Machine learning prognosis model for locally recurrent rectal cancer patients after radioactive

To develop and validate a multiscale radiomics prognostic tool for accurately predicting local control (LC) and overall survival (OS) in locally recurrent rectal cancer (LRRC) patients underwent CT-guided radioactive 125I seed implantation (RISI). 189 LRRC patients who treated with RISI were eligible for exploratory retrospective study and randomly divided into training and validation sets. Intra-and peri-tumoral handcrafted radiomics features (RFs) selection was performed using the univariate analysis and LASSO-Cox model. The deep learning RFs were also performed same procedures. The random survival forest (RSF) and Cox hazard regression (CHR) prognostic models were fitted with bootstrapping resampling and comprehensively evaluated by the concordance index (C-index), integrated brier score (IBS), and time-dependent area under the curve (tAUC). Among all peritumoral radscores (RS), the RSperi1mm and RSperi4mm demonstrated the best prediction for LC for OS in the validation set, respectively. The addition of deep learning radscores can also improve prediction efficiency. The combined RSF model demonstrated robust performance compared to CHR model for LC prediction, achieving a C-index (95%CI) of 0.78 (0.74–0.84) and an IBS of 0.13 (0.12–0.14). Similar results were observed in predicting OS with a C-index of 0.76 (0.75–0.77), an IBS of 0.11 (0.10–0.12). According to the RSF model predictions, the LRRC patients were significantly dichotomized into two different prognostic groups (p < 0.001). The RSF model could provide more accurate LC and OS prediction and remarkable prognostic stratification than the CHR model for LRRC patients after RISI treatment.