Enhancing Data Interoperability in Alzheimer's Disease Cohort Studies via Comprehensive Semantic and Statistical Analyses

Abstract

As a prominent neurodegenerative disorder, Alzheimer's disease causes extensive cognitive decline and remains a substantial public health concern. Despite extensive research to understand the pathology of the disease, no effective treatment has yet been discovered. This is potentially due to the heterogeneous nature of cohort studies used to assess Alzheimer’s disease and their respective biases. Additionally, Alzheimer’s research has been skewed toward a select few cohort studies that are not necessarily representative of the general population. This selection bias is likely driven by the open-access policies and availability of those cohort datasets. Finally, in general, cohort datasets are far from adhering to the Findable, Accessible, Interoperable, and Reusable data principles. This hinders cross-cohort investigations, leading to single-cohort analyses in the majority of research.<br /> In this work, we evaluate several cohort studies that are either overanalyzed or underrepresented in Alzheimer’s disease research. Through meticulous comparison, we identify the semantic and statistical differences that exist across cohorts and highlight their data limitations and biases. To improve the structuring of datasets in line with the Findable, Accessible, Interoperable, and Reusable data principles, we generate an open-access web application aimed at improving cohort findability. We develop tools focused on the statistical comparison of cohort studies to assess their interoperability. Additionally, we provide researchers in this field with ample information regarding the data contained in the investigated cohorts, including modality, granular variable levels, and sources for acquiring the data (i.e., findability). We further harmonize the cohorts on a semantic level, leading to the creation of a common data model and the development of a data harmonization tool. Subsequently, the tool and the common data model are presented through another web application to accelerate the reusability of the data. Furthermore, we leverage our findings and efforts to assess cohort datasets on a deeper level. By performing two distinct evaluations, we highlight the variation in disease progression patterns across cohorts. These findings further reveal participant selection biases within the datasets and the immediate need to validate findings from single cohorts. Similarly, within our second analysis, we highlight that previous, data-driven thresholds applied for participant disease profiling (based on the absence or presence of pathology) are cohort-dependent and not interchangeable across cohorts. Therefore, participant categorization based on such thresholds varies substantially, indicating a lack of robustness.