Answers are grounded in this paper and cite their sources.
This study proposes a fever prediction model that utilizes historical body temperature data to determine whether a child will continue to have or develop a fever on the following day. This model may assist clinicians in identifying at-risk patients and enhancing pediatric clinical decision-making. Body temperature records were collected from hospitalized pediatric patients aged 0 to 18 years between 2010 and 2023. Temperature measurements were organized into consecutive 24-hour segments, each defined from 8:00 AM to 7:59 AM the following day. Each data point consisted of a pair of consecutive segments. Statistical and temporal features were extracted from the first segment, while the second segment was used to assign a categorical outcome label of either “Fever” (maximum temperature ≥ 38.0 °C) or “Normal”. Machine learning algorithms were then used to develop predictive models based on the extracted features. Five machine learning algorithms, including Logistic Regression, Support Vector Machine, Random Forest, XGBoost, and a fully connected neural network, were evaluated. The best-performing model was selected based on sensitivity. A total of 214,584 data points were generated. Two datasets were constructed: an original dataset and an interpolated dataset. Models trained on the original dataset demonstrated performance comparable to those trained on the interpolated dataset. When sensitivity was designated as the primary performance metric, the SVM-based model trained on the original dataset achieved an overall sensitivity of 0.7166 (95% CI: 0.7014, 0.7327). In subgroup analyses, the sensitivity was 0.7368 (95% CI: 0.7206, 0.7535) among febrile patients and 0.6241 (95% CI: 0.5879, 0.6634) among afebrile patients. Because of the severe class imbalance among afebrile patients, the positive predictive value (PPV) was evaluated in both resampled and pre-resampled test sets. In the resampled test set, in which febrile cases accounted for 15.1% of the sample, the PPV was 0.2543 (95% CI: 0.2326, 0.2778). In the pre-resampled test set, in which febrile cases accounted for only 1.6% of the sample, the PPV was substantially lower at 0.0296 (95% CI: 0.0265, 0.0327), indicating a high proportion of false-positive alerts under the observed class distribution. The proposed model demonstrated promising sensitivity among hospitalized febrile children, indicating potential utility in this patient group. However, its predictive performance was poorer among afebrile patients, with more frequent false-positive alerts. Therefore, the clinical utility of the model in afebrile patients should be interpreted with caution.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: