178P - Explainable radiomics, machine and deep learning models to predict immune-checkpoint inhibitor treatment efficacy in advanced non-small cell lung cancer patients

Background

IO revolutionized NSCLC treatment paradigm, but only a relatively small proportion of pts shows benefits. Since available biomarkers demonstrated only limited predictive efficacy, novel models are urgently needed. Radiomics, through machine learning (ML) and deep learning (DL) techniques, allows to extract features from raw images to build predictive models. This study aimed to develop and test 2 type of radiomics models in predicting IO efficacy on advanced NSCLC, one ML-based, other through an end-to-end DL pipeline.

Methods

We collected CT scans and RWD from 295 consecutive pts with advanced NSCLC pts receiving any-line anti-PD(L)1 therapy either alone or in combination with CHT at our Institution from 23/04/2013 to 12/05/2022 within APOLLO 11 study. ML workflow consisted of image preprocessing, feature extraction, 3 step feature selection, correlation, algorithm training and evaluation of 6 ML classifiers. ML models were trained using radiomics +/- RWD. Models were evaluated on independent test set and SHAP values used to explain model predictions. We compared developed model with novel DL end-to-end model, a 3D convolutional neural network.

Results

15 image features were selected to build radiomic model, achieving ACC of 0.61 and AUC of 0.58. Combining it with additional RWD features, accuracy and AUC of LR reached 0.71 and 0.76. SHAP analysis for best-performing models showed RW (ECOG PS, therapy line, concomitant CHT, PDL1) and radiomics features (tumor shape and intensity/distribution of gray-level values) that most influenced models. DL model achieved an ACC of 0.69 and a Loss value of 0.773. A high proportion of higher GLSZM and a high value of large area size zones were associated to a higher response to therapy.

Conclusions

The developed and validated ML-based model includes radiomic and RW features, demonstrating ability to predict IO efficacy and potential future applicability for treatment selection. We firstly demonstrated that a radiomics-based DL model outperform the all radiomics ML-based models highlighting the importance of this approach in image data. SHAP analysis provided valuable insights to explore features that most influenced the models' predictions

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Fondazione IRCCS Istituto Nazionale Tumori di Milano.

Funding

Has not received any funding.

Disclosure

F.G.M. De Braud: Financial Interests, Personal, Advisory Role: Roche. A. Prelaj: Financial Interests, Funding: Roche, AstraZeneca, BMS. All other authors have declared no conflicts of interest.