Revisiting mortality in acromegaly using machine learning

Abstract

Objective To evaluate predictors of mortality in acromegaly using machine learning (ML) models.

Methods We analyzed 607 patients with acromegaly. A grid search was applied to compare five types of prediction models in 1,200 configurations. Each model predicted mortality based on 40 disease characteristics, assessed using the Matthews’ correlation coefficient (MCC) and the area under the curve (AUC) metrics. To ensure robust and reliable predictions, we constructed a virtual ensemble model by generating 100 Explainable Boosting Machine (EBM) variants and evaluating their performance across 50 random train–test splits.

Results The single EBM model fitted to all data achieved an MCC of 0.88 and an AUC of 0.99. Feature importance analysis identified hypertension, first-line transsphenoidal surgery (TSS), disease duration, baseline insulin-like growth factor 1 levels, hypopituitarism, and repeat TSS as key predictors of mortality. A virtual ensemble prediction model showed good generalization performance, with an MCC of 0.64 ± 0.08 and an AUC of 0.96 ± 0.015 across 50 random data splits. 58% of the cohort were female, and 66.9% had macroadenomas. The first-line treatment was TSS in 83.2% of cases. Additional treatments were repeat TSS (16.6%), somatostatin analogs (57%), cabergoline (35.6%), pegvisomant (10.9%), and radiotherapy (16.6%), achieving a 77.8% overall success rate. Hypertension (37.2%) and diabetes (36.4%) were the most common comorbidities. The mortality rate was 8.4%, mainly from cardiovascular disease (41.2%) and cancer (21.6%).

Conclusion ML demonstrated strong performance in classifying survival status and identifying mortality predictors in acromegaly. The ensemble model provides a reliable and interpretable tool, balancing accuracy with the robustness required for clinical decision-making.

Explore related subjects