J
AI for clinical trial design, clinical applications and future developments
Jean-Philippe Vert, PhD
Owkin
Conflict of Interest
DISCLOSURE
I have the following potential conflict(s) of interest to report
Type of affiliation / financial interest Name of commercial company
Employee and shareholder Owkin
President and shareholder Bioptimus
>90% of drugs fail in clinical trial
Probability of success for oncology vs non-oncology drugs in the U.S. in the different development phases from 2011 to 2020
AI can optimize every clinical phase
Phase 1 / 2
Phase 3
(Federated) external control arms
Prognostic & predictive enrichment
Phase 2 / 3
Data-driven covariate adjustment
Gain early estimates of efficacy for single-arm phase I/II clinical trials to inform phase transition decisions
Inform trial recruitment and better define patient populations
Better measure treatment effect by increasing clinical trial statistical power without adding more patients
AI can optimize every clinical phase
Phase 1 / 2
Phase 3
(Federated) external control arms
Prognostic & predictive enrichment
Phase 2 / 3
Data-driven covariate adjustment
Gain early estimates of efficacy for single-arm phase I/II clinical trials to inform phase transition decisions
Inform trial recruitment and better define patient populations
Better measure treatment effect by increasing clinical trial statistical power without adding more patients
Motivation
Standard RCT (phase 3)
External control arm (ECA)
Phase 1/2
RWD, other CT
External control arm (ECA)
Phase 1/2
RWD, other CT
Need to correct for lack of randomization
Inverse Probability of Treatment Weighting
(IPTW)
2) Use weights to compare arms
10
But… data are siloed
FedECA overcomes the challenge of pooling data
Real data: FedECA detects effect that individual study don’t
Center | n GEN+NAB | n FOLFIRINOX | HR (95% CI) | p |
FFCD | 55 | 92 | 1.13 (0.80, 1.61) | 0.480 |
IDIBGI | 144 | 22 | 0.74 (0.40, 1.37) | 0.342 |
PANCAN | 94 | 101 | 0.80 (0.55, 1.16) | 0.234 |
FedECA | 293 | 215 | 0.75 (0.62, 0.92) | 0.005 |
Pancan
FedECA
AI can optimize every clinical phase
Phase 1 / 2
Phase 3
(Federated) external control arms
Prognostic & predictive enrichment
Phase 2 / 3
Data-driven covariate adjustment
Gain early estimates of efficacy for single-arm phase I/II clinical trials to inform phase transition decisions
Inform trial recruitment and better define patient populations
Better measure treatment effect by increasing clinical trial statistical power without adding more patients
Motivation
What if we could accelerate the enrollment process and include the right patients in your phase II/III trial to increase success rates?
Prognostic enrichment in heart failure
Risk profile in past HF trials
Risk profile in HF cohorts
Risk profile in an hypothetical trial with prognostic enrichment
Hypothetical uniform risk profile population
Sample size needed
Biomarkers for patient selection
PoS is doubled with biomarkers
Ex: AI biomarkers from pathology slides
[*] Saillard C et al. “Pacpaint: a histology-based deep learning model uncovers the extensive intratumor molecular heterogeneity of pancreatic adenocarcinoma”. �Nature Comms 2023.
[**] Schmauch B et al. “A deep learning model to predict RNA-Seq expression of tumours from whole slide images”. Nature Comms 2020.
[***] Saillard et al. “Validation of MSIntuit as an AI-based pre-screening tool for MSI detection from colorectal cancer histology slides”. Nature Comms 2023.
HE2RNA model [**]�Predicts gene expression�from H&E images
H&E slide
Predicted genes expression
PACpAInt model [*]�Predicts molecular subtypes�from H&E images in pancreatic cancer
H&E slide
Neoplastic region segmentation
MSIntuit Dx tool [***]�Pre-screen MSI from H&E images in colorectal cancer
H&E slide�CRC
Next frontier: few-short learning with foundation models
AI can optimize every clinical phase
Phase 1 / 2
Phase 3
(Federated) external control arms
Prognostic & predictive enrichment
Phase 2 / 3
Data-driven covariate adjustment
Gain early estimates of efficacy for single-arm phase I/II clinical trials to inform phase transition decisions
Inform trial recruitment and better define patient populations
Better measure treatment effect by increasing clinical trial statistical power without adding more patients
Motivation
Covariate adjustment
Survival rate
Time (years)
Control arm Treatment arm
Challenge
Treatment effect signal is impacted by prognostic heterogeneity
Solution
Adjust for data-driven AI-discovered prognostic covariates in the statistical analysis of your trial
Result
Optimized treatment effect estimation
→ Increased statistical power
→ Better adjustment for covariates
Time (years)
Survival rate
Ex: HCCnet
A deep learning prognostic model for resected hepatocellular carcinoma
Saillard , C., et al. Predicting Survival After Hepatocellular Carcinoma Resection Using Deep Learning on Histological Slides. Hepatology 72, 6 (2020) https://doi.org/10.1002/hep.31207
Using HCCnet decreases by 12% the size of the CT needed to reach 80% power
EMA supports AI-based covariate adjustment
“A main potential strength of the artificial intelligence (AI) models is the gain in prognostic performance compared to an adjustment with covariates used in current practice in clinical trial settings. This gain in performance could translate to gains in statistical power.”
Conclusion
Phase 1 / 2
Phase 3
(Federated) external control arms
Prognostic & predictive enrichment
Phase 2 / 3
Data-driven covariate adjustment
Gain early estimates of efficacy for single-arm phase I/II clinical trials to inform phase transition decisions
Inform trial recruitment and better define patient populations
Better measure treatment effect by increasing clinical trial statistical power without adding more patients
Thank you