Alzheimer's disease (AD) is a significant global healthcare challenge, with limited effective treatments and high clinical trial failure rates. A major challenge in AD clinical trials is the substantial individual variation in the rate of cognitive decline. This variation leads to two critical problems: (1) trials may include a high proportion of slow decliners or non-decliners, weakening the observed treatment effect; and (2) randomization may create an allocation bias, where treatment and placebo groups have disproportionate numbers of fast and slow decliners, leading to over- or underestimation of treatment effects. Previous attempts to address this have focused on selecting participants or using individual risk factors (like ApoE ε4) for stratification in randomization, but these methods have limitations. This study proposes a novel approach using artificial intelligence (AI) to predict cognitive decline and stratify participants in clinical trials, aiming to mitigate allocation bias and improve trial efficiency.

The literature highlights the high failure rate of AD drug development programs and the challenges associated with using cognitive decline as an endpoint in clinical trials. Studies have explored various methods, including machine learning, to predict disease progression and potentially improve participant selection. However, simply selecting fast decliners introduces uncertainty about drug effectiveness in different decline groups. The impact of allocation bias in cognitive decline on clinical trial outcomes has been recognized, but solutions remain elusive. While some trials use risk factors like ApoE ε4 for stratification, these factors have limited predictive power. Recent research has suggested the potential of AI to address allocation bias by improving the prediction of cognitive decline.

This study developed a hybrid multimodal deep learning model to predict changes in the Clinical Dementia Rating-Sum of Boxes (CDR-SB) score, a common endpoint in AD trials. The model integrates T1-weighted MRI images (from hippocampal and anterior temporal lobe regions) with non-image data (demographics, cognitive test scores, and biomarkers) at baseline. A DenseNet3D convolutional neural network extracts image features, and these are combined with non-image data using linear support vector regression (SVR). A multi-task loss function (combining regression, classification, and image recovery losses) was employed during training to improve robustness. The model's predictions of CDR-SB changes were then used as a stratification index in a stratified randomization method. Simulations using the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset compared the proposed AI-based randomization with conventional non-stratified randomization and other stratification methods (using age, ApoE ε4, biomarkers, and cognitive scores). Simulations assessed allocation bias (measured by standard allocation error, SAE), the required sample size to achieve a desired range of prospective effect size (PES), and the impact on the detection of actual treatment effects. Simulations also considered a two-phase trial design, with sample size estimation for the second phase based on the first phase's outcome.

The AI model achieved a mean absolute error (MAE) of 1.065 and a correlation coefficient of 0.601 between predicted and actual CDR-SB changes. The AI-based stratified randomization method significantly reduced the SAE compared to non-stratified randomization (by 22.4%). This led to a 37% reduction in the sample size needed to achieve the same 95% range of PES. Compared to stratification using other individual clinical factors, the AI-based approach yielded a substantially larger reduction in SAE and sample size. Simulations of trials with actual treatment effects demonstrated that AI-based randomization improved the detection rate of larger treatment effects while suppressing the detection of smaller (potentially false-positive) effects. In a simulated two-phase trial, AI-based randomization improved the overall success rate when the actual treatment effect was moderate to large.

This study demonstrates the effectiveness of incorporating AI-based prediction of cognitive decline into the randomization procedure of AD clinical trials. The AI-based stratified randomization significantly reduces allocation bias, leading to substantial improvements in trial efficiency. The ability to improve detection of true treatment effects while suppressing false positives is particularly valuable, especially in early-phase trials with limited sample sizes. This approach offers a practical solution to a major challenge in AD clinical trials. The findings highlight the potential of AI to enhance the design and interpretation of clinical trials for AD.

This study introduces a novel AI-based stratified randomization method that significantly improves the efficiency of Alzheimer's disease clinical trials by reducing allocation bias and optimizing sample size. Future work should validate these findings in real-world trials and explore the application of this approach to more complex trial designs and other endpoints. Further refinement of the predictive model using larger datasets and more sophisticated methodologies could further enhance its accuracy and utility.

The study's findings are based on simulations using the ADNI dataset, which may limit the generalizability of results to other populations. While the ADNI dataset reflects a potential clinical trial population, validation with actual trial data is needed. The study focused on CDR-SB changes as the primary endpoint, and further research is required to adapt the method for other endpoints. The simulations considered only two-arm trials with a single stratification index, and further research should explore the approach's effectiveness in more complex trial designs.