A machine learning model based on preoperative multiparametric quantitative DWI can effectively predict the survival and recurrence risk of pancreatic ductal adenocarcinoma

Abstract

Purpose: To develop a machine learning (ML) model combining preoperative multiparametric diffusion-weighted imaging (DWI) and clinical features to better predict overall survival (OS) and recurrence-free survival (RFS) following radical surgery for pancreatic ductal adenocarcinoma (PDAC).

Materials and methods: A retrospective analysis was conducted on 234 PDAC patients who underwent radical resection at two centers. Among 101 ML models tested for predicting postoperative OS and RFS, the best-performing model was identified based on comprehensive evaluation metrics, including C-index, Brier scores, AUC curves, clinical decision curves, and calibration curves. This model's risk stratification capability was further validated using Kaplan-Meier survival analysis.

Results: The random survival forest model achieved the highest C-index (0.828/0.723 for OS and 0.781/0.747 for RFS in training/validation cohorts). Incorporating nine key factors-D value, T-stage, ADC-value, postoperative 7th day CA19-9 level, AJCC stage, tumor differentiation, type of operation, tumor location, and age-optimized the model's predictive accuracy. The model had integrated Brier score below 0.13 and C/D AUC values above 0.85 for both OS and RFS predictions. It also outperformed traditional models in predictive ability and clinical benefit, as shown by clinical decision curves. Calibration curves confirmed good predictive consistency. Using cut-off scores of 16.73/29.05 for OS/RFS, Kaplan-Meier analysis revealed significant prognostic differences between risk groups (p < 0.0001), highlighting the model's robust risk prediction and stratification capabilities.

Conclusion: The random survival forest model, combining DWI and clinical features, accurately predicts survival and recurrence risk after radical resection of PDAC and effectively stratifies risk to guide clinical treatment.

Critical relevance statement: The construction of 101 ML models based on multiparametric quantitative DWI combined with clinical variables has enhanced the prediction performance for survival and recurrence risks in patients undergoing radical resection for PDAC.

Key points: This study first develops DWI-based radiological-clinical ML models predicting PDAC prognosis. Among 101 models, RFS is the best and outperforms other traditional models. Multiparametric DWI is the key prognostic predictor, with model interpretations through SurvSHAP.

Keywords: Diffusion-weighted imaging; Machine learning; Pancreatic ductal adenocarcinoma; Prediction model; Prognosis.

Fig. 1

Flowchart for patient inclusion and model establishment

Fig. 2

DWI image (A), corresponding ADC map (B), D map (C), D^* map (D), f map (E), and DDC map (F) with corresponding ROI measurements

Fig. 3

Prediction models were developed using 101 ML algorithms, and the best-performing model was identified through evaluation. A Prediction models developed using 101 ML algorithms were evaluated for their performance in predicting OS and RFS in both the training and validation cohorts using the key metric, the C-index. The models were ranked based on the average C-index across cohorts, identifying the random survival forest model as the best-performing ML model for predictive capability. B The random survival forest model was constructed using the top 5–12 variables ranked by average C-index (OS train, OS valid, RFS train, and RFS valid cohorts). It was observed that when the top nine variables were included, the average C-index across the four cohorts was the highest, indicating the best predictive performance of the model. In both the training and validation cohorts, the random survival forest model, identified as the best-performing model, achieved Brier scores below 0.13 and C/D AUC curves exceeding 0.85 for predicting OS (C) and RFS (D) following radical resection of PDAC, demonstrating remarkable predictive accuracy

Fig. 4

Evaluation and explanation of the global and localized impact of the random survival forest model in predicting OS. A, B The global impact of each variable on the random survival forest model’s predictions for OS was assessed using Brier score loss and C/D AUC loss after permutation. The analyses revealed that variable importance evolves over time, with the Brier score and C/D AUC loss demonstrating a pronounced time-dependent effect. Notably, the D value consistently showed the greatest significance as survival time increased. C The variable importance summary plot ranks nine key features by their influence on OS. D–G Partial dependence survival profiles (PDPs) illustrate how changes in individual variables, such as D value, T stage, ADC value, and postoperative 7th day CA19-9 level, affect OS while holding other variables constant. Patients with lower D value experienced a more rapid decline in survival compared to those with higher D value, reflecting the significant protective impact of D value on postoperative survival. Narrow bands in the PDP plots indicate stable predictions, while wider bands suggest greater sensitivity to variable changes

Fig. 5

Evaluation and explanation of the global and localized impact of the random survival forest model in predicting RFS. A, B The global impact of each variable on the random survival forest model’s predictions for RFS was evaluated using Brier score loss and C/D AUC loss after permutation. Similar to OS, variable importance demonstrated a time-dependent effect, with the D value consistently identified as the most significant variable for predicting RFS over time. C The variable importance summary plot ranks nine key features by their influence on RFS. D–G Partial dependence survival profiles (PDPs) highlight how individual variables, such as D value, T stage, ADC value, and postoperative 7th day CA19-9 level, influence RFS predictions while keeping other variables constant. Patients with lower D values showed a more rapid decline in RFS compared to those with higher D values. The PDP plots show narrow and overlapping bands for certain variables, suggesting stable predictions, while wider bands indicate that even small changes in the variable values can significantly affect RFS predictions

Fig. 6

Performance evaluation of the radiological–clinical random survival forest model for predicting OS and RFS. A–C DCA comparing the radiological–clinical random survival forest model to four common models (radiological, clinical, TNM, and AJCC stage) for predicting OS in the training cohort at 12 months (A), 18 months (B), and 24 months (C). D–F DCA for the radiological–clinical model and the four common models for predicting RFS in the training cohort at 6 months (D), 12 months (E), and 18 months (F). The radiological–clinical model consistently outperformed the other models, demonstrating superior clinical utility. G, H Calibration curves for OS (G) and RFS (H) in the training cohort at corresponding time points (12 months, 18 months, 24 months for OS and 6 months, 12 months, 18 months for RFS). The curves align closely with the diagonal dashed line, indicating strong agreement between predicted and observed outcomes. I–L Kaplan–Meier survival curves stratifying patients into low-risk and high-risk groups for OS in the training cohort (I) and validation cohort (J), and for RFS in the training cohort (K) and validation cohort (L). The model effectively separated patients into distinct risk categories, with significant survival differences (p < 0.0001)