Multi-cohort expansion

Approximately 3,000 AP records across four cohorts.

PenuX-AP-Severity is being expanded from a single-dataset experiment into a cohort-aware validation framework. The objective is not simply to increase sample size, but to test whether an early XGBoost signal is stable across institutions, case mixes and data-collection systems.

Nominal N = 3,0174 source cohortsLabs + diagnosesXGBoostAtlanta-compatible SAP outcomeTarget sensitivity ≥98%

Important: 3,017 is a planning total of source records, not yet a verified count of unique eligible patients. The two Guilin cohorts may contain overlapping individuals, and eICU will lose records after eligibility/missingness/outcome-label checks.

Core cohort composition

public / repository

1,289

Guilin Multi-ML

204 SAP; ~60 predictors. Main tabular laboratory benchmark.

public / repository

722

Guilin LNN

137 SAP; ~107 predictors. Related Guilin cohort with a richer feature set.

public source

260

Hefei / Han et al.

200 development + 60 validation records with labs, clinical variables and SAP/NSAP labels.

credentialed public data

746*

eICU-CRD AP

Published candidate AP cohort from a 208-hospital US ICU database; final PenuX N must be re-derived.

Why the datasets must not simply be concatenated

Cohort	Setting	Strength	Main limitation	Recommended role
Guilin Multi-ML	Single institution, China	Largest directly usable labeled tabular cohort	Single-center; source-specific measurement patterns	Primary development / internal validation
Guilin LNN	Same institution, overlapping calendar period	More predictors and additional records	Possible patient overlap with Multi-ML cannot be excluded after de-identification	Sensitivity analysis; not an independent external cohort
Hefei / OSF	Different Chinese institution	Institutionally independent source with labs and severity labels	Smaller sample	External / transportability validation
eICU-CRD	208 US hospitals, ICU-enriched	Strong geographic and institutional shift; timestamped labs and diagnoses	Credentialed access; ICU spectrum differs from general AP; Atlanta-compatible SAP label must be derived	Multi-center stress test and time-aware validation

Validation architecture

Source-specific QC
units, ranges, missingness

→

Common lab schema
harmonized predictors

→

Development CV
XGBoost tuning

→

Locked model
98% sensitivity threshold

Hefei test
external site

eICU test
US multicenter shift

→

Per-cohort report
AUROC · AUPRC · calibration

→

Pooled summary
only after heterogeneity

Common laboratory feature set

A cross-cohort model should prioritize measurements that can be mapped reliably across datasets. Candidate common variables include WBC, hemoglobin/hematocrit, platelets, BUN/urea, creatinine, glucose, calcium, sodium, potassium, albumin, bilirubin, AST/ALT, LDH, coagulation tests and selected vital signs. Features that exist only in one cohort can be evaluated in cohort-specific secondary models but should not define the primary transportability model.

Leakage rule: predictors must be available before the prediction time. Persistent organ failure or later interventions used to define the future SAP outcome cannot re-enter the model as predictors.

Outcome compatibility

The Chinese labeled cohorts already provide SAP/non-SAP outcomes tied to acute-pancreatitis severity definitions. eICU does not provide a ready-made Revised Atlanta SAP column. For the primary severity analysis, PenuX should derive a persistent-organ-failure endpoint (>48 h) from time-stamped organ-failure measurements where feasible. ICU admission or mortality alone must not be relabeled as SAP.

For eICU, mortality can be reported as a separate secondary outcome. This preserves scientific interpretability and prevents incompatible labels from being pooled.

Additional auxiliary dataset

auxiliary · not counted in core N

538

MIMIC-IV-Ext pancreatitis cases

MIMIC-IV-Ext Clinical Decision Making contains 538 pancreatitis cases and extensive laboratory results. It is useful for laboratory-schema mapping, diagnosis tasks and representation/domain experiments, but it is not counted in the ~3,017 SAP pool until an outcome compatible with the PenuX severity definition is established.

PhysioNet source

What success would mean

Metric	Development goal	Why it matters
Sensitivity	≥98% at locked watch threshold	Minimize missed future severe cases in the research early-warning setting.
Specificity	Report, do not force	Shows the alert burden created by high sensitivity.
AUROC	Discrimination across cohorts	Measures rank discrimination but does not define the alert threshold.
AUPRC	Primary model-selection metric	More informative when SAP is the minority class.
Calibration	Slope/intercept + Brier + reliability	Required before interpreting XGBoost output as absolute risk.
External stability	No large collapse on Hefei/eICU	More important than a marginal gain on the original development cohort.