{
  "abstract": "Introduction Pancreatic neuroendocrine tumours (PanNET) are rare. This study aims to build a risk stratification machine learning (ML) model for PanNET and externally validate it.Methods A total of 8,322 patients with PanNET (2011-2022) were extracted from the SEER database (USA), and 3,612 patients (2012-2021) were extracted from the NCRAS database (England). Two Cox Survival XGBoost ML models were developed from the SEER cohort. Model 1 was developed based on TNM stage only. Model 2 was developed based on all available variables (age, sex, race, site, size, morphology, T stage, N stage, M stage, TNM stage, and grade). The top-performing models based on the concordance index (C-index) after hyperparameter tuning and 5-fold cross-validation were chosen. Model 2 was externally validated on the NCRAS cohort. Patients were assigned a risk score based on Model 2 and were grouped into 2 groups based on risk scores: Group 1 (low risk for overall mortality) and Group 2 (high risk for overall mortality). A decision tree ML model was built to help allocate patients into the 2 groups.Results In both cohorts, the median age was 62 years with a slight predominance of males (54-55%), a mean tumour size of 32-33 mm and a similar incidence in the pancreatic head (24-27%). The NCRAS cohort had more white patients (86% vs. 65%) and functional morphology (2.3% vs. 1.3%). The most predominant stage in SEER was Stage 1 (42.4%), while it was 19.4% in NCRAS. The C-index for Model 1 and Model 2 were 0.70 and 0.75 in SEER. The C-index for Model 2 after external validation in NCRAS was 0.77. The most important variables in Model 2 in descending order of importance were age, TNM stage, M stage, grade, sex, size, N stage, site, T stage, race, and functionality. AUROC, sensitivity and specificity of the decision tree ML model to categorise patients into high and low risk for overall mortality were 0.97, 0.90 and 0.94, respectively.Conclusion Including multiple variables in an ML model could improve risk stratification of patients with PanNET. This ML model is both internally and externally valid. Applying this ML model in the form of a decision tree or interactive app could be useful for risk stratification of patients with PanNET.",
  "authors": [
    {
      "affiliations": [
        "Hampshire Hospital NHS Trust, Basingstoke, United Kingdom"
      ],
      "name": "Mohamed Mortagy"
    },
    {
      "affiliations": [
        "Hampshire Hospital NHS Trust, Basingstoke, United Kingdom"
      ],
      "name": "Ker Shiong Tan"
    },
    {
      "affiliations": [
        "Hampshire Hospital NHS Trust, Basingstoke, United Kingdom"
      ],
      "name": "Aya Abdelhameed"
    },
    {
      "affiliations": [
        "Hampshire Hospital NHS Trust, Basingstoke, United Kingdom"
      ],
      "name": "Benjamin E White"
    },
    {
      "affiliations": [
        "King’s College Hospital NHS Trust, London, United Kingdom"
      ],
      "name": "Rajaventhan Srirajaskanthan"
    },
    {
      "affiliations": [
        "Hampshire Hospital NHS Trust, Basingstoke, United Kingdom"
      ],
      "name": "John Ramage"
    }
  ],
  "title": "O77 External validation of an explainable survival machine learning model for pancreatic neuroendocrine tumours: a population-based study of 11,934 patients",
  "uid": "0fa5fc58-215f-5204-8623-5ec29c5de575"
}