A quantitative and qualitative assessment of differential privacy's ability to support collaborative research using a real-world data analysis.

Objective: Sharing clinical data for research that is both collaborative and privacy-preserving remains a challenge. Differential privacy (DP) offers a solution by introducing noise to query results. Using the PrivateSQL DP platform, this study assesses the resulting utility of differentially private data at different levels of aggregation through analyses of COVID-19 pandemic associations with new cancer diagnosis counts (NCCs). Materials and Methods: Data from a multi-hospital system for adult (18-90 years) encounters from 2019-2021 with apparently new cancer diagnoses were extracted, then queried using standard SQL ("original") and DP, each with 1-week and 4-week aggregations. Analyses on the 4 datasets included NCCs by year and multivariate regression models of associations between COVID-19 positivity rates (by county) and change in NCCs between pre- and post-COVID-19 start. Results: NCCs dropped in 2020, rebounding in 2021. This same pattern was demonstrated in the 4-week, but not the 1-week, DP dataset. Confidence intervals were substantially narrower in regressions using original datasets compared to those using DP datasets, and narrower in DP dataset regressions using 4-week rather than 1-week aggregation. Post-hoc sensitivity analyses found significant associations with 2 variables of interest on the original datasets (though these have methodologic limitations), but not the DP datasets. Discussion: DP reduces analytic accuracy to protect data privacy, but aggregation mitigated this tradeoff. Strategies for using DP in healthcare research and potential opportunities to enhance the DP platform were identified. Conclusion: DP platform enhancements for hypothesis-driven medical studies may expand DP's ability to support fruitful, cross-institutional research collaborations.