Endocrine Abstracts

Aim: To identify the causes and predictors of mortality in acromegaly patients diagnosed, treated, and followed in the Pituitary Center of Istanbul University-Cerrahpasa over the last three decades using machine learning (ML) models.

Methods: The study sample consisted of 607 consecutive patients diagnosed with acromegaly at Cerrahpasa Faculty of Medicine during the period 1990 – 2023. The main inclusion criteria were age > 18 years and a clear-cut diagnosis of acromegaly according to standard criteria at the time of presentation. There were 556 alive patients (91.6%) and 51 deceased patients (8.4%), indicating a class imbalance. Synthetic Minority Over-sampling Technique (SMOTE) was applied due to imbalanced data. Minimum Redundancy Maximum Relevance (mRMR) and Recursive Feature Elimination (RFE) methods were used to select features. Logistic Regression, Random Forest, and XGBoost models were used to generate feature importance. The accuracy of the models were assessed through 4-fold cross validation and precision used as the performance metric. The importance of potential predictors were measured by SHAP values.

Results: Of 607 patients, 42% (n=257) were male. The median age at diagnosis and last visit were 40 [32–50] and 54 [45–64] years, respectively. The median follow up duration was 99 [40–150] months. Macroadenoma, cavernous invasion, and hypopituitarism was present in 66.9%, 8.5%, and 10.7% of the patients at diagnosis, respectively. Transsphenoidal surgery (TSS) was the initial treatment in 90.3% of the patients with a success of achieving remission in 28.7%. After TSS, 16.1% of patients received additional radiotherapy and 57.1% received postoperative drug treatment. Somatostatin receptor ligands (SLAR) was the most common drug and 36.1% of the patients were SLAR resistant. At last visit, 22% of the patients had active disease. Hypertension was the most common comorbidity (37.2%), followed by diabetes mellitus (36.4%), cardiovascular disease (CVD) (15.7%), colon polyps (12.2%), and malignancy (10%). Thyroid carcinoma was the most common malignity (5.03%), followed by gastrointestinal cancers (1.6%). The main cause of death was CVD (41.2%) and malignities (21.6%). XGBoost had the highest accuracy and precision scores. Top-ranked features in XGBoost model with decreasing importance were malignity and cardiovascular disease.

Conclusion: Malignity and cardiovascular diseases are important causes of morbidity and mortality in patients with acromegaly. ML models may a better understanding of patients’ mortality risk, assisting healthcare professionals to tailor management strategies.