Machine Learning-based Prediction Model for Treatment of Acromegaly With First-generation Somatostatin Receptor Ligands
Luiz Eduardo Wildemberg,Aline Helen da Silva Camacho,Renan Lyra Miranda,Paula C L Elias,Nina R de Castro Musolino,Debora Nazato,Raquel Jallad,Martha K P Huayllas,Jose Italo S Mota,Tobias Almeida,Evandro Portes,Antonio Ribeiro-Oliveira,Lucio Vilar,Cesar Luiz Boguszewski,Ana Beatriz Winter Tavares,Vania S Nunes-Nogueira,Tânia Longo Mazzuco,Carolina Garcia Soares Leães Rech,Nelma Veronica Marques,Leila Chimelli,Mauro Czepielewski,Marcello D Bronstein,Julio Abucham,Margaret de Castro,Leandro Kasuki,Mônica Gadelha
DOI: https://doi.org/10.1210/clinem/dgab125
2021-06-16
Abstract:Context: Artificial intelligence (AI), in particular machine learning (ML), may be used to deeply analyze biomarkers of response to first-generation somatostatin receptor ligands (fg-SRLs) in the treatment of acromegaly. Objective: To develop a prediction model of therapeutic response of acromegaly to fg-SRL. Methods: Patients with acromegaly not cured by primary surgical treatment and who had adjuvant therapy with fg-SRL for at least 6 months after surgery were included. Patients were considered controlled if they presented growth hormone (GH) <1.0 ng/mL and normal age-adjusted insulin-like growth factor (IGF)-I levels. Six AI models were evaluated: logistic regression, k-nearest neighbor classifier, support vector machine, gradient-boosted classifier, random forest, and multilayer perceptron. The features included in the analysis were age at diagnosis, sex, GH, and IGF-I levels at diagnosis and at pretreatment, somatostatin receptor subtype 2 and 5 (SST2 and SST5) protein expression and cytokeratin granulation pattern (GP). Results: A total of 153 patients were analyzed. Controlled patients were older (P = .002), had lower GH at diagnosis (P = .01), had lower pretreatment GH and IGF-I (P < .001), and more frequently harbored tumors that were densely granulated (P = .014) or highly expressed SST2 (P < .001). The model that performed best was the support vector machine with the features SST2, SST5, GP, sex, age, and pretreatment GH and IGF-I levels. It had an accuracy of 86.3%, positive predictive value of 83.3% and negative predictive value of 87.5%. Conclusion: We developed a ML-based prediction model with high accuracy that has the potential to improve medical management of acromegaly, optimize biochemical control, decrease long-term morbidities and mortality, and reduce health services costs.