Development and validation of a predictive model for feeding intolerance in intensive care unit patients with sepsis

Abstract

Background: Feeding intolerance in patients with sepsis is associated with a lower enteral nutrition (EN) intake and worse clinical outcomes. The aim of this study was to develop and validate a predictive model for enteral feeding intolerance in the intensive care unit patients with sepsis.

Methods: In this dual-center, retrospective, case-control study, a total of 195 intensive care unit patients with sepsis were enrolled from June 2018 to June 2020. Data of 124 patients for 27 clinical indicators from one hospital were used to train the model, and data from 71 patients from another hospital were used to assess the external predictive performance. The predictive models included logistic regression, naive Bayesian, random forest, gradient boosting tree, and deep learning (multilayer artificial neural network) models.

Results: Eighty-six (44.1%) patients were diagnosed with enteral feeding intolerance. The deep learning model achieved the best performance, with areas under the receiver operating characteristic curve of 0.82 (95% confidence interval = 0.74-0.90) and 0.79 (95% confidence interval = 0.68-0.89) in the training and external sets, respectively. The deep learning model showed good calibration; based on the decision curve analysis, the model's clinical benefit was considered useful. Lower respiratory tract infection was the most important contributing factor, followed by peptide EN and shock.

Conclusions: The new prediction model based on deep learning can effectively predict enteral feeding intolerance in intensive care unit patients with sepsis. Simple clinical information such as infection site, nutrient type, and septic shock can be useful in stratifying a septic patient's risk of EN intolerance.

Keywords: Deep learning; enteral feeding intolerance; predictive model; sepsis.

Figure 1

ROC curve of the training cohort. AUC of each model is 0.70 (95% confidence interval [CI]; 0.60–0.80) in Bayes Net model, 0.73 (95% CI 0.64–0.82) in Logistic regression model, 0.82 (95% CI 0.74–0.90) in Deep Learning model 0.92 (95% CI 0.87–0.97) in Random Forest model, and 0.94 (95% CI 0.90–0.99) in Gradient Boosting model.

Figure 2

ROC curve of the validation cohort. AUC of each model is 0.73 (95% CI 0.60–0.85) in Bayes Net model, 0.69 (95% CI 0.56–0.81) in Logistic regression model, 0.79 (95% CI 0.68–0.89) in Deep Learning model 0.63 (95% CI 0.50–0.76) in Random Forest model, and 0.60 (95% CI 0.47–0.74) in Gradient Boosting model.

Figure 3

The ranking of top 15 predictive variables in the deep learning model. COPD: chronic obstructive pulmonary disease

Figure 4

Calibration curve of deep learning model in overall datasets. The red line represents a perfect prediction by an ideal model. The black line represents the performance of the deep learning model. The calibration curve in overall cohorts shows the agreement between predicted (x-axis) and observed (y-axis) risk of feeding intolerance in ICU patients with sepsis.

Figure 5

Decision Curve analysis for the deep learning model. The y-axis measures the net benefit. The red line represents the deep learning model. The gray line represents the assumption that all the patients have feeding intolerance. The black line represents the assumption that no patients have feeding intolerance.

Figure 6

The deep learning-based model showed a risk of 19% for a patient with sepsis aged 69 years with a history of hypertension and diabetes, who had the following features: infection, pneumonia; APACHE II score, 22; albumin, 30.8 g/L; creatinine, 87 μmol/L; urea nitrogen, 8.75 mmol/L; lactic acid, 1.2 mmol/L; sedation and analgesia treatment; treatment with two types of antibiotics; intermittent feeding with short peptides; and intra-abdominal pressure, <12 mmHg.