Predicting Response to Somatostatin Analog Therapy in GEP-NET Liver Metastases: A Machine Learning Approach with ⁶⁸Ga-DOTATATE and ¹⁸F-FDG PET/CT.
Authors
Affiliations (5)
Affiliations (5)
- Department of Outpatient Chemotherapy, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang 150000, China (J.X., J.W., N.H., R.Q., J.L., H.L.).
- Department of PET-CT/MR, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang 150000, China (M.Y., K.W.).
- Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University, Jinan, Shandong 250117, China (J.Y.).
- Department of Radiology, Shandong Provincial Hospital Affiliated to Shandong First Medical University, Jinan, Shandong 250021, China (R.S.).
- Department of Outpatient Chemotherapy, Harbin Medical University Cancer Hospital, Harbin, Heilongjiang 150000, China (J.X., J.W., N.H., R.Q., J.L., H.L.). Electronic address: [email protected].
Abstract
Predicting response to somatostatin analogs (SSAs) in gastroenteropancreatic neuroendocrine tumors (GEP-NETs) with liver metastases (LMs) remains challenging. This study aimed to develop and validate an interpretable machine learning (ML) model based on dual-tracer (<sup>68</sup>Ga-DOTATATE and <sup>18</sup>F-FDG) PET/CT parameters to predict octreotide response and identify key imaging predictors using SHAP analysis. 32 GEP-NET patients with LMs who received first-line octreotide therapy were retrospectively enrolled. Quantitative PET/CT metrics, dual-tracer ratios, and lesion size were extracted from 202 hepatic lesions. Lesions were classified as disease-control (SD/PR) or disease-progression (PD) using size-change thresholds derived from RECIST 1.1. Patients were divided at the patient level into training and validation cohorts. LASSO regression was used for feature selection, followed by the construction of multiple ML models. Leave-one-patient-out cross-validation (LOPOCV) was performed to further assess model robustness, and SHAP analysis was applied for model interpretability. ⁶⁸Ga-DOTATATE TLR, rTLR, and primary size differed significantly between disease-control and disease-progression groups. LASSO identified six key predictive features. The XGBoost model showed the best performance in the patient-level validation cohort, with an AUC of 0.825 and an accuracy of 0.742. In LOPOCV, XGBoost achieved an AUC of 0.856 and an accuracy of 0.787. SHAP analysis revealed that ⁶⁸Ga-DOTATATE TLR, primary size, and ⁶⁸Ga-DOTATATE SUVmean were the three most important predictors. An interpretable ML model integrating dual-tracer PET/CT parameters can effectively predict octreotide response in GEP-NETs with LMs. This approach may provide a noninvasive tool for treatment-response stratification and individualized management.