640 - Improving Neonatal Weight Gain Modeling by Coupling Physiological Data to Nutritional Data

Publication Number: 1617.640

Will King, Medical Predictive Science Corporation, Charlottesville, VA, United States; Gustave H. Falciglia, Northwestern University The Feinberg School of Medicine, Chicago, IL, United States

Background: Nutrition intake is important for the growth and development of critically ill preterm and term infants, yet NICU clinicians lack reliable, real-time intake data. We previously reported success in calculating daily energy intake using a natural language processing engine that examined fluid intake data collected from the EHR; however, performance of growth models were unsatisfactory. We have recently added high-fidelity physiological data along with intake data.

Objective: We hypothesize that combining physiological data with nutritional data, and testing alternative modeling techniques, will improve performance of growth modeling.

Design/Methods: From infants at the Lurie Children's NICU with complete data, we extracted a daily summary of growth, fluid intake, nutrition orders, and physiological data via a real-time HL7 messaging interface with the EHR and monitoring network. For each patient-day, we calculated a smoothed growth velocity in g/kg/day using a locally estimated scatterplot smoothing (LOESS) function to reduce noise. After splitting the data into 50% training and 50% testing sets, we tested 12 modeling techniques and 28 candidate explanatory variables including measures of heart rate variability: mean daily R-R interval and mean daily HeRO Score. Using 5-fold cross-validation with 25 repeats and 5 tuning parameter iterations for each of the 12 modeling techniques with the caret package in R, we modeled weekly growth in the training set. The model with the highest correlation in the test set was selected. We evaluated model performance on the test set with odds ratio and AUROC, defining growth failure as < 8 g/kg/day.

Results: Only the partial least squares model avoided overfitting with apparent correlation of 0.73 and correlation on the test data of 0.70 (11 of 12 models had apparent correlation as high as 0.99 on the training data but inferior results on the test data). Test set patients modeled at < 8 g/kg/day had an odds ratio of 41 (95% CI: 21 - 90) for achieved growth failure. The AUC of the PLS model for achieved growth < 8 g/kg/day was 0.92 (95% CI: 0.89 - 0.94) in the test set. The strongest independent variables in the PLS model were demographic or fluid intake/output, followed by physiological.

Conclusion(s): We have improved performance to model growth by adding physiological variables and more sophisticated modeling techniques. We expect to further refine our models with the addition of high-fidelity respiratory rate and pulse oximetry, macro- and micronutrient, lab, and clinical data, and by taking steps to reduce overfitting by manually implementing k-fold cross-validation.

Table 1.
Patient demographics. Data are expressed as Median [IQR] or Count (Percent).

Figure 1.
Growth plot of patients in the test set. The dots represent smoothed weight (in grams), while the lines the beginning and end, respectively, of a 7-day weight gain prediction. The x-axis the post menstrual age. Importance of the top 20 explanatory variables is also included.

Figure 2.
Model Performance. Panel A (left): scatterplot of achieved versus modeled weekly weight gain. Dashed lines indicate the growth failure threshold for modeled (vertical line) and achieved (horizontal line). Panel B (right): ROC curve for sensitivity versus specificity for growth failure (achieved growth <8 g/kg/day) versus modeled weekly weight gain.