An innovative deep learning approach for ventilator-associated pneumonia (VAP) prediction in intensive care units - Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology (PREDICT)

Background Ventilator-associated pneumonia (VAP) remains a major complication in intensive care units (ICUs), affecting up to 40% of mechanically ventilated patients and significantly increasing morbidity, and healthcare burden. Current VAP diagnosis relies on retrospective clinical, radiological, and microbiological criteria, leading to delays in targeted treatment and an overuse of broad-spectrum antibiotics. Early and accurate prediction of VAP is essential to optimize patient outcomes and antimicrobial stewardship. This study aimed to develop and validate PREDICT (Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology), a novel deep learning algorithm for early VAP prediction in mechanically ventilated ICU patients. We hypothesized that temporal variations in vital signs could enable early detection of VAP before clinical suspicion arises, outperforming conventional machine learning (ML) models. Methods A retrospective cohort study was conducted using the MIMIC-IV database, including ICU patients requiring invasive mechanical ventilation for at least 48 hours. Vital signs (respiratory rate, SpO₂, heart rate, temperature, and mean arterial pressure) were extracted and structured into time-series windows. The PREDICT model, based on a Long Short-Term Memory neural network, was trained to predict VAP onset at 6, 12, and 24 hours in the future. Its performance was compared to traditional ML models (Random Forest, XGBoost, and Logistic Regression) using key metrics such as area under the precision-recall curve (AUPRC), sensitivity, specificity, and predictive values. Results: PREDICT model demonstrated superior predictive accuracy across all time horizons, achieving an AUPRC of 96.0%, 94.1%, and 94.7% for VAP prediction at 6, 12, and 24 hours, respectively. Sensitivity avec Predictive Positive Value remained consistently high (≥85%), ensuring robust early detection. Traditional ML models showed declining performance for longer prediction windows, underscoring the advantage of deep learning for time-series analysis. Model interpretability using Integrated Gradients revealed that respiratory rate, SpO₂, and temperature were the most influential features in VAP prediction. Conclusion: This study presents PREDICT, the first deep learning model tailored for VAP prediction in ICU, offering a reliable tool for early identification of at-risk patients. By enabling timely interventions, PREDICT could reduce unnecessary antibiotic use and improve patient outcomes.