Session Item

After (chemo)radiotherapy, patients with rectal cancer who achieve pathological complete response (pCR) have a significantly better prognosis and can avoid surgery, thus retaining a higher quality of life. The likelihood of pCR increases with the dose of radiation, but so do side effects. Therefore, it is of particular interest to predict which patients will achieve a near complete response, so these patients might be considered for treatment intensification. In this study, we have trained a model to predict near-complete pathological response using data from different clinics.

We used data extracted from two European radiation oncology clinics to develop a prediction model for near complete response (defined as a residual tumour diameter < 2cm and pT1-2N0M0). The variables included in the model based on availability and expert knowledge were: T, N and M stages, mesorectal fascia involvement, WHO status, tumour volume, distance to the anal junction, treatment, age, and gender. We imputed missing values using multiple imputation with chained equations for the training dataset. We trained a Bayesian network (BN) applying the nonparametric bootstrap using the hill climbing algorithm and the Bayesian Dirichlet equivalent score without sharing data, using federated learning. We asked three radiation oncologists to propose the structure of the BNs and then compared the performance of the BN in terms of area under the ROC curve (AUC) when the structure was determined by experts, when it was learnt from data, and when expert-provided structure was fine-tuned based on data. We split the data 70/30 for training and testing.

In total, we included data from 1325 patients, 77 of which achieved near complete response. The training and testing AUCs for the BN whose structure was learnt from data were 66% and 58% percent respectively. The BNs whose structure was determined by experts, achieved average AUCs of 76% and 54% for training and testing respectively. Finally, the BNs with expert provided structures fine-tuned with data achieved average AUCs of 66% and 66% for training and testing respectively. See Table 1 for more detailed results.

Expert-provided structures fine-tuned with data resulted in higher testing AUCs. Near complete response is a difficult outcome to predict, but our model could be used as an additional tool when determining a patient’s treatment plan as long as the uncertainty of the predictions is taken into account. Increasing the sample used for training would likely lead to performance improvements. Fortunately, the federated learning infrastructure we have set up allows for a straightforward addition of new clinics while circumventing issues with data sharing.