A likelihood-based convolution approach to estimate event occurrences in large longitudinal incomplete clinical data

Registry and clinical trials data have been the preferred data source for vast amounts of research effort. However, these data sources offer limited sample sizes, follow-up time, temporal granularity and often focus on a single disease, making them insufficient for modern research. Recent years has seen the emergence of commercial data sets, aggregated from multiple health system, including pharmacy, clinical and claims data. Unlike registry data, these large-scale data sources can contain inaccuracies and may be incomplete, often missing the dates of major health events. In this paper, we develop a method for estimating the dates of major health events that are defined by significant changes in physiology and treatment regimen before and after the event. Our methodology synthesizes information from multiple sources (labs, medications, procedures, billing and claims diagnoses) using a novel convolution-based change detection methodology. We apply our methodology to identifying liver transplant dates for patients who underwent liver transplant from the OptumLabs® Data Warehouse totaling 31,000 transplant patients. For a cohort of 4,000 patients with known transplant date, our method managed to estimate the transplant date within the two-week window for 92% of the cohort. On the vast majority in the data set (84%), the error (absolute difference between the predicted and true transplant dates in days) is less than one week and for almost the entire data set (92%), the errors fall within two weeks of the true date.