Predicting Nosocomial Infections in Hematologic Patients: A Machine Learning Model Based on Dynamic Body Temperature...

To identify body temperature dynamic patterns and develop a machine learning model for the early detection of nosocomial infections.

A retrospective and observational study of patients hospitalised in the haematology department of the Chinese People's Liberation Army General Hospital between January 2014 and December 2023.

A latent class trajectory model was used to discover patterns in patients' body temperatures over time. Machine learning models were then built to predict nosocomial infections and evaluated using standard metrics (AUROC, sensitivity, specificity). SHAP (SHapley Additive exPlanations) values were used to interpret the final model.

Among 6989 patients, we identified four distinct body temperature trajectories. Bloodstream infections were most common in patients exhibiting either a slow rise followed by a gradual decrease or a rapid rise followed by a quick decrease in body temperature. The XGBoost model showed excellent predictive performance (AUROC = 0.801), with balanced sensitivity (0.718) and specificity (0.701). The top five predictors of nosocomial infections were elevated procalcitonin, neutropenia, prolonged central venous catheter use and two specific temperature trajectories: 'stable and relatively high' and 'a rapid rise followed by a quick decrease'.

The XGBoost model effectively predicted nosocomial infections. Dynamic body temperature trajectories provided early, objective warning signs of infection. This predictive tool empowered nursing staff to proactively monitor nosocomial infection, allowing for timely, data-driven interventions in vulnerable hematologic patients.

The developed machine learning predictive tool can help clinical medical staff identify nosocomial infections as early as possible, facilitate personalised rehabilitation and health management plans, aligning with the philosophy of patient-centred precision nursing. Further, the four body temperature trajectory patterns identified provide nurses with objective, dynamic indicators for recognising potential infection subphenotypes, supporting a shift from experience-driven reactive care towards data-driven proactive nursing.

Previous studies suggested that body temperature could indicate the severity and prognosis of infections, but the pattern was unknown. In this study, we found that body temperature trajectories could signal infection subphenotypes, such as bloodstream infections being most common in patients with a slow rise followed by a gradual decrease in body temperature or with a rapid rise followed by a quick decrease. By integrating body temperature trajectories with key clinical biomarkers, the developed prediction model enables early and accurate identification of nosocomial infections in hematologic patients. The application of this tool may significantly shorten the time window between infection onset and intervention, potentially reducing infection-related complications, mortality and healthcare costs, thereby improving overall care quality and patient outcomes.

The study adhered to the relevant EQUATOR reporting guidelines, the TRIPOD Checklist for Prediction Model Development and Validation.

The research team included nursing staff and clinicians responsible for infection surveillance and control in the hospital, who contributed real-world insights into the definition of predictors, interpretation of temperature trajectories, clinical implications of the prediction model and preparation of the manuscript. Their expertise helped ensure that the study addressed relevant clinical questions and that the findings are interpretable and actionable in practice.