Predicting near-complete pathological response to (chemo)radiotherapy in patients with rectal cancer: A federated...

Organ preservation in patients with locally advanced rectal cancer has attracted interest due to improved quality of life and functional outcomes compared with total mesorectal excision. Hence, patients who achieve clinical complete response (cCR) after (chemo)radiotherapy are offered a watch-and-wait strategy. Those who are likely to fall short of the strict criteria of cCR and only achieve near complete response (NCR) might benefit from radiation boosting to avoid surgery.

To develop a prediction model that estimates the probability of NCR trained on data from different clinics.

We used data from two clinics (Maastro and Gemelli) on 1305 patients. We set up and used a federated learning infrastructure to leverage patient data from the clinics without transferring it. We used Bayesian networks for their capacity to combine expert knowledge with data and their ability to handle missing data. In this article, we propose a novel federated learning algorithm for Bayesian networks. In addition, we explore different approaches to handle missing data and train the models, combining expert elicited structures and those learnt from data. The discriminative performance of the models is reported using the area under the ROC curve (AUC).

The model trained on data from both clinics and a structure learnt from data performed well (AUC 0.77). When using a structure elicited from an expert, the performance of the model decreased (AUC 0.68). Fine-tuning the expert structure with data led to a middle ground performance (AUC 0.72). Models trained on data from a single clinic failed to generalize when tested on data from the other clinic (AUCs 0.50 and 0.59).

The model trained with federated learning showed good discriminative performance, which indicates that it could be useful to identify which patients with rectal cancer would benefit the most from a radiotherapy boost. This study shows that federated learning has the potential to lead to better models by allowing access to more data.