Development and validation of an interpretable machine learning model for predicting intraoperative HDI in PPGL based
Shurong Li1, Zhiqiang Zhang2, Yubing Zhang3
1Department of Radiology, The First Affiliated Hospital, Sun Yat-sen University, Guangzhou, Guangdong 510080, China.
Purpose:
Intraoperative hemodynamic instability (HDI) might lead to severe complications for pheochromocytoma and paraganglioma (PPGL) patients. This study aims to construct a machine learning (ML) model to predict HDI based on intratumoral and peritumoral CT radiomics.
Materials And Methods:
Totally, 223 patients diagnosed with PPGL from two centers were retrospectively included and analyzed. Intratumoral and peritumoral radiomic features were derived from preoperative computed tomography (CT) imaging and then selected through least absolute shrinkage and selection operator (LASSO) algorithm. Three predictive models including intratumoral, peritumoral and fusion radiomic models were constructed by applying eight ML methods. The performance of predictive models were evaluated by receiver operating characteristic curve (ROC) with area under the curve (AUC), calibration and decision curves. Furthermore, SHapley Additive exPlanations (SHAP) analysis was used for the interpretability of the optimal model by ranking features importance.
Results:
Enrolled patients were randomly divided into training cohort (n = 156) and testing cohort (n = 67). By the feature screening process, eight, fourteen and eighteen features were selected for intratumoral, peritumoral and fusion radiomic models, respectively. The fusion radiomic models exerted the best performance for predicting HDI with the AUC values of 0.840 (95 %CI: 0.779-0.901) and 0.800 (95 %CI: 0.693-0.907) in the training and testing cohort. Calibration curve indicated a high level of agreement between the predictive probabilities and observed outcomes, and further DCA showed the best clinical benefits among the three models.
Conclusion:
The fusion radiomic model including intratumoral and peritumoral features is promising nonivasive tool for predicting intraoperative HDI for PPGL patients, which might facilitate individualized evaluation and management in clinical practice.
