Establishing An Optimal Nutritional Support Model For Septic Patients Based On Machine Learning Techniques

Objective:Based on machine learning technology,we aim to establish a mortality dynamic prediction model relying on energy and protein intake in sepsis patients,in order to improve the prognosis of patients and reduce the fatality rate.Methods:We conduct a retrospective cohort study designed with patients from the Emergency Intensive Care Unit(EICU)of Sichuan Provincial People’s Hospital.Inclusion criteria:age ≥18 years(adult patients);APACHEII score>10 on the day of admission;receiving nutritional support therapy(including enteral and parenteral nutrition)for at least 5 days during ICU stay,and energy intake ≥10kcal/(kg·d).Exclusion criteria:age<18 years(adult patients);not receiving nutritional support or receiving for less than 5 days;patients treated with ECMO Extracorporeal Membrane Oxygenation(ECMO)or renal replacement therapy;women during pregnancy or lactation;patients enrolled in other clinical trials.For each enrolled patient,47 indicators were collected for 14 consecutive days after admission to the ICU,including basic conditions,physiological indicators,biochemical indicators,and nutritional indicators.After that,data cleaning and data pre-processing were performed first to make sure that the data suitable to build the model.According to four different periods of metabolism in critically ill patients,namely "metabolic instability period","catabolic period","anabolic period",and "recovery period",and formed four datasetswere available for analysis.Logistic regression,random forest(RF),and support vector machine(SVM)were used to develop prediction models for each period respectively.The study was ethically approved and obtained a registration number in the Chinese Clinical Trials Registry(clinical registration number:ChiCTR2200056316)Results:A total of 179 retrospective data samples hospitalized in the EICU of Sichuan Provincial People’s Hospital from September 2018 to January 2020 were included in this study,of which 78 patients died within 28 days after admission to the ICU.The analysis showed that the daily amino acid intake of patients in the survivor group were significantly higher than death group on days of 9,10,and 14(P=0.03,P=0.001,P=0.012)and the same as energy intake on days of 4,8,and 10(P=0.047.P=0.046,P=0.022)of admission to the ICU.In addition,the cumulative energy debit(energy debit)in survivor group was significantly lower in the first three periods were also significantly lower than death group(P<0.05).PLS-DA analysis indicated that some of enrolled characteristics show different importance in predicting the prognosis of sepsis patients,such as white blood cell count and neutrophil count were more important in the period of catabolic phase,while platelets were more important in the period of anabolic phase.Model prediction results showed that logistic regression models were the best in predicting prognosis in sepsis patients in the metabolic instability period,with AUROC of 0.82 in both energy and protein prediction.the best prediction was achieved by SVM in the recovery period,with AUROC of 0.96 in energy prediction AUROC of 0.94 in protein prediction.Conclusion:Elevated cumulative energy debt is an important factor in the mortality of sepsispatients,accompanied by a dynamic process of change in the key controlled indicators of energy and protein intake.Machine learning methods can bring good mortality predictor of sepsis patients with nutritional indicators.And It will be an important aid for clinicians and dietitians to make nutrition support programs.However,due tothe poor sample size and limitation of retrospective study in our study,this conclusion remains to be validated by further prospective clinical trials with large samples.