A Multicenter Machine Learning-Based Predictive Model of Acute Toxicity in Prostate Cancer Patients Undergoing Salvage Radiotherapy (ICAROS Study)

Abstract

Background: This study aimed to develop a predictive model for acute gastrointestinal (GI) and genitourinary (GU) toxicity in prostate cancer patients treated with salvage radiotherapy (SRT) post-prostatectomy, using machine learning techniques to identify key prognostic factors.

Methods: A multicenter retrospective study analyzed 454 patients treated with SRT from three Italian radiotherapy centers. Acute toxicity was assessed using Radiation Therapy Oncology Group criteria. Predictors of grade ≥ 2 toxicity were identified through Least Absolute Shrinkage and Selection Operator (LASSO) regression and Classification and Regression Tree (CART) modeling. The analyzed variables included surgical technique, clinical target volume (CTV) to planning target volume (PTV) margins, extent of lymphadenectomy, radiotherapy technique, and androgen-deprivation therapy (ADT).

Results: No patients experienced grade ≥ 4 toxicity, and grade 3 toxicity was below 1% for both GI and GU events. The primary determinant of acute toxicity was the surgical technique. Open prostatectomy was associated with significantly higher grade ≥ 2 GI (41.8%) and GU (35.9%) toxicity compared to laparoscopic/robotic approaches (18.9% and 12.2%, respectively). A CTV-to-PTV margin ≥ 10 mm further increased toxicity, particularly when combined with extensive lymphadenectomy. SRT technique and ADT were additional predictors in some subgroups.

Conclusions: SRT demonstrated excellent tolerability. Surgical technique, CTV-to-PTV margin, and treatment parameters were key predictors of toxicity. These findings emphasize the need for personalized treatment strategies integrating surgical and radiotherapy factors to minimize toxicity and optimize outcomes in prostate cancer patients.

Keywords: LASSO regression; acute toxicity; androgen-deprivation therapy; gastrointestinal toxicity; genitourinary toxicity; machine learning; planning target volume; predictive model; prostate cancer; salvage radiotherapy.

Figure 1

Overall flowchart of the study.

Figure 2

Variable selection using the least absolute shrinkage and selection operator (LASSO) binary logistic regression model. (a,c) Binomial deviation as a function of the tuning penalization parameter λ. The numbers at the top of the chart are the remaining coefficients at the corresponding log lambda values. (b,d) Correlation matrix heatmaps.

Figure 3

ROC curves for the training and validation cohorts of the five different machine algorithms.

Figure 4

Calibration curves for the five ML models on the validation set. The dotted line represents the perfect calibration curve; i.e., the predicted probability matches the true probability perfectly.

Figure 5

Predictive model of acute gastrointestinal toxicity (grade ≥ 2) obtained by the Decision Tree algorithm. The values highlighted close to the red bars denote the percentages of acute toxicity incidence.

Figure 6

Predictive model of acute genitourinary toxicity (grade ≥ 2) obtained by the Decision Tree algorithm. The values highlighted close to the red bars denote the percentages of acute toxicity incidence.