A machine learning algorithm for automatic tumour board recommendations in prostate cancer patients

Abstract

Background and objective

Multidisciplinary tumour boards (MTBs) play a critical role in prostate cancer management, but their time-intensive nature limits accessibility. This study evaluates machine learning (ML) algorithms for automating MTB recommendations in prostate cancer patients, focusing on multi-label classification for diagnostic and therapeutic decisions.

Methods

A retrospective dataset of 1929 MTB recommendations from 2020 to 2024 was used for model development and validation at a single academic centre. Three ML algorithms—Decision Tree, Random Forest and K-Nearest Neighbours (KNN)—were trained to predict recommendations for PSMA-PET, conventional imaging, active surveillance and local therapy (radical prostatectomy or radiotherapy). Model performance was assessed using accuracy, precision, recall and F1-score.

Key findings and limitations

The Random Forest model achieved the highest overall accuracy (66.3%, 95% CI 61.7–71%) and showed stable performance across most outcome categories. Predictions for local therapy were highly accurate (F1-score: 0.99), but model performance was lower for less frequent recommendations such as PSMA-PET and active surveillance, reflecting class imbalance and recent guideline changes. Limitations include moderate overall accuracy, retrospective single-centre design and the need for extensive manual data preprocessing. In addition, a high proportion of patients were eligible for multiple treatment options, which may limit the discriminatory value of certain outcomes.

Conclusions and clinical implications

This study demonstrates the potential of ML to replicate MTB decision patterns in prostate cancer with reasonable accuracy. However, the current model requires further optimization before it can be considered for clinical application. It should be regarded as a proof-of-concept that highlights both the opportunities and the challenges of algorithm-based decision support in oncology. Future work should focus on improving model performance through multi-institutional data, prospective validation and continuous adaptation to evolving clinical guidelines.