1362P - Evaluation of imaging-based prognostication (IPRO) for advanced non-small cell lung cancer (aNSCLC) using deep learning applied to computed tomography (CT)

Background

In aNSCLC, pretreatment CTs help determine TNM stage, an important prognostic factor in treatment decisions, as well as determination of clinical trial eligibility/stratification. However, CTs contain prognostic features beyond TNM stage that are not easily quantifiable or reproducible. We examine an imaging-based artificial intelligence algorithm to predict overall survival (OS) compared to TNM staging.

Methods

IPRO, a fully automated deep learning system extracting 5,341 features from chest CT (including tumor burden, and body/cardiovascular composition), was trained on an independent dataset to generate mortality risk scores. We retrospectively evaluate IPRO compared to TNM stage to predict OS from 2,894 subjects diagnosed with aNSCLC at 17 cancer centers in Alberta from 2008 - 2017. Kaplan-Meier and Cox regression methods are used to estimate covariate effects on OS including a model stratified by stage. We evaluate Harrell’s c-index and AIC as measures of goodness of fit comparing models with IPRO, TNM stage, and IPRO + stage (full model). Smaller values for AIC and larger values of the c-index indicate better models.

Results

Median OS was 4.3 months, 80.3% of patients were stage IV at diagnosis, and 52.6% were male. Hazard ratios for OS by IPRO within stage were 1.39 (IIIB) and 1.34 (IV), suggesting IPRO associations were similar by stage group. The c-index for IPRO was larger than for TNM staging. Removing IPRO from the full model resulted in a larger deterioration in AIC than removal of TNM staging.

Conclusions

IPRO provides better discrimination of OS compared to TNM stage based on the c-index. The addition of stage in the full model negligibly improves goodness of fit. This suggests IPRO is a better predictor of OS than stage in this population and may provide greater utility as a prognostic factor in clinical research. Future analyses will include biomarker status and incorporation of other anatomic regions into the model. Table: 1362P

Summary of goodness of fit statistics

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

Altis Labs, Inc.

Funding

Digital Supercluster, Altis Labs, Inc.

Disclosure

O.F. Khan: Financial Interests, Personal, Invited Speaker: Merck, Gilead, Novartis; Financial Interests, Personal, Advisory Board: Pfizer, Knight Therapeutics, AstraZeneca. J. Riskas:Financial Interests, Personal, Financially compensated role: Altis Labs, Inc.; Financial Interests, Personal, Stocks/Shares: Altis Labs, Inc. S.A. Haider, O. Samorodova, J. Hennessy, V. Sivan, D. Shah: Financial Interests, Personal, Full or part-time Employment: Altis Labs, Inc.; Financial Interests, Personal, Stocks/Shares: Altis Labs, Inc. A. Mitchell: Financial Interests, Personal, Full or part-time Employment: Altis Labs, Inc.; Financial Interests, Personal, Stocks or ownership: Altis Labs, Inc. F. Baldauf-Lenschen: Financial Interests, Personal, Ownership Interest: Altis Labs, Inc.; Financial Interests, Personal, Full or part-time Employment: Altis Labs, Inc. S. Raman: Financial Interests, Institutional, Research Grant: AstraZeneca, Knight Therapeutics; Financial Interests, Personal, Speaker, Consultant, Advisor: Bayer, AstraZeneca, Tersera, Sanofi, Verity Pharma. N.B. Leighl: Financial Interests, Personal, Other, CME/independent lectures: MSD, BMS; Financial Interests, Personal, Invited Speaker, independent lectures: Novartis, Takeda; Financial Interests, Personal, Invited Speaker, independent CME lecture: BeiGene; Financial Interests, Personal, Invited Speaker, Independent CME lectures: Janssen; Financial Interests, Institutional, Research Grant: AstraZeneca, Lilly, MSD, Pfizer, Janssen, Inivata/NeoGenomics, Novartis; Other, Travel funding to deliver independent CME lectures: AstraZeneca; Other, Travel funding to deliver independent CME lecture: Janssen, MSD, Sanofi, Guardant Health. All other authors have declared no conflicts of interest.