Record Linkage Using Metagenomic Microbiome Profiles

Abstract

Personal microbiome identification (PMI) links individuals by matching microbiome samples collected at different points in time, providing a means of record linkage. Among previous works, three PMI approaches include the metagenomic-code method, GePMI, and the nearest-neighbour method. These approaches differ in their feature abundance thresholds, feature representations, and matching mechanisms. Although the studies introducing these approaches demonstrated their feasibility, subsequent comparisons have included only subsets of them. No study has evaluated all proposed PMI approaches within a unified setting, leaving their relative performance unclear. In this thesis, we evaluated all proposed PMI approaches within a unified setting that more closely reflects real-world conditions than previous studies. In terms of the samples included, the evaluation setting involved a more challenging study population, greater longitudinal diversity, and a higher number of true negatives. This allowed us to identify the matching scenarios in which the different PMI approaches perform best and inform their future application. A further focus of this thesis is the association between linkage errors and PMI approaches. At the sample level, we analysed sample interactions and method-related factors and used the findings to identify strategies to improve PMI performance. The mutual nearest-neighbour (MNN) method without rematching increased the average F1 score by approximately 2% compared with the nearest-neighbour method in evaluation scenarios including true negatives. At the feature level, we analysed feature contribution patterns and identified the features that frequently dominated each linkage outcome. As a final focus, given the privacy concerns arising from the unintended or unauthorised use of PMI, we evaluated privacy-preserving techniques, including data synthetisation and data anonymisation, as safeguards for microbiome data shared for research purposes. Taken together, this thesis provides a comparative evaluation of PMI approaches, investigates patterns underlying linkage errors, and evaluates privacy-preserving techniques for microbiome data.