Abstract

Matteo Pepa¹, Mattia Zaffaroni¹, Stefania Volpe^1,2, Giulia Marvaso^1,2, Johannes Lars Isaksson¹, Simona Barzaghi³, Federica Benigni³, Marta Callegari³, Alessia Gismundi³, Francesco Maria La Fauci⁴, Giulia Corrao^1,2, Matteo Augugliaro¹, Federica Cattani⁴, Guido Baroni³, Elena De Momi³, Roberto Orecchia⁵, Barbara Alicja Jereczek-Fossa^1,2

¹IEO European Institute of Oncology IRCCS, Division of Radiation Oncology, Milan, Italy; ²University of Milan, Department of Oncology and Hemato-Oncology, Milan, Italy; ³Politecnico di Milano, Department of Electronics Information and Bioengineering (DEIB), Milan, Italy; ⁴IEO European Institute of Oncology IRCCS, Medical Physics Unit, Milan, Italy; ⁵IEO European Institute of Oncology IRCCS, Scientific Direction, Milan, Italy

In the last decades, radiotherapy (RT) treatments have become safer and more effective, allowing for dose escalation to the target volume without jeopardizing the sparing of surrounding organs at risk (OARs). However, effective toxicity prediction tools are essential in the era of tailored treatments. The purpose of the study is to test machine learning (ML)-based predictive models of toxicity in prostate cancer (PCa) patients (pts) treated with ultra-hypofractionated RT regimens.

Two cohorts of 61 and 186 non-metastatic low-intermediate risk PCa pts (from AIRC IG-13218 prospective trial and “Give me Five” retrospective trial, respectively), who underwent ultra- hypofractionated RT (35 Gy/5 fractions), were considered (Ethics Committee Notification UID 2410). Dosimetric and clinical features were used to train different ML models to predict genitourinary (GU) and gastrointestinal (GI) acute toxicities. The area under the receiver operating characteristic curve (AUC) was used to compare the model performances.

Separate analyses of the three groups (61, 186 and 247) were carried out. Bagged trees outperformed all the others on the 61 cohort (Fig. 1a), with an AUC of 0.75, while SVM resulted the best algorithm on the 186 (Fig 1b) and 247 cohort (Fig 1c), with an AUC of 0.94 and 0.66, respectively. Overall, the best performing algorithm was LR, since, among the 8 best results, 4 were achieved with this one. The most predictive features were found to be T-stage, age, OARs volume and DVH punctual values. No significant correlation with the outcome was found for sub-areas under the DVH. Overall, the models achieved better AUC values when the two subgroups of pts were considered separately. The study presents some limitations, such as the relatively low occurrence of toxicity events, often resulting in a scarce capability of identifying true positives, and low AUC values in certain configurations.

Figure 1

In the era of personalised medicine and tailored treatments, a model that accurately predicts toxicity could represent a useful clinical tool for better patients’ selection. The study shows that the performance of the predictive models is highly dependent on the choice of classifier, features and training/testing sets. Differences between groups can be due to the different characteristics of trials. Further analyses should focus on features and classifiers selection, to improve the performances and generalizability of the models.