Obtaining personalized predictions from a randomized controlled trial on Alzheimer's disease

Area of Science:

• Biostatistics
• Clinical Trials Methodology
• Data Science

Background:

• Randomized controlled trials (RCTs) provide population-level evidence but struggle with patient-level outcome inference.
• Missing data due to treatment discontinuation or unobserved treatment outcomes pose significant challenges in RCTs.
• Developing methods to infer individual patient outcomes is crucial for advancing precision medicine.

Purpose of the Study:

• To infer patient-level outcomes from population-level randomized controlled trials (RCTs).
• To address missing data in RCTs, specifically unrecorded outcomes from treatment discontinuation and unobserved outcomes from unassigned treatments.
• To introduce and evaluate the Synthetic Nearest Neighbors (SNN) estimator for these purposes.

Main Methods:

• Utilized the Synthetic Nearest Neighbors (SNN) estimator, a method that leverages information across patients to impute missing data.
• Applied SNN to impute outcomes for patients who discontinued treatment, thereby de-biasing RCTs.
• Used SNN to simulate "synthetic RCTs" by imputing outcomes for unassigned treatments, predicting individual patient responses.
• Evaluated SNN performance using Phase 3 clinical trial data for Alzheimer's Disease patients.

Main Results:

• The SNN estimator demonstrated strong performance in imputing missing outcomes for both scenarios (treatment discontinuation and unassigned treatments).
• Empirical findings showed SNN outperformed several standard methods in these applications.
• SNN proved to be interpretable, transparent, and causally justified across various missing data conditions.

Conclusions:

• The SNN estimator offers a robust solution for handling patient dropouts in clinical trials.
• SNN serves as a valuable new tool for developing precision medicine by enabling individual patient outcome prediction.
• The methodology shows potential for broader generalization to diverse real-world healthcare applications.