Building an interpretable machine learning prognosis prediction model-based on baseline examinations of patients with esophageal cancer undergoing surgery.

To create an interpretable machine learning model based on non-invasive biomarkers for the early diagnosis and improved prognostic value of esophageal cancer. We gathered a private dataset at Sichuan Cancer Hospital, comprising 3204 esophageal cancer patients who underwent surgery. Baseline markers and preoperative biochemical blood tests were thoroughly reviewed. The necessary factors were identified using an elastic net, and 27 machine learning methods were used to construct prediction models. The random forest model was chosen for its high performance, and additional optimization was performed using the White Shark Optimizer (WSO) algorithm. To improve the model's interpretability, the Shapley Additive exPlanations (SHAP) technique was used. We proposed an interpretable machine learning framework for exploratory early risk stratification of postoperative esophageal squamous cell carcinoma patients using routinely available preoperative baseline examinations. The random forest (RF) model achieved 0.74 accuracy and 0.72 F1 score after 1 year. It was adjusted to 0.69 in the Area Under the Curve (AUC). Gender and blood magnesium levels were the best six characteristics that affected the 1-year and 5-year survival rates. The model with the highest discriminative ability accurately predicted the prognosis of esophageal cancer patients in the test population. Our research led to the creation of a machine learning model that accurately predicts the prognosis of esophageal cancer and identifies early indicators of adverse outcomes. SHAP-improved interpretability facilitates physician-patient interaction and trust and promotes individualized therapy.