Machine Learning to Target Procalcitonin Testing

ABSTRACT

Objectives: Procalcitonin (PCT) testing is widely used in infection management but may be overused in routine clinical practice. This study developed and validated machine learning (ML) models to predict elevated PCT levels (≥ 0.50 ng/mL) using routine, low-cost laboratory parameters to support diagnostic stewardship.

Study Design: We conducted a retrospective observational study using real-world clinical laboratory data from a tertiary hospital to develop and temporally validate predictive ML models for PCT positivity.

Methods: Four ML algorithms—logistic regression (LR), random forest (RF), eXtreme Gradient Boosting (XGBoost), and support vector machine—were trained using routine complete blood count indices, C-reactive protein, and serum creatinine. Patient-level data splitting and isotonic probability calibration were used. Performance was evaluated through internal and temporal validation (n = 1265) using AUROC, accuracy, and decision curve analysis.

Results: In temporal validation, LR demonstrated the highest discrimination (AUROC, 0.90; 95% CI, 0.88-0.92) and strong accuracy (84%). RF performed similarly (AUROC, 0.89; accuracy, 84%), and XGBoost achieved an AUROC of 0.88. When optimized for a rule-out clinical strategy to minimize missed cases, the RF model achieved a sensitivity of 96% and a negative predictive value of 96% in the temporal validation cohort. Decision curve analysis confirmed that ML-guided testing provides higher net clinical benefit than universal testing. Shapley Additive Explanations analysis identified composite features as the most impactful predictors.

Conclusions: ML models using inexpensive routine laboratory tests predicted PCT positivity with high discrimination. Implementing these calibrated models may reduce unnecessary PCT testing and support diagnostic stewardship.

Am J Manag Care. 2026;32(11):632-640

doi:10.37765/ajmc.2026.90042

Takeaway Points

Procalcitonin testing is costly and often overused, and health systems need practical ways to identify which patients are likely to benefit from it.

• Using inexpensive, routinely collected laboratory data, machine learning models predicted procalcitonin positivity with high discrimination in a validation cohort.
• The models identified patients unlikely to require procalcitonin testing, reducing modeled testing volume by approximately 27% while maintaining a similar per-patient cost.
• By flagging only high-risk patients for testing, the models support diagnostic stewardship.
• With prospective validation, hospitals, insurers, and health systems could use this approach to manage diagnostic test utilization and promote value-based care.

Procalcitonin (PCT) is a widely used biomarker for differentiating bacterial from viral infections and assessing disease severity.^1,2 It supports clinical decision-making across emergency and inpatient settings and, depending on the selected cutoff, has high sensitivity for detecting bacteremia, which can help identify patients at low risk.^3,4 When combined with clinical tools such as the Shapiro-Procalcitonin algorithm, PCT reduces unnecessary blood culture use without compromising diagnostic accuracy.^5,6 PCT-guided algorithms also support antibiotic stewardship, particularly in lower respiratory tract infections, where antibiotic initiation may be discouraged at PCT concentrations below 0.25 ng/mL and encouraged at concentrations of 0.50 ng/mL or higher, depending on the clinical context.^2,7,8

Despite strong clinical utility, PCT testing is costly and often overused, particularly in high-volume emergency departments.⁹ Machine learning (ML)–based clinical decision support (CDS) tools may help reduce unnecessary PCT orders by leveraging routinely collected and inexpensive tests.^10-13 These tools can refine diagnostic reasoning, reduce variability in ordering practices, and promote cost-effective use of resources. Previous ML studies in this field have primarily focused on predicting sepsis or bacterial infections, often incorporating PCT as a predictor rather than treating PCT itself as the target outcome.^14,15 In contrast, relatively few studies have investigated whether PCT positivity or negativity can be predicted from routinely available laboratory parameters to guide test-ordering decisions. Agnello et al explored an ML-based strategy to improve the appropriateness of PCT requests, highlighting the potential of this approach for diagnostic stewardship.¹⁶ Furthermore, although C-reactive protein (CRP) and serum creatinine (SCr) are known to be associated with PCT concentrations, their combined predictive contributions remain insufficiently characterized in ML-based PCT prediction. These gaps underscore the need for models that directly predict PCT positivity using inexpensive routine laboratory tests and evaluate their performance through temporal validation and clinically relevant decision thresholds.

In this study, we developed ML-based clinical decision support models using routinely available, low-cost laboratory tests such as complete blood count (CBC), CRP, and SCr to predict PCT positivity (≥ 0.50 ng/mL) in both pediatric and adult populations. Because these tests are substantially less expensive than PCT, the proposed approach enables cost-efficient identification of patients unlikely to benefit from PCT testing. Integration of these models into clinical workflows may reduce unnecessary PCT orders, support more targeted diagnostic evaluation, and indirectly support antimicrobial stewardship.

METHODS

This study was approved by the Izmir City Hospital Clinical Research Ethics Committee, with informed consent waived. ML model development followed current methodological guidelines.^12,17

Data Collection and Preprocessing

The study analyzed CBC results and PCT, CRP, and SCr levels for all inpatients and outpatients at Izmir City Hospital from January 1 to August 31, 2024. To enable temporal validation, the data set was divided chronologically: Encounters collected between January 1 and May 31, 2024, were used for model development, and encounters collected between June 1 and August 31, 2024, were reserved as an independent temporal validation cohort. Laboratory records were extracted from the laboratory information system and linked at the encounter level using the patient identifier and accession information to ensure that predictor tests and PCT originated from the same clinical episode. If duplicate records occurred within the same encounter, only 1 result was retained for analysis. We included PCT measurements accompanied by CBC, CRP, and SCr results from the same clinical episode and excluded records with missing required laboratory measurements. During the study period, 111,980 PCT tests were requested, of which 108,974 were successfully processed and reported. Of these, 24,760 had accompanying CBC, CRP, and SCr requests, excluding 84,214 records without the required combination of tests. After matching results by patient, date, and time, 18,154 records remained. Subsequently, repeated measurements from the same patient were excluded, retaining 1 unique record per patient. The final data set comprised 7047 patients in the development cohort and 1265 patients in the temporal validation cohort; 3375 (47.9%) in the development cohort were positive based on a PCT cutoff of 0.50 ng/mL. The PCT selection process is summarized in Figure 1. PCT results were not selected based on age, sex, diagnosis, or other clinical criteria.

The data set included 25 predictive variables encompassing inflammatory markers, hematological indices, and renal function parameters, as detailed in Table 1.

Missing values were handled using median imputation implemented within a scikit-learn pipeline fitted on only the training data, with identical transformations applied unchanged to the internal test and temporal data sets. No forward-fill or test-informed preprocessing was applied.

CRP and SCr were measured on a cobas c 702 module (Roche Diagnostics) using immunoturbidimetric and kinetic colorimetric (Jaffé) methods, respectively. PCT was analyzed by electrochemiluminescence immunoassay (Elecsys BRAHMS PCT; Roche Diagnostics) on a cobas e 801 platform (Roche Diagnostics). Hematology parameters were obtained using an XN-1000 analyzer (Sysmex Corporation).

Feature Engineering

The unit of analysis was the patient-level record, with 1 eligible encounter retained per patient. Composite variables were prespecified based on clinical rationale and included interaction terms between absolute neutrophil count and CRP (ANC × CRP), serum creatinine and CRP (SCr × CRP), and platelet large cell ratio and CRP (P-LCR × CRP) to capture combined inflammatory and renal-function signals. The neutrophil-to-lymphocyte ratio (NLR) was calculated by dividing the absolute neutrophil count by the absolute lymphocyte count. Distributions of predictive features in the training and test data sets, stratified by PCT positivity, are presented in eAppendix Figure 1 (eAppendix available at ajmc.com).

Model Development

The analysis was performed in Python 3.10 (Python Software Foundation) using the scikit-learn library. PCT results were binarized using a cutoff of 0.50 ng/mL (PCT ≥ 0.50 ng/mL positive).¹⁸ The development data set was split at the patient level into a training/calibration set (80%) and an independent internal test set (20%), ensuring that all observations from a given patient were contained entirely within either the training or test set. The training/calibration set was further partitioned into model-training (80%) and calibration (20%) subsets. The calibration data set was defined as this separate, held-out subset of the development data, which was not used for model fitting or hyperparameter optimization. The ML models considered in this study were logistic regression (LR), random forest (RF), eXtreme Gradient Boosting (XGBoost), and support vector machine (SVM). All 4 models underwent probability calibration using isotonic regression fitted on the calibration data set. This data set was also used for decision-threshold selection. The internal test set and temporal validation cohort were not used for fitting the probability calibrators. Hyperparameter tuning was performed using patient-level grouped cross-validation (GroupKFold). Model interpretation was performed using Shapley Additive Explanations (SHAP) version 0.46.0 (SHAP developers).

Performance Evaluation and Explainability

Model performance was evaluated using several metrics, including area under the receiver operating characteristic curve (AUROC), accuracy, sensitivity, specificity, confusion matrix, positive predictive value (PPV), negative predictive value (NPV), and F1 score (the harmonic mean of PPV and sensitivity). Discrimination was assessed using ROC analysis, and calibration was evaluated using the Brier score and calibration curves. These metrics and visualizations were used to assess model suitability for predicting PCT positivity.

Feature Importance and Interpretability

Model-based feature importance scores were used to quantify the contribution of individual predictors to model output. SHAP values were calculated to support model interpretability and visualize the direction and magnitude of feature effects. SHAP values represent the marginal contribution of each feature to the prediction, ensuring both local accuracy and consistency. Collectively, these techniques were used to describe the contribution of individual features, including the composite interaction terms, to model predictions.

Temporal Validation Cohort

To evaluate model generalizability, we used an independent temporal validation cohort collected at the same tertiary care institution. The validation data set included 1265 encounters collected between June 1 and August 31, 2024, of which 379 (30.0%) had PCT levels of 0.50 ng/mL or higher. This data set was not used for model training, calibration, threshold selection, or hyperparameter tuning. Independence between data sets was ensured using patient-level identifiers, confirming that no patients from the development cohort appeared in the validation cohort.

Health Economic and Clinical Utility Analysis

To evaluate implementation impact, we performed decision curve analysis (DCA) and cost-weighted utility modeling from a managed care perspective. DCA was performed across probability thresholds from 0.05 to 0.95 using model-predicted probabilities and compared against universal testing (test all) and no testing (test none). Two threshold-based deployment strategies were evaluated: a high-sensitivity rule-out strategy prioritizing sensitivity and NPV to minimize missed PCT-positive cases, and a balanced strategy selecting thresholds based on balanced accuracy to balance sensitivity and specificity. The balanced strategy generally provided higher specificity than the rule-out strategy.

Cost-weighted utility analysis used investigator-defined assumptions of US$27 per PCT test, US$150 per day of unnecessary antibiotic exposure (assuming 1 day per false-positive prediction), and US$3000 per missed PCT-positive case as a hypothetical penalty for potential downstream complications. Published economic evaluations provided context for these assumptions: Kip et al reported €15 (approximately US$19.92) per PCT test and €126.66 (approximately US$168.22) per antibiotic day in 2013 euros, whereas Voermans et al reported US$95.81 per PCT test and US$174.02 per antibiotic day in 2019 US dollars.^19,20 The model inputs were hypothetical US-dollar assumptions rather than directly sourced costs, with no standardized cost year. A limited managed care perspective included modeled testing costs, unnecessary antibiotic costs, and missed-case penalties, excluding other health care expenditures and patient expenses. Total modeled costs and costs per patient were calculated for each strategy and compared with universal testing across decision thresholds.

RESULTS

In this study, we analyzed a data set of 7047 encounters. The general characteristics of the data set, including the median and IQR values for each variable, are summarized in Table 1. PCT positivity was higher in the development cohort (47.9% vs 30.0%), whereas clinical setting distributions were similar, with 6279 of 7047 encounters (89.1%) and 1132 of 1265 encounters (89.5%) occurring in inpatients, respectively. The temporal cohort reflected routine clinical practice and included both inpatient and outpatient encounters. Demographic characteristics were also comparable, with male patients accounting for 43.0% and 41.4% of the development and temporal validation cohorts, respectively. Pediatric patients represented 16.8% and 17.3% of the respective cohorts. Among pediatric patients, the median age (IQR) was 9 (3–13) years in the development cohort and 7 (2–14) years in the temporal validation cohort. Among adults, the corresponding median ages (IQR) were 63 (48–75) and 64 (48–74) years, respectively. Distributions of predictive features in the validation data set by PCT status are shown in eAppendix Figure 2.

The median (IQR) CRP level was 68.40 mg/L (27.80-137.10 mg/L), and the median (IQR) PCT level was 0.40 ng/mL (0.21-1.20 ng/mL).Exploratory correlation analysis revealed predominantly weak associations between PCT levels and individual biomarkers, including positive correlations with CRP and creatinine and weak inverse correlations between PCT levels and lymphocyte percentages (Figure 2).

Feature importance analyses identified SCr×CRP as the highest-ranked predictor in both RF and XGBoost models (Figure 3). Other important features included P-LCR × CRP, CRP, and SCr in RF, and neutrophil percentage (NE%) and platelet count (PLT) in XGBoost. The interaction between CRP and SCr emerged as a key contributor to model predictions of PCT positivity.

SHAP analysis further supported these findings, identifying SCr × CRP and ANC × CRP as the leading contributors to model predictions, followed by SCr, P-LCR × CRP, and CRP (Figure 4A). Composite features were among the most influential predictors, highlighting their substantial contribution to model output. The SHAP summary plot further demonstrated that features with larger absolute SHAP values, particularly SCr × CRP and ANC × CRP, had the strongest impact on predictions. Pairwise comparisons of SHAP values illustrated the relationships between ANC × CRP and SCr × CRP (Figure 4B), and between SCr × CRP and P-LCR × CRP (Figure 4C), with points colored according to CRP concentrations. These results highlight the importance of combining inflammatory and renal biomarkers through composite features to capture complex biological relationships and improve PCT prediction.

Performance of ML Algorithms

Model performance was evaluated using discrimination, accuracy, and probability calibration across internal and temporal data sets (eAppendix Table 1; eAppendix Table 2). On the internal test set, ensemble methods demonstrated superior discrimination and calibration; RF achieved the highest AUROC (0.93) and the lowest Brier score (0.112), followed by XGBoost (AUROC, 0.91; Brier score, 0.120) (eAppendix Figure 3; eAppendix Figure 4A). In contrast, SVM and LR showed lower internal performance (AUROC, 0.85 and 0.78, respectively) and poorer calibration (Brier scores, 0.152 and 0.188).

However, temporal validation revealed a substantial shift in model robustness. LR (AUROC, 0.90) and RF (AUROC, 0.89) emerged as the top performers, both achieving 84% accuracy in the independent data set. This difference in LR performance may reflect variations in case mix and predictor distributions between the cohorts rather than changes in model training or preprocessing. Although nonlinear ensemble models effectively captured complex patterns during development, the calibrated LR model also demonstrated competitive performance in temporal validation (Brier score, 0.125). These findings suggest that although ensemble methods maximize internal metrics, properly calibrated linear models may offer improved transportability for clinical deployment without compromising predictive power.

Threshold Selection and Trade-Off Analysis

Threshold selection was performed using the calibration data set to prevent optimization bias. Candidate probability thresholds from 0.05 to 0.95 were evaluated, and sensitivity, specificity, PPV, NPV, accuracy, F1 score, Youden index, and balanced accuracy were calculated.

Because the primary intended application is the safe reduction of unnecessary PCT, thresholds were first evaluated using a rule-out framework prioritizing sensitivity and NPV. For RF, a rule-out threshold of 0.25 achieved a sensitivity of 96.9%, a specificity of 49.6%, a PPV of 63.6%, and an NPV of 94.6%, supporting deferral of PCT testing in predicted-negative cases. At this threshold, RF achieved a net benefit of 0.374, exceeding that of universal testing. Similarly, XGBoost at a threshold of 0.25 achieved a sensitivity of 94.2% and an NPV of 91.8% (specificity, 59.1%), with a net benefit of 0.378—representing a 0.075 improvement over universal testing.

These findings confirm that rule-out deployment provides measurable clinical utility while preserving diagnostic safety.

For routine clinical deployment, thresholds maximizing balanced accuracy were also evaluated on the internal calibration data set. The balanced operating threshold for RF was 0.50 (sensitivity, 84.1%; specificity, 84.1%; NPV, 85.3%), yielding a net benefit of 0.318 and a 0.365 improvement over universal testing. For XGBoost, the balanced operating threshold was 0.55 (sensitivity, 73.1%; specificity, 92.7%; NPV, 79.1%), with a net benefit of 0.301 and a 0.465 improvement compared with universal testing (eAppendix Figure 4B).

Operating thresholds were selected using the internal calibration dataset. The selected thresholds were subsequently applied to the internal test set to calculate threshold-specific performance metrics.

Economic Evaluation of Threshold-Based Testing Strategies

In the internal test set (n = 1410), the RF model at a probability threshold of 0.25 reduced the number of PCT tests from 1410 under universal testing to 1023 (27.4% reduction). The estimated total cost decreased from US$148,770 to US$146,421, corresponding to a reduction in cost per patient from US$106 to US$104 (eAppendix Figure 4D). At this threshold, the model achieved a sensitivity of 96.9% and a specificity of 49.6%, with 21 missed PCT-positive cases. In comparison, XGBoost at the same threshold reduced the number of tests to 935 but resulted in 39 missed PCT-positive cases and a higher estimated cost per patient of US$133.01.

These findings demonstrate that threshold selection can be tailored to institutional priorities, with stewardship-focused settings favoring high-sensitivity rule-out thresholds and cost-constrained settings favoring higher-specificity thresholds.

Temporal validation demonstrated consistent performance across balanced and rule-out strategies (eAppendix Table 2). LR achieved the highest overall discrimination (AUROC, 0.90; 95% CI, 0.88-0.92), followed by RF (AUROC, 0.89; 95% CI, 0.87-0.91) and XGBoost (AUROC, 0.88; 95% CI, 0.86-0.90). Calibration was acceptable across models, with lower Brier scores for XGBoost (0.122) and LR (0.125) than for RF (0.129) and SVM (0.200).

Under balanced operating thresholds, LR achieved an accuracy of 84.1% (sensitivity, 77.6%; specificity, 86.9%) and RF demonstrated comparable overall accuracy (84.3%) with slightly lower sensitivity (73.1%) and higher specificity (89.2%). XGBoost showed modestly lower balanced performance (accuracy, 81.2%; sensitivity, 76.3%; specificity, 83.3%). Under rule-out thresholds, sensitivity increased across models, reaching 0.96 for RF and 0.92 for XGBoost, with corresponding NPVs of 0.96 and 0.94. As expected, rule-out thresholds reduced specificity and overall accuracy.

Overall, ensemble and calibrated linear models demonstrated robust temporal performance while threshold selection enabled tailoring between diagnostic safety (rule-out strategy) and overall performance optimization (balanced strategy).

DISCUSSION

In this study, ML algorithms were developed and validated using routinely available laboratory parameters (CBC, CRP, and SCr) to predict PCT positivity. By integrating standard biomarkers with biologically motivated composite features, RF and XGBoost achieved strong internal diagnostic performance and high discrimination (AUROC, 0.925 and 0.907, respectively), whereas LR achieved the highest discrimination in temporal validation (AUROC, 0.902). Calibration analysis demonstrated reliable probability estimation (Brier score, 0.112-0.120), supporting probability-based CDS deployment.

Our findings are consistent with those of Agnello et al, who evaluated ML models for predicting PCT of 0.50 ng/mL or higher.¹⁶ In that study, RF and XGBoost achieved specificities of 88% and 91%, respectively,¹⁶ comparable with those observed in our models. A key distinction of our study is the inclusion of SCr, which influences PCT concentrations through renal clearance.²¹ In addition, composite features (SCr × CRP, ANC× CRP, and P-LCR × CRP) integrated inflammatory, renal, and platelet-related signals, capturing clinically relevant biological interactions. These features demonstrated strong predictive importance and were confirmed by SHAP analysis. SHAP summary and scatter plots (eAppendix Figure 4C) highlight the dominant role of CRP in modulating the predictive effects of ANC, SCr, and P-LCR, improving model interpretability and supporting PCT-based stewardship applications. Although we independently identified SCr × CRP as a critical predictor of PCT positivity, Chen et al reported its moderate predictive value for postoperative cardiovascular events.²² This parallel underscores the versatility of such composite features in capturing clinically relevant interactions across diverse contexts.

The models rely exclusively on routine laboratory tests, supporting practical integration as CDS tools. Two complementary deployment strategies are feasible: balanced-performance deployment and stewardship-focused rule-out deployment. Balanced strategies maintain overall discrimination while optimizing testing efficiency, whereas rule-out strategies prioritize high sensitivity and NPV to safely reduce unnecessary PCT testing. Importantly, threshold selection can be tailored to institutional priorities and resource constraints.

It is important to emphasize that the model predicts the probability of PCT positivity, not infection or sepsis directly. By enabling probability threshold–based deployment tailored to institutional priorities, ML-guided CDS tools may support more consistent and cost-efficient PCT utilization across health care settings. Several guidelines do not recommend PCT as a stand-alone tool for initiating antibiotics, and elevated PCT may reflect renal dysfunction or inflammatory states unrelated to bacterial infection.^21,23,24 Therefore, this CDS approach is not intended to guide antibiotic initiation decisions or replace clinical judgment. Instead, it aims to support diagnostic stewardship by identifying patients unlikely to benefit from PCT testing. Final clinical decisions should incorporate patient presentation, comorbidities, imaging, microbiology, and physician assessment.

Variability in PCT usage—stemming from the lack of standardized guidelines and inconsistent thresholds for diagnostic applications—raises concerns about unnecessary or repetitive testing.²⁵ Although the 2012 Surviving Sepsis Campaign Guidelines introduced PCT as a diagnostic marker for sepsis, its widespread adoption has been limited, partly due to cost-related barriers.^24,26,27 Nonetheless, a study conducted in the Netherlands found that using PCT-guided algorithms resulted in a cost reduction of €3503 (US$4652) per patient, for an estimated annual savings of €46 million (US$61.1 million) for the country’s health care system.¹⁹ Such disparities in PCT utilization not only impede its broader clinical implementation but also contribute to substantial financial costs, as evidenced by Salinas et al, who reported an annual expenditure of US$1,346,406 for 142,644 PCT tests across institutions with significant variation in test demand.⁹ Against this backdrop, our ML models present a viable solution. By targeting PCT testing more effectively, these models have the potential to reduce redundant diagnostics and alleviate the financial burden associated with inconsistent PCT utilization.

An important characteristic of PCT testing is that negative results often provide greater clinical utility than positive results, particularly for supporting antibiotic de-escalation or avoiding unnecessary antimicrobial therapy.² In this context, the rule-out deployment strategy evaluated in our study aligns with real-world clinical use of PCT by prioritizing high sensitivity and NPV. Rather than focusing only on patients likely to have a positive PCT result, the proposed CDS framework aims to identify patients unlikely to benefit from PCT testing, reducing unnecessary testing. However, this involves a potential trade-off, as withholding testing may eliminate negative PCT results that could inform clinical decisions, particularly antibiotic de-escalation or discontinuation. Our economic analysis highlighted the trade-off between testing volume and estimated downstream costs. In the internal test set, RF at a threshold of 0.25 reduced PCT testing by approximately 27% while maintaining 96.9% sensitivity and slightly lowering estimated per-patient costs (eAppendix Figure 4D).

These findings suggest that rule-out deployment may represent a high-value stewardship strategy for health care systems prioritizing diagnostic safety and resource optimization. However, the magnitude and direction of the net cost change depend on local prevalence and on assumptions regarding test costs and the downstream costs of false-negative and false-positive decisions. When scaled to large health care systems, these utilization differences could translate into meaningful system-level cost impacts.

Limitations

This study has several limitations. First, it was conducted at a single tertiary center, which may limit generalizability. Although temporal validation was performed, multicenter prospective validation is needed to confirm robustness across diverse health care systems.

Second, the data set included all eligible inpatient and outpatient encounters without stratification by clinical indication for PCT ordering. The specific diagnostic context (eg, suspected sepsis, pneumonia, postoperative monitoring) was not available, and infection end points were not adjudicated. Therefore, the models predict laboratory-defined PCT positivity rather than confirmed infection or the need for antibiotics. This design reflects real-world laboratory utilization but limits direct inference regarding infection outcomes.

Third, this model must not be interpreted as restricting infection treatment or delaying the initiation of appropriate antibiotics. Instead, it supports institutional stewardship strategies aimed at reducing unnecessary PCT orders in low-risk scenarios. However, the clinical utility of a negative PCT result may differ between initial diagnostic evaluation and monitoring during antibiotic therapy. Deferring PCT testing in patients predicted to have negative results may eliminate clinically useful information, particularly when serial PCT measurements are used to support antibiotic de-escalation or discontinuation. Future studies incorporating clinical diagnoses, antibiotic exposure, and patient outcomes would strengthen evaluation within antimicrobial stewardship frameworks. Prospective validation integrating laboratory and clinical variables will be essential to further refine threshold strategies and assess impact on patient-centered outcomes.

CONCLUSIONS

Although most ML studies focus on directly predicting sepsis, this study adopts a laboratory stewardship perspective by predicting PCT positivity using inexpensive, routine biomarkers.^14,15,28 The LR, RF, and XGBoost models predicted PCT positivity with high discrimination in validation (AUROC, 0.88-0.90), with LR performing best overall and RF best suited to a rule-out strategy. Integrating these tools into clinical workflows could enhance decision-making, improve resource utilization, and reduce PCT testing, indirectly supporting antimicrobial stewardship without replacing clinician judgment.²⁹ Prospective, multicenter validation is needed before clinical implementation.

Author Affiliations: Department of Medical Biochemistry, Izmir City Hospital (EI, DIT, TS, IK), Izmir, Turkey; Department of Infectious Diseases and Clinical Microbiology, Izmir City Hospital (SC), Izmir, Turkey.

Source of Funding: This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Authorship Information: Concept and design (EI); acquisition of data (DIT, TS, IK); analysis and interpretation of data (DIT, TS, SC); drafting of the manuscript (DIT); critical revision of the manuscript for important intellectual content (EI, DIT, IK, SC); statistical analysis (EI, TS); administrative, technical, or logistic support (IK, SC); and supervision (EI, SC).

Address Correspondence to: Emre Ispir, MD, Izmir City Hospital, Central Laboratory, Sevket Ince Neighborhood, 2148/11 St, No. 1/11, 35540 Bayrakli, Izmir, Türkiye. Email: emreispir@gmail.com.

REFERENCES

1. Simon L, Gauvin F, Amre DK, Saint-Louis P, Lacroix J. Serum procalcitonin and C-reactive protein levels as markers of bacterial infection: a systematic review and meta-analysis. Clin Infect Dis. 2004;39(2):206-217. doi:10.1086/421997

2. Schuetz P, Albrich W, Mueller B. Procalcitonin for diagnosis of infection and guide to antibiotic decisions: past, present and future. BMC Med. 2011;9:107. doi:10.1186/1741-7015-9-107

3. Hoeboer SH, van der Geest PJ, Nieboer D, Groeneveld ABJ. The diagnostic accuracy of procalcitonin for bacteraemia: a systematic review and meta-analysis. Clin Microbiol Infect. 2015;21(5):474-481. doi:10.1016/j.cmi.2014.12.026

4. van Nieuwkoop C, Bonten TN, van’t Wout JW, et al. Procalcitonin reflects bacteremia and bacterial load in urosepsis syndrome: a prospective observational study. Crit Care. 2010;14(6):R206. doi:10.1186/cc9328

5. Laukemann S, Kasper N, Kulkarni P, et al. Can we reduce negative blood cultures with clinical scores and blood markers? results from an observational cohort study. Medicine (Baltimore). 2015;94(49):e2264. doi:10.1097/MD.0000000000002264

6. Wyss G, Berger S, Haubitz S, et al. The Shapiro-Procalcitonin algorithm (SPA) as a decision tool for blood culture sampling: validation in a prospective cohort study. Infection. 2020;48(4):523-533. doi:10.1007/s15010-020-01423-6

7. Christ-Crain M, Müller B. Biomarkers in respiratory tract infections: diagnostic guides to antibiotic prescription, prognostic markers and mediators. Eur Respir J. 2007;30(3):556-573. doi:10.1183/09031936.00166106

8. Jeon K, Suh JK, Jang EJ, et al. Procalcitonin-guided treatment on duration of antibiotic therapy and cost in septic patients (PRODA): a multi-center randomized controlled trial. J Korean Med Sci. 2019;34(14):e110. doi:10.3346/jkms.2019.34.e110

9. Salinas M, López-Garrigós M, Flores E, Uris J, Leiva-Salinas C; Pilot Group of the Appropriate Utilization of Laboratory Tests (REDCONLAB) working group. Procalcitonin in the emergency department: a potential expensive over-request that can be modulated through institutional protocols. Am J Emerg Med. 2018;36(1):158-160. doi:10.1016/j.ajem.2017.07.033

10. Vasey B, Ursprung S, Beddoe B, et al. Association of clinician diagnostic performance with machine learning-based decision support systems: a systematic review. JAMA Netw Open. 2021;4(3):e211276. doi:10.1001/jamanetworkopen.2021.1276

11. Xu Q, Xie W, Liao B, et al. Interpretability of clinical decision support systems based on artificial intelligence from technological and medical perspective: a systematic review. J Healthc Eng. 2023;2023:9919269. doi:10.1155/2023/9919269

12. Çubukçu HC, Topcu DIŁ, Yenice S. Machine learning-based clinical decision support using laboratory data. Clin Chem Lab Med. 2024;62(5):793-823. doi:10.1515/cclm-2023-1037

13. Valente F, Paredes S, Henriques J, Rocha T, de Carvalho P, Morais J. Interpretability, personalization and reliability of a machine learning based clinical decision support system. Data Min Knowl Discov. 2022;36(3):1140-1173. doi:10.1007/s10618-022-00821-8

14. Steinbach D, Ahrens PC, Schmidt M, et al. Applying machine learning to blood count data predicts sepsis with ICU admission. Clin Chem. 2024;70(3):506-515. doi:10.1093/clinchem/hvae001

15. Choi JS, Trinh TX, Ha J, et al. Implementation of complementary model using optimal Combination of hematological parameters for sepsis screening in patients with fever. Sci Rep. 2020;10(1):273. doi:10.1038/s41598-019-57107-1

16. Agnello L, Vidali M, Ciaccio AM, et al. A machine learning strategy to mitigate the inappropriateness of procalcitonin request in clinical practice. Heliyon. 2024;10(5):e26556. doi:10.1016/j.heliyon.2024.e26556

17. Master SR, Badrick TC, Bietenbeck A, Haymond S. Machine learning in laboratory medicine: recommendations of the IFCC Working Group. Clin Chem. 2023;69(7):690-698. doi:10.1093/clinchem/hvad055

18. de Jong E, van Oers JA, Beishuizen A, et al. Efficacy and safety of procalcitonin guidance in reducing the duration of antibiotic treatment in critically ill patients: a randomised, controlled, open-label trial. Lancet Infect Dis. 2016;16(7):819-827. doi:10.1016/S1473-3099(16)00053-0

19. Kip MMA, Kusters R, IJzerman MJ, Steuten LMG. A PCT algorithm for discontinuation of antibiotic therapy is a cost-effective way to reduce antibiotic exposure in adult intensive care patients with sepsis. J Med Econ. 2015;18(11):944-953. doi:10.3111/13696998.2015.1064934

20. Voermans AM, Mewes JC, Broyles MR, Steuten LMG. Cost-effectiveness analysis of a procalcitonin-guided decision algorithm for antibiotic stewardship using real-world U.S. hospital data. OMICS. 2019;23(10):508-515. doi:10.1089/omi.2019.0113

21. Chun K, Chung W, Kim AJ, et al. Association between acute kidney injury and serum procalcitonin levels and their diagnostic usefulness in critically ill patients. Sci Rep. 2019;9(1):4777. doi:10.1038/s41598-019-41291-1

22. Chen Y, Teng Y, Peng X, et al. Combination of creatinine with inflammatory biomarkers (PCT, CRP, hsCRP) for predicting postoperative ICU admissions for elderly patients. Adv Ther. 2024;41(7):2776-2790. doi:10.1007/s12325-024-02874-1

23. Lee CC, Porta L, Liu Y, et al. GRADE-based procalcitonin guideline for emergency departments. Am J Emerg Med. 2025;89:109-123. doi:10.1016/j.ajem.2024.11.093

24. Evans L, Rhodes A, Alhazzani W, et al. Surviving Sepsis Campaign: international guidelines for management of sepsis and septic shock 2021. Crit Care Med. 2021;49(11):e1063-e1143. doi:10.1097/CCM.0000000000005337

25. Cotter JM, Hardee I, Moss A, Dempsey A, Ambroggio L. Procalcitonin use: variation across hospitals and trends over time. Hosp Pediatr. 2021;e2021005992. doi:10.1542/hpeds.2021-005992

26. Balk RA, Kadri SS, Cao Z, Robinson SB, Lipkin C, Bozzette SA. Effect of procalcitonin testing on
health-care utilization and costs in critically ill patients in the United States. Chest. 2017;151(1):23-33. doi:10.1016/j.chest.2016.06.046

27. Dellinger RP, Levy MM, Rhodes A, et al. Surviving sepsis campaign: international guidelines for management of severe sepsis and septic shock: 2012. Crit Care Med. 2013;41(2):580-637. doi:10.1097/CCM.0b013e31827e83af

28. Adams R, Henry KE, Sridharan A, et al. Prospective, multi-site study of patient outcomes after implementation of the TREWS machine learning-based early warning system for sepsis. Nat Med. 2022;28(7):1455-1460. doi:10.1038/s41591-022-01894-0

29. Dellit TH, Owens RC, McGowan JE Jr, et al. Infectious Diseases Society of America and the Society for Healthcare Epidemiology of America guidelines for developing an institutional program to enhance antimicrobial stewardship. Clin Infect Dis. 2007;44(2):159-177. doi:10.1086/510393