191 Development of a Novel Individualized Interpretable Transformer-Based Deep Learning Model for Predicting Early Postoperative Hypokalemia in Pituitary Adenomas

INTRODUCTION: Hypokalemia is a common complication following pituitary adenoma surgery that can have negative effects on patient prognosis, hospitalization duration, and rehospitalization risk. The recognition of factors that predict early postoperative hypokalemia is advantageous for the management of pituitary tumor patients after surgery. METHODS: This retrospective cohort study analyzed the association between clinical factors of pituitary adenoma and hypokalemia in 280 patients with PAs between October 2018 and October 2020 who were treated at the Department of Neurosurgery at Nanjing Drum Tower Hospital, Nanjing, China. A ViT model was modified as a predictive model and compared with XGBoost. The performance of the models was evaluated using Accuracy, Precision, Recall, F1, NPV, PPV, Sensitivity, Specificity, ROC curve and PR curve. Five-fold cross-validation was performed. XGBoost can automatically provide estimates of feature importance from a trained predictive model that does not change with the patient's condition. In the Transformer model, Grad_cam is used to visualize the importance of the input clinical parameters, which is personalized for each patient. RESULTS: The performance metrics of the proposed transformer-based model, are as follows: Accuracy of 0.8357±0.0387, Precision of 0.8704±0.0158, Recall of 0.8357±0.0387, F1-score of 0.8357±0.0387,NPV of 0.9478±0.0028, PPV of 0.5967±0.0715, Sensitivity of 0.8385±0.0071, Specificity of 0.8351±0.0484, AUC-ROC of 0.8578±0.0335, and AUC-PR of 0.7014±0.0817. The preoperative factors including gender, hypertension, serum potassium concentration, and disease duration are significantly predictive of postoperative hypokalemia. CONCLUSIONS: Our proposed model demonstrates superior predictive performance on early postoperative hypokalemia in pituitary adenomas when compared to XGBoost. The attention maps derived from the Transformer-based predictive model allows for individualized interpretation of prediction outcomes.