2025 Proffered Presentations

 

← Back to Previous Page

 

S042: USE OF MACHINE-LEARNING TO PREDICT GROSS TOTAL RESECTION OF NON-FUNCTIONING PITUITARY ADENOMAS: A SINGLE INSTITUTION SERIES
Benjamin B Fixman, BS; Ishan Shah, BS; Apurva Prasad, BA; Gage A Guerra, BS; Jeffrey J Feng, MD; Robert G Briggs, MD; Gabriel Zada, MD, MS; University of Southern California

Pituitary adenomas are a common intracranial neoplasm for which surgical resection remains the definitive treatment. Though the decision to operate is multifactorial, the likelihood of GTR is a particularly notable consideration. The rapid advancement of machine learning (ML) presents an opportunity to determine the likelihood of GTR based on pre-operative characteristics and aid in preoperative planning for these patients.

A prospectively maintained database of 219 endoscopic operations completed by one surgeon at our institution between 2013-2023 was first analyzed with a logistic regression model for statistical inference of potentially predictive features. 4 features, including gender, prior operation, tumor diameter, and Knosp score were selected for model training and prediction. Next, the dataset was stratified by outcome and split 80:20 (training/validation:testing). Various models, including Logistic Regression, K-Nearest Neighbors, Random Forest, and Gradient Boosting were trained, 5-fold cross-validated, and hyperparameter tuned to optimize classification accuracy using mlr3 in R. Finally, the optimal model was trained on the full training/validation dataset and used to make predictions on the test set. 

Gradient-boosting, which employs successive iterations of decision trees trained on the residuals of previous trees, proved optimal at minimizing cross-validation error (0.27). On an independent test set, this model achieved an error: 0.23, sensitivity: 0.83, specificity: 0.64, AUC: 0.76, PPV: 0.83, NPV: 0.64.

The use of ML to pre-operatively predict probability of GTR may help surgeons and patients make informed decisions regarding risk and benefit of surgical resection. Our models performed well on an independent test set, achieving 77% accuracy. The PPV and NPV highlight their utility.

Table 1. Data characteristics

	Gender	Reoperation	Diameter	Knosp Score	GTR	Patients
N (%) [SD]	M: 113 (51.6)	No: 175 (79.9)	Median: 25mm [9.93]	0: 33 (15.1) Low:109 (49.8)   High: 77 (35.2)	Yes: 147 (67.1)	219

Table 2. Inference model

Variable	Coefficient	P Value
Gender	-0.59	0.098
Reoperation	-0.34	0.40
Maximal Tumor Diameter	-0.071	0.0010
Knosp (Low)	0.58	0.27
Knosp (High)	-1.03	0.061

Table 3. 5-Fold Cross Validated Scoring Metrics of Tuned Models

Model	Error	Sensitivity	Specificity	AUC
Logistic Regression	0.27	0.89	0.43	0.78
K Nearest Neighbors	0.29	0.90	0.34	0.74
Random Forest	0.30	0.89	0.32	0.75
Gradient Boosting	0.27	0.87	0.49	0.76

Table 4: Test Performance of Gradient Boosting

Model	Error	Sensitivity	Specificity	AUC	PPV	NPV
Gradient Boosting	0.23	0.83	0.64	0.76	0.83	0.64

 

← Back to Previous Page