A machine learning model trained on longitudinal body temperature measurements can predict whether a hospitalized child will develop or continue to have a fever on the following day, according to a study published in BMC Medical Informatics and Decision Making. Evaluated across body temperature records from pediatric patients aged 0 to 18 years, the primary support vector machine model achieved an overall sensitivity of 0.7166, with a 95 percent confidence interval ranging from 0.7014 to 0.7327. The research team designed the predictive approach to assist clinicians in identifying at-risk pediatric inpatients and to support clinical decision-making on hospital wards.
The investigators formulated the prediction task using historical inpatient temperature records gathered between 2010 and 2023. Each segment spanned from 8:00 AM on a given morning to 7:59 AM on the subsequent day. Predictive data points were then created by linking pairs of these consecutive 24-hour segments to determine whether patterns in the initial window could forecast thermal trajectories in the subsequent window.
From the first 24-hour segment of each pair, the researchers extracted both statistical and temporal features reflecting the longitudinal temperature variations of the hospitalized children. The second 24-hour segment was utilized solely to define the categorical clinical outcome. If the maximum recorded body temperature during that second segment reached or exceeded 38.0 degrees Celsius, the window was assigned an outcome label of fever. If temperatures remained strictly below that threshold, the outcome was categorized as normal. In total, the authors assembled 214,584 paired data points across the patient population.
To systematically evaluate automated prediction methods, the researchers trained and compared five distinct machine learning algorithms: Logistic Regression, Support Vector Machine, Random Forest, XGBoost, and a fully connected neural network. The investigative team designated sensitivity as the primary performance metric for model selection, prioritizing the identification of pediatric patients who would experience fever. In parallel, two separate data representations were constructed and analyzed: an original dataset and an interpolated dataset. Models developed using the original dataset demonstrated predictive performance comparable to those developed using the interpolated dataset.
When evaluated across patient subgroups, the best-performing model—a Support Vector Machine trained on the original dataset—exhibited distinct performance characteristics depending on the child’s baseline febrile status. Among pediatric patients who were already febrile during the initial segment, the model demonstrated a sensitivity of 0.7368, with a 95 percent confidence interval between 0.7206 and 0.7535. In contrast, among hospitalized children who were afebrile in the initial window, the model demonstrated a lower sensitivity of 0.6241, with a 95 percent confidence interval spanning from 0.5879 to 0.6634.
The researchers also documented significant challenges associated with class imbalance, particularly within the afebrile patient cohort where subsequent fever events were infrequent. To assess this limitation, the authors analyzed the positive predictive value across both resampled and pre-resampled test environments. In the resampled test set, where febrile cases accounted for 15.1 percent of the evaluation cohort, the model achieved a positive predictive value of 0.2543, with a 95 percent confidence interval of 0.2326 to 0.2778.
Under realistic clinical conditions in the pre-resampled test set, however, febrile cases accounted for only 1.6 percent of the sample. In this natural distribution, the positive predictive value dropped markedly to 0.0296, with a 95 percent confidence interval from 0.0265 to 0.0327. This substantial decrease reflected a high proportion of false-positive alerts when screening hospitalized children who did not currently have a fever. Consequently, the study authors concluded that while the model exhibits promising sensitivity among febrile children, its clinical utility among afebrile patients remains constrained by false alarms and should be interpreted with caution.
Subject of Research: Machine learning prediction of pediatric fever using longitudinal body temperature data
Article Title: Predicting future fever in hospitalized children using longitudinal body temperature data and machine learning
Article References: Chen, J.-J., Wu, J. C.-H., Yang, J.-M., & Lu, H. H.-S. (2026). Predicting future fever in hospitalized children using longitudinal body temperature data and machine learning. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03963-7
Image Credits: AI Generated
DOI: 10.1186/s12911-026-03963-7
Keywords: pediatric fever prediction, machine learning in medicine, body temperature records, support vector machine, hospitalized children, longitudinal temperature data, pediatric decision-making, logistic regression, random forest, xgboost, neural network models, fever sensitivity, positive predictive value, class imbalance, pediatric inpatient care, false-positive alerts
News Source: Blake Davidson. (November 19, 2026). Support Vector Machine Predicts Next-Day Fever in Hospitalized Children. Scienmag.



