Enhancing mortality prediction in intensive care units: improving APACHE II, SOFA, and SAPS II scoring systems using long short-term memory.

Traditional scoring systems, such as APACHE II, SOFA, and SAPS II, predict ICU mortality using logistic regression (LR) models based on the worst values of time-series variables. This methodology often disregards temporal dynamics, limiting predictive accuracy. This study aimed to enhance mortality prediction in ICU patients by utilizing Long Short-Term Memory (LSTM) algorithms, which capture the full temporal information from time-series data collected within the first 24 h of ICU admission. A total of 1336 ICU patients from the Medical Information Mart for Intensive Care-IV database were included in this study. Time-series variables from the APACHE II, SOFA, and SAPS II scoring systems during the initial 24 h of each patient were collected retrospectively to develop LSTM-based prediction models for hospital mortality. Three distinct data representations were employed for each scoring system: (1) original hourly data over the 24-h period, (2) categorized hourly data derived from the cutoff points defined in the scoring systems, and (3) total scores calculated on an hourly basis. These data representations were used as inputs for separate LSTM models. Model performance was evaluated using a range of metrics, including the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, F1 score, and Brier score. The LSTM-optimized models significantly outperformed LR model using conventional scoring systems. For APACHE II, the LSTM model achieved an AUC of 0.898 compared to 0.777 for LR (P < 0.001). Similarly, the SOFA and SAPS II LSTM models achieved AUCs of 0.861 and 0.897, respectively, compared to 0.715 (P < 0.001) and 0.708 (P = 0.004) for their LR counterparts. While all three LSTM models performed well, the model utilizing original hourly values consistently demonstrated the highest accuracy, particularly for APACHE II (P = 0.021). Incorporating time-series data into LSTM-based prediction models significantly improves the accuracy of ICU mortality prediction compared to traditional LR-based scoring systems. This approach has the potential to redefine clinical decision support tools in critical care settings by leveraging temporal dynamics in patient data.