# Machine Learning Approach for Differentiation of Pheochromocytoma from Adrenocortical Cancer and Non-Functioning Adrenal Adenomas

## TL;DR

This paper presents a machine learning model that helps doctors distinguish pheochromocytoma from other adrenal tumors using clinical and radiological features.

## Contribution

The study introduces an interpretable machine learning model validated for differentiating pheochromocytoma from adrenocortical carcinomas and non-functioning adrenal adenomas.

## Key findings

- Machine learning models achieved strong predictive performance with AUC over 0.8 for differentiating pheochromocytoma from other adrenal tumors.
- Key clinical features like systolic blood pressure and palpitation were identified as important discriminators for pheochromocytoma.
- Tumor density and size were found to be important markers for differentiating pheochromocytoma from non-functioning adrenal adenomas and adrenocortical carcinomas.

## Abstract

Background: The differentiation of pheochromocytoma (PCC) from other adrenal lesions, particularly in incidentalomas with non-benign radiological characteristics (size > 4 cm or density > 10 HU), remains a clinical challenge. The study aimed to develop and validate an interpretable machine learning (ML) model for pairwise differentiation of PCC from adrenocortical carcinomas (ACCs) and non-functioning adrenal adenomas (NAAs) and to identify the most important clinical features. Methods: We analyzed a dataset of 50 clinical, laboratory, and radiological parameters from 123 patients with histologically verified adrenal tumors (63 PCC, 30 ACC, 30 NAA). Four classifiers—Logistic Regression (LR), Random Forest (RF), Linear Discriminant Analysis (LDA), and Extreme Gradient Boosting (XGBoost)—were trained for binary classification tasks (PCC vs. ACC, PCC vs. NAA, ACC vs. NAA) using a robust nested stratified cross-validation pipeline to ensure generalizability and avoid overfitting. Results: All four models showed strong predictive performance, with discrimination (AUC) more than 0.8. Our analysis, based on the interpretable LR model, identified the key discriminators differentiated PCC from both ACC and NAA: maximum systolic blood pressure, grade 3 hypertension, headache, palpitation, tachycardia, male sex, and concomitant gastric and duodenal ulcers. In contrast, lower back pain and general weakness were strong signs of lower probability of PCC. The tumor density specifically differentiated PCC from NAA, whereas tumor size was an important marker for distinguishing PCC and ACC. Conclusions: We developed robust ML models capable of accurately differentiating PCC from other adrenal tumors in complex cases. The models provide a clinically actionable tool for pre-surgical decision support. Furthermore, the identification of key discriminative features enhances the clinical understanding of PCC and facilitates its differential diagnosis prior to histological verification.