Introduction

Pancreatic Ductal Adenocarcinoma (PDAC) remains one of the deadliest solid tumors, largely due to its late-stage diagnosis, limited therapeutic responsiveness, and poor overall prognosis. Despite ongoing advances in imaging techniques and molecular characterization, effective tools for early prognostic stratification are still lacking. Recent research suggests that integrating machine learning (ML) with clinicopathologic variables and tumor microenvironment components, particularly cancer-associated fibroblasts (CAFs, may significantly enhance predictive accuracy. Leveraging these computational methods could facilitate earlier intervention and guide personalized treatment decisions, offering a potential paradigm shift in PDAC management.

Aims & Methods

This study aims to develop and validate machine learning models that incorporate clinicopathologic parameters and cancer-associated fibroblast (CAF) expression profiles to enhance prognostic stratification in patients with PDAC. A retrospective analysis was conducted on a cohort of 76 patients diagnosed with PDAC through endoscopic ultrasound-guided fine-needle biopsy (EUS-FNB). Data collected included alcohol consumption, smoking status, type II diabetes, C-reactive protein (CRP), albumin, neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), CRP/albumin ratio, carcinoembryonic antigen (CEA), carbohydrate antigen 19-9 (CA 19-9), CA 125, and expression levels of CAF markers S100A4 and alpha-smooth muscle actin (α-SMA). The primary endpoint was the classification of tumor stage into early (T1–T2) and advanced (T3–T4) categories. Four machine learning algorithms—Latent Dirichlet Allocation (LDA), Quadratic Discriminant Analysis (QDA), Logistic Regression (LR), and K-Nearest Neighbors (K-NN)—were applied and their performance evaluated using independent validation datasets.

Results

All four machine learning models demonstrated strong predictive performance, with accuracy scores ranging from 0.77 to 1.00. Quadratic Discriminant Analysis (QDA) emerged as the most accurate classifier. Latent Dirichlet Allocation (LDA) achieved 91% accuracy and identified the neutrophil-to-lymphocyte ratio (NLR) as a key feature associated with early-stage (T1–T2) disease. In contrast, CRP/albumin ratio and PLR were more predictive of advanced-stage (T3–T4) tumors. Logistic Regression (LR), with an 83% accuracy rate, also highlighted PLR, CRP/albumin, and CAF markers (S100A4 and α-SMA) as significant contributors to advanced staging. These findings underscore the potential of integrating clinicopathologic and stromal biomarkers into ML-driven prognostic models.

Conclusion

Integrating machine learning with clinicopathologic and cancer-associated fibroblast (CAF) expression data offers a promising strategy for improving prognostic assessment in PDAC. By accurately distinguishing early- from advanced-stage disease, these models can support more personalized risk stratification and treatment planning. The results underscore the value of AI-driven approaches in precision oncology, with potential to enhance clinical decision-making and ultimately improve outcomes for patients with pancreatic cancer.

Introduction

Chronic liver disease is a major global health concerns, contributing to over 1.4 million annual deaths worldwide¹. Its primary fatal outcomes, hepatic complications^2,3, arise from liver cirrhosis, hepatocellular carcinoma and hepatic decompensation. The early, reversible stages of chronic liver disease are frequently asymptomatic and overlooked, leading to delayed diagnoses and irreversible complications that treatment options are limited.^4,5. Therefore, reliable tools to identify whether a person will develop hepatic complications in the future have become a public health priority, enabling early diagnosis and timely intervention in high-risk populations.

Aims & Methods

This study aimed to employ machine learning to develop proteomics-based models for the early identification of hepatic complications. In this study, 37,178 adults without hepatic complications at baseline were analyzed from a prospective cohort from United Kingdom, with plasma levels of 2,736 proteins measured using Olink technology. Cox regression was used to identify significant proteins associated with incident hepatic complications. Models based on importance-top-ranking proteins were constructed using LightGBM learners of Automated Machine Learning (AutoML) and internal validated through 5-fold cross validation in the derivation set (n = 37,178). Independent validation was then conducted in a geographically distinct UKB cohort (n = 15,767).

Results

Over a median follow-up of 13.8 years, 338 participants developed hepatic complications. Cox regression identified 593 proteins significantly associated with incident hepatic complications. Among them, GDF15, PROC, GGT1, ANXA10, IL18, MFAP4, COL4A1, CDCP1 and IL6 were ranked highest in protein importance ordering. The protein model showed a considerable predictive accuracy for hepatic complications under different time scenarios in the UKB validation set: all-year (area under the curve [AUC] = 0.89), within 5 years (AUC = 0.88), within 10 years (AUC = 0.91), over 10 years (AUC = 0.81), and achieved superior predictive performance compared to models with demographic predictors, laboratory indicators, polygenic risk score (PRS) and traditional risk score models (Protein 0.89 vs. Demographics 0.83, Laboratory 0.85, PRS 0.49; Fibrosis-4 index [FIB-4] 0.76, Aspartate aminotransferase to platelet ratio index [APRI] 0.78; all P < 0.05). Combined with clinical data and PRS, the predictor performance is further enhanced: all-year (AUC = 0.91), within 5 years (AUC = 0.91), within 10 years (AUC = 0.90), over 10 years (AUC = 0.83). Individuals in the high-risk group stratified by the protein model were 16.27 times more likely to develop hepatic complications.

Conclusion

Our study constructed a novel protein model for hepatic complications risk stratification and could be used to predict individuals who will develop hepatic complications up to 16 years in advance in the general population.

References

1. Tham EKJ, Tan DJH, Danpanichkul P, et al. The Global Burden of Cirrhosis and Other Chronic Liver Diseases in 2021. Liver Int 2025;45:e70001.
2. Åberg F, Asteljoki J, Männistö V, et al. Combined use of the CLivD score and FIB-4 for prediction of liver-related outcomes in the population. Hepatol Baltim Md 2024;80:163–172.
3. Innes H, Morling JR, Buch S, et al. Performance of routine risk scores for predicting cirrhosis-related morbidity in the community. J Hepatol 2022;77:365–376.
4. Karlsen TH, Sheron N, Zelber-Sagi S, et al. The EASL–Lancet Liver Commission: protecting the next generation of Europeans against liver disease complications and premature mortality. The Lancet 2022;399:61–116.
5. Serra-Burriel M, Juanola A, Serra-Burriel F, et al. Development, validation, and prognostic evaluation of a risk score for long-term liver-related outcomes in the general population: a multicohort study. Lancet Lond Engl 2023;402:988–996.