Interpretable machine learning integrating intra-/peritumoral CT radiomics and serum indicators for predicting poorly differentiated esophageal squamous cell carcinoma.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, Luzhou People's Hospital, Luzhou, China.
- Department of Pathology, Luzhou People's Hospital, Luzhou, China.
Abstract
To develop an interpretable ensemble learning model integrating multiple machine learning algorithms, intratumoral and peritumoral CT radiomics, and serum biomarkers for predicting poorly differentiated esophageal squamous cell carcinoma (ESCC). This retrospective study enrolled 261 ESCC patients, who were randomly allocated to training (n=183) and validation (n=78) cohorts. Radiomics features were extracted from intratumor, peritumoral 0.3 cm and intra-peritumoral 0.3 cm areas of enhanced CT arterial phase. Serum neutrophils (NEU) and alkaline phosphatase (ALP) were collected. Following feature dimensionality reduction, ensemble radiomics score (ENs) was calculated for each region. Seven individual machine learning models and an ensemble model (ENML) were subsequently constructed. Model performance was assessed using the area under the curve (AUC), confidence interval (CI) and decision curve analysis (DCA), while interpretability was evaluated via SHAP, correlation, and restricted cubic spline (RCS) analyses. In the validation cohort, ENML achieved the highest AUC (0.799), F1 score (0.810), recall (0.870), and Brier score (0.166), and demonstrated clinical net benefit across threshold probabilities ranging from 10% to 75%. Bootstrap resampling with 1,000 iterations confirmed its stable performance in the full cohort (AUC: 0.849, 95% CI: 0.790-0.891). Spearman correlation analyses revealed strong positive associations between intratumoral and peritumoral radiomic features (r = 0.54, 0.49, and 0.51, respectively; all P < 0.001). RCS analyses identified significant nonlinear relationships between intratumor, intra-peritumoral 0.3 cm, ENs, and ALP levels and the increased incidence of poorly differentiated ESCC (Overall P < 0.05; Nonlinear P < 0.05). Additionally, both ALP and NEU exhibited significant inflection point effects in predicting poorly differentiated ESCC. SHAP analyses identified the IntraPeri3mm_wavelet.LLL_glszm_SmallAreaLowGrayLevelEmphasis and SVM as the primary contributors to the fusion model. The interpretable ENML model exhibits favorable discriminative capability for predicting poorly differentiated ESCC; however, further multicenter external validation is warranted to confirm its generalizability.