Abstract 5619: Radiomics-based machine learning models to predict progression and biomarker status in non-small cell lung cancer (NSCLC) patients treated with immunotherapy

Abstract

Abstract Background: Radiomics is an emerging tool that involves the extraction of high-throughput features from medical images. These quantitative values can be used to develop predictive models for clinical characteristics and treatment outcomes. We evaluated radiomic features-based models as imaging biomarkers in NSCLC patients. Methods: 71 patients with NSCLC treated with immunotherapy who had pretreatment CT chest with contrast were retrospectively evaluated. The main tumor and 1cm-thick peritumoral space surrounding the tumor were manually segmented using LIFEx software (IMIV/CEA, Orsay, France) by four physicians. Of 255 radiomic features collected, those with >0.4 of Fleiss’ kappa coefficient were selected. The Random Forest (RF) algorithm with mixed effects was used to develop multi-reader models and assess feature importance. The dataset was divided into a training set (75%) and a test set (25%). Bootstrapping with 1,000 iterations was conducted to estimate the model performance. Durable disease control was defined as having no progression of diseases per RECIST 1.1 up to 24 weeks from starting immunotherapy. Results: Among 71 patients, 35 (49.3%) are female and 36 (50.7%) are male. The median age was 66. 48 (67.6%) adenocarcinoma, 13 (18.3%) squamous cell carcinoma, and 10 (14.1%) other histologic types were included. 22 radiomic features were included based on importance in the prediction models from both the tumor and peritumoral space. Each model is trained to predict patients’ durable disease control, TTF1 expression, PD-L1 expression, histology (adenocarcinoma or not), and Neutrophils Lymphocyte Ratio (NLR; greater than 5 or not) status. The statistical results from the models to predict clinical outcomes are shown in Table. Conclusion: The radiomic features-based models lack accuracy in predicting clinical characteristics and outcomes. Further validation with larger cohorts is warranted. Statistics of radiomics-based models in predicting clinical characteristics and treatment outcomes Durable Disease Control(Yes/No)(n=64) TTF1 expression(Yes/No)(n=62) Histology(Adeno/Other)(n=71) NLR(>=5/<5)(n=71) PD-L1 expression(Yes/No)(n=52) Patient Number(%) 33 (51.56%)/31 (48.44%) 37 (59.68%)/25 (40.32%) 48 (67.61%)/23 (32.39%) 28 (39.44%)/43 (60.56%) 35 (67.31%)/17 (32.69%) Sensitivity (95% CI) 0.63 (0.58, 0.72) 0.62 (0.56, 0.74) 0.69 (0.56, 0.82) 0.55 (0.47, 0.61) 0.57 (0.48, 0.65) Specificity (95% CI) 0.46 (0.37, 0.52) 0.68 (0.58, 0.76) 0.22 (0.12, 0.34) 0.60 (0.56, 0.68) 0.36 (0.30, 0.45) Positive Predictive Value(95% CI) 0.52 (0.49, 0.57) 0.44(0.37, 0.60) 0.62 (0.59, 0.64) 0.69 (0.67, 0.74) 0.72 (0.68, 0.77) Negative Predictive Value(95% CI) 0.58 (0.54, 0.63) 0.79 (0.74, 0.88) 0.28 (0.22, 0.32) 0.46 (0.39, 0.51) 0.25 (0.21, 0.28) Citation Format: Jisang Yu, Yury Velichko, Hyeonseon Kim, Moataz Soliman, Nicolo Gennnaro, Leeseul Kim, Youjin Oh, Trie Arni Djunadi, Jeeyeon Lee, Liam Il-Young Chung, Sungmi Yoon, Zunairah Shah, Soowon Lee, Cecilia Nam, Timothy Hong, Rishi Agrawal, Pascale Aouad, Young Kwang Chae. Radiomics-based machine learning models to predict progression and biomarker status in non-small cell lung cancer (NSCLC) patients treated with immunotherapy. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5619.