The Silent Killer That Speaks Early

Pancreatic cancer kills approximately 51,000 Americans each year. Caught while still localized, the five-year survival rate is 44 percent. Once it has spread, fewer than 4 percent of patients survive five years.¹ It carries the most dramatic survival gap of any cancer on this list. It also leaves one of the most specific pre-diagnostic metabolic signals in the standard blood panel.

44% survival when caught early, but under 4% once spread. No cancer has a wider gap. No cancer has a more specific pre-diagnostic blood signal.

The Signature in the Blood

The primary signal is glucose metabolism disruption. Pancreatic tumors impair islet cell function years before causing symptoms, producing a pattern of new-onset diabetes or accelerating glucose instability. A landmark study of 28,137 pancreatic ductal adenocarcinoma (PDAC) cases confirmed that HbA1c diverges from controls two to three years before diagnosis, while liver enzymes (bilirubin, alkaline phosphatase [ALP], alanine transaminase [ALT], gamma-glutamyl transferase [GGT]) show a rapid increase approximately one year before diagnosis, and white blood cell (WBC) and platelet counts rise in the final year.²

The glucose signal has been quantified with precision. A study of fasting blood glucose trajectories in PDAC patients showed that hyperglycemia diverged from matched controls 30 to 36 months before diagnosis and correlated directly with tumor volume.³ New-onset diabetes precedes pancreatic cancer diagnosis in 52 percent of cases, making it a specific metabolic fingerprint detectable from any glucose or HbA1c measurement on the comprehensive metabolic panel (CMP).

The Machine Learning Case

An XGBoost model trained on routine clinical data in patients with new-onset diabetes achieved an area under the curve (AUC) of 0.80 for identifying which new-onset diabetics would develop pancreatic cancer within one year, outperforming the END-PAC (AUC 0.63) and Boursi (AUC 0.68) models, and capturing 75 percent of cancers at 70 percent specificity.⁴ A second XGBoost model trained on electronic health record (EHR) data achieved an AUC of 0.84 and detected 58 percent of late-stage patients a median of 24 months before their actual diagnosis.⁵

What This Proves

The glucose-disruption signal for pancreatic cancer is measurable on a standard CMP two to three years before diagnosis. The XGBoost methodology is validated. What has not been built is the routine primary care algorithm that flags new-onset diabetes patients for pancreatic risk stratification based on the metabolic trajectory already in the EHR. The science is established. The deployment is the missing step.

Endnotes

1.National Cancer Institute. “Cancer Stat Facts: Pancreatic Cancer.” SEER, 2024.

SEER data reporting five-year relative survival of 43.6% for localized pancreatic cancer versus approximately 3.2% for distant-stage disease—the most extreme survival differential across all cancer types. Only 15% of cases are diagnosed at the localized stage.

2.Tan PS, Garriga C, Clift A, et al. “Temporality of body mass index, blood tests, comorbidities and medication use as early markers for pancreatic ductal adenocarcinoma: a nested case-control study.” Gut. 2023;72(3):512–521.

Population-based nested case-control of 28,137 PDAC cases and 261,219 matched controls in England. HbA1c diverged two to three years before diagnosis; liver enzymes rose sharply one year before; WBC and platelet counts rose in the final year. Establishes the multi-marker pre-diagnostic CMP and CBC timeline.

3.Sharma A, et al. “Fasting Blood Glucose Levels Provide Estimate of Duration and Progression of Pancreatic Cancer Before Diagnosis.” Gastroenterology. 2018;155(2):490–500.

Fasting blood glucose in PDAC patients diverged from controls 30-36 months before diagnosis and correlated with tumor volume. Post-surgical resection improved FBG, confirming a causal relationship between the tumor and the metabolic CMP signal.

4.Khan S, Bhushan B. “Machine Learning Predicts Patients With New-onset Diabetes at Risk of Pancreatic Cancer.” J Clin Gastroenterol. 2024;58(7):681–691.

XGBoost model trained on new-onset diabetes patients, achieving AUC 0.80 for identifying future pancreatic cancer, outperforming END-PAC (0.63) and Boursi (0.68). Key features: age, weight-loss trajectory, and glucose-destabilization patterns from routine chemistry.

5.Chen Q, Cherry DR, Nalawade V, et al. “Clinical Data Prediction Model to Identify Patients With Early-Stage Pancreatic Cancer.” JCO Clin Cancer Inform. 2021;5:279–287.

XGBoost model on 3,322 early-stage and 25,908 late-stage pancreatic cancer cases from Optum EHR database. AUC 0.84; detected 58% of late-stage patients a median of 24 months before their actual diagnosis. Demonstrates that machine learning applied to standard clinical data identifies pancreatic cancer at a stage where survival outcomes improve dramatically.