ESTRO 2024 - Abstract Book

S4562

Physics - Machine learning models and clinical applications

ESTRO 2024

This study highlighted the feasibility of employing an automatable ML-based tool for the replanning of patients at risk of low deliverability. The replanning process, focusing on specific parameters (ASC and MUlimit), effectively produces plans with improved deliverability while maintaining consistent quality. This approach holds the potential to enhance the quality of care delivered to patients undergoing radiotherapy.

Keywords: machine learning, PSQA, autoplanning

References:

[1] Lambri et al., Multicentric evaluation of a machine learning model to streamline the radiotherapy patient specific quality assurance process

2782

Digital Poster

Autoplanning for mediastinal lymphoma: patient specific solution with machine learning approach

Christian Fiandra 1 , Stefania Zara 2 , Greta De Giorgi 1 , Celislami Ada 1 , Ilaria Bonavero 1 , Sara Bartoncini 3 , Chiara Cavallin 3 , Umberto Ricardi 1 , Mario Levis 1 1 University of Turin, Oncology, Turin, Italy. 2 University of Turin, Doctoral school, Turin, Italy. 3 Città della Salute e della Scienza, Oncology, Turin, Italy

Purpose/Objective:

To use clinical characteristics, i.e. age, gender or previous disease (involving heart and/or lung) for predicting a patient specific solution in terms of assigned priorities to the desired trade-offs between all objectives for automatic planning.

Material/Methods:

Eleven clinical properties including age, gender, type of chemotherapy (ABVD 4/ABVD 6/RCHOP 6), total dose of anthracyclines (200/300 mg/m2), diabetes (yes/no), hypertension (yes/no), hypercholesterolemia (yes/no), familiarity for cardiovascular disease (yes/no), familiarity for breast neoplasia (yes/no), obesity (yes/no) and smoke (yes/no/past) were collected for 100 patients. Four different classes of patients based on previous clinical input were recognized by clinicians: i) low risk (LR), ii) heart risk (HR), iii) second cancer risk (SCR), iv) second cancer risk with cardio congestive heart failure (SCR_CHF). Two classifiers were trained and evaluated for ability to correctly predict the clinician’s patient class: random forest (RF) and constrained logistic regression (LR) classifiers. Model building was performed with nested cross-validation with an outer and an inner loop and prediction performance was assessed by calculating mean values and standard errors of the Accuracy, area under the receiver-operator characteristic curve (AUC) and Cohen's kappa coefficient (κ) for the ten outer loop models. Importance of included features for predictions were investigated to better understand clinician’s choices and reported in figure for both classifiers.