Early prediction of fever and hypothermia in adult ICU patients using machine learning: A multicenter study

Résumé fourni par la source

Background Core body temperature (CBT) plays a pivotal role in determining the prognosis of patients with neurological impairments. This study aimed to develop and evaluate a machine learning (ML) algorithm capable of forecasting CBT, as well as predicting fever and hypothermia, several hours in advance. Methods We conducted a multicenter retrospective observational study in three mixed intensive care units (ICUs): two university hospitals in France (Hôpital de la Timone and Nord), and one teaching military hospital in France (Sainte-Anne). We evaluated several prediction methods, including Neural Networks, Gradient Boosting, Random Forests, linear regression models, LSTM, and XGBoost. Inputs consisted of past temperature measurements, blood pressure, heart rate, and time of day. Results Data from 10,189 ICU patients were analyzed. A training cohort (n = 5,146) from two ICUs was used to develop the models, and an independent evaluation cohort (n = 5,043) from the third ICU was used for testing. XGBoost consistently demonstrated the highest predictive performance for both fever and hypothermia. Sensitivity for fever (and hypothermia) prediction was 93.2 % (90.2 %) at 1 h, 86.6 % (82.3 %) at 2 h, and 76.8 % (70.3 %) at 4 h. Prediction of CBT values yielded Root Mean Square Errors of 0.19 °C, 0.31 °C, and 0.46 °C at 1, 2, and 4 h, respectively. Conclusion This is the first large multicenter study to evaluate the contribution of ML to CBT prediction in ICU patients. Our findings show that fever and hypothermia can be reliably detected up to four hours before their occurrence, paving the way for more proactive and personalized patient management.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.