Computational Model
XGBoost equations rendered with TeX, learned gain weights from the public cohort, target coding and high-sensitivity threshold design.
Acute pancreatitis · open research framework
Early deterioration prediction from routine hospital data.
PenuX-AP-Severity studies whether admission and first-24-hour laboratory data can predict a later Atlanta-defined severe acute pancreatitis outcome. XGBoost is the principal machine-learning model; Revised Atlanta remains the reference outcome definition.
XGBoostInternal Medicine WardTarget sensitivity ≥98%~3,000 nominal source recordsHTML + TeXExternal validation
Research use only. This site is not a validated diagnostic, triage, ICU-transfer, discharge or treatment system.
Research site
Simulator
Interactive HTML/JavaScript demonstration of score, sigmoid and threshold mechanics, explicitly separated from the trained model.
Findings
Calculated feature importance, published benchmark evidence, PenuX targets and current validation status.
Datasets
Guilin Multi-ML, Guilin LNN, Hefei/OSF and eICU planning cohorts with provenance, label and overlap caveats.
Validation
OOF threshold locking, calibration, bootstrap uncertainty, leave-one-cohort-out evaluation and time-aware deterioration endpoints.
Internal Medicine Ward
The intended research scenario: identify future deterioration while the acute-pancreatitis patient is still on a general/internal medicine ward.
~3,000 Multi-cohort Plan
Scale the evidence base without treating nominal source counts as guaranteed unique patients or using a naive pooled random split.
Model Weights JSON
Machine-readable normalized XGBoost gain importances generated from development data with the corrected SAP label mapping.
Source Code
Open the PenuX-AP-Severity implementation and validation tools on GitHub.
Current computational snapshot
1,289
records in the current public gain-weight calculation
15.8%
SAP prevalence after correct source-label normalization
Ca
highest XGBoost gain feature in the current development model
Label quality-control finding.
The Multi-ML source encodes raw Diagnostic Result as 0=SAP and 1=non-SAP. PenuX explicitly converts this to the internal convention 0=non-SAP, 1=SAP before training.