Bridging the Gap in Early Abdominal Cancer Detection
A guideline-anchored clinical AI pipeline empowering clinicians with decision support for pancreatic, gastric, and esophageal cancers — at first presentation, before symptoms become undeniable.
⚕️ Research Prototype Only. Not validated for clinical use. Not approved by FDA, CE, or any regulatory authority. Do not use for actual patient care decisions.
The Platform
One pipeline. Two entry points.
OncosenseAI meets patients and clinicians where they are — from self-triage to automated referral documentation.
⚠️ Methodological note: Survival and outcome metrics are derived directly from SEER registry variables (age, stage, treatment, histology). Symptom-level features in Module 1 are calibrated from published NICE NG12 clinical prevalence data, as SEER does not capture primary care symptom presentation. This distinction is fully disclosed in the preprint.
👤 For Patients
Self-Triage & Early Awareness
Conversational symptom assessment built on NICE NG12 standards. No medical background needed.
Visual biomarker prompts for visible clinical signs
Tiered next-step recommendations to prioritise urgent care
Accessible from any device, in any setting
🩺 For Clinicians
SEER-Validated Risk Scoring
Ensemble ML risk scores with SHAP explainability and automated 2WW referral documentation.
Logistic Regression model (AUROC 0.790) — best individual performer
SHAP explainability showing which features drove each risk tier
One-click PDF referral packets with NICE NG12 justification
Treatment matching via NCCN/ESMO/NICE with ClinicalTrials.gov integration
SAMPLE RISK OUTPUT — RESEARCH PROTOTYPE
HIGH RISK
2WW Referral Indicated
Patient — 64F · Pancreatic suspicion
NICE NG12 Assessment · Composite Score 14/18
● ALARM Unexplained weight loss
● ALARM Jaundice (scleral)
● PRESENT Epigastric discomfort
○ ABSENT Rectal bleeding
Clinical Evidence
Built on real population data
All survival and outcome metrics derived directly from SEER registry variables via 5-fold cross-validated models.
0.790
AUROC — Logistic Regression
Best individual model · 5-fold CV
0.753
AUROC — Ensemble (LR+RF+GB)
Probability averaging · stable across folds
0.853
NPV — Ensemble Model
Primary safety-netting metric
100%
NG12 Alarm Sensitivity
All alarm cases → urgent 2WW tier
26.4%
Surgery — Top SHAP Feature
Highest feature importance in RF model
EQUITY FINDING
A 2-Month Survival Gap — Identified in the Data
Analysis of the SEER cohort revealed a statistically meaningful racial disparity in median survival: Black patients showed median survival of 7 months versus 9 months for White patients — a gap that OncosenseAI is designed to surface and address at the point of primary care triage.
Figure 1: ROC curves (LR AUROC=0.790, Ensemble=0.753), Random Forest feature importance, and survival by treatment received. All results from 5-fold cross-validation on SEER n=500 cohort.
Figure 2: Kaplan-Meier survival analysis. Left: overall survival by cancer site. Centre: survival by stage at diagnosis. Right: survival by race/ethnicity — illustrating the equity gap OncosenseAI targets.
Figure 3: Deep-dive survival analysis. Site-stratified stage curves, stage distribution by cancer type, median survival heatmap, and 1- and 2-year survival rates across the SEER cohort.
📋 Limitations — Stated Transparently
Retrospective SEER data does not capture the full clinical narrative of a primary care consultation.
Analytic cohort (n=500) is modest — larger independent validation required for generalisability.
SEER does not contain primary care symptom data. Module 1 symptom features are calibrated from published NICE NG12 prevalence distributions — not directly measured.
Modules 2, 3, and 4 performance metrics are not yet reported and require separate validation.
This is a research prototype. It must not be used for actual patient care decisions.
Pipeline Architecture
Four modules. One pipeline.
Deterministic guideline logic feeding ML inference, visual diagnostics, treatment matching, and automated reporting.
MODULE 01
🧠
Symptom Intelligence Engine
Three-tier NICE NG12-aligned risk stratification. Dual-layer: deterministic alarm logic + ensemble ML composite scoring across 19 clinical features. The model cannot override a guideline-mandated referral.
✓ Validated
MODULE 02
🔭
Visual Diagnostics Support
OpenCV-based image analysis for endoscopic and radiological imaging features. Designed for jaundice detection, mucosal changes, and lesion flagging. Performance metrics pending.
⏳ In Validation
MODULE 03
🧬
Treatment Matcher
NCCN/ESMO/NICE-aligned treatment recommendations with live ClinicalTrials.gov API integration. Cross-references cancer type, stage, and genomic marker profile.
● Active
MODULE 04
📑
Clinical Report Generator
Automated PDF generation producing structured referral summaries — risk tier, alarm features, composite score, SHAP rationale, and NICE NG12 justification for 2WW referrals.
● Active
Research
The Preprint
A multi-cancer retrospective validation study using SEER population data. Targeting MedRxiv, Q3 2026.
OncosenseAI: A NICE NG12-Aligned Clinical Decision Support Pipeline for Early Detection of Abdominal Cancers
FDA 510(k) pre-submission · EU MDR 2017/745 pathway
The detection window is narrow. The opportunity is now.
OncosenseAI is seeking primary care collaborators for prospective IRB-sponsored validation. Co-authorship open to researchers with cohort access or IRB infrastructure.
