Clinical Machine Learning Model for Predicting Pathological Complete Response in Patients with Esophageal and Gastroesophageal Junction Adenocarcinoma After Trimodality Therapy

BACKGROUND: Accurate prediction of pathological complete response (pCR) after preoperative chemoradiation therapy, followed by surgery (trimodality therapy) in esophageal adenocarcinoma (EAC) and gastroesophageal junction adenocarcinoma (GEJAC) may improve clinical decision-making and patient counseling before esophagectomy. This study aimed to develop predictive models for pCR after trimodality using machine learning (ML) approaches. PATIENTS AND METHODS: A total of 569 patients with EAC and GEJAC who received trimodality therapy at MD Anderson Cancer Center between 2002 and 2022 were included. Clinicopathological characteristics and survival benefit of patients who achieved a pCR were reviewed via descriptive and survival analyses. Subsequently, ML models based on clinical variables were employed to predict pCR, including BART, random forest, and XGBoost, logistic regression, and LASSO. RESULTS: pCR was achieved in 132 patients (23.2%). Poorly differentiated tumors, tumors with signet ring cell component, higher T stage, higher clinical stage, residual tumor on biopsy after chemoradiation, and higher SUVmax on positron emission tomography-contract tomography (PET-CT) after chemoradiation were significantly associated with non-pCR. pCR patients had significantly longer overall survival (OS) and relapse free survival (RFS) compared with non-pCR patients (median OS, 10.40 versus 4.42 years, log-rank p = 0.0041; median RFS, 10.40 versus 2.35 years, log-rank p < 0.0001). The random forest model showed the highest predictive ability for pCR with an AUC value of 0.702 among the employed models. CONCLUSIONS: This first exploratory study supports the validity and potential utility of ML-based models for predicting pCR after trimodality therapy in EAC and GEJAC. Further validation is warranted before clinical application.