68/100

Moderate evidence

Cite
More
EmailLinkedInX

Study of 454 prostate cancer patients finds AI can predict radiotherapy side effects

Study title: Machine learning explainability of acute toxicity in salvage prostate cancer radiotherapy: A multicentric investigation

Evidence type
Controlled or validated comparison
Sample
n = 454
Sources
1 sourcePeer-reviewed
Published
2 min read

Study title: Machine learning explainability of acute toxicity in salvage prostate cancer radiotherapy: A multicentric investigation

Researchers built a machine learning model to predict which prostate cancer patients will develop grade 2 or higher bowel or urinary side effects from salvage radiotherapy after prostatectomy. They used data from 454 patients treated at three Italian centres.

Read the study abstract

To develop and validate an explainable machine learning (ML) model for early prediction of grade ≥ 2 acute gastrointestinal (GI) and genitourinary (GU) toxicities in prostate cancer patients treated with salvage radiotherapy (SRT) post-prostatectomy. Patient-, surgical- and treatment-related data from 454 patients treated with SRT in three Italian centers were collected and analyzed. 271 patients from two centers served as the training and internal validation set, while 183 patients from the third institution we…

Machine learning explainability of acute toxicity in salvage prostate cancer radiotherapy: A multicentric investigation
"20250808 Intensive care unit bed" by Ismael Olea, CC BY 4.0.

The gain

In a study of 454 post-prostatectomy patients at three Italian centres, a machine learning model predicted grade 2 or higher genitourinary toxicity from salvage radiotherapy with an AUC of 0.855 on an external test set of 183 patients.

In brief

On an external test set of 183 patients from the third centre, the model reached an AUC of 0.706 for gastrointestinal toxicity and 0.855 for genitourinary toxicity. Predictions held up better for urinary than bowel effects, and the authors say the model could help identify high-risk patients early and plan individual treatment.

Main points

  1. The study included 454 patients who received salvage radiotherapy after prostatectomy at three Italian centres.
  2. 271 patients from two centres formed the training and internal validation set; 183 patients from the third centre formed the external test set.
  3. Grade 2 or higher toxicity occurred in 28.2% of patients for the gastrointestinal tract and 21.8% for the genitourinary tract.
  4. Internal validation AUCs were 0.784 for gastrointestinal and 0.911 for genitourinary toxicity.
  5. External test AUCs fell to 0.706 for gastrointestinal and 0.855 for genitourinary toxicity.
  6. The strongest predictors were surgical technique, CTV to PTV margin, type of androgen deprivation therapy, extent of lymphadenectomy and dose fractionation.

The full story

The team used a light gradient boosting machine algorithm trained on patient, surgical and treatment data. SHAP analysis ranked which features drove predictions, making the model explainable.

Performance dropped on the independent centre compared with internal validation, most clearly for gastrointestinal toxicity (AUC 0.784 to 0.706). Genitourinary predictions stayed higher (0.911 to 0.855). The calibration curve showed high agreement between predicted and observed risks.

Across both toxicity types, the most significant determinants were surgical technique, the CTV to PTV margin, the type of androgen deprivation therapy, the extent of lymphadenectomy and dose fractionation. The authors conclude that combining patient, surgical and radiotherapy factors can help minimise adverse effects.

Sources

1 on file
  1. Physica Medicadoi:10.1016/j.ejmp.2026.107215
    Peer-reviewed2026-11-19

Talk it over

Three questions to start on, two more if you have time
The external test set of 183 patients came from the third centre, and AUCs fell from 0.784 to 0.706 for gastrointestinal toxicity. What does that drop suggest about using the model at a new hospital?

Start with a question

Study of 454 prostate cancer patients finds AI can predict radiotherapy side effectshttps://truvace.com/item/machine-learning-explainability-of-acute-toxicity-in-salvage-prostate-cancer-rad

  1. The external test set of 183 patients came from the third centre, and AUCs fell from 0.784 to 0.706 for gastrointestinal toxicity. What does that drop suggest about using the model at a new hospital?

    Answer in the thread
  2. Surgical technique, CTV to PTV margin, type of androgen deprivation therapy, extent of lymphadenectomy and dose fractionation were top predictors. Which of these would you want clinicians to change first if the model flagged a patient as high risk?

    Answer in the thread
  3. If you were a patient after prostatectomy told the model gave a high risk of grade 2 or higher toxicity, what would you ask your care team before agreeing to salvage radiotherapy?

    Answer in the thread
If you have more time
  • Look up what grade 2 or higher gastrointestinal and genitourinary toxicity means for daily life. How would that change how much accuracy you want from the model?

    Answer in the thread
  • Which patients at other centres, with different surgery or radiotherapy practices, might the 454 patients not represent?

    Answer in the thread

What readers say