Background Pituitary adenomas occur in a heterogeneous patient population with varying degrees of preoperative morbidity and perioperative risk factors. Moreover, risk factors that reliably predict poor postoperative outcomes have not been identified. To date, no predictive models have been developed to risk stratify pituitary adenoma patients for poor early postoperative outcomes.
Objective We aim to (1) identify risk factor for early postoperative complications and (2) develop and validate a machine learning model to predict early (<30 days) postoperative outcomes after pituitary adenoma surgery.
Methods We performed a retrospective chart review of 300 consecutive adult pituitary adenoma patients treated via an endoscopic endonasal approach. Patient demographics, preoperative characteristics, and perioperative course were recorded. We defined a poor early postoperative outcome as (1) postoperative inpatient length of stay > 10 days or any of the following within 30 days of discharge, (2) emergency room admission, (3) inpatient readmission, (4) reoperation, or (5) death. Using 27 potential predictive risk factors (e.g., tumor type, age, body mass index [BMI], and postoperative serum sodium), we evaluated five machine learning algorithms (multivariate logistic regression, elastic net regression, random forest classification, support vector machines, and multilayer perceptron) for predicting early postoperative outcomes after pituitary adenoma resection. After model training and leave-one-out cross-validation for hyperparameter tuning, model testing was performed on a prospective 100 patient cohort.
Results Of the 400 total patients, 60% had nonfunctioning adenomas; GH-secreting adenomas were the most common functioning adenoma (22.3%) followed by ACTH-secreting adenomas (13%). Mean age was 53.9 ± 16.4 years with nonfunctioning adenomas presenting later (59.0 ± 14.5 vs. 46.4 ± 16.1, p = 0.001). Preoperative morbidity was significant for a 58% obesity rate with ACTH-secreting adenoma patients representing the highest rate at 75% (mean BMI: 36.0 ± 9.6). For postoperative complications, ACTH-secreting adenomas were a major risk factor for deep venous thrombosis/pulmonary embolism (odds ratio [OR] 13.8 95% CI: 2.5–113.9, p = 0.003), diabetes insipidus (OR: 3.2 95% CI: 1.6–6.1, p = 0.001), and extended length of stay (mean: 4.4 ± 3.41 vs. 2.3 ± 2.2 days, p < 0.000). CSF leaks and symptomatic hyponatremia were not associated with tumor type, sex, age, or BMI. Poor early postoperative outcomes occurred in 85 patients (21.3%), including 4 deaths. Of our machine learning models to predict poor postoperative outcomes, elastic net regression had the best performance, achieving an 87.4% prediction accuracy and an area under the curve (AUC) of 0.92 on receiver operating characteristic analysis for the prospective 100 patient testing cohort. Most predictive variables included lowest postoperative sodium level, ACTH-secreting adenoma, postoperative CSF leak, and body mass index.
Conclusion Using a machine learning approach, early postoperative outcomes after pituitary adenoma surgery can be predicted with high accuracy using well-defined risk factors. Accurate risk stratification of patients with pituitary adenomas can improve preoperative decision making and postoperative care.
