Predicting intensive care unit severity: A multilayer perceptron approach to simplified acute physiology score–I (SAPS-I)

OpenAlex İngilizce

Precise severity testing of illnesses in the Intensive Care Unit (ICU) is vital in the management of patients and allocation of resources. The conventional scores such as the Simplified Acute Physiology Score (SAPS-I) are based on linear models and they might not exhaust the most complicated physiological associations. Machine Learning (ML) is a promising alternative to model these non-linearities. The current study aims to construct and validate a Multilayer Perceptron (MLP) neural network to achieve the prediction of the SAPS-I score on the basis of routine clinical data gathered during the first 24 hours of ICU admission. The design was performed as a retrospective cohort study based on the use of the data created by the Death Classification ICU (n=3.600). An MLP model with one hidden layer was trained on 31 standardized variables, consisting demographics, vital signs, and laboratory parameters. The data was divided into two subsets training (70%) and testing (30%) to develop a model and to evaluate it unbiasedly. The model assessment were performed using the Sum of Squares Error and the Relative Error. The MLP model represented strong generalization, with comparable error metrics between the training (Relative Error: 0.526) and testing (Relative Error: 0.535) sets. The Sequential Organ Failure Assessment (SOFA) score was found to be the salient predictor variable by the importance analysis (100% normalized importance), followed by White Blood Cell count (63.3%) and Respiratory Rate (59.3%). Predicting the SAPS-I score with an MLP is feasible because the network captures complex nonlinear patterns in clinical data. The model’s internal validation aligns with established knowledge that organ dysfunction, inflammation, and respiratory status drive severity. This matches the broader potential of ML in critical care for improved automated severity assessment, while still requiring external validation and stronger interpretability.