The PANEN nomogram: clinical decision support for patients with metastatic pancreatic neuroendocrine neoplasm referred for peptide receptor radionuclide therapy

Abstract

Introduction: Patients with pancreatic neuroendocrine neoplasms (P-NEN) may benefit from peptide receptor radionuclide therapy (PRRT). Prediction of overall survival (OS) using statistical models has the potential to guide treatment decisions. In this study, we have generated a clinicopathological and imaging parameter-based internally validated nomogram of patients who received PRRT for metastatic P-NEN to facilitate treatment decision support for the clinical management of such patients.

Patients and methods: We reviewed 447 pancreatic NEN patients treated with PRRT. Clinical variables for the prediction of overall survival (OS) included age, gender, Karnofsky performance score (KPS), weight loss, hepatomegaly, time from diagnosis to first PRRT (days), tumor functionality, presence of Hedinger syndrome, presence of liver metastases, presence of bone metastases, presence of lung metastases, alkaline phosphatase, 2-deoxy-2-[18F]fluoro-D-glucose ([₁₈F]FDG) positron emission tomography (PET) scan positivity, erythrocytes, platelets, creatinine clearance, leucocytes, and histologic grade of tumor differentiation based on KI-67 staining. A random survival forests (RSF) method was used to construct a model with an optimal number of clinical variables. The model was developed on 80% of the data and tested on the remaining 20% of the data. Performance of prediction was calculated using the c-index, a generalization of the area under the ROC curve (AUC) for survival models.

Results: Median follow up time was 2045 days (min 136 days, max 10329 days). Time from diagnosis to 1^st PRRT, alkaline phosphatase, KPS, hepatomegaly, weight loss, [₁₈F]FDG-PET scan positivity, Ki-67% derived histologic grade, lung metastases, age, presence of bone metastases, platelet count, erythrocyte count, creatinine clearance, hemoglobin, presence of functioning tumor, creatinine, and gender, were in order of importance, all independent predictors for overall survival. The development set c-index was 0.86, while the test set c-index was 0.82. A nomogram was constructed based on the optimal number of clinical parameters selected in the RSF model.

Conclusion: This study proposes an internally validated nomogram (PANEN-N) to accurately predict overall survival for P-NEN patients following PRRT, which could be used for patient counseling to facilitate informed and shared decision support in daily clinical practice as well as for generating new hypotheses.

Keywords: clinical decision support nomogram; machine learning; pancreatic neuroendocrine neoplasm; pedictict overall survival; peptide receptor radionuclide therapy.

Figure 1

Schematic overview of the Random Survival Forest (RSF) model development process used in this study.

Figure 2

(A) Web-based survival rate calculator (Dynamic Nomogram (shinyapps.io)) to predict the overall survival of metastatic P-NEN patients treated with PRRT. Time_diagnosis_to_treatment refers to time from diagnosis to first PRRT treatment (in days). Alkaline phosphatase (ALP) values are shown in µkat/L, weight difference in kg, platelet count in G/L, erythrocyte count in T/L, creatinine clearance in mL/min/1.73 m² and creatinine in µmol/L. (B) Nomogram for prediction of overall survival (OS) in metastatic pancreatic NEN treated with PRRT. The nomogram is based on a cox proportional hazards model and is used by drawing a vertical line from each predictor value to the score scale at the ‘top-points’. After manually summing up the individual scores, the ‘total points’ correspond to the probability (prob) of overall survival, which are estimated by drawing a vertical line from this value to the bottom scale ‘2-year survival prob’ or ‘5-year survival prob’ to estimate overall survival.

Figure 3

Variable importance and model error rate with increase in number of trees.