Identify deterioration risk while the patient is still on the ward.
PenuX-AP-Severity is designed as a research framework for patients with acute pancreatitis admitted to an internal medicine ward who do not yet have a fully established severe course. Early laboratory and bedside data are used by XGBoost to estimate the probability of later Atlanta-defined severe acute pancreatitis or organ-failure deterioration.
Clinical-research scenario
internal medicine
0–24 h
risk estimation
low / watch / high
Atlanta-defined SAP
What information enters the model?
Laboratory data
Candidate variables include WBC, CRP, BUN/urea, creatinine, glucose, calcium, sodium, potassium, albumin, LDH, amylase, hemoglobin/hematocrit, platelets, bilirubin, AST/ALT and coagulation variables where available.
Basic clinical context
Age, sex, selected vital signs, timing from presentation, etiology and other variables may be included only when they are available at or before the intended prediction time.
What must stay out
Future organ-failure measurements, later interventions, ICU events and any variable that directly reveals the outcome must not be used as early predictors.
What the output means
| Output | Meaning in the research workflow | What it does not mean |
|---|---|---|
| Calibrated risk probability | Estimated probability of a later severe outcome conditional on the early data. | It is not a diagnosis and does not establish SAP at the prediction time. |
| Watch threshold | Development-locked threshold targeting ≥98% sensitivity. | It is not an approved bedside escalation threshold. |
| High-risk band | A more selective research stratum with higher predicted risk. | It does not automatically imply ICU transfer or a particular treatment. |
| Atlanta outcome | Reference standard used to determine the subsequent severity state. | Atlanta is not a competing machine-learning model. |
Why an internal medicine ward is an important setting
The research question is most meaningful before severe deterioration is obvious. If a patient is already experiencing established persistent organ failure, the problem is no longer early prediction of that event. The intended evaluation therefore focuses on patients who are still eligible for prediction at admission or during the first prespecified ward window and asks whether their early pattern predicts a later severe course.
Validation required before any bedside claim
Performance must be reported separately for each cohort and hospital context. The primary analysis should include AUROC, AUPRC, sensitivity, specificity, PPV, NPV, calibration and alert burden at the locked threshold. Independent-site validation is required because ward populations can differ markedly by institution and case mix.
The current multi-cohort plan combines source evidence from Guilin Multi-ML, Guilin LNN, Hefei/OSF and a credentialed eICU acute-pancreatitis cohort, while retaining cohort identity and avoiding a naive pooled random split.