Inclusion criteria.

All adult patients having attended the APUH ED between January 1^st, 2021, and December 31^st, 2024, will be screened. The primary outcome label for model development is RAUC + , which is defined as an adult patient with an index ED visit, subsequent admission for gastrointestinal surgery, and the performance of emergency gastrointestinal surgery within 72 hours of the index ED visit (i.e., by analogy with the RAUC-AMIENS trial’s inclusion criteria). Encounters not meeting all of these criteria will be labelled as “RAUC-”. Emergency gastrointestinal surgery is defined as any urgent surgical intervention on the gastrointestinal tract (e.g., appendectomy, cholecystectomy, bowel obstruction surgery, perforation repair, etc.) performed because of an acute condition (i.e., as opposed to elective or scheduled surgery). In practice, these patients are typically admitted via the ED and taken to the operating theatre within hours or days of admission. We shall use admission records and financial logs to track admissions. Thus, the RAUCisable cohort will effectively consist of all ED patients during the study period, with a subclass identified as “RAUC+” (i.e., those who underwent emergency gastrointestinal surgery). The classification models will be developed to detect this class.

Exclusion criteria.

Encounters involving direct transfer from the ED to the operating room or direct admission to the ICU will not be included in the primary modelling cohort. This choice reflects the model’s intended use (i.e., to facilitate the early identification of potentially eligible patients in standard ED pathways. These situations might correspond to a distinct subset of patients who are often recognized clinically at a very early stage, and their inclusion might (at least in part) shift the prediction target away from the present study’s main objective. Patients who have formally objected to the research use of their personal health data will also be excluded. There are no exclusion criteria related to incomplete datasets at this stage; however, missing data will be accounted for during the analysis.

Pseudonymization and confidentiality.

In compliance with data protection regulations and to protect patient privacy, all data will be pseudonymized prior to analysis outside the hospital’s secure network. Each patient’s identifying information (e.g., name, surname, exact date of birth, address, etc.) will be stripped out of the dataset. Instead of real identifiers, we shall use a coded unique study ID for each patient: the permanent patient identifier (PPI). The mapping between patient identities (i.e., hospital IDs) and PPIs will be maintained in a separate encrypted file at the hospital and will not be accessible to the analysis team on external platforms. The patient’s unique hospital identifier will be transformed using a combination of a one-way cryptographic hash function (SHA-256) and a secret salt stored at the hospital. For example, a patient with a “123456789” PPI might be hashed (with a secret salt) into an irreversible but reproducible code like “43a3d35bc82f…908cf6”. A patient who is readmitted will have the same PPI and will not be considered to be a new patient. This ensures that no directly identifiable information is present in the analysis dataset but allows potential linkage back to the source records by authorized personnel, if needed for validation. Free-text documents will automatically scanned for typical name patterns; using an internal dictionary, any occurrences of patient names, addresses, or contact details in the clinical notes will be redacted. Thus, the dataset for analysis will not contain any direct identifiers.

Transfer to a high-performance computing platform.

All pseudonymized data intended for model training will be transferred via an encrypted channel (SSH) to the MatriCS secure computing platform (Jules Verne University of Picardie, Amiens, France) for analysis. The data archive will be encrypted with strong (GPG RSA 4096-bit) encryption. Only authorized study investigators will have access to data on the analysis platform, and even that dataset will be a pseudonymized version. The original data will remain on the hospital’s servers and will be accessible only to personnel bound by a duty of medical confidentiality. If the source data have to be checked (i.e., for quality control or audit purposes), the correspondence table (i.e., the correspondence between the study ID and the PPI) will be used within the hospital by authorized staff and will never leave the secure hospital environment. All study personnel with data access are trained in patient confidentiality and have a duty of confidentiality.

Data processing.

Given the retrospective nature of the study, data quality and preparation are critical steps. To handle missing values, outliers, and inconsistencies, we shall clean the data. We do not plan to delete visits that have some missing data; indeed, data not purposely collected might be missing at random [24]. Continuous variables (e.g., vital signs and lab test results) will be checked for physiologically implausible values and extreme outliers; these values may be omitted if they are clearly erroneous and will replaced by a null value. Categorical variables (e.g., triage categories and diagnosis codes) will be standardized (e.g., one-hot-encoded). If certain key variables have a high missingness rate, appropriate imputation methods will be used (e.g., multiple imputation or the use of “unknown” categories for categorical data) and carefully documented in all derived publications. We shall also create new derived variables (i.e., indicating the missingness status of the value) that may improve prediction.

The dataset is organized at the visit level, meaning that each variable is specific for the index visit. Subsequent events (e.g., 30-day readmission and 30-day mortality) are identified using the PPI and expressed as time intervals relative to the index visit, or coded as missing when no subsequent event is found in the database. There will not be any data leakage between the training set and the test sets.

Natural language processing (NLP).

To extract useful features, free-text clinical notes will be analyzed using NLP techniques. Prior to analysis, these texts will be pseudonymized, and any personal identifiers will be removed. Once the text has been de-identified, we plan to use text mining to identify keywords or concepts indicative of severity (e.g., “sepsis”, “rigid abdomen”, “shock”, etc.) or specific diagnoses. We might employ algorithms such as spaCy or similar NLP libraries to parse the text, or a bag-of-words/embedding approach to incorporate text data into the model. One approach will involve the generation of binary features for the presence of certain terms (e.g., a predefined lexicon of important words related to emergency abdominal conditions) and possibly the use of word embedding or topic modelling to capture more complex patterns in the narratives. These text-derived features will complement the structured data in the predictive models.

Outcomes and definitions.

Since this is an observational, data-centred study, our “outcomes” correspond to both the events we are seeking to predict and our predictive models’ performance metrics. The key outcomes of interest are as follows:

• RAUC pathway eligibility (the primary classification outcome): this is a binary outcome indicating whether a given index ED visit meets the RAUC+ rule, i.e., ED attendance, subsequent admission to hospital for gastrointestinal surgery, and emergency gastrointestinal surgery within 72 hours. It serves as the ground truth for training the classification model that will distinguish RAUC-eligible patients from other patients. In the context of the RAUCisable study, this outcome is determined retrospectively for each case on the basis of whether emergency surgery was performed.
• Time to surgery: a continuous outcome measured in hours. For RAUC-eligible patients, this is defined as the time interval between ED admission (i.e., the triage timestamp) and the surgical procedure’s start time. For patients having undergone two or more emergency operations, only the earliest will be considered. This outcome will be used to train a regression model (or possibly a survival model) for predicting how quickly surgery is needed. In practice, we might also categorize this outcome into clinically relevant strata (e.g., surgery within 6 hours, within 12 hours, or within 24 hours) for decision-making, in line with the RAUC triage post-operative categories suggested (<12 h vs. < 24 h vs. longer). However, to retain granularity, the model will probably treat this prediction as a continuous variable.
• LOS: a continuous outcome defined as the total length (in days) of the hospital stay during which the emergency surgery took place (i.e., from admission or surgery to hospital discharge). We shall predict LOS as a proxy marker of resource utilization and recovery speed. If a patient is transferred to another facility, LOS will be censored at transfer. Patients who die during the hospital stay will have LOS recorded up to the time of death, and death will be accounted for separately as the mortality outcome.
• 30-day readmission: a binary outcome indicating whether the patient was unexpectedly readmitted to any acute care hospital within 30 days of discharge from the index hospital stay (i.e., the stay with emergency surgery). This outcome will be determined by reviewing the hospital’s records and (potentially) any linked regional health system data. However, since this is a single-centre study, we shall capture readmissions to our hospital and assume that most patients return to the same hospital if issues arise; we note this as a study limitation. Unscheduled readmissions will be a primary measure of the RAUC programme’s success. The prediction of the readmission risk will help us to identify patients who might benefit from closer post-discharge follow-up (e.g., telehealth check-ups).
• 30-day mortality: a binary outcome indicating whether the patient died within 30 days of the emergency surgery. This includes in-hospital death during the initial stay or any post-discharge death within 30 days. The outcome will be determined from hospital records and by checking the civil vital status registry, if available: in France, hospitals are allowed to query civil registries for death notifications. This outcome will be relatively rare but is crucial for the identification of high-risk patients.

In addition to these patient-centred outcomes, an important study output will be the set of the predictive models’ performance metrics. The primary performance measure for classification will be the area under the receiver operating characteristic curve (AUROC), along with the sensitivity, specificity, and positive predictive value at chosen probability thresholds. For regression outcomes (i.e., time to surgery and LOS), we shall evaluate metrics such as the mean squared error. For binary outcomes like readmission and mortality, we shall probably also use classification models (e.g., the AUROC). Furthermore, we shall compute measures of feature importance from the models (e.g., using permutation importance or Shapley additive explanation (SHAP) values to identify the variables that contribute most strongly to predictions. These importance rankings serve as secondary outcomes and might highlight key, clinically informative, predictive factors.

Overall, the RAUCisable study’s success will be measured by our ability to develop models with good discriminant ability (e.g., an AUROC >0.8 for the primary classifications) and to generate clinically interpretable insights that feed into the design of prospective RAUC interventions.

Sample size considerations.

During the 4-year study window, we expect to observe between 240,000 and 300,000 visits to the APUH adult ED (i.e., 60,000–75,000 visits per year). This very large sample provides high statistical power for detecting even faint patterns and is particularly advantageous for training data-intensive machine learning models. We expect the number of RAUC-eligible patients (i.e., for emergency gastrointestinal surgery) to be substantially smaller; based on hospital records, this will be a few hundred per year. For example, if ~2% of ED visits culminate in emergency gastrointestinal surgery, there will be 2,000–3,000 RAUC-eligible cases over 4 years; this is an estimate, and exact numbers will be determined from the data. This number is sufficient to train robust, predictive models – in fact, one reason for using the full 4-year span is to ensure that there are enough outcome events, especially for rarer outcomes like 30-day mortality that might account for a few percent of surgical patients). Conventional sample size “power” calculations are less applicable here because we are building predictive models and not testing a specific hypothesis via inference. Nonetheless, the large N will allow us to (i) split data into ample training and validation sets and (ii) include a high-dimensional feature space (with many variables) without overfitting because the rule-of-thumb of requiring a certain number of events per predictor will be satisfied. In summary, the dataset size is determined by data availability and the need for comprehensive machine learning training and is sufficiently large to meet the study’s objectives.

Analysis overview.

The analysis will proceed in four main steps: descriptive statistics, predictive model development, model performance evaluation, and explainability.

Descriptive statistics: We shall first characterize the overall ED cohort and the subset of RAUC-eligible surgical patients. Continuous variables will be summarized as the mean ± standard deviation or the median [interquartile range], depending on distribution. Categorical variables will be summarized as the count (percentage). We shall compare the RAUC-eligible group with other groups, in order to highlight differences (using t-tests for continuous variables and χ² tests for categorical variables). Key baseline characteristics to be reported include age distribution, sex, common presenting diagnoses, vital sign abnormalities, proportion admitted, etc. For the RAUC-eligible group specifically, we shall detail the indication for surgery (e.g., appendicitis, perforation, etc.), the surgical procedures performed, the median operating time, complication rates, mean LOS, and readmission and mortality rates. This provides context and might also guide feature selection (e.g., if certain variables differ greatly from one group to another). We shall also evaluate the completeness of data and any patterns of missingness by applying imputation or sensitivity analyses, if needed.

Predictive model development: For the primary classification task (i.e., identifying RAUC-eligible patients at the ED triage stage), we shall use supervised machine learning classification techniques. Potential algorithms include logistic regression (with L1/L2 regularization, if needed), random forests, gradient boosted trees (e.g., XGBoost or LightGBM), and (if appropriate) modern methods like deep learning (although interpretability will always be considered). We shall probably begin with simpler, interpretable models (logistic regression) and then move to more complex models if they offer performance gains. Feature selection could be performed through a combination of expert knowledge (including variables known to be relevant to surgical decision-making) and algorithmic methods (like tree-based feature importance or recursive feature elimination on the training set). We shall use k-fold cross-validation on the training data to tune model hyper-parameters and prevent overfitting. The dataset will be split chronologically or randomly into a training set (e.g., 80% of encounters) and an independent hold-out validation set (20%) for the evaluation of final performance. To simulate prospective performance in the latest year, we might prefer a chronological split: for instance, the use of 2021–2023 data for training and 2024 data for validation.

For continuous outcome predictions (e.g., time to surgery and LOS), we shall use regression models. We might opt for algorithms like linear regression after log-transforming skewed outcomes like LOS, or regression variants of the above machine learning methods (e.g., random forest and regression). Alternatively, survival analysis techniques (like Cox proportional hazards models and survival forests) might be used for time-to-surgery if we consider patients who did not undergo surgery immediately as being censored beyond a certain timeframe. However, given that every RAUC-eligible patient ultimately undergoes surgery (by definition) we shall know the actual times, so standard regression suffices.

For binary outcomes (e.g., readmission and mortality), we shall again use classification models. These could potentially be combined with the primary classification: for instance, a multi-output model that first identifies surgical patients and simultaneously predicts their risk outcomes. However, it may be more straightforward to first identify the surgical cohort and then develop separate predictive models within that cohort for readmission and mortality risks, using the data available at end of their initial hospital stay or (if we aim to predict early) even data available on admission).

Throughout the model development phase, we shall incorporate the unstructured text features (extracted via NLP) alongside the structured data. For example, we might include indicators like “peritonitis_mentioned=Yes/No” as a feature. If advanced NLP (e.g., word embeddings) is used, we could integrate these features into a deep learning model. Given the complexity, an iterative approach will be used to ensure that the text data adds value to the structured data.

To avoid time leakage, predictor availability will be anchored to predefined clinical timepoints. The primary classification model will use only information available up to triage completion or the first clinical assessment in the ED. Variables collected later (including laboratory and imaging results obtained after triage, specialty consultation notes, operative reports, discharge summaries, and disposition variables) will not be used as predictors for the primary endpoint.

Assessment of the model’s performance: We shall evaluate the final model’s level of performance on the hold-out validation set or, if appropriate, via the cross-validation results. For the RAUC eligibility classifier, we shall plot the receiver operating characteristic (ROC) curve and report the AUROC. We shall also select a probability threshold that might be used in practice (perhaps one that is sensitive enough to capture all cases of surgery, while maintaining reasonable specificity) and report the confusion matrix at that threshold (i.e., true positives, false negatives, etc.), as well as precision and recall. For regression outcomes, we shall report the mean squared error and mean absolute error on validation data, and possibly R-squared for the variance explained. In clinical time-to-event predictions, R² might be low; in that case, we shall focus on the magnitude of the error. For binary outcomes like readmission, we shall again use the AUROC or (perhaps) precision-recall curve for class imbalance.

We shall employ techniques to guard against overfitting – notably by keeping the test set fully separate and by using cross-validation. For ease of implementation in the future, we might also consider simplifying the models if the difference in performance is small.

To assess any observed difference in performance between models, the latter will be trained on several random seeds applied to a dataset split and on the weight initiations, and compared statistically using the Wilcoxon signed-rank test [25].

Explainability of Artificial Intelligence: Lastly, we shall examine the feature importance in the best-performing model. For tree-based models, feature importance scores (e.g., Gini importance or permutation importance) are readily available. For logistic regression, we shall look at the predictor’s odds ratio and thus identify the most predictive factors. These findings will be used to shape the RAUC pathway (e.g., the refinement of inclusion criteria, or targeted interventions for high-risk features).

We shall perform post-hoc explainable AI analyses on the best-performing models. Overall interpretability will be addressed using partial dependence plots (PDP) to visualise the marginal effect of individual predictors on the model’s predictions across the dataset. For local interpretability and case-level explanations of model output, we shall apply SHAP and local interpretable model-agnostic explanations (LIME). SHAP values will be computed for each prediction, in order to attribute the contribution of each feature and thus offer a unified measure of feature importance across instances. LIME will be used to approximate the local behaviour of complex models (e.g., tree ensembles) by fitting an interpretable surrogate model around a given prediction; this should help clinicians to understand why a model produces a given risk score for a given patient. These techniques will be particularly useful for identifying potential biases, validating clinical plausibility, and facilitating trust in the model when deployed in clinical environments.

Study timeline.

The overall timeline for RAUCisable is as follows (see also Fig 2):

• Data extraction: this began in January 2026, immediately after the end of the inclusion period (2024). This phase includes querying the ED and hospital databases for all required data and building the merged dataset. The expected duration is ~ 2–3 months (by February 2026), including data cleaning.
• Data processing and management: during and after extraction, roughly 3 more months (up until June 2026) will be needed for pseudonymization, NLP feature extraction, and the preparation of analysis files.
• Analysis: model development and statistical analysis will take approximately 6–8 months (from June to December 2026). This includes iterative model training, validation, and writing up the results. Given the volume of data, substantial computing time will be allocated on the MatriCS platform.
• Study completion: the planned study end date is December 2026, by which time a final study report will have been prepared. We anticipate having preliminary results earlier (in the autumn of 2026) to guide the design of the RAUC-ALGO trial.
• Dissemination: manuscripts (including the current protocol article and subsequent articles on results) and conference presentations will follow. The RAUCisable results will also be shared with the RAUC project steering committee, for integration into the RAUC-AMIENS and RAUC-ALGO trial logistics.

No interim analyses are planned because this is not an interventional trial and all the data must be collected prior to analysis. However, if data extraction reveals unexpected and significantly different numbers or major data issues, the plan may be adjusted (e.g., extension of the extraction to early June 2026 to increase the sample size or the refine inclusion criteria).

Ethical and regulatory considerations.

This study is subject to the French legislation on noninterventional, retrospective observational analyses of existing health data. Accordingly, the RAUCisable study did not require the provision of formal written informed consent from each patient; instead, it falls under the “Reference Methodology 004” (MR004) provisions of the French National Data Protection Commission (Commission Nationale de l’Informatique et des Libertés, Paris, France) for the research use of routinely collected health data. The APUH’s registered, overarching commitment to MR004 (registration number 2208336, dated 9 Oct 2018) covers the RAUCisable study. In compliance with MR004 requirements, patients whose data are included have been informed of the study through public notices and have been given the opportunity to opt out. Specifically, information about RAUCisable, its objectives, procedures and datasets was made available via the hospital website and posters in the hospital, and patients could exercise their right to refuse the use of their data and thus be excluded from the study.

The study protocol was nevertheless submitted to the local independent ethics committee for confirmation that it meets the criteria for non-interventional research (reference: PI2025_843_0015) and has been registered at ClinicalTrials.Gov (NCT07037719) [26].

Dissemination plan.

The results of the RAUCisable study will be disseminated through several channels. We plan to publish the main findings in a peer-reviewed journal, with a focus on the predictive models’ development and performance. Furthermore, insights into key predictive features will be shared with the clinical community (possibly in a surgery or emergency medicine conference), to inform clinicians about any important risk factors identified. Importantly, the RAUCisable outcomes will feed directly into the design of the RAUC-ALGO trial; for example, the algorithmic trigger for pathway activation in RAUC-ALGO will be based on RAUCisable’s model. We shall include our algorithm (or a simplified score derived from it) in the RAUC-ALGO intervention. The RAUCisable findings will also be disseminated to relevant APUH departments (i.e., the ED, surgical wards, the anaesthesia unit, and quality improvement) because they highlight the hospital’s current performance and areas for improvement in emergency surgical care (e.g., the typical time interval to surgery and its impact). There is also potential to share data or collaborate with other centres (i.e., RAUC-OUVERT will externally validate the approach at another hospital), and an aggregated multicentre analysis might follow. All individual-level data will remain secure; only de-identified, aggregated data or possibly a trained model (which does not contain patient data) would ever be shared beyond the institution.

Future directions.

If RAUCisable’s models prove to be accurate, the next step (RAUC-ALGO) will implement them prospectively. In the RAUC-ALGO trial, we shall be able to measure whether or not using the algorithm-driven triage actually changes outcomes (e.g., lowers complication rates or speeds up care), compared with the standard of care. Success in the RAUC-ALGO trial might pave the way for the wider integratation of AI tools into EDs. Furthermore, the combination of telemedicine/e-health with the identification of high-risk patients could ensure that the latter are closely followed up after discharge, in order to potentially preventing readmissions. In essence, RAUCisable provides the “who” and “what to expect”, RAUC-AMIENS provides the “how to provide care,” and RAUC-ALGO will test the combination of “who and how” in improving outcomes.