A machine learning model integrating clinical, histopathologic, and epigenetic markers improved prediction of postoperative recurrence in patients with nonfunctioning pituitary adenomas compared with a model based solely on clinical variables, according to research presented at the 2026 AACE Annual Meeting in Las Vegas.

Researchers analyzed data from a national registry of 501 patients in Uzbekistan with nonfunctioning pituitary adenomas. Among these patients, 227 underwent transnasal and/or transcranial adenomectomy, with subsequent classification into recurrent (37%) and stable (63%) disease courses. 

The analysis incorporated clinical presentation, hormonal status, tumor morphology, immunohistochemical markers, and circulating microRNAs.  Associations with postoperative recurrence were evaluated using correlation analysis, receiver operating characteristic analysis, proportional hazards modeling, and multivariable predictive modeling. A risk algorithm and computerized calculator were subsequently developed from the final model. 

The researchers found that patients with recurrent tumors were more likely to have endocrine dysfunction, visual impairment, and systemic manifestations than those without recurrence. 

Among histopathological markers, Ki-67 demonstrated the strongest discriminatory performance, with an area under the curve (AUC) of 0.90. A cutoff value of 4.5% yielded a sensitivity of 76% and a specificity of 88% for recurrence prediction; p53 expression achieved an AUC of 0.81, while chromogranin A achieved an AUC of 0.77. 

Among epigenetic biomarkers, microRNA-128 demonstrated strong predictive performance, with an AUC of 0.86, sensitivity of 67%, and specificity of 98%. By comparison, microRNA-155 demonstrated limited utility, with an AUC of 0.61. 

Model performance improved as additional biological markers were incorporated. A clinical model alone achieved an AUC of 0.74. Adding histopathological markers, including Ki-67 and p53, increased the AUC to 0.94. Incorporation of microRNA-128 further improved performance, resulting in an AUC of 0.98. The final model produced a Brier score of 0.06. The researchers reported that incorporation of histopathological and epigenetic markers improved predictive performance compared with clinical variables alone.  Internal cross-validation yielded AUC values of approximately 0.89 and 0.94.  

Expert Insight

AACE Endocrine AI invited lead researcher Sevara Anvarova, a doctoral student with the Neuroendocrinology Department, Republican Specialized Scientific-Practical Medical Center of Endocrinology named after Academician Ya. Kh. Turakulov in Tashkent, Uzbekistan, to elaborate on the findings.

Why does this study matter? 

Predicting recurrence after surgery remains one of the major challenges in the management of nonfunctioning pituitary neuroendocrine tumors (PitNETs). While some patients remain disease-free for many years, others experience recurrence despite apparently favorable clinical features. Current risk assessment relies largely on imaging and a limited number of pathological markers, which do not fully capture the biological complexity of these tumors. 

We explored whether combining clinical characteristics, histopathological markers, circulating microRNAs, and machine-learning–based risk modeling could improve prediction of postoperative recurrence. More accurate risk stratification could help clinicians identify patients who require closer surveillance while reducing unnecessary follow-up in lower-risk individuals. This is particularly important in healthcare settings where access to specialized care and repeated imaging may be limited. 

When you had all the data in front of you, was there a finding, or perhaps more than one, that surprised you? 

The most surprising finding was the additional predictive value provided by microRNA-128. As expected, Ki-67 remained one of the strongest predictors of recurrence. However, microRNA-128 substantially improved prediction beyond traditional histopathological markers alone, suggesting that circulating biomarkers may capture important aspects of tumor biology that are not fully reflected by tissue pathology. 

At the same time, another candidate biomarker, microRNA-155, contributed much less predictive information than anticipated. This was an important reminder that promising biomarkers must be rigorously validated before they can be considered clinically useful. 

How might the findings influence clinical endocrinology practice? 

Although these findings require external validation before routine clinical implementation, they support a more personalized approach to postoperative management. In the future, integrated prediction models could help endocrinologists tailor surveillance strategies according to an individual patient's risk profile rather than applying the same follow-up schedule to everyone. 

Such approaches may be particularly valuable in regions where access to frequent MRI monitoring or specialized endocrine services is constrained. More precise risk assessment could direct healthcare resources to patients most likely to benefit from intensive surveillance while maintaining high-quality long-term care.