Jean-Baptiste Woillard1,2,3, Charlotte Salmon Gandonnière4, Alexandre Destere5,6,7, Stephan Ehrmann4,8, Hamid Merdji9,10, Armelle Mathonnet11, Pierre Marquet5,6,7, Chantal Barin-Le Guellec6,12,13. 1. Faculté de Médecine de Limoges, University of Limoges, IPPRITT, 2 rue du docteur Marcland, 87025, Limoges cedex, France. jean-baptiste.woillard@unilim.fr. 2. INSERM, IPPRITT, U1248, 87000, Limoges, France. jean-baptiste.woillard@unilim.fr. 3. Department of Pharmacology and Toxicology, CHU Limoges, 87000, Limoges, France. jean-baptiste.woillard@unilim.fr. 4. Médecine Intensive Réanimation, INSERM CIC 1415, CRICS-TriggerSep Research Network, CHRU de Tours, 37044, Tours, France. 5. Faculté de Médecine de Limoges, University of Limoges, IPPRITT, 2 rue du docteur Marcland, 87025, Limoges cedex, France. 6. INSERM, IPPRITT, U1248, 87000, Limoges, France. 7. Department of Pharmacology and Toxicology, CHU Limoges, 87000, Limoges, France. 8. Centre D'étude Des Pathologies Respiratoires INSERM U1100, Faculté de médecine, Université de Tours, Tours, France. 9. Faculté de Médecine, Hôpitaux universitaires de Strasbourg, Nouvel Hôpital Civil, Service de réanimation, Université de Strasbourg (UNISTRA), Strasbourg, France. 10. UMR 1260, Regenerative Nano Medecine, INSERM, Fédération de Médecine Translationnelle de Strasbourg (FMTS), Université de Strasbourg, Strasbourg, France. 11. Médecin Intensive Réanimation, Centre Hospitalier Régional D'Orléans, Orléans, France. 12. Laboratoire de Biochimie Et de Biologie Moléculaire, CHU de Tours, 37044, Tours, France. 13. Université de Tours, 37044, Tours, France.
Abstract
OBJECTIVE: This work aims to evaluate whether a machine learning approach is appropriate to estimate the glomerular filtration rate in intensive care unit patients based on sparse iohexol pharmacokinetic data and a limited number of predictors. METHODS: Eighty-six unstable patients received 3250 mg of iohexol intravenously and had nine blood samples collected 5, 30, 60, 180, 360, 540, 720, 1080, and 1440 min thereafter. Data splitting was performed to obtain a training (75%) and a test set (25%). To estimate the glomerular filtration rate, 37 candidate potential predictors were considered and the best machine learning approach among multivariate-adaptive regression spline and extreme gradient boosting (Xgboost) was selected based on the root-mean-square error. The approach associated with the best results in a ten-fold cross-validation experiment was then used to select the best limited combination of predictors in the training set, which was finally evaluated in the test set. RESULTS: The Xgboost approach yielded the best performance in the training set. The best combination of covariates comprised iohexol concentrations at times 180 and 720 min; the relative deviation from these theoretical times; the difference between these two concentrations; the Simplified Acute Physiology Score II; serum creatinine; and the fluid balance. It resulted in a root-mean-square error of 6.2 mL/min and an r2 of 0.866 in the test set. Interestingly, the eight patients in the test set with a glomerular filtration rate < 30 mL/min were all predicted accordingly. CONCLUSIONS: Xgboost provided accurate glomerular filtration rate estimation in intensive care unit patients based on two timed blood concentrations after iohexol intravenous administration and three additional predictors.
OBJECTIVE: This work aims to evaluate whether a machine learning approach is appropriate to estimate the glomerular filtration rate in intensive care unit patients based on sparse iohexol pharmacokinetic data and a limited number of predictors. METHODS: Eighty-six unstable patients received 3250 mg of iohexol intravenously and had nine blood samples collected 5, 30, 60, 180, 360, 540, 720, 1080, and 1440 min thereafter. Data splitting was performed to obtain a training (75%) and a test set (25%). To estimate the glomerular filtration rate, 37 candidate potential predictors were considered and the best machine learning approach among multivariate-adaptive regression spline and extreme gradient boosting (Xgboost) was selected based on the root-mean-square error. The approach associated with the best results in a ten-fold cross-validation experiment was then used to select the best limited combination of predictors in the training set, which was finally evaluated in the test set. RESULTS: The Xgboost approach yielded the best performance in the training set. The best combination of covariates comprised iohexol concentrations at times 180 and 720 min; the relative deviation from these theoretical times; the difference between these two concentrations; the Simplified Acute Physiology Score II; serum creatinine; and the fluid balance. It resulted in a root-mean-square error of 6.2 mL/min and an r2 of 0.866 in the test set. Interestingly, the eight patients in the test set with a glomerular filtration rate < 30 mL/min were all predicted accordingly. CONCLUSIONS: Xgboost provided accurate glomerular filtration rate estimation in intensive care unit patients based on two timed blood concentrations after iohexol intravenous administration and three additional predictors.