate on sepsis risk. Partial effect plots from the Generalized Additive Model showing non-linear associations between predictors and sepsis probability. Threshold effects for age (>78 years), CRP (>70 mg/L) and BUN (>35 mg/dL) are apparent.

LASSO regression achieved an AUC of 0.9421 (95% CI 0.9103–0.9738). The model retained the key predictors identified in logistic regression: age, BUN, CRP, systolic blood pressure, respiratory rate and temperature. The coefficient paths across different regularization strengths (Fig. 3) illustrate how predictor importance shifts as the penalty increases, reinforcing the robustness of the selected features. The decision tree model identified clear stratification thresholds, with CRP ≥70 mg/L, respiratory rate ≥22 breaths/min, and BUN ≥35 mg/dL emerging as the strongest predictors (Fig. 4). The structure of the decision tree provides an intuitive, rule-based approach to sepsis risk prediction. Each branching point represents a clinical cutoff, guiding risk assessment in a stepwise manner. For instance, patients with CRP levels above 70 mg/L were categorized as high risk, while those with lower CRP but elevated respiratory rates and BUN levels were also identified as at-risk subgroups. This visualization supports clinical interpretation without requiring complex calculations. The model had an AUC of 0.9152 (95% CI 0.8842–0.9462).

Fig. 3.Coefficient paths in least absolute shrinkage and selection operator (LASSO) regression. Each colored line represents a predictor variable. As the regularization strength increases (log lambda moves right), less informative coefficients are progressively shrunk toward zero and excluded from the model. This visualization illustrates the variable selection process and helps identify which predictors remain robust across a range of penalization levels. LASSO: Least Absolute Shrinkage and Selection Operator.

Fig. 4.Decision tree model for sepsis prediction. Each node represents a decision point based on a clinical variable and threshold. The top value in each box is the predicted class (0 = no sepsis, 1 = sepsis). The middle value represents the predicted probability of sepsis for patients in that node, and the bottom percentage indicates the proportion of the total population falling into that category. Green boxes represent terminal nodes predicting sepsis (class 1), and blue boxes represent terminal nodes predicting no sepsis (class 0).

3.3 Machine learning performance

Comparative analysis using DeLong’s test revealed statistically significant differences in AUC performance across the models. GAM significantly outperformed logistic regression (p < 0.001), as did LASSO (p < 0.001). Additionally, GAM demonstrated a statistically significant improvement over the decision tree model (p = 0.046). However, no significant difference was observed between LASSO and the decision tree (p = 0.235) or between GAM and LASSO (p = 0.548). These results suggest that GAM and LASSO are both superior to logistic regression, while GAM also holds a slight edge over the decision tree. For calibration assessment, Brier scores were computed, with LASSO exhibiting the best calibration (Brier Score = 0.0240), followed by GAM (0.0331), logistic regression (0.0367) and the decision tree (0.0379). While GAM had the highest AUC, LASSO demonstrated superior specificity (0.986 vs. 0.959), positive predictive value (PPV) (0.894 vs. 0.742), and F1 score (0.900 vs. 0.801), suggesting improved performance in identifying true positive cases with fewer false positives.

Net Reclassification Improvement (NRI) and Integrated Discrimination Improvement (IDI) analyses further demonstrated the superior predictive value of GAM compared to logistic regression. The NRI for GAM versus logistic regression was 0.1397 (13.97%), indicating an improved classification of sepsis cases. The decision tree model, while useful, showed a modest improvement with an NRI of 0.0441 (4.41%). These findings suggest that GAM offers the most substantial enhancement in predictive accuracy for early sepsis detection in elderly patients with UTI.