TITLE:
Enhanced Predictive Modelling for Delirium in Intensive Care Using Simplified Deep Learning Architecture with Attention Mechanism
AUTHORS:
Rocco de Filippis, Abdullah Al Foysal
KEYWORDS:
Delirium Prediction, ICU Monitoring, Deep Learning, Attention Mechanism, LSTM Model, Time-Series Analysis, Risk Stratification
JOURNAL NAME:
Open Access Library Journal,
Vol.12 No.1,
January
23,
2025
ABSTRACT: Delirium is a common yet critical condition among Intensive Care Unit (ICU) patients, characterized by acute cognitive disturbances that can lead to severe complications, prolonged hospital stays, and increased mortality rates. Early detection and proactive management of delirium are essential to mitigate its adverse effects. This study introduces a novel deep learning model designed to predict the onset of delirium in ICU patients, aiming to assist healthcare professionals in identifying high-risk individuals at an early stage. Our approach integrates both static and dynamic patient data—capturing baseline characteristics and real-time physiological trends—to provide a comprehensive risk assessment framework. The model architecture employs a simplified Long Short-Term Memory (LSTM) network enhanced with an attention mechanism, which enables the model to focus on critical time points in dynamic patient data. This allows for a more interpretable and effective prediction of delirium risk by weighing the most relevant signals in the patient’s physiological timeline. Static features such as age, APACHE-II score, and comorbidity levels are combined with dynamic features, including vital signs monitored over time (e.g., heart rate, blood pressure, and respiratory rate), to form a holistic representation of each patient’s health status. Synthetic data was generated to simulate realistic ICU scenarios, with clear patterns introduced to mimic the physiological changes associated with delirium onset. The model was evaluated using a comprehensive set of metrics, including accuracy, Area Under the Receiver Operating Characteristic (ROC-AUC), and Area Under the Precision-Recall Curve (PR-AUC). The results demonstrate that the model achieved near-perfect performance on synthetic data, with an AUC of 1.00 in both ROC and PR curves, and a classification report indicating 100% precision, recall, and F1-score for both classes. The model’s rapid convergence and stable validation metrics further confirm its robustness and effectiveness in identifying delirium cases within a simulated ICU dataset. While these results highlight the model’s potential for real-time delirium prediction, it is important to recognize the limitations of synthetic data in capturing the full complexity of clinical settings. Future work will focus on validating the model with real-world ICU data to assess its generalizability and adaptability to diverse patient populations. Additionally, improvements to the synthetic data generation process, such as introducing variability within each class and incorporating potential confounding factors, will enhance the realism of simulated ICU data for further testing. The attention-based LSTM model presented in this study represents a promising tool for ICU delirium risk stratification, providing a foundation for further advancements in AI-driven healthcare solutions aimed at improving patient outcomes through early intervention).