WITHDRAWN Real-World Surveillance of COVID-Related Neurological Symptoms and Vaccine-Protective Effects Through Data-Adaptive Targeted Learning

OBJECTIVES:

The ongoing COVID-19 pandemic challenged the global health infrastructure and profoundly impacted populations, with an ongoing and evolving understanding of the total health implication of infection. In addition to impacts on the respiratory system, emerging evidence implicates viral infection on developing neurological symptoms. The open questions about covid-19 and the (potentially protective) effect of vaccines on developing neurological symptoms, such as encephalitis, and anxiety disorders, outpace the feasibility of conducting randomized clinical trials (RCT). The ongoing impact motivates a modern sentinel-like surveillance system to enhance knowledge about the disease.

METHODS:

This study has two stages: In stage 1, we’ll generate synthetic data, establish and evaluate a surveillance analytic approach to estimate the potential protective effects of the covid-19 vaccination on neurological symptoms, and compare estimates to (generated) protective effects. The second stage will implement and evaluate the surveillance system within a large healthcare system.

For stage 1, we simulated a plausible real-world data generation process (DGP), including varying sampling and analysis approaches (traditional vs. tmle), on estimating the effects of vaccines on subsequent neurological symptom progression.

Data applications for stage 2 are pending.

RESULTS:

Targeted learning approaches produced estimates closer [avg. bias: 9.8%] to simulated ‘true’ values than produced by (unadjusted) conditional estimates [avg. bias: 42.8%] and traditional propensity score estimates [avg. bias: 27.6%]. Results were most sensitive to potential sampling bias, indicating the need for special care in causal considerations for bias correction.

CONCLUSIONS:

Targeted learning provides a valuable framework for estimating effects in real-world data, both in focusing on causal structural considerations and in analytic effect de-biasing. However, much care still needs to go into cohort and outcome definitions for targeted learning to support scientific discovery. Future work may incorporate longitudinal data and decision opportunity sequences (dynamic and stochastic treatment strategies).