PT07 - PHENOTYPIC AND ETIOLOGICAL CRITERIA FOR MALNUTRITION AND MORTALITY RISK ASSESSMENT USING MACHINE LEARNING

PT07

PHENOTYPIC AND ETIOLOGICAL CRITERIA FOR MALNUTRITION

AND MORTALITY RISK ASSESSMENT USING MACHINE LEARNING

 

A. Garcia-Grimaldo^1,2,*, C. A. Galindo-Martin³, N. C. Rodriguez-Moguel¹, C. M. Hernandez-Cardenas⁴, J. D. Cadeza-Aguilar⁵, M. Godinez-Victoria², I. A. Osuna- Padilla¹

¹Clinical Nutrition Department, National Institute of Respiratory Diseases, ²Seccion de Estudios de Posgrado e Investigacion, Escuela Superior de Medicina, Instituto Politecnico Nacional, ³Nutrition Service, Hospital San Angel Inn Universidad, ⁴General Direction, ⁵Critical Areas Department, National Institute of Respiratory Diseases, Mexico City, Mexico

 

Rationale: The aim of this study is to develop and validate a risk prediction model for all-cause mortality using machine learning in critically ill respiratory patients.

Methods: This is prospective cohort of critically ill respiratory patients. Machine learning algorithms were applied to selected variables including clinical data (age, SOFA), phenotypic criteria (mid-arm circumference (MAC) and calf circumference (CC) adjusted to BMI), and etiological criteria (C-reactive protein (CRP)). Hyper parameters were 10-fold cross-validation, 500 maximum iterations and model selection by precision. Statistical analyses were performed using R software. The Neuralnet function of caret package was used to generate artificial NN. Performance was evaluated using a confusion matrix and corresponding calculations, including accuracy, sensitivity (SE), specificity (SP), negative predictive value (NPV), and positive predictive value (PPV). A Receiver Operating Characteristics (ROC) curve was generated as a process to evaluate the performance of the artificial NN. The curve and the area under the curve (AUC) were reported.

Results: A total of 226 critically ill patients were included. Of these, 70% (n = 159) were assigned to the training set and 30% (n = 67) comprised the test set for mortality prediction, with no differences in clinical and demographics characteristics (Table 1).

Age, disease severity assessed by SOFA score, serum levels of CRP, CC and MAC were included in the model (Figure 1). 

Image:

 

Conclusion: The optimal predictive model developed using NN showed a moderate ability to predict mortality in critically ill patients. The variables included are easy to obtain in low-middle resource settings.

Despite the moderate capacity, this study remarks the importance of nutritional variables in determining the likelihood of survival in critically-ill patients with respiratory disease.

Disclosure of Interest: None declared