PenuX — Early Prediction of Severe Acute Pancreatitis

How It Works

A four-step pipeline from admission to risk stratification, fully automated and EHR-integrated.

Patient Admitted

Patient presents with acute pancreatitis. Routine admission labs are drawn (standard of care — no extra tests needed).

Lab Values Extracted

WBC, CRP, creatinine, glucose, LDH, AST, hematocrit and 12 other markers are pulled from the EHR via FHIR R4 or HL7.

Model Scores Risk

The PenuX gradient-boosted classifier returns a probability (0–1) and risk tier: Low / Moderate / High within seconds.

Clinical Decision Support

High-risk patients are flagged for ICU triage, aggressive fluid resuscitation, and early organ-failure monitoring.

Model Performance

11 models evaluated on a Chinese AP cohort (n=722, 585 severe / 137 mild) with 5-fold stratified cross-validation against Atlanta 2012 labels.

Best model: Random Forest — AUC=0.877, threshold=0.535 (F1-optimal)

0.877
AUROC — Random Forest
Best across 11 models

96.8%

Sensitivity

T=0.535 · TP=566, FN=19

38.7%

Specificity

T=0.535 · TN=53, FP=84

0.917

F1 Score

Random Forest best

87.1%

PPV

Positive Predictive Value

0.777

Best LSTM AUROC

CNN-LSTM architecture

Optimal-Threshold Confusion Matrices (all 11 models)

Model	Type	Best Threshold	TP	FP	FN	TN	Sensitivity	Specificity	PPV	F1	AUC
Logistic Regression	ML	0.575	549	77	36	60	93.8%	43.8%	87.7%	0.907	0.817
Random Forest ★	ML	0.535	566	84	19	53	96.8%	38.7%	87.1%	0.917	0.877
Gradient Boosting	ML	0.350	568	85	17	52	97.1%	38.0%	87.0%	0.918	0.874
MLP (3-layer)	DL	0.282	567	103	18	34	96.9%	24.8%	84.6%	0.909	0.836
Residual MLP	DL	0.203	572	98	13	39	97.8%	28.5%	85.4%	0.912	0.804
Attention MLP	DL	0.418	575	105	10	32	98.3%	23.4%	84.6%	0.909	0.784
LSTM	LSTM	0.300	585	137	0	0	100.0%	0.0%	81.0%	0.895	0.699
Stacked LSTM	LSTM	0.489	575	121	10	16	98.3%	11.7%	82.6%	0.898	0.705
Bidirectional LSTM	LSTM	0.318	582	129	3	8	99.5%	5.8%	81.9%	0.898	0.715
LSTM + Attention	LSTM	0.172	585	136	0	1	100.0%	0.7%	81.1%	0.896	0.684
CNN-LSTM ★ LSTM	LSTM	0.082	585	135	0	2	100.0%	1.5%	81.2%	0.897	0.777

Comparison to Existing Severity Scores

Score	AUROC	Sensitivity	Specificity	Timing	Requires
PenuX-AP-Severity (RF) This work	0.877	96.8%	38.7%	<4 hours	106 routine admission labs · Chinese cohort n=722
BISAP	0.77	0.68	0.89	24 hours	BUN, mental status, SIRS, age, pleural effusion
APACHE-II	0.74	0.65	0.85	24 hours	12 physiological parameters + GCS
Ranson Criteria	0.71	0.62	0.83	48 hours	11 criteria over 2 time points
CTSI (imaging)	0.79	0.70	0.88	CT required	Contrast-enhanced CT scan

* Comparative figures are approximate values from published literature. PenuX figures are from 5-fold stratified CV on the Chinese AP cohort (n=722). No external test set — results are from in-sample cross-validation only.

Key Features

Designed for real clinical workflows — fast, interpretable, and privacy-preserving.

⚡

Instant Prediction

Returns a risk score in <200 ms. No CT scan or 24-hour observation required.

🧪

Routine Labs Only

Uses only standard admission labs — no special tests, no imaging, no extra cost.

🔗

FHIR R4 Native

Direct integration with any FHIR-compliant EHR. Supports Camelion, Epic, Cerner, and OpenEMR.

🔒

Privacy by Design

Patient identifiers (TZ, MRN) are stripped before scoring. No PII stored or logged.

📖

Fully Open Source

MIT licence. Full training code, model weights, and documentation on GitHub.

🌍

International Collaboration

Seeking clinical validation partners across Europe, USA, Asia, and Africa.

REST API

Three endpoints — plain JSON, FHIR R4 Bundle, or HL7 v2 message.

# POST /predict  — plain JSON admission data
curl -X POST https://api.penux.uk/predict \
  -H "Content-Type: application/json" \
  -d '{
    "age":         62,
    "sex":         "M",
    "wbc":         18.5,
    "crp":         220,
    "creatinine": 1.8,
    "glucose":    180,
    "ldh":         450,
    "ast":         90,
    "hematocrit": 42
  }'

# POST /fhir/predict  — FHIR R4 Bundle (Patient + Observations)
curl -X POST https://api.penux.uk/fhir/predict \
  -H "Content-Type: application/fhir+json" \
  -d '{
    "resourceType": "Bundle",
    "type": "collection",
    "entry": [
      {
        "resource": {
          "resourceType": "Patient",
          "birthDate": "1962-03-15",
          "gender": "male"
        }
      },
      {
        "resource": {
          "resourceType": "Observation",
          "code": {"coding": [{"system": "http://loinc.org", "code": "1988-5"}]},
          "valueQuantity": {"value": 220, "unit": "mg/L"}
        }
      }
    ]
  }'

# POST /hl7/predict  — raw HL7 v2.x message (text/plain)
curl -X POST https://api.penux.uk/hl7/predict \
  -H "Content-Type: text/plain" \
  --data-raw 'MSH|^~\&|LIS|HOSPITAL|PENUX|API|20240601||ORU^R01|001|P|2.5
PID|1||MRN123||DOE^JOHN||19620315|M
OBR|1|||AP_PANEL
OBX|1|NM|1988-5^CRP^LN||220|mg/L|0-5||||F
OBX|2|NM|6690-2^WBC^LN||18.5|10*3/uL|4-11||||F
OBX|3|NM|2160-0^CREATININE^LN||1.8|mg/dL|0.7-1.2||||F'

# Response (all endpoints return same structure)
{
  "severe_ap_probability": 0.782,
  "risk_group":            "High",
  "recommendation":       "Consider ICU/HDU admission. Early aggressive fluid resuscitation.",
  "fields_used":          ["wbc", "crp", "creatinine", "glucose", "ldh", "ast"],
  "missing_fields":       ["bun", "calcium", "albumin"],
  "model_version":        "1.0.0",
  "atlanta_classification": "Potentially Severe"
}

📚 Full API Docs 🐙 View on GitHub

EHR Integration

Plug-and-play integration with major hospital information systems.

🏥

Camelion

Native JSON adapter + FHIR R4 Bundle endpoint. Hebrew & English field names supported.

Supported

⚕️

Epic

SMART on FHIR R4 with Epic LOINC and vendor LIS codes. Uses Epic's open.epic endpoint.

Supported

🔬

Cerner / Oracle

HL7 v2.x OBX segments + FHIR R4 via Cerner Millennium API.

Supported

📋

Generic FHIR R4

Any FHIR R4 compliant system. Patient + Observation resources with LOINC codes.

Supported

📡

HL7 v2.x

Classic ORU^R01 messages with MSH, PID, OBR, OBX segments from any LIS.

Supported

🔧

Custom / Other

CSV upload, direct JSON, or contact us to build a custom adapter.

On Request

Supported Lab Parameters

19 routine admission values with their standard LOINC codes.

Parameter	LOINC Code	Unit	Clinical Relevance
WBC (White Blood Cells)	`6690-2`	10³/µL	Inflammatory response
CRP (C-Reactive Protein)	`1988-5`	mg/L	Primary severity marker
Creatinine	`2160-0`	mg/dL	Renal failure indicator
BUN (Blood Urea Nitrogen)	`3094-0`	mg/dL	Renal / hydration status
Glucose	`2345-7`	mg/dL	Metabolic stress
LDH (Lactate Dehydrogenase)	`2532-0`	U/L	Tissue necrosis marker
AST	`1920-8`	U/L	Hepatic / pancreatic injury
ALT	`1742-6`	U/L	Biliary pancreatitis
Hematocrit	`20570-8`	%	Haemoconcentration / necrosis
Calcium	`17861-6`	mg/dL	Ranson / APACHE criterion
Albumin	`1751-7`	g/dL	Nutritional / severity marker
Total Bilirubin	`1975-2`	mg/dL	Biliary obstruction
Lipase	`3040-3`	U/L	Diagnostic confirmation
Amylase	`1798-8`	U/L	Diagnostic confirmation

Early Prediction of
Severe Acute Pancreatitis
in Under 4 Hours

How It Works

Patient Admitted

Lab Values Extracted

Model Scores Risk

Clinical Decision Support

Model Performance

Optimal-Threshold Confusion Matrices (all 11 models)

Comparison to Existing Severity Scores

Key Features

Instant Prediction

Routine Labs Only

FHIR R4 Native

Privacy by Design

Fully Open Source

International Collaboration

REST API

EHR Integration

Camelion

Epic

Cerner / Oracle

Generic FHIR R4

HL7 v2.x

Custom / Other

Supported Lab Parameters

Interested in Collaboration?