Novo índice clínico-laboratorial para predição de deterioração em adultos hospitalizados com modelo híbrido de aprendizagem profunda

Abstract

INTRODUCTION: Clinical deterioration in patients, which may lead to increased mortality risk, particularly in intensive care units, is associated with a dynamic process of physiological decline characterized by the progression of organ dysfunction. This process can be monitored over time through measurable indicators obtained from patients’ clinical records. Assessing clinical deterioration is a demanding task due to the complex and variable clinical behavior of each patient. The growing volume of clinical data within Electronic Health Records (EHRs) is characterized as longitudinal patient health information. These data enable the application of clinical deterioration protocols capable of observing and tracking physiological progression in hospitalized patients. The use of EHRs allows electronic records to generate insights and predictions of future clinical events. This study is justified by the need to develop a novel index grounded in hybrid artificial intelligence capable of processing heterogeneous temporal clinical data derived from electronic health records. OBJECTIVE: To develop and test an index composed of clinical and laboratory data for predicting deterioration in hospitalized adult patients using machine learning algorithms. METHODS: This study has an exploratory experimental technological design. A total of 1,100,000 patients were selected from two datasets originally extracted from the MIMIC-III database. The study was structured into three stages: (i) a scoping review, (ii) development of an adult patient deterioration index based on the most relevant clinical and laboratory time-series data treated statistically, and (iii) computational experiments using a hybrid approach combining convolutional neural networks (CNNs) and long short-term memory (LSTM) networks to predict the deterioration index at future time steps. Mathematical models were constructed to calculate the weights of variables composing the deterioration index. RESULTS: The hybrid neural network model was evaluated across 10 different scenarios with hyperparameter tuning (Seed, Past_H, Future_H, Val_Split, Epochs, and Batch_Size). The model demonstrated the ability to predict future events using retrospective time-series records to forecast 6-hour, 12-hour, 18-hour, and 24-hour horizons. Global performance metrics achieved were MAE ranging from 9.36 to 9.89 and MAPE from 24.19% to 25.96%. CONCLUSIONS: The hybrid model showed consistent performance under similar hyperparameters across different patient cohorts from MIMIC-III and MIMIC-IV. The best results were observed for 6-hour-ahead predictions. The study presents innovative potential and offers scientific contributions to the academic community, society, and healthcare delivery processes. The findings suggest that the proposed approach can anticipate clinical events and provide opportunities for timely interventions and clinical decision-making in hospital settings, thereby enhancing patient safety.