Machine Learning-Based Model to Predict Long-Term Tumor Control and Additional Interventions following Transsphenoidal Surgery for Acromegaly Patients

Objective: Patients with acromegaly secondary to growth hormone hypersecretion suffer from various symptoms, which can ultimately lead to life-threatening condition if left untreated. This study aimed to establish a supervised machine learning (ML) model to predict long-term biochemical outcomes and additional intervention-free survival after endonasal transsphenoidal surgery (ETS) for acromegaly patients.

Methods: The authors reviewed the medical records of patients who underwent ETS for acromegaly patients at our institution between 2013 and 2022. Data on patients’ baseline characteristics, intervention details, histopathology, surgical outcomes, and postoperative neurological and endocrine functions were collected. The primary outcome of this study was the intervention-free survival (IFS) rate, and the therapeutic outcomes were labeled as “under control” or “treatment failure,” depending on whether additional therapeutic interventions were required. Multiple ML models, including decision tree, random forest, support vector machine, and K-neighbors were tested to predict long-term tumor recurrence ([Fig. 1]).

Shapley additive explanations (SHAP) were calculated for feature importance using the best-performing model.

Results: Data on 122 ETSs for 100 patients with a median (range) follow-up period of 64 (1–130) months were extracted. In the entire cohort, 32 patients (32%) required additional interventions for persistent or recurrent acromegaly. Consequently, IFS rates following primary ETS alone were 70% at 3 years and 67% at 5 years in the entire cohort ([Fig. 2]).

Multivariable Cox proportional hazard analysis demonstrated that higher Knosp-Steiner (KS) grade (hazard ratio [HR], 9.52; 95% confidence interval [CI], 3.28–27.59; p = 0.001) and lower tumor resection rates (continuous, HR, 1.02; 95% confidence interval, 1.00–1.03; p = 0.001) were the significant risk factors associated with worsen IFS. In the decision tree analysis, when patients had a gross total resection (GTR) of the tumor and patients’ age was 25 or older, they had a 79% risk of additional intervention (accuracy 81%, [Fig. 3]).

ML using SHAP showed that tumor size <9 mm, tumor resection rate of 100%, patient’s age <65, and Knosp grade 0 had the highest impact on better IFS ([Fig. 4]).

Conclusion: Based on this supervised ML model, KS grade, tumor resection rates, tumor size, and patient age were important predictors of acromegaly IFS. These data provide insight into at-risk acromegaly patients who may require additional interventions on follow-up.

Publikationsverlauf

Artikel online veröffentlicht:
07. Februar 2025