2025 Poster Presentations
P235: PROGNOSTICATION OF BIOCHEMICAL REMISSION FOLLOWING PITUITARY ADENOMA RESECTION IN CUSHING'S DISEASE USING POSTOPERATIVE CORTISOL TRAJECTORIES
Rushmin Khazanchi1; Sachin Govind1; Joshua Vignolles-Jeong2; Vikas Munjal2; Stephen T Magill1; Daniel M Prevedello2; 1Northwestern University; 2Ohio State University
Background: Prognostication of a patient’s biochemical status following transsphenoidal pituitary surgery for Cushing’s Disease holds significant clinical importance. Prior studies have used machine learning (ML) approaches to complete this task. However, no studies have examined the predictive value of serially monitored cortisol lab values in the immediate postoperative period. Understanding whether the morphology of the postoperative cortisol trajectory impacts future biochemical outcome could serve to enhance prediction models and inform postoperative patient monitoring strategies. The goal of this study was to develop and assess a ML model trained on cortisol morphologic features.
Methods: This is a retrospective study of patients undergoing pituitary adenoma resection for Cushing’s Disease from 2009-2019. Predictor variables of interest included demographics, tumor characteristics, and postoperative serum cortisol values collected at 6-hour intervals for up to 96-hours postoperatively. The outcome of interest was biochemical remission within 1 year.
A random forest ML model was trained and evaluated on various feature subsets (all features, solely postoperative cortisol, solely demographic and tumor characteristics) and time points using 50-fold random cross validation. Both AUC-ROC and AUC-PR metrics were used to assess model performance. An optimal model and time point was chosen based on these performance metrics, and important features were extracted through aggregate SHAP values.
Results: Of 55 patients, 47 (85%) achieved cure within 1 year (Table 1). At all evaluated time points, models that used cortisol data outperformed models that did not use cortisol data in terms of AUC-ROC (p < 0.01), and there was no significant difference between models that used postoperative cortisol only versus postoperative cortisol in conjunction with demographic and tumor characteristics (Figure 1). These trends were largely preserved with respect to AUC-PR. Optimal performance was at 60 hours with the cortisol-only model achieving an average AUC-ROC of 0.791.
Analysis of the cortisol-only model demonstrates that lower cortisol nadir values as well as a more negative 48-hour running slope of the cortisol curve were positively associated with biochemical remission (Figure 2).
Conclusion: Postoperative cortisol trajectory is highly predictive of biochemical cure status. ML models constructed with solely postoperative cortisol morphology features significantly exceeded the performance ML models built without these features. Future prediction algorithms should consider incorporation of these features into their models.