(P 290) ARTIFICIAL INTELLIGENCE-BASED COMPARISON OF ADULT AND ADOLESCENT AND YOUNG ADULTS (AYA) SARCOMAS: A PILOT STUDY

Objective: The AYA (Adolescent and Young Adult) group represents a population of patients aged between 16 and 39 years of age. Sarcomas represent 8% of all cancer affecting this age group. The increasing incidence of cancer in AYA, along with a lack of improvement in outcomes over the past decades, has raised worldwide concern. We conducted a retrospective monocentric study utilizing the innovative tool of artificial intelligence (AI), in particular the development and validation of a multimodal machine learning method that could provide a prognostic model in AYA patients affected by sarcomas. Our overall aim is to explore the clinical, radiomics and pathological features that might be enlightening of the disease outcome. We report here on a pilot study of AYA sarcoma patients treated at our institute.

Methods: We retrospectively collected data on all 645 patients affected by sarcomas and treated in our institute from 01/01/2015 to 31/12/2022. We excluded patients with histotypes rare in the AYA population, patients with diagnoses made before the time range of our study, and patients for whom there were missing data. We focused on extremity soft tissue sarcomas (STS), being the most common presentation site, to improve consistency of the analysis. Patients were required to have had magnetic resonance imaging (MRI) to avoid discrepancies in the radiomics features from different imaging modalities. We examined clinical characteristics including demographic variables, diagnostic delay (if present), disease stage and characteristics, treatments, and radiomic features extracted from MRI, with the aim of finding statistically significant differences between AYA and non-AYA patients. To incorporate histopathological images in our multimodal analysis, AI techniques were used to extract relevant histological latent features. Finally, clinical, radiological and histopathological characteristics were used to match similar patient profiles in the two groups by logistic regression propensity scores. Matched patients’ disease-free survival was described using Kaplan-Meier methodology.

Results: Our pilot study involved an initial cohort of 22 patients. The final cohort includes 13 non-AYA and 9 AYA patients, from whom clinical, demographic and pre-treatment MRI data were collected. For 10 patients, comprising 3 AYA and 7 non-AYA, histopathological images were also retrieved. All patients analyzed were treated similarly at the same center with similar treatments for localized cancer at the time of diagnosis. Our analysis revealed no statistically significant correlation between clinical and radiological features that could explain the differences in prognosis between the two groups, arguably due to the small cohort size of our pilot analysis. The AI analysis, per Kaplan-Meier, demonstrated that AYA patients had a worse prognosis than non-AYA patients (p < 0.05), confirming by deep machine learning for the first time, to our knowledge, what is observed in clinical practice (Figure 1- 2).

Conclusion: Our study highlights the unique challenges and poorer prognosis observed in certain sarcoma histotypes among the AYA population compared to adults. Despite the limitation of our pilot study, we confirmed our ability to conduct an AI-based analysis of AYA sarcoma patients. Our findings provide valuable insights into the different prognosis characterizing these two groups. While AI holds promise in uncovering hidden characteristics, larger, prospective studies are necessary to validate these findings. We will expand our analysis to a larger number of patients, and also will incorporate biological markers to our model, in order to explore therapeutic targets further.