Evaluating Imaging Artificial Intelligence (AI) Matching Real-World Digital Twins (rwDTs) Into an External Control Arm (ECA) for MYSTIC: A Phase 3 Clinical Trial in Metastatic Non-Small Cell Lung Cancer (mNSCLC)
Author(s)
Omar F. Khan, MBA, MD1, John Riskas, MSc, MBA2, Shahid Haider, PhD2, Oleksandra Samorodova, MD2, Jay Hennessy, MEng2, Vignesh Sivan, MASc2, Felix Baldauf-Lenschen, BA2, Harish RaviPrakash, PhD3, Qin Li, PhD4, Kedar Patwardhan, PhD5;
1University of Calgary, Calgary, AB, Canada, 2Altis Labs, Inc., Toronto, ON, Canada, 3AstraZeneca, Machine Learning and Artificial Intelligence, Oncology Biometrics, Waltham, MA, USA, 4AstraZeneca, Waltham, MA, USA, 5AstraZeneca, Oncology Research & Development, Waltham, MA, USA
1University of Calgary, Calgary, AB, Canada, 2Altis Labs, Inc., Toronto, ON, Canada, 3AstraZeneca, Machine Learning and Artificial Intelligence, Oncology Biometrics, Waltham, MA, USA, 4AstraZeneca, Waltham, MA, USA, 5AstraZeneca, Oncology Research & Development, Waltham, MA, USA
Presentation Documents
OBJECTIVES: ECAs created from real-world data can assist with understanding treatment effects. However, inherent limitations in data capture can result in biased or inefficient analyses lacking prognostic utility. We explore how prognostic, AI-derived spatial imaging biomarkers (SIBs) from 3D computed tomography (CT) imaging may enable matching of real-world digital twins (rwDTs) to clinical trial subjects to generate ECAs.
METHODS: Eligibility criteria from MYSTIC (NCT02453282) were applied to a real-world imaging, clinical, and outcomes database (rwICO) to identify patients receiving standard of care (SOC) chemotherapy used in the MYSTIC control arm. Over 7,000 SIBs were generated from each MYSTIC baseline (BL) CT scan and used to identify rwDTs in rwICO via cosine similarity. Kaplan-Meier analyses and hazard ratios (HRs) compared median overall survival (mOS) across trial arms: SOC, Durvalumab-only (D), Durvalumab and Tremelimumab (DT), SOC-matched ECA (ECA-SOC), SOC effect in D (ECA-D) and SOC effect in DT (ECA-DT).
RESULTS: 672 MYSTIC subjects with available BL CT scans and consent to this research were used to match rwDTs. mOS outcomes were similar between SOC and ECA-SOC (11.7 vs 10.7 months, HR 0.92, Table 1). Replacing SOC with rwDT ECA-SOC resulted in similar mOS outcomes compared to D (D vs SOC 14.4 vs 11.7 months, HR 0.89; D vs ECA-D 14.4 vs. 10.3 months, HR 0.79) as well as DT (DT vs SOC 11.5 vs. 11.7 months, HR 0.97; DT vs ECA-DT 11.5 vs 10.5 months, HR 0.98), highlighting the ability of rwDT to be used as a “matched” control in propensity score analysis.
CONCLUSIONS: rwDTs matched using AI-derived SIBs from BL CT scans generated ECAs that accurately emulated the SOC and the observed OS treatment effect. Future analysis will evaluate the improvement in statistical power and potential impact on sample size.
METHODS: Eligibility criteria from MYSTIC (NCT02453282) were applied to a real-world imaging, clinical, and outcomes database (rwICO) to identify patients receiving standard of care (SOC) chemotherapy used in the MYSTIC control arm. Over 7,000 SIBs were generated from each MYSTIC baseline (BL) CT scan and used to identify rwDTs in rwICO via cosine similarity. Kaplan-Meier analyses and hazard ratios (HRs) compared median overall survival (mOS) across trial arms: SOC, Durvalumab-only (D), Durvalumab and Tremelimumab (DT), SOC-matched ECA (ECA-SOC), SOC effect in D (ECA-D) and SOC effect in DT (ECA-DT).
RESULTS: 672 MYSTIC subjects with available BL CT scans and consent to this research were used to match rwDTs. mOS outcomes were similar between SOC and ECA-SOC (11.7 vs 10.7 months, HR 0.92, Table 1). Replacing SOC with rwDT ECA-SOC resulted in similar mOS outcomes compared to D (D vs SOC 14.4 vs 11.7 months, HR 0.89; D vs ECA-D 14.4 vs. 10.3 months, HR 0.79) as well as DT (DT vs SOC 11.5 vs. 11.7 months, HR 0.97; DT vs ECA-DT 11.5 vs 10.5 months, HR 0.98), highlighting the ability of rwDT to be used as a “matched” control in propensity score analysis.
CONCLUSIONS: rwDTs matched using AI-derived SIBs from BL CT scans generated ECAs that accurately emulated the SOC and the observed OS treatment effect. Future analysis will evaluate the improvement in statistical power and potential impact on sample size.
Conference/Value in Health Info
2025-05, ISPOR 2025, Montréal, Quebec, CA
Value in Health, Volume 28, Issue S1
Code
MSR147
Topic
Methodological & Statistical Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics
Disease
SDC: Oncology, STA: Personalized & Precision Medicine