Abstract 103P
Background
RECIST (version 1.1) is used to anticipate overall survival (OS) differences among subjects in clinical trials. However, it is prone to inter-reader variability and lacks strong association with clinical outcomes. AI-derived quantifications from serial CT scans show promise in predicting OS differences beyond RECIST.
Methods
IPRO-Δ, a deep learning model trained on serial imaging data and survival outcomes of real-world aNSCLC patients, generated scores from baseline (BL), week 6 (W6), and week 12 (W12) CT scans in the control arm of MYSTIC (in which consented subjects received first-line chemotherapy). RECIST assessments were used to categorize subjects as complete or partial response (CR/PR), stable disease (SD), and progressive disease (PD) at W6 and W12, respectively. IPRO-Δ scores were thresholded into 3 similar categories (CR/PR-like/SD-like/PD-like) by matching the distribution of subjects across the RECIST categories. We used Kaplan-Meier methods and hazard ratios (HRs) to compare OS differences between categories at W6 and W12.
Results
173 and 150 subjects had BL+W6 and BL+W12 scans, respectively, for comparative analysis (Table). OS HR predictions from IPRO-Δ for responders (CR/PR-like relative to SD-like) outperformed RECIST v1.1 (CR/PR relative to SD) at W6 (HR 0.54 vs 1.10) and W12 (HR 0.61 vs. 0.91). HRs between PD vs SD did not change substantially in the comparison of PD-like vs SD-like.
Table 103PAnalysis of OS by measure and category at W6 and W12
| Week 6 | Week 12 | |||||
| n (%) | mOS (mths) | OS HR (95% CI) | n (%) | mOS (mths) | OS HR (95% CI) | |
| RECIST v1.1 Categories | ||||||
| CR/PR | 25 | 13.6 | 1.1 | 42 | 12.5 | 0.91 |
| (14.4) | (0.70, 1.77) | (28.0) | (0.59, 1.39) | |||
| SD | 127 | 12.2 | ref | 78 | 11.9 | ref |
| (73.4) | (52.0) | |||||
| PD | 21 (12.1) | 4.0 | 2.75 (1.69,4.47) | 30 (20.0) | 5.7 | 1.49 (0.93,2.39) |
| IPRO-Δ Categories | ||||||
| CR/PR-like | 25 (14.4) | 15.2 | 0.54 (0.31, 0.92) | 42 (28.0) | 13.9 | 0.61 (0.39, 0.96) |
| SD-like | 127 (73.4) | 11.2 | ref | 78 (52.0) | 10.1 | ref |
| PD-like | 21 (12.1) | 4.9 | 2.14 (1.31, 3.5) | 30 (20.0) | 5.5 | 1.38 (0.87, 2.19) |
Conclusions
IPRO-Δ provided better prognostic separation than RECIST at the same time point (particularly for prediction of responders to treatment) and may offer greater utility to predict long-term clinical benefit when used in conjunction with RECIST.
Legal entity responsible for the study
The authors.
Funding
Altis Labs, Inc., AstraZeneca.
Disclosure
O.F. Khan: Financial Interests, Personal, Invited Speaker: AstraZeneca, Novartis, Merck; Financial Interests, Personal, Advisory Board: Pfizer, Knight Therapeutics, Gilead Sciences; Financial Interests, Institutional, Invited Speaker, Local PI: Cogent Biosciences; Financial Interests, Institutional, Invited Speaker, Coordinating PI: Altis Labs, Inc.; Non-Financial Interests, Personal, Principal Investigator: Altis Labs, Inc., Breast Cancer Canada. J. Riskas, S.A. Haider, O. Samorodova, J. Hennessy, V. Sivan: Financial Interests, Personal, Full or part-time Employment: Altis Labs, Inc.; Financial Interests, Personal, Stocks/Shares: Altis Labs, Inc. F. Baldauf-Lenschen: Financial Interests, Personal, Officer: Altis Labs, Inc.; Financial Interests, Personal, Stocks/Shares: Altis Labs, Inc. K.A. Patwardhan, Q. Li, H. Ravi Prakash: Financial Interests, Personal, Full or part-time Employment: AstraZeneca.