T-16: Generating Synthetic Control Patients Using Machine Learning for Alzheimer’s Disease Clinical Trials
Poster Presenter
Yannick Pouliot
Sr. Computational Biologist
Unlearn.AI United States
Objectives
Simulating the disease progression of individual synthetic control patients (SCPs) as assessed by multiple metrics in order to diminish reliance on actual control patients in Alzheimer’s Disease clinical trials.
Method
We developed a machine learning model of Alzheimer’s Disease progression trained with data from 1,335 patients from 24 clinical trial control arms involving early or moderate Alzheimer’s Disease. The model generates values for 44 variables for each SCP at three month intervals over 18 months.
Results
We validated our model by comparing predicted ADAS-Cog scores with the scores of patients from a diverse set of clinical trials that were not involved in training the model. Predicted scores were indistinguishable from observed scores. We further compared the ADAS-Cog scores of SCPs computed by our unsupervised model based on the Conditional Restricted Boltzmann Machine (CRBM) algorithm with those generated by several supervised machine learning algorithms of different types, trained on the same data. All algorithms exhibited a similar error profile over time, with CRBM exhibiting the best performance. These comparisons demonstrate that our model is capable of making accurate and precise simulations based on unsupervised learning from extant trials.
Conclusion
Recently, increased attention has been focused on the benefits of using SCPs clinical trials as a mechanism to accelerate trials and diminish their cost. Machine learning models are one approach for comprehensively simulate the evolution of multiple metrics simultaneously for individual SCPs, thus diminishing reliance on actual control patients. Our work demonstrates the potential for the unsupervised Conditional Restricted Boltzmann Machine algorithm used here to generate SCPs indistinguishable from actual control patients. e discuss how SCP models may be used in clinical trial design, as well as to supplement patients in the control arms of Alzheimer’s Disease clinical trials.
