gms | German Medical Science

Background: Sepsis is the leading cause of hospital readmission and death in the developed world [1]. Cancer patients are at high risk due to their underlying disease and therapies. Meanwhile, antimicrobial resistance increases steadily, warranting a more careful administration of antibiotics [2]. In this work, the potential of machine learning algorithms to predict fever persistence 48 hours after initiation of antibiotics based on electronic health records is explored.

Methods: All cancer patients of the University Hospital Essen (UHE) between 1/1/2020-1/1/2024, receiving broad-spectrum i.v. antibiotics, and having persistent fever after 48 hours of first administration were extracted from UHE’s Fast Health Interoperability Resources clinical database. Vital signs, laboratory values and derived lag features as well as patient characteristics were included (Figure 1 [Fig. 1]). The binary target to predict was prevalence of body temperature ≥38 °C after 48-72 hours after initiation of antibiotics therapy. On average, 5.8 (+-1.9) body temperature measurements were recorded per patient daily during routine care (e.g., tympanic or axillary temperatures).

Data quality was assessed based on established practices in handling EHR, including checks for completeness, duplicates, plausibility and consistency [3]. Various gradient boosting algorithms, a deep learning neural network architecture, logistic regression with optimized l1 and l2 regularization and random forest were selected as algorithms based on their properties and reported applicability on tabular electronic health records for clinical prediction tasks [4], [5]. To leverage model diversity and improve generalizability and robustness of predictions, different voting ensemble strategies were implemented. Hyperparameter tuning and model validation were performed through nested 10-fold stratified cross validation (90/10% train/test split) over 50 trials.

Results: Of UHE’s patient records in hematology/oncology wards, 335 patients had persistent fever after 48 hours, of which 119 patients had solid cancers, 132 leukemia, and 87 other hematologic malignancies. 72 hours after antibiotic therapy 218 patients (64.5%) still had persistent fever. The voting ensemble model with weighted voting and optimized model weights had the highest predictive performance. The model showed an area under the precision-recall curve (AUPRC) of 0.87, a sensitivity of 0.88, a specificity of 0.64, a Negative Predictive Value (NPV) of 0.77 and a Positive Predictive Value (PPV) of 0.80 on a cross-validated test set (Table 1 [Tab. 1]). Across employed models, statistical features of lagged body temperatures had the most impact on predictions. Among laboratory values, Leucocytes were the most predictive variable.

Conclusions: The model performance based on internal validation indicates potential of machine learning algorithms to predict fever evolution during antibiotics therapy from EHR and assist in clinical decision-making. This could contribute to better patient care by preventing unnecessary antibiotics escalations and CT scans. However, because external validation of the model is still pending, its generalization to different clinical practices, documentation standards and patient cohorts is unclear, which prevents more conclusive assessments of a potential clinical impact. Respective results from external validation, as well as error analysis and covariate importance, will be presented at the conference.

The Python code for replicating the feature engineering and training pipeline will be made available under the following GitHub-repository after completion of the external validation of the model: https://github.com/gernotpuc/fever_model.git

Open access to the patient data used for training and validating the model is not possible due to data privacy reasons.

The authors declare that they have no competing interests.

The authors declare that a positive ethics committee vote has been obtained.