ESTRO 2024 - Abstract Book

S2935

Interdiscplinary - Other

ESTRO 2024

1871

Digital Poster

Generative artificial intelligence for toxicity detection in radiotherapy

Nicola Dinapoli 1 , Francesco Esposito 2 , Martina D'Antoni 2 , Vito Lanzotti 2 , Luca Tagliaferri 1 , Mariangela Massaccesi 1 , Francesco Miccichè 3 , Vincenzo Valentini 1 , Maria Antonietta Gambacorta 1 1 Fondazione Policlinico Universitario Agostino Gemelli IRCCS, Radioterapia, Roma, Italy. 2 KBMS Data Force, KBMS, Roma, Italy. 3 Ospedale Isola Tiberina Gemelli Isola, Radioterapia Oncologica, Roma, Italy

Purpose/Objective:

Monitoring toxicity in patients undergoing radiotherapy is a crucial aspect of follow-up and daily patient treatment checks. In many centres today, the recording of patient visits and related reports is often still done using descriptively entered data, without a structured approach to collection. Certainly, in many centres, there still is a need to access data from previously treated patients, whose details have been stored in unstructured textual datasets. The aim of this work is to develop a method for analysing unstructured textual data, to derive databases containing data organised into predefined categories (such as toxicity levels) to facilitate analysis without the need to manually consult individual files. Our primary goal is to extract information regarding cutaneous and mucosal toxicity in head and neck cancer patients from the examination of reports using an innovative approach that use generative artificial intelligence by a Large Language Model (LLM) able to extract structured information hidden within medical examination reports. Patients’ files previously analysed in a protocol that got the authorization from internal ethical committee have been used in an OpenAI GPT3.5 based application, as the LLM, hosted in the Microsoft Azure cloud and its European version, and a technological stack based on Python has been developed to tackle this task. To ensure patient privacy, the use of the Microsoft Presidio tool has been integrated into the pipeline to remove any personally identifiable information (PII) from the medical reports. The process of toxicity extraction has been refined through a hybrid approach. Initially, a team of experts manually extracted toxicity-related data from a sample of patients' training data, using CTCAE v5 criteria definition, thereby creating a customized training dataset. These training data have been instrumental in dynamically generating "few-shot" examples to use as prompts for the automatic extraction of toxicity. At the core of this approach lies the "function calling" technique of Azure OpenAI GPT3.5, which enables specific functions within the model to be invoked. Using this functionality, we have guided the model through prompts, demonstrating how to accurately identify and extract toxicity information from the reports. Material/Methods:

Results:

A series of 200 patients’ record has been splitted into a training set (40 patients) and a testing set (160 patients). The total number of reports was 380. Detected toxicity grade are summarized in table 1. The records of treatments visits for acute toxicity monitoring have been analysed, trying to address the maximum grade of acute toxicity for skin and oropharyngeal mucosa seen during the radiotherapy course. At the end of the analysis 141 patients over 160 (88,1%) have been classified correctly with respect to the manual classification, in both skin and oral mucosa toxicity, 9 (5,6%) patients showed only one of the two classifications as correct, 10 patients (6,3%) have been misclassified.