Machine learning for mortality prediction using nutritional and inflammatory markers in critically ill patients
A García-Grimaldo1, C A Galindo-Martín2, N C Rodriguez-Moguel3
1Departamento de Nutrición Clínica, Instituto Nacional de Enfermedades Respiratorias, Mexico City, Mexico; Sección de Estudios de Posgrado e Investigación, Escuela Superior de Medicina, Instituto Politécnico Nacional. Mexico City, Mexico.
Introduction:
Malnutrition and muscle loss are key determinants of outcomes in critically ill patients, yet conventional ICU mortality scores (e.g., APACHE II, SOFA) do not incorporate nutritional status. This study aimed to develop a machine learning model integrating clinical and nutritional indicators associated with malnutrition diagnosis to predict in-hospital mortality in critically ill patients with respiratory diseases.
Methods:
The cohort was divided into training (70%) and test (30%) sets. An artificial neural network (NN) was trained using age, SOFA score, serum C-reactive protein (CRP), calf circumference (CC) adjusted for body mass index, and mid-upper arm circumference (MUAC), selected as nutritional and inflammatory indicators. Patients were not formally classified as malnourished; instead, individual phenotypic and etiologic indicators were used as predictors. Ten-fold cross-validation and random oversampling addressed class imbalance, and hyperparameter tuning was based on precision. Model performance was evaluated using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and area under the receiver operating characteristic curve (AUC).
Results:
We analyzed 226 critically ill patients. According to variable importance, the three most influential predictors were age, BMI-adjusted CC and SOFA score. The test set achieved an AUC of 0.77 (95% CI: 0.62-0.92), accuracy of 0.67, positive predictive 0.35 value and NPV of 0.89. Inclusion of nutritional and inflammatory indicators improve predictive performance.
Conclusion:
This NN-based model provides a feasible, clinically relevant tool for predicting survival in critically ill patients, integrating nutritional and inflammatory markers. Its reliance on simple, routinely collected variables enables implementation in resource-limited environments. Further studies evaluating the response to nutritional support using this model are warranted.