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)

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) | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article 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) Geoffray AGARD, Christophe ROMAN, Christophe GUERVILLY, Jean Marie FOREL, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6151630/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract 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. Ventilator-associated pneumonia artificial intelligence deep learning predictive modeling intensive care machine learning MIMIC-IV time-series analysis Long Short-Term memory Figures Figure 1 Figure 2 Figure 3 Figure 4 Background Ventilator-associated pneumonia (VAP) is one of the most common complication in intensive care units (ICU). Five to 40% of patients under invasive mechanical ventilation (MV) are likely to develop at least one VAP during their stay. VAP increase duration of mechanical ventilation (MV) and ICU length of stay, leading to a potential increase in mortality of up to 50% 1 . Diagnosis of VAP is based on a combination of three criteria: clinical suspicion, apparition or worsening of radiological infiltrates and positivity of a respiratory tract culture 2 . Clinical suspicion, the key element of diagnosis, has very low sensitivity and specificity 3,4 . In addition, VAP diagnosis is retrospective by nature, linked to the positivity of respiratory samples. Current recommendations therefore call for respiratory sampling and probabilistic antibiotic therapy while awaiting microbiological results to confirm or invalidate the diagnosis 1,2 . Although this approach limits the low sensitivity of markers of clinical suspicion, it exposes the patient to increased consumption of broad-spectrum antibiotics in the ICU. Early diagnosis of VAP is a critical priority, as delayed or missed diagnosis can lead to prolonged infections and worse patient outcomes, while premature empirical treatments increase the risks of antibiotic resistance and adverse drug effects. The ability to anticipate VAP onset before clinical suspicion arises could enable targeted interventions, reducing unnecessary antibiotic exposure and improving survival rates. In order to improve the earliness of VAP diagnosis, it is important to compute and merge information contained in various indicators. The recent advent of artificial intelligence (AI), a set of technologies designed to simulate human cognitive abilities, could make it possible to improve the earliness of VAP diagnosis. To date, literature on systems for VAP prediction using AI is scarce. Most of these systems use machine learning (ML) algorithms, a branch of AI that uses statistical algorithms to learn from data. Samadani et al. used a ML model for VAP prediction within 24 hours with demographic data, vital constants, biology and mechanical ventilation data for training 5 . Although AUROC (75.6%) for their algorithm appeared better than CPIS score, sensitivity (68%) and specificity (67%) remained unsatisfactory. Meanwhile, Deep learning (DL), a specialized branch of machine learning (ML), uses advanced structures called deep neural networks to analyze and interpret data 6,7 . Unlike traditional methods that often require manual selection of relevant features, DL algorithms can automatically identify and learn important patterns directly from raw data. Additionally, DL is particularly well-suited for processing time-based information, as it can recognize long-term trends and complex sequences in data, making it an ideal tool for tasks that involve changes over time. 8 . There is currently no work dealing with the application of deep learning to the prediction of VAPs in the ICU. The objective of this study was to develop and validate PREDICT ( Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology ), a deep learning algorithm designed to facilitate the early diagnosis of VAP in ICU patients who have been mechanically ventilated for more than 48 hours. Our main hypothesis was that variations in vital signs, in particular respiratory rate and SpO2, could enable a DL tool to early detect VAP occurrence. Additionally, the study sought to demonstrate that deep learning outperforms traditional machine learning methods in accurately predicting VAP. Materials and Methods Study design This was a retrospective cohort study conducted using the publicly available MIMIC-IV (Medical Information Mart for Intensive Care IV) database version 2.2. MIMIC-IV contains de-identified information from ICU admissions at the Beth Israel Deaconess Medical Center between 2008 and 2019 in United-States 9 . This database provides a rich dataset including demographic details, vital signs, laboratory results, and treatment information. In this work, to simplify the implementation of a VAP prediction tool in ICU, we chose to focus solely on vital signs data, which can be directly obtained from monitoring systems. Notably, similar approaches have shown promising results in developing sepsis prediction algorithms using this limited data set 10,11 . The study design followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD + AI) guidelines 12 ( additional file n°8 ). The primary objective was to develop and validate a deep learning algorithm for the early prediction of VAP at 6, 12 and 24 hours in the future using vital signs data. Patient Population Patients were included in the study if they : Were aged 18 years or older Required invasive MV for more than 48 hours Had complete records of vital signs during the study period. Exclusion criteria were defined as : The presence of community-acquired pneumonia (as identified by ICD-10 codes or documentation of respiratory infection prior to ICU admission) MV duration shorter than 48 hours Incomplete or missing key variables required for modeling The VAP risk period was defined as starting 48 hours after the initiation of MV and continuing until 72 hours post-extubation. This window was chosen to capture nosocomial pneumonia cases while excluding early-onset pneumonia likely acquired before ICU admission. Data Collection For the purposes of this study, a MV episode was defined as any continuous period during which a patient received invasive mechanical ventilation, either through an endotracheal tube or a tracheostomy. The start of an MV episode was marked by the initiation of invasive ventilation, and the end was defined as the moment the patient was extubated or transitioned to non-invasive ventilation for more than 48 hours 13 . Episodes with interruptions of less than 24 hours were considered part of the same continuous ventilation period to account for temporary weaning or procedural pauses commonly seen in ICU settings. Only episodes lasting more than 48 hours were included in the analysis. These episodes were extracted from the MIMIC-IV records using specific Structured Query Language (SQL) queries. Vital signs extracted for this study included respiratory rate, heart rate, mean arterial pressure, body temperature, and oxygen pulsed saturation (SpO2). These variables were also extracted from the MIMIC-IV database's time-series records of patient monitoring with SQL queries. Additional demographic and clinical variables, such as age, gender, Sepsis-related Organ Failure Assessment (SOFA) and Simplified Acute Physiology Score (SAPS-II) at ICU admission, were extracted from corresponding patient records. ICD-10 codes were used to classify and exclude patients with community-acquired pneumonia and to identify relevant comorbidities, such as chronic obstructive pulmonary disease (COPD), ischemic heart disease, diabetes mellitus, chronic kidney insufficiency, and active cancer. Data on ICU admission sources, ventilation type (invasive or tracheostomy), and duration of MV were obtained from structured database fields. Ventilation Free Days (VFDs) D28 refers to the number of days within the first 28 days after the start of MV during which a patient is alive and not dependent on mechanical ventilation. Outcomes The primary outcome of this study was the accurate early prediction of VAP within predefined time windows (6, 12, and 24 hours) following the observation period of 24 hours. Performance was measured using metrics such as the Area Under the Precision-Recall Curve (AUPRC), sensitivity, positive predictive value (PPV), specificity, and negative predictive value (NPV) which are defined in additional file. Secondary outcomes included a comparison of the model performance, assessing the superiority of the deep learning (DL) algorithm (PREDICT) over traditional ML models, including Random Forest, XGBoost, and Logistic Regression, in terms of predictive accuracy and robustness across all prediction horizons. Annotation of VAP events VAP events were identified based on a standardized methodology using clinical and microbiological criteria, consistent with international guidelines (IDSA/ATS 14 and SFAR/SRLF 15 ) and the infection labelling method initially proposed by Seymour et al. 16 and adopted by Samadani et al. 5 in a previous work on VAP prediction ( Additional file n°2 - Supplemental figure 1 ). First, suspected VAP cases were identified by screening MV episodes longer than 48 hours and during the VAP risk period previously defined. During this period, respiratory microbiological cultures obtained through bronchial aspirates, bronchoalveolar lavage (BAL), or tracheal aspirates were reviewed for evidence of infection. A positive culture was required to support a diagnosis of pneumonia. Clinical interventions were then assessed in conjunction with microbiological findings. New antibiotic regimens targeting respiratory pathogens were identified if initiated within a defined temporal window (72 hours before or 24 hours after the collection of the microbiological sample). Additionally, episodes of VAP occurring less than 48 hours apart were considered as part of the same infection event and were not treated as independent episodes. The onset of a VAP episode was defined as the earliest timepoint of either the microbiological sample collection or the initiation of new antibiotics. Selection of comparator patients without VAP Patients without ventilator-associated pneumonia (VAP) were included as comparators using a 1:1 matching strategy to ensure robust comparisons and minimize confounding. Potential controls were identified among patients with at least one MV episode larger than 48 hours without meeting VAP diagnostic criteria. For each VAP episode, a ventilation episode from a patient without VAP was matched. Matching between these two episodes was stratified on age, sex, SAPS-II score at ICU admission, and duration of MV. This selection process provided negative examples essential for the machine learning algorithm to distinguish between VAP and non-VAP episodes, improving its specificity. Data preprocessing Preprocessing included several steps to ensure the dataset was suitable for training a predictive model ( Additional n°3 - Supplemental figure 2 ). 1) Data resampling and cleaning : Vital signs were resampled at hourly scale to standardize time intervals and reduce measurement errors. Missing values, were handled using linear interpolation to preserve the continuity and integrity of the time-series data. 2) Normalization : Each vital sign value was standardized by subtracting the mean and dividing by the standard deviation. This normalization step ensured comparability across variables, preventing any single variable with a larger numerical range (e.g., mean arterial pressure) from disproportionately influencing the algorithm. 3) Temporal windows creation : To allow the algorithm to analyze time-dependent patterns in patient data, we divided the continuous flow of vital signs into temporal windows. A temporal window is a defined time segment that contains patient data recorded over a specific period. In this study, each temporal window consisted of 24 hours of continuous vital sign recordings. Each temporal window was composed of two key parts. The observation window represent a 24-hour period during which vital signs were collected each hour. This is the input data used by the algorithm to identify physiological changes. The prediction window is the period following the observation window during which the algorithm predicts the occurrence of VAP. For this study, three prediction windows were defined at 6, 12 and 24 hours after the observation window. Temporal windows were created using a sliding window approach. This means that after constructing a 24-hour observation window, we moved it forward by 1 hour to create the next window. This overlap ensures the algorithm captures granular temporal patterns without missing key changes in vital signs. Each temporal window was labeled based on whether a VAP episode occurred within the associated prediction window. If a VAP episode occurred within the prediction window (e.g., at 4 hours for a 6-hour prediction horizon), this windows was labelled as positive. Otherwise, the time window was labelled as negative. This labeling process allowed the algorithm to differentiate between patterns leading to VAP and those not associated with infection. 4) Balancing the dataset : Because VAP events were rare in this dataset (less than 1% of temporal windows), the Synthetic Minority Oversampling Technique (SMOTE) 17 was applied ( Additional file n°4 - Supplemental figure 3 ). This method artificially generated synthetic examples of VAP-positive temporal windows while preserving the structure of the original data. It ensured the model was exposed to sufficient positive examples, enhancing its sensitivity and specificity 18 . Details on SMOTE technique are available in additional file. Data splitting for algorithm training and evaluation The dataset was divided into three subsets: Training set (60%) used to train the algorithm, validation set (20%) used to fine-tune hyperparameters and prevent overfitting and test set (20%) held back for final evaluation of model performance ( Additional file n°3 - Supplemental figure 2 ). Stratified sampling was employed to ensure that the proportions of VAP-positive and VAP-negative windows were consistent across all subsets. This step minimized bias and ensured the generalizability of the results. Algorithm development PREDICT model was based on Long Short-Term Memory (LSTM) networks. This type of deep learning architecture is particularly suited for sequential data, as it captures temporal dependencies, such as trends or variations in vital signs over time 19 . The baseline model architecture included 3 LSTM layers, each with 50 cells, followed by a fully connected layer with a sigmoid activation function. A dropout regularization (5–10%) was applied to prevent overfitting. Output of the model was the VAP probability on the prediction window considered. Training was conducted using a cross-entropy loss function and the Adam optimizer with adaptive learning rates and used a stratified cross-validation method with 5-folds ( Additional file n°5 ). Hyperparameters such as the number of LSTM cells, dropout rate, and learning rate were optimized using the validation set. The best configuration was selected based on area under the precision-recall curve (AUPRC), a metric particularly suited for imbalanced datasets 20 ( Additional file n°1 ). Comparator Models To evaluate the added value of deep learning, we trained several traditional machine learning models, including Random Forest, XGBoost, and logistic regression. These models were trained and tested using the same dataset and preprocessing pipeline. Their performance was compared to the PREDICT algorithm using metrics such as sensitivity, specificity, precision, and AUPRC. We compared PREDICT‘s AUPRC with other models AUPRC using a bootstrap resampling approach (n=100 resamples) to determine statistical difference in AUPRC. Model Explainability To enhance clinical interpretability, we applied the Integrated Gradients technique 21,22 to analyze the model's decision-making process. This approach quantifies the contribution of each input variable to the prediction, enabling us to identify which vital signs were most influential in detecting VAP. Statistical Analysis Continuous variables were summarized as medians with interquartile ranges and compared using the Wilcoxon rank-sum test. Categorical variables were presented as counts and percentages, with differences assessed using Pearson’s χ² test or Fisher’s exact test. Metrics such as area under the curve of receiver-operating characteristic (AUROC), precision (Predictive Positive Value), recall (Sensitivity), area under the curve of precision-recall curve (AUPRC), and Youden-index (definitions in supplementary) were calculated for each model with 95% confidence intervals using bootstrapping (n=100 iterations). AUPRC was chosen as the primary metric because it better accounts for imbalanced datasets than AUROC 20 . Deep learning models were constructed using tensorflow (v 2.15) and machine learning models were built with scikit-learn (v 1.5.1) packages . All analysis have been performed in Python language ( v 3.11.9 ) with also use of pandas (v2.2.2) and numpy (v 1.26.4) packages. Results VAP episodes We identified 38 750 invasive MV episodes with 9 849 sessions greater than 48 hours (25.4%) concerning 7 871 patients in MIMIC-IV dataset between 2008 and 2019 (Fig. 1 ). Our VAP annotation algorithm identified 452 VAP episodes on 397 patients and 404 MV episodes > 48 hours (4.1%). During their stay in ICU, 351 patients presented 1 VAP episode, 41 patients two VAP, 1 patient three VAP and 4 patient presented four VAP episodes. Median time from ICU admission to first VAP episode was 6 days (IQR 4–12) and also 6 days (IQR 3.7–11.3) from MV initiation to the first VAP episode. Population characteristics Despite stratification, men were significantly more at risk for VAP during their stay (67 vs 59% - P = 0.011). Patients in the VAP group had less ischemic heart disease, chronic kidney insufficiency and active cancer but more COPD patients than no-VAP group. In the population, 20% of patients were admitted with sepsis, 11% with trauma and 8.3% with stroke. The main characteristics of population are presented in Table 1 . Outcomes - Model performance PREDICT was able to predict VAP six hours before onset with a sensitivity of 89.7%, a predictive positive value of 89.8% and a predictive negative value of 99.7%. The sensitivity and positive predictive value exceeded 85% for all prediction thresholds (Table 2 ) including 24-hour prediction (Sensitivity 85.1% - Specificity 99.2%). Figure 2 shows how the probability of VAP onset at 12 hours evolves as a function of the vital signs of a patient who presented several VAP during his ICU stay. Table 1 Patients characteristics. 1 n (%); Median [25%-75%] 2 Pearson's Chi-squared test; Wilcoxon rank sum - COPD : Chronic Obstructive Pulmonary Disease – ICU : Intensive Care Unit Overall 1 , N = 904 VAP 1 , N = 452 No VAP 1 , N = 452 p-value 2 Sex (Male) 573 (63%) 305 (67%) 268 (59%) 0.011* Age (years) 64.2 [52.1–75.3] 63.9 [50.7–74.3] 65.1 [52.9–76.1] 0.2 Pre-existing Diseases n(%) Hypertension 450 (50%) 217 (48%) 233 (52%) 0.3 Ischemic heart disease 306 (34%) 131 (29%) 175 (39%) 0.002* Diabetes mellitus 216 (24%) 106 (23%) 110 (24%) 0.8 Chronic renal failure 218 (24%) 95 (21%) 123 (27%) 0.029* Obstructive sleep apnea 145 (16%) 69 (15%) 76 (17%) 0.5 Active cancer 134 (15%) 56 (12%) 78 (17%) 0.039* COPD 66 (7.3%) 43 (9.5%) 23 (5.1%) 0.011* Active hematological malignancy 31 (3.4%) 18 (4.0%) 13 (2.9%) 0.4 Source of admission to ICU 0.9 Emergency ward 665 (74%) 331 (73%) 334 (74%) Medical ward 150 (17%) 74 (16%) 76 (17%) Elective surgery 89 (9.8%) 47 (10%) 42 (9.3%) SOFA – admission 2.0 [0.0–4.0] 1.0 [0.0–4.0] 2.0 [0.0–4.0] 0.2 SAPS-II on admission 42.0 [32.0–54.0] 42.0 [31.0–53.0] 43.0 [34.0–55.0] 0.053 Time from ICU admission to initiation of MV (hours) 7.2 [1.7–67.3] 8.2 [1.8–73.0] 6.0 [1.7–60.4] 0.4 Reason for ICU admission Sepsis 179 (20%) 91 (20%) 88 (19%) Trauma 97 (11%) 73 (16%) 24 (5.3%) Hemorrhagic or ischemic stroke 75 (8.3%) 44 (9.7%) 31 (6.9%) Acute malignancy 46 (5.1%) 17 (3.8%) 29 (6.4%) ARDS 40 (4.4%) 34 (7.5%) 6 (1.3%) Pneumonia 34 (3.8%) 20 (4.4%) 14 (3.1%) Myocardial infarction 26 (2.9%) 8 (1.8%) 18 (4.0%) Comparators Models Comparatively with other machine learning algorithm we trained, PREDICT offers the best AUPRC with respectively 96.0%, 94.1% and 94.7% for VAP prediction at 6, 12 and 24 hours. The XGBoost and Random Forest algorithms provide viable alternatives for early predictions at 6 h with an AUPRC of 94.1% and 91.7% respectively. However, performance rapidly declines for long-range prediction thresholds until it becomes completely worthless for 24-hour predictions. Figure 3 shows the AUPRC and AUROC curves for the PREDICT model compared of concurrent ML algorithms. The comparison of AUPRC between PREDICT and traditional machine learning models revealed a statistically significant difference ( p < 0.001 ) for all comparisons. Table 2 PREDICT model performance for different VAP prediction threshold. Best classification threshold was calculated for Sensibility = PPV (Precision = Recall). 95% confidence intervals are available in additional file. AUPRC : Area Under Precision Recall Curve – PPV : Predictive Positive Value – NPV : Predictive Negative Value VAP prediction Best threshold AUPRC (%) Sensibility (%) Specificity (%) PPV (%) NPV (%) PREDICT Model 6 h 0,53 96,0 89,7 99,7 89,8 99,7 12 h 0,52 94,1 85,9 99,6 85,6 99,6 24 h 0,43 94,7 85,1 99,2 85,0 99,2 The best structure of the PREDICT algorithm selected after HPO was: 3 hidden layers of 50 LSTM cells with a dropout rate of 10% for VAP predictions at 6 and 12 hours and 5% for predictions at 24 hours ( Additional file n°5 ). Training curves and confusion matrix for each prediction threshold are provided in additional files (Additional file n°6) as well as all models performance metrics with 95 confidence intervals ( Additional file n°7 ). PREDICT model explainability The integrated gradients analysis of the PREDICT model reveals how each feature influences predictions over time (Fig. 4 ). On temporal window scale, heart rate had a steady, positive contribution, while mean arterial pressure showed significant early fluctuations. SpO2 had minimal impact, remaining close to zero, whereas temperature shifted from a negative to a positive influence toward the end of sequences. Respiratory rate gained relevance progressively over time. Considering overall scale, the most critical features for prediction were SpO2, temperature, and respiratory rate, with mean arterial pressure and heart rate playing smaller roles. Patients’ prognosis Patients in the VAP group had higher hospital length of stay (25.9 vs 16.3 days – P < 0.001) and ICU (18.9 vs 8.4 days – P < 0.001) with also an increased mortality in ICU. Duration of MV was increased in the VAP group and Ventilation Free Days at day 28 were also significantly lower. Outcomes for population are described in Table 3 . Table 3 Patient prognosis outcomes − 1 n (%); Median [25%-75%] 2 Pearson's Chi-squared test; Wilcoxon rank sum test Overall 1 , N = 904 VAP 1 , N = 452 No VAP 1 , N = 452 p-value 2 Length of stay – Hospital (days) 21.4 [12.9–35.2] 25.9 [17.1–39.0] 16.3 [9.7–28.3] < 0.001* Length of stay – ICU (days) 14.0 [7.8–22.9] 18.9 [13.0–30.0] 8.4 [5.3–15.7] < 0.001* Time from ICU admission to death (days) 38.0 [13.5-117.5] 42.5 [18.3-103.3] 30.1 [7.7-147.8] 0.007* Duration of mechanical ventilation (days) 8.4 [4.0-15.4] 13.6 [8.4–21.7] 4.4 [3.0-8.3] < 0.001* Ventilation free days D28 (days) 14.5 [0.0-22.6] 9.1 [0.0-17.6] 20.9 [0.0-24.5] < 0.001* ICU mortality 188 (21%) 110 (24%) 78 (17%) 0.009* In-hospital mortality 261 (29%) 139 (31%) 122 (27%) 0.2 Discussion In this study, we developed and validated PREDICT, a new innovative deep learning model for the early detection of VAP at 6, 12 and 24 hours in the future. PREDICT demonstrated excellent diagnostic performance, with an AUROC of 99% for all prediction thresholds. Sensitivity (Se) and positive predictive value (PPV) ranged from 85.1–89.7%, while the AUPRC exceeded 94%, indicating high predictive capability for both the presence and absence of VAP. PREDICT represents the first application of deep learning to VAP prediction in ICUs. Previous studies using machine learning algorithms achieved lower performance. Liang et al. developed a RandomForest model to predict VAP within 24 hours of intubation using around thirty demographic, biological, and physiological variables 23 . While this model showed satisfactory performance (AUROC 84%, Se 74%, specificity 71%), its reliance on diverse and inconsistent data sources limited its ICU applicability. Similarly, Samadani et al. used an XGBoost model, incorporating a wide range of variables, to predict VAP within 24 hours 5 . Although encouraging (AUROC 76%, AUPRC 75%), this model had low sensitivity (67.5%) and PPV (68.5%), restricting its clinical utility. In contrast, PREDICT outperformed these models across all prediction horizons and maintained stable predictive performance even for longer intervals, highlighting the advantage of deep learning in modeling complex relationships without extensive manual feature engineering 7 . To ensure comprehensive evaluation, we compared PREDICT with Random Forest and XGBoost models. While these traditional algorithms closely matched PREDICT’s performance for short prediction tasks (6h), with AUROCs of 99% and AUPRCs between 92% and 94%, their accuracy declined for longer prediction horizons, unlike PREDICT, which remained robust. The stability of PREDICT’s predictions illustrates the strength of deep learning in extracting subtle patterns from time-series data, offering more reliable and generalizable predictions. PREDICT offers physicians a valuable tool for the early identification of VAP, allowing intervention 6 to 24 hours before clinical suspicion. This early detection can enable timely initiation of targeted antibiotics, potentially reducing the progression to severe infections, organ dysfunction, and prolonged ICU stays 24 . Gaining 6 to 24 hours is crucial in critically ill patients, where early intervention can prevent progression to severe infections, organ dysfunction, and prolonged ICU stays 26 . PREDICT high negative predictive value can also help in optimizing antimicrobial stewardship by avoiding unnecessary broad-spectrum antibiotics, thus reducing antibiotic resistance risks 25 . In such cases, it can prompt the physician to investigate for another origin of sepsis than lungs. The fact that PREDICT only uses vital constants available on ICU monitors gives us hope that in the future, such technology can be integrated directly into these monitors, making it much easier for clinicians to use. A key barrier to deep learning adoption in medicine has been the lack of explainability. Trust in AI systems depends on understanding the reasoning behind predictions specifically in medical field 27 . Recent advancements, such as the integrated gradient technique 21 , have improved interpretability. By applying this method, we highlighted the contribution of individual variables to predictions, identifying temperature and respiratory rate changes as key indicators of VAP. These findings align with earlier studies showing the predictive value of these variables in machine learning models 5,23 . It may seem paradoxical that PREDICT, despite using fewer variables, outperforms models incorporating dozens of features. Complex models often suffer from overfitting 28 , capturing noise along with trends, which hampers their generalization to new data. Moreover, multicollinearity between variables can dilute the predictive signal, further reducing performance 29 . Similar observations were made by Desautels et al. who achieved promising sepsis prediction results (Se 80%, Sp 80%, AUROC 88%) using only vital signs, SpO₂, Glasgow score, and age, demonstrating that simpler models can sometimes yield superior results 10 . Our work also stands out due to its annotation technique, adapted from Seymour et al and previously used by Samadani et al. which adheres to international recommendations 5,14–16 . This method ensures microbiological accuracy and resolves key challenges in annotating VAP, such as variability in definitions. ICD codes, often used for annotation, lack both sensitivity (59%) and specificity (PPV 27–42%), making them unsuitable for precise classification 30 . By employing time windows to pinpoint VAP diagnosis, we avoided the limitations of retrospective ICD coding, which fails to capture the exact timing of clinical events. A primary limitation of this study lies in its monocentric design, relying exclusively on the MIMIC-IV database, which may limit the generalizability of PREDICT to diverse ICU settings. Additionally, the data available in MIMIC-IV may not fully capture the complexity of clinical decisions. To overcome these challenges, future work should include retraining the model with data from multiple hospitals, incorporating diverse patient populations and clinical practices. Finally, while PREDICT has shown potential to improve early VAP detection, its clinical benefits must be confirmed through prospective randomized controlled trials to assess its impact on clinical outcomes, such as mortality, duration of ventilation, and antibiotic exposure. Conclusion In this work we developed and internally validated a new deep-learning model for VAP prediction using the MIMIC-IV cohort. Using a small number of clinical variables, our PREDICT algorithm demonstrated a high capacity for detecting both short- and long-term VAP. The superiority of the deep-learning approach was confirmed by comparison with other machine learning algorithms trained in the same way. An external validation of the algorithm and its implementation in a prospective trial strategy should be considered in future work. Abbreviations - AI : Artificial Intelligence - ARDS : Acute Respiratory Distress Syndrome - AUPRC : Area Under the Precision-Recall Curve - AUROC : Area Under the Receiver Operating Characteristic Curve - BAL : Bronchoalveolar Lavage - CDC : Centers for Disease Control and Prevention - COPD : Chronic Obstructive Pulmonary Disease - CPIS : Clinical Pulmonary Infection Score - HPO : Hyperparameter Optimization - ICD-10 : International Classification of Diseases, 10th Revision - ICU – Intensive Care Unit - IDSA/ATS : Infectious Diseases Society of America / American Thoracic Society - IQR : Interquartile Range - LSTM : Long Short-Term Memory - ML : Machine Learning - MIMIC-IV – Medical Information Mart for Intensive Care IV - MV : Mechanical Ventilation - NPV : Negative Predictive Value - PAVM : Pneumonie Associée à la Ventilation Mécanique - PREDICT : Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology - PPV : Positive Predictive Value - ROC : Receiver Operating Characteristic - SAP-II : Simplified Acute Physiology Score II - SFAR/SRLF : Société Française d’Anesthésie et de Réanimation / Société de Réanimation de Langue Française - SMOTE : Synthetic Minority Oversampling Technique - SOFA : Sepsis-related Organ Failure Assessment - SQL : Structured Query Language - TRIPOD+AI : Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (with AI guidelines) - VAP : Ventilator-Associated Pneumonia - VFD D28 : Ventilation-Free Days at Day 28 Declarations Ethics approval and consent to participate Not applicable. This study was conducted using the publicly available MIMIC-IV database, which contains de-identified health records of ICU patients from the Beth Israel Deaconess Medical Center (BIDMC). The database is maintained by the Massachusetts Institute of Technology (MIT) Laboratory for Computational Physiology. The use of MIMIC-IV data is approved by the Institutional Review Board (IRB) of BIDMC and follows the ethical guidelines outlined by the Massachusetts Institute of Technology (MIT). As all data in MIMIC-IV are de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA), this study was exempt from additional ethical approval. All researchers accessing MIMIC-IV data must complete the Collaborative Institutional Training Initiative (CITI) program on Human Research Ethics, ensuring compliance with ethical guidelines for research involving human data. Since this study is based on a secondary analysis of fully anonymized data, informed consent from patients was not required, as determined by the BIDMC IRB. Ethics approval reference: BIDMC IRB Protocol #2001P001699 Consent for publication Not applicable. No identifiable personal data or patient information was used in this study. Availability of data and materials The dataset used in this study is publicly available through the MIMIC-IV database, hosted by the Massachusetts Institute of Technology (MIT) at https://physionet.org/content/mimiciv/. Access to MIMIC-IV requires completion of the appropriate data use agreement and certification in research ethics training. Code for this project is available on request from the first author (GA). Competing interests The authors declare that they have no competing interests. Funding This study did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Authors' contributions GA (Geoffray Agard) conceptualized the study, designed the methodology, performed data extraction and analysis, and drafted the manuscript. SH (Sami Hraiech) contributed to study and methodology design, clinical interpretation of results, and manuscript revision. LB (Laurent Boyer) provided statistical guidance and contributed to manuscript review. CR (Christophe Roman) and MO (Mustapha Ouladsine) provided expertise in machine learning and deep learning methodologies. JMF (Jean-Marie Forel), CG (Christophe Guervilly) and DB (Damien Barrau) contributed to manuscript revision. All authors read and approved the final manuscript. Acknowledgements The authors would like to thank the PhysioNet team and the Beth Israel Deaconess Medical Center for maintaining and providing access to the MIMIC-IV database. We also acknowledge the support of CERESS (Centre d'Etudes et de Recherches sur les Services de Santé et qualité de vie EA 3279), Aix-Marseille Université and the AP-HM Intensive Care Unit for providing an academic environment conducive to research in artificial intelligence and critical care medicine. Many thanks also to the Dr François Antonini and Mr Dorian GROUSSET for their help and support on this project. References Papazian L, Klompas M, Luyt CE. Ventilator-associated pneumonia in adults: a narrative review. Intensive Care Med. mai 2020;46(5):888‑906. CDC. Ventilator-associated Pneumonia Basics [Internet]. Ventilator-associated Pneumonia (VAP). 2024. Available on: https://www.cdc.gov/ventilator-associated-pneumonia/about/index.html Ramírez-Estrada S, Lagunes L, Peña-López Y, Vahedian-Azimi A, Nseir S, Arvaniti K, et al. Assessing predictive accuracy for outcomes of ventilator-associated events in an international cohort: the EUVAE study. Intensive Care Med. 1 août 2018;44(8):1212‑20. Jansson M, Ala-Kokko T, Ahvenjärvi L, Karhu J, Ohtonen P, Syrjälä H. What Is the Applicability of a Novel Surveillance Concept of Ventilator-Associated Events? Infect Control Hosp Epidemiol. août 2017;38(8):983‑8. Samadani A, Wang T, van Zon K, Celi LA. VAP risk index: Early prediction and hospital phenotyping of ventilator-associated pneumonia using machine learning. Artif Intell Med. déc 2023;146:102715. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. mai 2015;521(7553):436‑44. Schmidhuber J. Deep learning in neural networks: An overview. Neural Netw. 1 janv 2015;61:85‑117. Lim B, Zohren S. Time-series forecasting with deep learning: a survey. Philos Trans R Soc Math Phys Eng Sci. 15 févr 2021;379(2194):20200209. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 3 janv 2023;10(1):1. Desautels T, Calvert J, Hoffman J, Jay M, Kerem Y, Shieh L, et al. Prediction of Sepsis in the Intensive Care Unit With Minimal Electronic Health Record Data: A Machine Learning Approach. JMIR Med Inform. 30 sept 2016;4(3):e28. Calvert JS, Price DA, Chettipally UK, Barton CW, Feldman MD, Hoffman JL, et al. A computational approach to early sepsis detection. Comput Biol Med. 1 juill 2016;74:69‑73. Collins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD + AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 16 avr 2024;385:e078378. Richard J, Beydon L, Cantagrel S, Cuvelier A, Fauroux B, Garo B, et al. Sevrage de la ventilation mécanique (à l’exclusion du nouveau-né et du réveil d’anesthésie). Réanimation. déc 2001;10(8):699‑705. Guidelines for the Management of Adults with Hospital-acquired, Ventilator-associated, and Healthcare-associated Pneumonia. Am J Respir Crit Care Med. 15 févr 2005;171(4):388‑416. Leone M, Bouadma L, Bouhemad B, Brissaud O, Dauger S, Gibot S, et al. Pneumonies associées aux soins de réanimation. Anesth Réanimation. sept 2018;4(5):421‑41. Seymour CW, Liu VX, Iwashyna TJ, Brunkhorst FM, Rea TD, Scherag A, et al. Assessment of Clinical Criteria for Sepsis: For the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 23 févr 2016;315(8):762‑74. Lemaıtre G, Nogueira F. Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning. Belhaouari SB, Islam A, Kassoul K, Al-Fuqaha A, Bouzerdoum A. Oversampling techniques for imbalanced data in regression. Expert Syst Appl. 15 oct 2024;252:124118. Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Comput. 15 nov 1997;9(8):1735‑80. Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. In: Proceedings of the 23rd international conference on Machine learning [Internet]. New York, NY, USA: Association for Computing Machinery; 2006. p. 233‑40. (ICML ’06). Available on: https://doi.org/10.1145/1143844.1143874 Sundararajan M, Taly A, Yan Q. Axiomatic Attribution for Deep Networks [Internet]. arXiv; 2017. Available on: http://arxiv.org/abs/1703.01365 Shrikumar A, Greenside P, Shcherbina A, Kundaje A. Not Just a Black Box: Learning Important Features Through Propagating Activation Differences [Internet]. arXiv; 2017. Available on: http://arxiv.org/abs/1605.01713 Liang Y, Zhu C, Tian C, Lin Q, Li Z, Li Z, et al. Early prediction of ventilator-associated pneumonia in critical care patients: a machine learning model. BMC Pulm Med. 25 juin 2022;22(1):250. Gattarello S, Lagunes L, Vidaur L, Solé-Violán J, Zaragoza R, Vallés J, et al. Improvement of antibiotic therapy and ICU survival in severe non-pneumococcal community-acquired pneumonia: a matched case–control study. Crit Care. déc 2015;19(1):335. Timsit JF, Bassetti M, Cremer O, Daikos G, De Waele J, Kallil A, et al. Rationalizing antimicrobial therapy in the ICU: a narrative review. Intensive Care Med. févr 2019;45(2):172‑89. Kumar A, Roberts D, Wood KE, Light B, Parrillo JE, Sharma S, et al. Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human septic shock. Crit Care Med. juin 2006;34(6):1589‑96. Moazemi S, Vahdati S, Li J, Kalkhoff S, Castano LJV, Dewitz B, et al. Artificial intelligence for clinical decision support for monitoring patients in cardiovascular ICUs: A systematic review. Front Med. 2023;10:1109411. Hastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning [Internet]. New York, NY: Springer New York; 2009. (Springer Series in Statistics). Available on: http://link.springer.com/10.1007/978-0-387-84858-7 Dormann CF, Elith J, Bacher S, Buchmann C, Carl G, Carré G, et al. Collinearity: a review of methods to deal with it and a simulation study evaluating their performance. Ecography. 2013;36(1):27‑46. Wolfensberger A, Meier AH, Kuster SP, Mehra T, Meier MT, Sax H. Should International Classification of Diseases codes be used to survey hospital-acquired pneumonia? J Hosp Infect. mai 2018;99(1):81‑4. Ioffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift [Internet]. arXiv; 2015. Available on: http://arxiv.org/abs/1502.03167 Santurkar S, Tsipras D, Ilyas A, Madry A. How Does Batch Normalization Help Optimization? [Internet]. arXiv; 2019. Available on: http://arxiv.org/abs/1805.11604 Krogh A, Hertz J. A Simple Weight Decay Can Improve Generalization. In: Advances in Neural Information Processing Systems [Internet]. Morgan-Kaufmann; 1991. Available on: https://proceedings.neurips.cc/paper/1991/hash/8eefcfdf5990e441f0fb6f3fad709e21-Abstract.html Sutskever I, Martens J, Dahl G. On the importance of initialization and momentum in deep learning. Additional Declarations No competing interests reported. Supplementary Files Additionalfile1Metricsdefinition.docx Additionalfile2VAPdiagnosisalgorithm.docx Additionalfile3Dataprocessingpipeline.docx Additionalfile4Dataoversampling.docx Additionalfile5Modeldeveloppement.docx Additionalfile6PREDICTtrainingcurvesandconfusionmatrix.docx Additionalfile7Algorithmmetricscomparaison.docx Additionalfile8TRIPODAIchecklist.docx VisualAbstract.png Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6151630","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":425199684,"identity":"056f2138-3f47-4259-87fc-bc18135d1fa5","order_by":0,"name":"Geoffray AGARD","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABBElEQVRIiWNgGAWjYJCCDxVwJpsNkOABYhu8OhhnnEFoSYNqSSNey2HCWuTdmw82HNxjl9jAfvzi54qy8/LmDLxHNzAk3MOpxfDMscSGA8+SExt4coolz5y7bbizgS/tBkNCMW4tM3LMH384wJzYwJCTINnYdptxw/03ZjcYfyTg1jL/jWHDgQP1iQ38b5J/Nrads99wgMcMaAtuLfISPCAthxMbJNKPAW05kEhQiwFPGtAvB44bt0m8YbNsOJecvLMBqCUBny3thw8CtVTL9vOnP77ZUGZnu50BqOUDPlsOQGjHNgYeA4gIiMCtAWhLA4S2Z2Bgf4DQMgpGwSgYBaMACQAAuBBgQk/0srwAAAAASUVORK5CYII=","orcid":"","institution":"Hôpital Nord","correspondingAuthor":true,"prefix":"","firstName":"Geoffray","middleName":"","lastName":"AGARD","suffix":""},{"id":425199685,"identity":"8bf7959d-3fe5-4b43-9575-e9db88578f09","order_by":1,"name":"Christophe ROMAN","email":"","orcid":"","institution":"Aix-Marseille Université","correspondingAuthor":false,"prefix":"","firstName":"Christophe","middleName":"","lastName":"ROMAN","suffix":""},{"id":425199686,"identity":"7a639b17-660d-465f-bf9d-7a650e333d45","order_by":2,"name":"Christophe GUERVILLY","email":"","orcid":"","institution":"Hôpital Nord","correspondingAuthor":false,"prefix":"","firstName":"Christophe","middleName":"","lastName":"GUERVILLY","suffix":""},{"id":425199687,"identity":"7dfd5c4f-d11f-46b5-a481-fa78042e5e7a","order_by":3,"name":"Jean Marie FOREL","email":"","orcid":"","institution":"Hôpital Nord","correspondingAuthor":false,"prefix":"","firstName":"Jean","middleName":"Marie","lastName":"FOREL","suffix":""},{"id":425199688,"identity":"cc8f0117-3e19-4db4-bc16-3d993f4522e7","order_by":4,"name":"Veronica ORLEANS","email":"","orcid":"","institution":"Hôpital de la Conception","correspondingAuthor":false,"prefix":"","firstName":"Veronica","middleName":"","lastName":"ORLEANS","suffix":""},{"id":425199689,"identity":"d2a601a0-5b81-48f3-bfc6-22885a255610","order_by":5,"name":"Damien BARRAU","email":"","orcid":"","institution":"Hôpital Nord","correspondingAuthor":false,"prefix":"","firstName":"Damien","middleName":"","lastName":"BARRAU","suffix":""},{"id":425199690,"identity":"6ae4457d-aa25-412d-9021-79ad708aa9db","order_by":6,"name":"Pascal AUQUIER","email":"","orcid":"","institution":"Aix-Marseille Université","correspondingAuthor":false,"prefix":"","firstName":"Pascal","middleName":"","lastName":"AUQUIER","suffix":""},{"id":425199691,"identity":"ed04ef40-e55f-4137-8486-1494e9dd8d59","order_by":7,"name":"Mustapha OULADSINE","email":"","orcid":"","institution":"Aix-Marseille Université","correspondingAuthor":false,"prefix":"","firstName":"Mustapha","middleName":"","lastName":"OULADSINE","suffix":""},{"id":425199692,"identity":"cab10375-cf50-45e1-897a-67408b174e32","order_by":8,"name":"Laurent BOYER","email":"","orcid":"","institution":"Aix-Marseille Université","correspondingAuthor":false,"prefix":"","firstName":"Laurent","middleName":"","lastName":"BOYER","suffix":""},{"id":425199693,"identity":"7dce48c4-0239-4b74-ad24-f6644bf14a44","order_by":9,"name":"Sami HRAIECH","email":"","orcid":"","institution":"Hôpital Nord","correspondingAuthor":false,"prefix":"","firstName":"Sami","middleName":"","lastName":"HRAIECH","suffix":""}],"badges":[],"createdAt":"2025-03-04 07:08:06","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6151630/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6151630/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":78229356,"identity":"9dfc5df6-6122-435a-977f-d626a6180242","added_by":"auto","created_at":"2025-03-11 07:20:37","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":105886,"visible":true,"origin":"","legend":"\u003cp\u003eFlow chart\u003c/p\u003e","description":"","filename":"Figure1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/ccba086f446f3d0d34f0837a.jpeg"},{"id":78230757,"identity":"7f6ff241-3a18-470c-bbf1-d9283f981c58","added_by":"auto","created_at":"2025-03-11 07:28:37","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":103185,"visible":true,"origin":"","legend":"\u003cp\u003ePatient vital signs (top) and the predicted probability of VAP 12h in the future (bottom) over time for a patient with VAP during the stay. The orange shaded area represents the 24-hour observation window, the green shaded area marks the 12-hour prediction window. Red dashed lines indicate the clinical VAP diagnosis times. The algorithm predict a future VAP occurrence when the probability cross the threshold (orange line).\u003c/p\u003e","description":"","filename":"Figure2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/e792da561947e9252e640c0e.jpeg"},{"id":78229376,"identity":"a6c94d72-e39c-4f07-aa6d-43822f66be0c","added_by":"auto","created_at":"2025-03-11 07:20:38","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":193922,"visible":true,"origin":"","legend":"\u003cp\u003eROC and Precision-recall curves for all algorithms. \u003cstrong\u003eA)\u003c/strong\u003e On left, Receiver Operating Characteristic (ROC) curves with corresponding Area under Curve (AUROC) for all algorithms with VAP prediction 6 h in the future. On right, Precision-recall curves with corresponding Area under Curve (AUPRC) for all algorithms with VAP prediction 6 h in the future. \u003cstrong\u003eB)\u003c/strong\u003e VAP prediction 12 h in the future. \u003cstrong\u003eC)\u003c/strong\u003e VAP prediction 24 h in the future.\u003c/p\u003e","description":"","filename":"Figure3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/fcd3eb0af0950ad3ddd8afc5.jpeg"},{"id":78229358,"identity":"57f108e5-d5a7-4e6c-9aaf-9aec146c96d1","added_by":"auto","created_at":"2025-03-11 07:20:37","extension":"jpeg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":131623,"visible":true,"origin":"","legend":"\u003cp\u003ePREDICT model explainability. \u003cstrong\u003eA)\u003c/strong\u003e Integrated gradient attributions in temporal window scale. Integrated attributions measure the importance of input features for model prediction. High attribution values indicate that the feature had a significant impact on the prediction of the target class. \u003cstrong\u003eB)\u003c/strong\u003eFeature importance in model overall decision.\u003c/p\u003e","description":"","filename":"Figure4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/318188c422d9ee057961feae.jpeg"},{"id":78231122,"identity":"bf2d7170-8696-42cb-8d1c-405673450147","added_by":"auto","created_at":"2025-03-11 07:36:38","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1755691,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/cc01824b-9ef0-44c0-8fe0-a2f6e508ea07.pdf"},{"id":78229357,"identity":"6d0e3b60-d7bb-45de-8cc3-7cb596023337","added_by":"auto","created_at":"2025-03-11 07:20:37","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":32850,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile1Metricsdefinition.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/d364d69f44d57d95af48228c.docx"},{"id":78229360,"identity":"8667f250-10d6-4746-8ad1-4dea15969b8d","added_by":"auto","created_at":"2025-03-11 07:20:37","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":141940,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile2VAPdiagnosisalgorithm.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/4c886f884873a805d5e46704.docx"},{"id":78229369,"identity":"2c4aa2d4-6c6c-4db3-8a41-df9fa461b3d6","added_by":"auto","created_at":"2025-03-11 07:20:38","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":342648,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile3Dataprocessingpipeline.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/d7ae11a5b91366b45d105818.docx"},{"id":78230761,"identity":"4ed7318b-3514-491c-a17e-874c5217e4f2","added_by":"auto","created_at":"2025-03-11 07:28:38","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":391756,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile4Dataoversampling.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/51a1ed27a253658773693bd2.docx"},{"id":78229372,"identity":"0abd2d99-3890-4f6e-9360-f34ae4748b8e","added_by":"auto","created_at":"2025-03-11 07:20:38","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":215733,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile5Modeldeveloppement.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/4080e8fa3c271afa56680a60.docx"},{"id":78229362,"identity":"57543c3b-708f-464e-b1cc-33b1b6453c86","added_by":"auto","created_at":"2025-03-11 07:20:38","extension":"docx","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":434997,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile6PREDICTtrainingcurvesandconfusionmatrix.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/81ca50d4611aea9d103b5d70.docx"},{"id":78229378,"identity":"2c3f3826-7f01-4ad9-9881-f830e3528b99","added_by":"auto","created_at":"2025-03-11 07:20:41","extension":"docx","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":34142,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile7Algorithmmetricscomparaison.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/5761a1113a078d866b2b517a.docx"},{"id":78229373,"identity":"e9f16206-a370-4471-9413-20d05b00c38d","added_by":"auto","created_at":"2025-03-11 07:20:38","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":44651,"visible":true,"origin":"","legend":"","description":"","filename":"Additionalfile8TRIPODAIchecklist.docx","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/76ca5628a5618bc60f583bac.docx"},{"id":78230762,"identity":"5f01bf15-e86e-4d1b-855c-8a4f7aed26c3","added_by":"auto","created_at":"2025-03-11 07:28:38","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":903126,"visible":true,"origin":"","legend":"","description":"","filename":"VisualAbstract.png","url":"https://assets-eu.researchsquare.com/files/rs-6151630/v1/834f71e728b68b17c57d0597.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"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)","fulltext":[{"header":"Background","content":"\u003cp\u003eVentilator-associated pneumonia (VAP) is one of the most common complication in intensive care units (ICU). Five to 40% of patients under invasive mechanical ventilation (MV) are likely to develop at least one VAP during their stay. VAP increase duration of mechanical ventilation (MV) and ICU length of stay, leading to a potential increase in mortality of up to 50%\u003csup\u003e1\u003c/sup\u003e. Diagnosis of VAP is based on a combination of three criteria: clinical suspicion, apparition or worsening of radiological infiltrates and positivity of a respiratory tract culture \u003csup\u003e2\u003c/sup\u003e. Clinical suspicion, the key element of diagnosis, has very low sensitivity and specificity \u003csup\u003e3,4\u003c/sup\u003e. In addition, VAP diagnosis is retrospective by nature, linked to the positivity of respiratory samples. Current recommendations therefore call for respiratory sampling and probabilistic antibiotic therapy while awaiting microbiological results to confirm or invalidate the diagnosis \u003csup\u003e1,2\u003c/sup\u003e. Although this approach limits the low sensitivity of markers of clinical suspicion, it exposes the patient to increased consumption of broad-spectrum antibiotics in the ICU. Early diagnosis of VAP is a critical priority, as delayed or missed diagnosis can lead to prolonged infections and worse patient outcomes, while premature empirical treatments increase the risks of antibiotic resistance and adverse drug effects. The ability to anticipate VAP onset before clinical suspicion arises could enable targeted interventions, reducing unnecessary antibiotic exposure and improving survival rates.\u003c/p\u003e \u003cp\u003eIn order to improve the earliness of VAP diagnosis, it is important to compute and merge information contained in various indicators. The recent advent of artificial intelligence (AI), a set of technologies designed to simulate human cognitive abilities, could make it possible to improve the earliness of VAP diagnosis. To date, literature on systems for VAP prediction using AI is scarce. Most of these systems use machine learning (ML) algorithms, a branch of AI that uses statistical algorithms to learn from data. Samadani et al. used a ML model for VAP prediction within 24 hours with demographic data, vital constants, biology and mechanical ventilation data for training \u003csup\u003e5\u003c/sup\u003e. Although AUROC (75.6%) for their algorithm appeared better than CPIS score, sensitivity (68%) and specificity (67%) remained unsatisfactory. Meanwhile, Deep learning (DL), a specialized branch of machine learning (ML), uses advanced structures called deep neural networks to analyze and interpret data\u003csup\u003e6,7\u003c/sup\u003e. Unlike traditional methods that often require manual selection of relevant features, DL algorithms can automatically identify and learn important patterns directly from raw data. Additionally, DL is particularly well-suited for processing time-based information, as it can recognize long-term trends and complex sequences in data, making it an ideal tool for tasks that involve changes over time.\u003csup\u003e8\u003c/sup\u003e. There is currently no work dealing with the application of deep learning to the prediction of VAPs in the ICU.\u003c/p\u003e \u003cp\u003eThe objective of this study was to develop and validate PREDICT (\u003cem\u003ePneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology\u003c/em\u003e), a deep learning algorithm designed to facilitate the early diagnosis of VAP in ICU patients who have been mechanically ventilated for more than 48 hours. Our main hypothesis was that variations in vital signs, in particular respiratory rate and SpO2, could enable a DL tool to early detect VAP occurrence. Additionally, the study sought to demonstrate that deep learning outperforms traditional machine learning methods in accurately predicting VAP.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cstrong\u003eStudy design\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis was a retrospective cohort study conducted using the publicly available MIMIC-IV (Medical Information Mart for Intensive Care IV) database version 2.2. MIMIC-IV contains de-identified information from ICU admissions at the Beth Israel Deaconess Medical Center between 2008 and 2019 in United-States\u003csup\u003e9\u003c/sup\u003e. This database provides a rich dataset including demographic details, vital signs, laboratory results, and treatment information. In this work, to simplify the implementation of a VAP prediction tool in ICU, we chose to focus solely on vital signs data, which can be directly obtained from monitoring systems. Notably, similar approaches have shown promising results in developing sepsis prediction algorithms using this limited data set\u003csup\u003e10,11\u003c/sup\u003e. The study design followed the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD + AI) guidelines\u003csup\u003e12\u003c/sup\u003e (\u003cem\u003eadditional file n\u0026deg;8\u003c/em\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe primary objective was to develop and validate a deep learning algorithm for the early prediction of VAP at 6, 12 and 24 hours in the future using vital signs data.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePatient Population\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePatients were included in the study if they :\u0026nbsp;\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eWere aged 18 years or older\u003c/li\u003e\n \u003cli\u003e\u0026nbsp;Required invasive MV for more than 48 hours\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eHad complete records of vital signs during the study period.\u0026nbsp;\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eExclusion criteria were defined as :\u0026nbsp;\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eThe presence of community-acquired pneumonia (as identified by ICD-10 codes or documentation of respiratory infection prior to ICU admission)\u003c/li\u003e\n \u003cli\u003eMV duration shorter than 48 hours\u0026nbsp;\u003c/li\u003e\n \u003cli\u003eIncomplete or missing key variables required for modeling\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe VAP risk period was defined as starting 48 hours after the initiation of MV and continuing until 72 hours post-extubation. This window was chosen to capture nosocomial pneumonia cases while excluding early-onset pneumonia likely acquired before ICU admission.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Collection\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFor the purposes of this study, a MV episode was defined as any continuous period during which a patient received invasive mechanical ventilation, either through an endotracheal tube or a tracheostomy. The start of an MV episode was marked by the initiation of invasive ventilation, and the end was defined as the moment the patient was extubated or transitioned to non-invasive ventilation for more than 48 hours\u003csup\u003e13\u003c/sup\u003e. Episodes with interruptions of less than 24 hours were considered part of the same continuous ventilation period to account for temporary weaning or procedural pauses commonly seen in ICU settings. Only episodes lasting more than 48 hours were included in the analysis. These episodes were extracted from the MIMIC-IV records using specific Structured Query Language (SQL) queries.\u003c/p\u003e\n\u003cp\u003eVital signs extracted for this study included respiratory rate, heart rate, mean arterial pressure, body temperature, and oxygen pulsed saturation (SpO2). These variables were also extracted from the MIMIC-IV database\u0026apos;s time-series records of patient monitoring with SQL queries. Additional demographic and clinical variables, such as age, gender, Sepsis-related Organ Failure Assessment (SOFA) and Simplified Acute Physiology Score (SAPS-II) at ICU admission, were extracted from corresponding patient records. ICD-10 codes were used to classify and exclude patients with community-acquired pneumonia and to identify relevant comorbidities, such as chronic obstructive pulmonary disease (COPD), ischemic heart disease, diabetes mellitus, chronic kidney insufficiency, and active cancer. Data on ICU admission sources, ventilation type (invasive or tracheostomy), and duration of MV were obtained from structured database fields. Ventilation Free Days (VFDs) D28 refers to the number of days within the first 28 days after the start of MV during which a patient is alive and not dependent on mechanical ventilation.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eOutcomes\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe primary outcome of this study was the accurate early prediction of VAP within predefined time windows (6, 12, and 24 hours) following the observation period of 24 hours. Performance was measured using metrics such as the Area Under the Precision-Recall Curve (AUPRC), sensitivity, positive predictive value (PPV), specificity, and negative predictive value (NPV) which are defined in additional file.\u003c/p\u003e\n\u003cp\u003eSecondary outcomes included a comparison of the model performance, assessing the superiority of the deep learning (DL) algorithm (PREDICT) over traditional ML models, including Random Forest, XGBoost, and Logistic Regression, in terms of predictive accuracy and robustness across all prediction horizons.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAnnotation of VAP events\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eVAP events were identified based on a standardized methodology using clinical and microbiological criteria, consistent with international guidelines (IDSA/ATS\u003csup\u003e14\u003c/sup\u003e and SFAR/SRLF\u003csup\u003e15\u003c/sup\u003e) and the infection labelling method initially proposed by \u003cem\u003eSeymour et al.\u003c/em\u003e\u003csup\u003e16\u003c/sup\u003e and adopted by \u003cem\u003eSamadani et al.\u003c/em\u003e\u003csup\u003e5\u003c/sup\u003e in a previous work on VAP prediction (\u003cem\u003eAdditional file n\u0026deg;2 -\u0026nbsp;\u003c/em\u003e\u003cem\u003eSupplemental figure 1\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003eFirst, suspected VAP cases were identified by screening MV episodes longer than 48 hours and during the VAP risk period previously defined. During this period, respiratory microbiological cultures obtained through bronchial aspirates, bronchoalveolar lavage (BAL), or tracheal aspirates were reviewed for evidence of infection. A positive culture was required to support a diagnosis of pneumonia. Clinical interventions were then assessed in conjunction with microbiological findings. New antibiotic regimens targeting respiratory pathogens were identified if initiated within a defined temporal window (72 hours before or 24 hours after the collection of the microbiological sample).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAdditionally, episodes of VAP occurring less than 48 hours apart were considered as part of the same infection event and were not treated as independent episodes. The onset of a VAP episode was defined as the earliest timepoint of either the microbiological sample collection or the initiation of new antibiotics.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSelection of comparator patients without VAP\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePatients without ventilator-associated pneumonia (VAP) were included as comparators using a 1:1 matching strategy to ensure robust comparisons and minimize confounding. Potential controls were identified among patients with at least one MV episode larger than 48 hours without meeting VAP diagnostic criteria. For each VAP episode, a ventilation episode from a patient without VAP was matched. Matching between these two episodes was stratified on age, sex, SAPS-II score at ICU admission, and duration of MV. This selection process provided negative examples essential for the machine learning algorithm to distinguish between VAP and non-VAP episodes, improving its specificity.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData preprocessing\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePreprocessing included several steps to ensure the dataset was suitable for training a predictive model (\u003cem\u003eAdditional n\u0026deg;3 -\u0026nbsp;\u003c/em\u003e\u003cem\u003eSupplemental figure 2\u003c/em\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e1) Data resampling and cleaning\u003c/strong\u003e \u003cstrong\u003e:\u003c/strong\u003e Vital signs were resampled at hourly scale to standardize time intervals and reduce measurement errors. Missing values, were handled using linear interpolation to preserve the continuity and integrity of the time-series data.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e2) Normalization\u003c/strong\u003e \u003cstrong\u003e:\u003c/strong\u003e Each vital sign value was standardized by subtracting the mean and dividing by the standard deviation. This normalization step ensured comparability across variables, preventing any single variable with a larger numerical range (e.g., mean arterial pressure) from disproportionately influencing the algorithm.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e3) Temporal windows creation\u003c/strong\u003e : To allow the algorithm to analyze time-dependent patterns in patient data, we divided the continuous flow of vital signs into temporal windows. A temporal window is a defined time segment that contains patient data recorded over a specific period. In this study, each temporal window consisted of 24 hours of continuous vital sign recordings.\u003c/p\u003e\n\u003cp\u003eEach temporal window was composed of two key parts. The observation window represent a 24-hour period during which vital signs were collected each hour. This is the input data used by the algorithm to identify physiological changes. The prediction window is the period following the observation window during which the algorithm predicts the occurrence of VAP. For this study, three prediction windows were defined at 6, 12 and 24 hours after the observation window. Temporal windows were created using a sliding window approach. This means that after constructing a 24-hour observation window, we moved it forward by 1 hour to create the next window. This overlap ensures the algorithm captures granular temporal patterns without missing key changes in vital signs.\u003c/p\u003e\n\u003cp\u003eEach temporal window was labeled based on whether a VAP episode occurred within the associated prediction window. If a VAP episode occurred within the prediction window (e.g., at 4 hours for a 6-hour prediction horizon), this windows was labelled as positive. Otherwise, the time window was labelled as negative. This labeling process allowed the algorithm to differentiate between patterns leading to VAP and those not associated with infection.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e4) Balancing the dataset\u003c/strong\u003e \u003cstrong\u003e:\u003c/strong\u003e Because VAP events were rare in this dataset (less than 1% of temporal windows), the Synthetic Minority Oversampling Technique (SMOTE)\u003csup\u003e17\u003c/sup\u003e was applied (\u003cem\u003eAdditional file n\u0026deg;4 -\u0026nbsp;\u003c/em\u003e\u003cem\u003eSupplemental figure 3\u003c/em\u003e). This method artificially generated synthetic examples of VAP-positive temporal windows while preserving the structure of the original data. It ensured the model was exposed to sufficient positive examples, enhancing its sensitivity and specificity\u003csup\u003e18\u003c/sup\u003e. Details on SMOTE technique are available in additional file.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData splitting for algorithm training and evaluation\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset was divided into three subsets: Training set (60%) used to train the algorithm, validation set (20%) used to fine-tune hyperparameters and prevent overfitting and test set (20%) held back for final evaluation of model performance (\u003cem\u003eAdditional file n\u0026deg;3 -\u0026nbsp;\u003c/em\u003e\u003cem\u003eSupplemental figure 2\u003c/em\u003e). Stratified sampling was employed to ensure that the proportions of VAP-positive and VAP-negative windows were consistent across all subsets. This step minimized bias and ensured the generalizability of the results.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAlgorithm development\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePREDICT model was based on Long Short-Term Memory (LSTM) networks. This type of deep learning architecture is particularly suited for sequential data, as it captures temporal dependencies, such as trends or variations in vital signs over time\u003csup\u003e19\u003c/sup\u003e. The baseline model architecture included 3 LSTM layers, each with 50 cells, followed by a fully connected layer with a sigmoid activation function. A dropout regularization (5\u0026ndash;10%) was applied to prevent overfitting. Output of the model was the VAP probability on the prediction window considered. Training was conducted using a cross-entropy loss function and the Adam optimizer with adaptive learning rates and used a stratified cross-validation method with 5-folds (\u003cem\u003eAdditional file n\u0026deg;5\u003c/em\u003e). Hyperparameters such as the number of LSTM cells, dropout rate, and learning rate were optimized using the validation set. The best configuration was selected based on area under the precision-recall curve (AUPRC), a metric particularly suited for imbalanced datasets\u003csup\u003e20\u003c/sup\u003e (\u003cem\u003eAdditional file n\u0026deg;1\u003c/em\u003e).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eComparator Models\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo evaluate the added value of deep learning, we trained several traditional machine learning models, including Random Forest, XGBoost, and logistic regression. These models were trained and tested using the same dataset and preprocessing pipeline. Their performance was compared to the PREDICT algorithm using metrics such as sensitivity, specificity, precision, and AUPRC. We compared PREDICT\u0026lsquo;s AUPRC with other models AUPRC using a bootstrap resampling approach (n=100 resamples) to determine statistical difference in AUPRC.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eModel Explainability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo enhance clinical interpretability, we applied the Integrated Gradients technique\u003csup\u003e21,22\u003c/sup\u003e to analyze the model\u0026apos;s decision-making process. This approach quantifies the contribution of each input variable to the prediction, enabling us to identify which vital signs were most influential in detecting VAP.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eContinuous variables were summarized as medians with interquartile ranges and compared using the Wilcoxon rank-sum test. Categorical variables were presented as counts and percentages, with differences assessed using Pearson\u0026rsquo;s \u0026chi;\u0026sup2; test or Fisher\u0026rsquo;s exact test. Metrics such as area under the curve of receiver-operating characteristic (AUROC), precision (Predictive Positive Value), recall (Sensitivity), area under the curve of precision-recall curve (AUPRC), and Youden-index (definitions in supplementary) were calculated for each model with 95% confidence intervals using bootstrapping (n=100 iterations). AUPRC was chosen as the primary metric because it better accounts for imbalanced datasets than AUROC\u003csup\u003e20\u003c/sup\u003e. Deep learning models were constructed using \u003cem\u003etensorflow (v 2.15)\u003c/em\u003e and machine learning models were built with \u003cem\u003escikit-learn (v 1.5.1)\u0026nbsp;\u003c/em\u003epackages\u003cem\u003e.\u0026nbsp;\u003c/em\u003eAll analysis have been performed in Python language (\u003cem\u003ev 3.11.9\u003c/em\u003e) with also use of \u003cem\u003epandas\u003c/em\u003e (v2.2.2) and \u003cem\u003enumpy\u003c/em\u003e (v 1.26.4) packages. \u0026nbsp;\u0026nbsp;\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eVAP episodes\u003c/h2\u003e \u003cp\u003eWe identified 38 750 invasive MV episodes with 9 849 sessions greater than 48 hours (25.4%) concerning 7 871 patients in MIMIC-IV dataset between 2008 and 2019 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Our VAP annotation algorithm identified 452 VAP episodes on 397 patients and 404 MV episodes\u0026thinsp;\u0026gt;\u0026thinsp;48 hours (4.1%). During their stay in ICU, 351 patients presented 1 VAP episode, 41 patients two VAP, 1 patient three VAP and 4 patient presented four VAP episodes. Median time from ICU admission to first VAP episode was 6 days (IQR 4\u0026ndash;12) and also 6 days (IQR 3.7\u0026ndash;11.3) from MV initiation to the first VAP episode.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003ePopulation characteristics\u003c/h2\u003e \u003cp\u003eDespite stratification, men were significantly more at risk for VAP during their stay (67 vs 59% - \u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.011). Patients in the VAP group had less ischemic heart disease, chronic kidney insufficiency and active cancer but more COPD patients than no-VAP group. In the population, 20% of patients were admitted with sepsis, 11% with trauma and 8.3% with stroke. The main characteristics of population are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eOutcomes - Model performance\u003c/h2\u003e \u003cp\u003ePREDICT was able to predict VAP six hours before onset with a sensitivity of 89.7%, a predictive positive value of 89.8% and a predictive negative value of 99.7%. The sensitivity and positive predictive value exceeded 85% for all prediction thresholds (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) including 24-hour prediction (Sensitivity 85.1% - Specificity 99.2%). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e shows how the probability of VAP onset at 12 hours evolves as a function of the vital signs of a patient who presented several VAP during his ICU stay.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatients characteristics. \u003csup\u003e1\u003c/sup\u003en (%); Median [25%-75%] \u003csup\u003e2\u003c/sup\u003ePearson's Chi-squared test; Wilcoxon rank sum - COPD : Chronic Obstructive Pulmonary Disease \u0026ndash; ICU : Intensive Care Unit\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall \u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;904\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eVAP\u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;452\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo VAP \u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;452\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ep-value\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSex (Male)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e573 (63%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e305 (67%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e268 (59%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.011*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAge\u003c/b\u003e \u003cem\u003e(years)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e64.2 [52.1\u0026ndash;75.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e63.9 [50.7\u0026ndash;74.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e65.1 [52.9\u0026ndash;76.1]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePre-existing Diseases n(%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eHypertension\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e450 (50%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e217 (48%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e233 (52%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eIschemic heart disease\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e306 (34%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e131 (29%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e175 (39%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.002*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eDiabetes mellitus\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e216 (24%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e106 (23%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e110 (24%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eChronic renal failure\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e218 (24%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95 (21%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e123 (27%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.029*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eObstructive sleep apnea\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e145 (16%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e69 (15%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eActive cancer\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e134 (15%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56 (12%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.039*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eCOPD\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e66 (7.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e43 (9.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e23 (5.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.011*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eActive hematological malignancy\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31 (3.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18 (4.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e13 (2.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSource of admission to ICU\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.9\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eEmergency ward\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e665 (74%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e331 (73%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e334 (74%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eMedical ward\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e150 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e74 (16%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e76 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eElective surgery\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e89 (9.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e47 (10%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e42 (9.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSOFA \u0026ndash; admission\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e2.0 [0.0\u0026ndash;4.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.0 [0.0\u0026ndash;4.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2.0 [0.0\u0026ndash;4.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSAPS-II on admission\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e42.0 [32.0\u0026ndash;54.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42.0 [31.0\u0026ndash;53.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e43.0 [34.0\u0026ndash;55.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.053\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTime from ICU admission to initiation of MV\u003c/b\u003e\u003cem\u003e(hours)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e7.2 [1.7\u0026ndash;67.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8.2 [1.8\u0026ndash;73.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6.0 [1.7\u0026ndash;60.4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.4\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eReason for ICU admission\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eSepsis\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e179 (20%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e91 (20%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e88 (19%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eTrauma\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e97 (11%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e73 (16%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e24 (5.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eHemorrhagic or ischemic stroke\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e75 (8.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e44 (9.7%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e31 (6.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eAcute malignancy\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e46 (5.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 (3.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e29 (6.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eARDS\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e40 (4.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e34 (7.5%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e6 (1.3%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003ePneumonia\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e34 (3.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e20 (4.4%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14 (3.1%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eMyocardial infarction\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e26 (2.9%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (1.8%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e18 (4.0%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eComparators Models\u003c/h2\u003e \u003cp\u003eComparatively with other machine learning algorithm we trained, PREDICT offers the best AUPRC with respectively 96.0%, 94.1% and 94.7% for VAP prediction at 6, 12 and 24 hours. The XGBoost and Random Forest algorithms provide viable alternatives for early predictions at 6 h with an AUPRC of 94.1% and 91.7% respectively. However, performance rapidly declines for long-range prediction thresholds until it becomes completely worthless for 24-hour predictions. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the AUPRC and AUROC curves for the PREDICT model compared of concurrent ML algorithms. The comparison of AUPRC between PREDICT and traditional machine learning models revealed a statistically significant difference (\u003cem\u003ep\u0026thinsp;\u0026lt;\u0026thinsp;0.001\u003c/em\u003e) for all comparisons.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePREDICT model performance for different VAP prediction threshold. Best classification threshold was calculated for Sensibility\u0026thinsp;=\u0026thinsp;PPV (Precision\u0026thinsp;=\u0026thinsp;Recall). 95% confidence intervals are available in additional file. AUPRC : Area Under Precision Recall Curve \u0026ndash; PPV : Predictive Positive Value \u0026ndash; NPV : Predictive Negative Value\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eVAP prediction\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eBest\u003c/p\u003e \u003cp\u003ethreshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAUPRC (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSensibility (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSpecificity (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003ePPV (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNPV (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\" morerows=\"2\" rowspan=\"3\"\u003e \u003cp\u003e\u003cb\u003ePREDICT Model\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0,53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e96,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e89,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e99,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e89,8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e99,7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0,52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e85,9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e99,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e85,6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e99,6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24 h\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0,43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e94,7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e85,1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e99,2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e85,0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e99,2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe best structure of the PREDICT algorithm selected after HPO was: 3 hidden layers of 50 LSTM cells with a dropout rate of 10% for VAP predictions at 6 and 12 hours and 5% for predictions at 24 hours (\u003cem\u003eAdditional file n\u0026deg;5\u003c/em\u003e). Training curves and confusion matrix for each prediction threshold are provided in additional files (Additional file n\u0026deg;6) as well as all models performance metrics with 95 confidence intervals (\u003cem\u003eAdditional file n\u0026deg;7\u003c/em\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003ePREDICT model explainability\u003c/h2\u003e \u003cp\u003eThe integrated gradients analysis of the PREDICT model reveals how each feature influences predictions over time (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). On temporal window scale, heart rate had a steady, positive contribution, while mean arterial pressure showed significant early fluctuations. SpO2 had minimal impact, remaining close to zero, whereas temperature shifted from a negative to a positive influence toward the end of sequences. Respiratory rate gained relevance progressively over time. Considering overall scale, the most critical features for prediction were SpO2, temperature, and respiratory rate, with mean arterial pressure and heart rate playing smaller roles.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003ePatients\u0026rsquo; prognosis\u003c/h2\u003e \u003cp\u003ePatients in the VAP group had higher hospital length of stay (25.9 vs 16.3 days \u0026ndash; \u003cem\u003eP\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;0.001) and ICU (18.9 vs 8.4 days \u0026ndash; \u003cem\u003eP\u0026thinsp;\u0026lt;\u003c/em\u003e\u0026thinsp;0.001) with also an increased mortality in ICU. Duration of MV was increased in the VAP group and Ventilation Free Days at day 28 were also significantly lower. Outcomes for population are described in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003ePatient prognosis outcomes \u0026minus;\u0026thinsp;\u003csup\u003e1\u003c/sup\u003en (%); Median [25%-75%] \u003csup\u003e2\u003c/sup\u003ePearson's Chi-squared test; Wilcoxon rank sum test\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOverall \u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;904\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eVAP\u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;452\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eNo VAP \u003csup\u003e1\u003c/sup\u003e, \u003c/p\u003e \u003cp\u003eN\u0026thinsp;=\u0026thinsp;452\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ep-value\u003csup\u003e2\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLength of stay \u0026ndash; Hospital\u003c/b\u003e \u003cem\u003e(days)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e21.4 [12.9\u0026ndash;35.2]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e25.9 [17.1\u0026ndash;39.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e16.3 [9.7\u0026ndash;28.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLength of stay \u0026ndash; ICU\u003c/b\u003e \u003cem\u003e(days)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.0 [7.8\u0026ndash;22.9]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18.9 [13.0\u0026ndash;30.0]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e8.4 [5.3\u0026ndash;15.7]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTime from ICU admission to death\u003c/b\u003e \u003cem\u003e(days)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38.0 [13.5-117.5]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e42.5 [18.3-103.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e30.1 [7.7-147.8]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.007*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDuration of mechanical ventilation\u003c/b\u003e \u003cem\u003e(days)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e8.4 [4.0-15.4]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13.6 [8.4\u0026ndash;21.7]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e4.4 [3.0-8.3]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVentilation free days D28\u003c/b\u003e \u003cem\u003e(days)\u003c/em\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e14.5 [0.0-22.6]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e9.1 [0.0-17.6]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e20.9 [0.0-24.5]\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eICU mortality\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e188 (21%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e110 (24%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e78 (17%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003e0.009*\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eIn-hospital mortality\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e261 (29%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e139 (31%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e122 (27%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we developed and validated PREDICT, a new innovative deep learning model for the early detection of VAP at 6, 12 and 24 hours in the future. PREDICT demonstrated excellent diagnostic performance, with an AUROC of 99% for all prediction thresholds. Sensitivity (Se) and positive predictive value (PPV) ranged from 85.1\u0026ndash;89.7%, while the AUPRC exceeded 94%, indicating high predictive capability for both the presence and absence of VAP.\u003c/p\u003e \u003cp\u003ePREDICT represents the first application of deep learning to VAP prediction in ICUs. Previous studies using machine learning algorithms achieved lower performance. \u003cem\u003eLiang et al.\u003c/em\u003e developed a RandomForest model to predict VAP within 24 hours of intubation using around thirty demographic, biological, and physiological variables\u003csup\u003e23\u003c/sup\u003e. While this model showed satisfactory performance (AUROC 84%, Se 74%, specificity 71%), its reliance on diverse and inconsistent data sources limited its ICU applicability. Similarly, \u003cem\u003eSamadani et al.\u003c/em\u003e used an XGBoost model, incorporating a wide range of variables, to predict VAP within 24 hours\u003csup\u003e5\u003c/sup\u003e. Although encouraging (AUROC 76%, AUPRC 75%), this model had low sensitivity (67.5%) and PPV (68.5%), restricting its clinical utility. In contrast, PREDICT outperformed these models across all prediction horizons and maintained stable predictive performance even for longer intervals, highlighting the advantage of deep learning in modeling complex relationships without extensive manual feature engineering\u003csup\u003e7\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo ensure comprehensive evaluation, we compared PREDICT with Random Forest and XGBoost models. While these traditional algorithms closely matched PREDICT\u0026rsquo;s performance for short prediction tasks (6h), with AUROCs of 99% and AUPRCs between 92% and 94%, their accuracy declined for longer prediction horizons, unlike PREDICT, which remained robust. The stability of PREDICT\u0026rsquo;s predictions illustrates the strength of deep learning in extracting subtle patterns from time-series data, offering more reliable and generalizable predictions.\u003c/p\u003e \u003cp\u003ePREDICT offers physicians a valuable tool for the early identification of VAP, allowing intervention 6 to 24 hours before clinical suspicion. This early detection can enable timely initiation of targeted antibiotics, potentially reducing the progression to severe infections, organ dysfunction, and prolonged ICU stays\u003csup\u003e24\u003c/sup\u003e. Gaining 6 to 24 hours is crucial in critically ill patients, where early intervention can prevent progression to severe infections, organ dysfunction, and prolonged ICU stays\u003csup\u003e26\u003c/sup\u003e. PREDICT high negative predictive value can also help in optimizing antimicrobial stewardship by avoiding unnecessary broad-spectrum antibiotics, thus reducing antibiotic resistance risks\u003csup\u003e25\u003c/sup\u003e. In such cases, it can prompt the physician to investigate for another origin of sepsis than lungs. The fact that PREDICT only uses vital constants available on ICU monitors gives us hope that in the future, such technology can be integrated directly into these monitors, making it much easier for clinicians to use.\u003c/p\u003e \u003cp\u003eA key barrier to deep learning adoption in medicine has been the lack of explainability. Trust in AI systems depends on understanding the reasoning behind predictions specifically in medical field\u003csup\u003e27\u003c/sup\u003e. Recent advancements, such as the integrated gradient technique\u003csup\u003e21\u003c/sup\u003e, have improved interpretability. By applying this method, we highlighted the contribution of individual variables to predictions, identifying temperature and respiratory rate changes as key indicators of VAP. These findings align with earlier studies showing the predictive value of these variables in machine learning models\u003csup\u003e5,23\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eIt may seem paradoxical that PREDICT, despite using fewer variables, outperforms models incorporating dozens of features. Complex models often suffer from overfitting\u003csup\u003e28\u003c/sup\u003e, capturing noise along with trends, which hampers their generalization to new data. Moreover, multicollinearity between variables can dilute the predictive signal, further reducing performance\u003csup\u003e29\u003c/sup\u003e. Similar observations were made \u003cem\u003eby Desautels et al.\u003c/em\u003e who achieved promising sepsis prediction results (Se 80%, Sp 80%, AUROC 88%) using only vital signs, SpO₂, Glasgow score, and age, demonstrating that simpler models can sometimes yield superior results\u003csup\u003e10\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eOur work also stands out due to its annotation technique, adapted from \u003cem\u003eSeymour et al\u003c/em\u003e and previously used by \u003cem\u003eSamadani et al.\u003c/em\u003e which adheres to international recommendations\u003csup\u003e5,14\u0026ndash;16\u003c/sup\u003e. This method ensures microbiological accuracy and resolves key challenges in annotating VAP, such as variability in definitions. ICD codes, often used for annotation, lack both sensitivity (59%) and specificity (PPV 27\u0026ndash;42%), making them unsuitable for precise classification\u003csup\u003e30\u003c/sup\u003e. By employing time windows to pinpoint VAP diagnosis, we avoided the limitations of retrospective ICD coding, which fails to capture the exact timing of clinical events.\u003c/p\u003e \u003cp\u003eA primary limitation of this study lies in its monocentric design, relying exclusively on the MIMIC-IV database, which may limit the generalizability of PREDICT to diverse ICU settings. Additionally, the data available in MIMIC-IV may not fully capture the complexity of clinical decisions. To overcome these challenges, future work should include retraining the model with data from multiple hospitals, incorporating diverse patient populations and clinical practices. Finally, while PREDICT has shown potential to improve early VAP detection, its clinical benefits must be confirmed through prospective randomized controlled trials to assess its impact on clinical outcomes, such as mortality, duration of ventilation, and antibiotic exposure.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this work we developed and internally validated a new deep-learning model for VAP prediction using the MIMIC-IV cohort. Using a small number of clinical variables, our PREDICT algorithm demonstrated a high capacity for detecting both short- and long-term VAP. The superiority of the deep-learning approach was confirmed by comparison with other machine learning algorithms trained in the same way. An external validation of the algorithm and its implementation in a prospective trial strategy should be considered in future work.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003e- AI : Artificial Intelligence\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- ARDS : Acute Respiratory Distress Syndrome\u003c/p\u003e\n\u003cp\u003e- AUPRC : Area Under the Precision-Recall Curve\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- AUROC : Area Under the Receiver Operating Characteristic Curve\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- BAL : Bronchoalveolar Lavage\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- CDC : Centers for Disease Control and Prevention\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- COPD : Chronic Obstructive Pulmonary Disease\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- CPIS : Clinical Pulmonary Infection Score\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- HPO : Hyperparameter Optimization\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- ICD-10 : International Classification of Diseases, 10th Revision\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- ICU \u0026ndash; Intensive Care Unit\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- IDSA/ATS : Infectious Diseases Society of America / American Thoracic Society\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- IQR : Interquartile Range\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- LSTM : Long Short-Term Memory\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- ML : Machine Learning\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- MIMIC-IV \u0026ndash; Medical Information Mart for Intensive Care IV \u0026nbsp;\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- MV : Mechanical Ventilation\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- NPV : Negative Predictive Value\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- PAVM : Pneumonie Associ\u0026eacute;e \u0026agrave; la Ventilation M\u0026eacute;canique\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- PREDICT : Pneumonia Risk Evaluation and Diagnostic Intelligence via Computational Technology\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- PPV : Positive Predictive Value\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- ROC : Receiver Operating Characteristic\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- SAP-II : Simplified Acute Physiology Score II\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- SFAR/SRLF : Soci\u0026eacute;t\u0026eacute; Fran\u0026ccedil;aise d\u0026rsquo;Anesth\u0026eacute;sie et de R\u0026eacute;animation / Soci\u0026eacute;t\u0026eacute; de R\u0026eacute;animation de Langue Fran\u0026ccedil;aise\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- SMOTE : Synthetic Minority Oversampling Technique\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- SOFA : Sepsis-related Organ Failure Assessment\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- SQL : Structured Query Language\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- TRIPOD+AI : Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (with AI guidelines)\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- VAP : Ventilator-Associated Pneumonia\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e- VFD D28 : Ventilation-Free Days at Day 28\u0026nbsp;\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This study was conducted using the publicly available MIMIC-IV database, which contains de-identified health records of ICU patients from the Beth Israel Deaconess Medical Center (BIDMC). The database is maintained by the Massachusetts Institute of Technology (MIT) Laboratory for Computational Physiology.\u003c/p\u003e\n\u003cp\u003eThe use of MIMIC-IV data is approved by the Institutional Review Board (IRB) of BIDMC and follows the ethical guidelines outlined by the Massachusetts Institute of Technology (MIT). As all data in MIMIC-IV are de-identified in compliance with the Health Insurance Portability and Accountability Act (HIPAA), this study was exempt from additional ethical approval.\u003c/p\u003e\n\u003cp\u003eAll researchers accessing MIMIC-IV data must complete the Collaborative Institutional Training Initiative (CITI) program on Human Research Ethics, ensuring compliance with ethical guidelines for research involving human data.\u003c/p\u003e\n\u003cp\u003eSince this study is based on a secondary analysis of fully anonymized data, informed consent from patients was not required, as determined by the BIDMC IRB.\u003c/p\u003e\n\u003cp\u003eEthics approval reference: BIDMC IRB Protocol #2001P001699\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. No identifiable personal data or patient information was used in this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study is publicly available through the MIMIC-IV database, hosted by the Massachusetts Institute of Technology (MIT) at https://physionet.org/content/mimiciv/. Access to MIMIC-IV requires completion of the appropriate data use agreement and certification in research ethics training.\u003c/p\u003e\n\u003cp\u003eCode for this project is available on request from the first author (GA).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cul type=\"disc\"\u003e\n \u003cli\u003e\u003cstrong\u003eGA\u003c/strong\u003e (Geoffray Agard) conceptualized the study, designed the methodology, performed data extraction and analysis, and drafted the manuscript.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eSH\u003c/strong\u003e (Sami Hraiech) contributed to study and methodology design, clinical interpretation of results, and manuscript revision.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eLB\u003c/strong\u003e (Laurent Boyer) provided statistical guidance and contributed to manuscript review.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eCR\u003c/strong\u003e (Christophe Roman) and \u003cstrong\u003eMO\u003c/strong\u003e (Mustapha Ouladsine) provided expertise in machine learning and deep learning methodologies.\u003c/li\u003e\n \u003cli\u003e\u003cstrong\u003eJMF\u003c/strong\u003e (Jean-Marie Forel), \u003cstrong\u003eCG\u003c/strong\u003e (Christophe Guervilly) \u0026nbsp;and \u003cstrong\u003eDB\u003c/strong\u003e (Damien Barrau) contributed to manuscript revision.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eAll authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank the PhysioNet team and the Beth Israel Deaconess Medical Center for maintaining and providing access to the MIMIC-IV database. We also acknowledge the support of CERESS (Centre d\u0026apos;Etudes et de Recherches sur les Services de Sant\u0026eacute; et qualit\u0026eacute; de vie EA 3279), Aix-Marseille Universit\u0026eacute; and the AP-HM Intensive Care Unit for providing an academic environment conducive to research in artificial intelligence and critical care medicine. Many thanks also to the Dr Fran\u0026ccedil;ois Antonini and Mr Dorian GROUSSET for their help and support on this project.\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003ePapazian L, Klompas M, Luyt CE. Ventilator-associated pneumonia in adults: a narrative review. Intensive Care Med. mai 2020;46(5):888‑906.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCDC. Ventilator-associated Pneumonia Basics [Internet]. Ventilator-associated Pneumonia (VAP). 2024. Available on: https://www.cdc.gov/ventilator-associated-pneumonia/about/index.html\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRam\u0026iacute;rez-Estrada S, Lagunes L, Pe\u0026ntilde;a-L\u0026oacute;pez Y, Vahedian-Azimi A, Nseir S, Arvaniti K, et al. Assessing predictive accuracy for outcomes of ventilator-associated events in an international cohort: the EUVAE study. Intensive Care Med. 1 ao\u0026ucirc;t 2018;44(8):1212‑20.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJansson M, Ala-Kokko T, Ahvenj\u0026auml;rvi L, Karhu J, Ohtonen P, Syrj\u0026auml;l\u0026auml; H. What Is the Applicability of a Novel Surveillance Concept of Ventilator-Associated Events? Infect Control Hosp Epidemiol. ao\u0026ucirc;t 2017;38(8):983‑8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSamadani A, Wang T, van Zon K, Celi LA. VAP risk index: Early prediction and hospital phenotyping of ventilator-associated pneumonia using machine learning. Artif Intell Med. d\u0026eacute;c 2023;146:102715.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeCun Y, Bengio Y, Hinton G. Deep learning. Nature. mai 2015;521(7553):436‑44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchmidhuber J. Deep learning in neural networks: An overview. Neural Netw. 1 janv 2015;61:85‑117.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLim B, Zohren S. Time-series forecasting with deep learning: a survey. Philos Trans R Soc Math Phys Eng Sci. 15 f\u0026eacute;vr 2021;379(2194):20200209.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data. 3 janv 2023;10(1):1.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDesautels T, Calvert J, Hoffman J, Jay M, Kerem Y, Shieh L, et al. Prediction of Sepsis in the Intensive Care Unit With Minimal Electronic Health Record Data: A Machine Learning Approach. JMIR Med Inform. 30 sept 2016;4(3):e28.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalvert JS, Price DA, Chettipally UK, Barton CW, Feldman MD, Hoffman JL, et al. A computational approach to early sepsis detection. Comput Biol Med. 1 juill 2016;74:69‑73.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCollins GS, Moons KGM, Dhiman P, Riley RD, Beam AL, Van Calster B, et al. TRIPOD\u0026thinsp;+\u0026thinsp;AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ. 16 avr 2024;385:e078378.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRichard J, Beydon L, Cantagrel S, Cuvelier A, Fauroux B, Garo B, et al. Sevrage de la ventilation m\u0026eacute;canique (\u0026agrave; l\u0026rsquo;exclusion du nouveau-n\u0026eacute; et du r\u0026eacute;veil d\u0026rsquo;anesth\u0026eacute;sie). R\u0026eacute;animation. d\u0026eacute;c 2001;10(8):699‑705.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGuidelines for the Management of Adults with Hospital-acquired, Ventilator-associated, and Healthcare-associated Pneumonia. Am J Respir Crit Care Med. 15 f\u0026eacute;vr 2005;171(4):388‑416.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLeone M, Bouadma L, Bouhemad B, Brissaud O, Dauger S, Gibot S, et al. Pneumonies associ\u0026eacute;es aux soins de r\u0026eacute;animation. Anesth R\u0026eacute;animation. sept 2018;4(5):421‑41.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeymour CW, Liu VX, Iwashyna TJ, Brunkhorst FM, Rea TD, Scherag A, et al. Assessment of Clinical Criteria for Sepsis: For the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3). JAMA. 23 f\u0026eacute;vr 2016;315(8):762‑74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLemaıtre G, Nogueira F. Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBelhaouari SB, Islam A, Kassoul K, Al-Fuqaha A, Bouzerdoum A. Oversampling techniques for imbalanced data in regression. Expert Syst Appl. 15 oct 2024;252:124118.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Comput. 15 nov 1997;9(8):1735‑80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavis J, Goadrich M. The relationship between Precision-Recall and ROC curves. In: Proceedings of the 23rd international conference on Machine learning [Internet]. New York, NY, USA: Association for Computing Machinery; 2006. p. 233‑40. (ICML \u0026rsquo;06). Available on: https://doi.org/10.1145/1143844.1143874\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSundararajan M, Taly A, Yan Q. Axiomatic Attribution for Deep Networks [Internet]. arXiv; 2017. Available on: http://arxiv.org/abs/1703.01365\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShrikumar A, Greenside P, Shcherbina A, Kundaje A. Not Just a Black Box: Learning Important Features Through Propagating Activation Differences [Internet]. arXiv; 2017. Available on: http://arxiv.org/abs/1605.01713\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiang Y, Zhu C, Tian C, Lin Q, Li Z, Li Z, et al. Early prediction of ventilator-associated pneumonia in critical care patients: a machine learning model. BMC Pulm Med. 25 juin 2022;22(1):250.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGattarello S, Lagunes L, Vidaur L, Sol\u0026eacute;-Viol\u0026aacute;n J, Zaragoza R, Vall\u0026eacute;s J, et al. Improvement of antibiotic therapy and ICU survival in severe non-pneumococcal community-acquired pneumonia: a matched case\u0026ndash;control study. Crit Care. d\u0026eacute;c 2015;19(1):335.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTimsit JF, Bassetti M, Cremer O, Daikos G, De Waele J, Kallil A, et al. Rationalizing antimicrobial therapy in the ICU: a narrative review. Intensive Care Med. f\u0026eacute;vr 2019;45(2):172‑89.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKumar A, Roberts D, Wood KE, Light B, Parrillo JE, Sharma S, et al. Duration of hypotension before initiation of effective antimicrobial therapy is the critical determinant of survival in human septic shock. Crit Care Med. juin 2006;34(6):1589‑96.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoazemi S, Vahdati S, Li J, Kalkhoff S, Castano LJV, Dewitz B, et al. Artificial intelligence for clinical decision support for monitoring patients in cardiovascular ICUs: A systematic review. Front Med. 2023;10:1109411.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHastie T, Tibshirani R, Friedman J. The Elements of Statistical Learning [Internet]. New York, NY: Springer New York; 2009. (Springer Series in Statistics). Available on: http://link.springer.com/10.1007/978-0-387-84858-7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDormann CF, Elith J, Bacher S, Buchmann C, Carl G, Carr\u0026eacute; G, et al. Collinearity: a review of methods to deal with it and a simulation study evaluating their performance. Ecography. 2013;36(1):27‑46.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWolfensberger A, Meier AH, Kuster SP, Mehra T, Meier MT, Sax H. Should International Classification of Diseases codes be used to survey hospital-acquired pneumonia? J Hosp Infect. mai 2018;99(1):81‑4.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIoffe S, Szegedy C. Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift [Internet]. arXiv; 2015. Available on: http://arxiv.org/abs/1502.03167\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSanturkar S, Tsipras D, Ilyas A, Madry A. How Does Batch Normalization Help Optimization? [Internet]. arXiv; 2019. Available on: http://arxiv.org/abs/1805.11604\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKrogh A, Hertz J. A Simple Weight Decay Can Improve Generalization. In: Advances in Neural Information Processing Systems [Internet]. Morgan-Kaufmann; 1991. Available on: https://proceedings.neurips.cc/paper/1991/hash/8eefcfdf5990e441f0fb6f3fad709e21-Abstract.html\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSutskever I, Martens J, Dahl G. On the importance of initialization and momentum in deep learning.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Ventilator-associated pneumonia, artificial intelligence, deep learning, predictive modeling, intensive care, machine learning, MIMIC-IV, time-series analysis, Long Short-Term memory","lastPublishedDoi":"10.21203/rs.3.rs-6151630/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6151630/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003cbr\u003e\n \u003c/strong\u003eVentilator-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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003cbr\u003e\n \u003c/strong\u003eA 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults:\u003cbr\u003e\n \u003c/strong\u003ePREDICT 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.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusion:\u003cbr\u003e\n \u003c/strong\u003eThis 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.\u003c/p\u003e","manuscriptTitle":"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)","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-03-11 07:20:32","doi":"10.21203/rs.3.rs-6151630/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8ddf93ab-b49f-496f-acad-7c4953952e78","owner":[],"postedDate":"March 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-03-11T07:20:32+00:00","versionOfRecord":[],"versionCreatedAt":"2025-03-11 07:20:32","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6151630","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6151630","identity":"rs-6151630","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}