Deep Learning-based Propensity Scores for Confounding Control in Comparative Effectiveness Research: A Large-scale,...

Area of Science:

• Machine learning applications in healthcare research
• Real-world data analysis and confounding control

Background:

• Real-world data (RWD) analysis requires balancing prognostic factors due to non-randomized designs.
• Propensity scores (PSs) are commonly used to address confounding in RWD studies.

Purpose of the Study:

• To investigate the utility of autoencoders, a type of unsupervised deep learning architecture, for computing propensity scores.
• To evaluate the performance of autoencoder-based PS computation against established methods.

Main Methods:

• Trained an autoencoder on patient-level data from 128,368 first-line cancer patients to learn patient representations for PS computation.
• Conducted a simulation study to assess balancing and adjustment performance using metrics like standardized mean differences and RMSE.
• Emulated the PRONOUNCE trial in an observational setting to illustrate the application of autoencoder-based PS.

Main Results:

• Autoencoder-based PS, along with LASSO and other methods, achieved well-balanced cohorts (average standardized mean differences <0.1).
• LASSO showed the lowest estimation deviation (RMSE 0.0205), closely followed by the autoencoder approach (RMSE 0.0248).
• Sensitivity analyses indicated comparable performance between autoencoder and LASSO methods.

Conclusions:

• Autoencoder-based PS computation is a viable method for confounding control in RWD studies.
• The autoencoder approach demonstrated feasibility but did not surpass the performance of established methods like LASSO.