Evaluation of a Binary Semi-supervised Classification Technique for Probabilistic Record Linkage.

ABSTRACT
 ObjectivesWhile record linkage has become a strategic research priority within Australia and internationally, legal and administrative issues prevent data linkage in some situations due to privacy concerns. Even current best practices in record linkage carry some privacy risk as they require the release of personally identifying information to trusted third parties. Application of record linkage systems that do not require the release of personal information can overcome legal and privacy issues surrounding data integration. Current conceptual and experimental privacy-preserving record linkage (PPRL) models show promise in addressing data integration challenges but do not yet address all of the requirements for real-world operations. This paper aims to identify and address some of the challenges of operationalising PPRL frameworks.
 ApproachTraditional linkage processes involve comparing personally identifying information (name, address, date of birth) on pairs of records to determine whether the records belong to the same person. Designing appropriate linkage strategies is an important part of the process. These are typically based on the analysis of data attributes (metadata) such as data completeness, consistency, constancy and field discriminating power. Under a PPRL model, however, these factors cannot be discerned from the encrypted data, so an alternative approach is required. This paper explores methods for data profiling, blocking, weight/threshold estimation and error detection within a PPRL framework.
 ResultsProbabilistic record linkage typically involves the estimation of weights and thresholds to optimise the linkage and ensure highly accurate results. The paper outlines the metadata requirements and automated methods necessary to collect data without compromising privacy. We present work undertaken to develop parameter estimation methods which can help optimise a linkage strategy without the release of personally identifiable information. These are required in all parts of the privacy preserving record linkage process (pre-processing, standardising activities, linkage, grouping and extracting).
 ConclusionsPPRL techniques that operate on encrypted data have the potential for large-scale record linkage, performing both accurately and efficiently under experimental conditions. Our research has advanced the current state of PPRL with a framework for secure record linkage that can be implemented to improve and expand linkage service delivery while protecting an individual’s privacy. However, more research is required to supplement this technique with additional elements to ensure the end-to-end method is practical and can be incorporated into real-world models.

IntroductionResearch data combined with administrative data provides a robust resource capable of answering unique research questions. However, in cases where personal health data are encrypted, due to ethics requirements or institutional restrictions, traditional methods of deterministic and probabilistic record linkages are not feasible. Instead, privacy-preserving record linkages must be used to protect patients’ personal data during data linkage.ObjectivesTo determine the feasibility and validity of a deterministic privacy preserving data linkage protocol using homomorphically encrypted data.MethodsFeasibility was measured by the number of records that successfully matched via direct identifiers. Validity was measured by the number of records that matched with multiple indirect identifiers. The threshold for feasibility and validity were both set at 95%. The datasets shared a single, direct identifier (health card number) and multiple indirect identifiers (sex and date of birth). Direct identifiers were encrypted in both datasets and then transferred to a third-party server capable of linking the encrypted identifiers without decrypting individual records. Once linked, the study team used indirect identifiers to verify the accuracy of the linkage in the final dataset.ResultsWith a combination of manual and automated data transfer in a sample of 8,128 individuals, the privacy-preserving data linkage took 36 days to match to a population sample of over 3.2 million records. 99.9% of the records were successfully matched with direct identifiers, and 99.8% successfully matched with multiple indirect identifiers. We deemed the linkage both feasible and valid.ConclusionsAs combining administrative and research data becomes increasingly common, it is imperative to understand options for linking data when direct linkage is not feasible. The current linkage process ensured the privacy and security of patient data and improved data quality. While the initial implementations required significant computational and human resources, increased automation keeps the requirements within feasible bounds.

This thesis presents a methodology of data preparation with probabilistic record linkage and information fusion for improving and enriching information visualizations of biomedical citation data. The problem of record linkage of citation databases where only non-unique identifiers such as author names and document titles are available as common identifiers to be linked was investigated. This problem in citation data parallels problems in clinical data and Knowledge Discovery in Databases (KDD) methods from clinical data mining are evaluated. Probabilistic and deterministic (exact-match) record linkage models were developed and compared through the use of a gold standard or truth dataset. Empirical comparison with ROC analysis of record linkage models showed a significant difference (p=.000) in performance of a probabilistic model over deterministic models. The methodology was evaluated with probabilistic linkage of records from the Web of Science, Medline, and CINAHL citation databases in the knowledge domains of medical informatics, HIV/AIDS, and nursing informatics. Data quality metrics for datasets prepared with probabilistic record linkage and information fusion showed improvement in completeness of key variables and reduction in sample bias. The resulting visualizations offered a richer information space for users through an increase in terms entering the visualization. The significant contributions of this work include the development of a novel model of probabilistic record linkage for biomedical citation databases which improves upon existing deterministic models. In addition a methodology for improving and enriching knowledge domain visualizations though a data preparation approach has been validated with analyses of multiple citation databases and knowledge domains. The data preparation methodology of probabilistic record linkage with information fusion offers a remedy for data quality problems, and the opportunity to enrich visualizations with added content for user exploration, which in turn improves the utility of knowledge domain visualizations as a medium for assessing available evidence and forming hypotheses.

Record linkage represents the process of identifying records that refer to the same real-world entity across multiple sources. In the recent past, the large explosion of data from organizations and individuals has created critical challenges in record linkage process, leading the scientific community to develop a rich series of techniques. Among these recent developments, the design of record linkage solutions for medical data is a cornerstone for health-care systems, as personal identifying information, such as name and date of birth, is often used for linkage. In this setting, record linkage solutions are expected to preserve individual patients’ privacy in addition to be effective and efficient. In this chapter, we provide a broad review of the recent works in privacy preserving record linkage (PPRL). From a privacy-centric perspective, we summarize a comprehensive framework of the PPRL process and describe the privacy assurance techniques adopted for each step in the framework. With the applications in biomedical domain in mind, we identify several challenges for PPRL and discuss future research directions.

BackgroundData on newborns is regularly linked for epidemiological research. However, hospital data often suffers from incomplete data. We report on a linkage of two population-covering administrative health databases containing neonatal and perinatal data without unique personal identifiers and with incomplete information in standard patient identifiers.
 GoalTo study the effects of a policy-induced change from linking a national database without standard patient identifiers to a privacy-preserving Record Linkage method, we compare the linkage system in use to clear-text and privacy-preserving Record Linkage techniques. We expected large proportions of missing identifiers since they are not needed for clinical practice. Therefore, we expected missing links caused by missing identifiers. To study the impact of these missing identifiers on these successful links, we compared several linkage methods. Furthermore, we study the variations of linkage success between hospitals.
 MethodsPerinatal and neonatal data from population-covering real-world administrative databases was linked using several variants of state of the art methods, including Privacy-preserving Record Linkage (PPRL) techniques such as multiple match keys and Bloom filter methods.
 Results We report on the variation of linkage results between the hospitals and give possible explanations for the differences.
 The resulting linkage success is reported for each method. The impact of incomplete data on linkage success for each method is documented.
 Finally, we report on the relative performance of the modified techniques compared to standard linkage procedures used in practice.
 ConclusionImplementing a record linkage system based on identifiers not required for clinical practice caused a large number of missing identifiers. Since this information is essential for successful clear-text and private linkage methods, emphasizing the need for documenting patient identifiers, especially in cases where auxiliary information (such as stable addresses, date of birth or health insurance numbers) are missing, is of central importance for implementing a privacy-preserving Record Linkage system.

Many learning techniques have been applied to identify disease-associated genes. At the early, they usually approached this problem as a binary classification, where training set is comprised of positive and negative samples. In which, positive samples are constructed from known disease genes, whereas negative samples are the remaining which are not known to be associated with diseases. This is the limitation of the binary classification-based solutions, since the negative training set should be actual non-disease genes; however, construction of this set is nearly impossible in biomedical researches. Therefore, to reduce this uncertainty, more realistic classification-based methods have been proposed. For instance, unary classification technique based on one-class SVM method was proposed by learning from only positive samples. In addition, the remaining set may contain unknown disease genes; therefore, semi-supervised methods such as binary semi-supervised and positive and unlabeled (PU) learning classifications have been proposed. In particular, PU learning methods, which learn from both known disease genes and the remaining genes, were shown to outperform others. In these studies, data sources are usually represented by vectorial form for binary classifiers, while they are in kernel matrices for unary and PU learning ones. The kernel-based data fusion is only suitable for data with different types and it seems unfair for the comparison based on different data representations. Therefore, in this study, we compared different classification techniques for the disease gene prediction based on vectorial representation of samples. The simulation result showed that the unary classification technique, which combines both density and class probability estimation strategies, achieved the best performance, whereas it is worst for the one-class SVM-based method. Interestingly, performance of the best binary classification technique is comparable with that of biased SVM-based PU learning and binary semi-supervised classification methods. And, they are all better than the multi-level SVM-based one.

ABSTRACTObjectiveIn most European settings, record linkage across different institutions has to be based on personal identifiers such as names, birthday or place of birth. To protect the privacy of research subjects, the identifiers have to be encrypted. In practice, these identifiers show error rates up to 20% per identifier, therefore linking on encrypted identifiers usually implies the loss of large subsets of the databases. In many applications, this loss of cases is related to variables of interest for the subject matter of the study. Therefore, this kind of record-linkage will generate biased estimates.
 These problems gave rise to techniques of Privacy Preserving Record Linkage (PPRL). Many different PPRL techniques have been suggested within the last 10 years, very few of them are suitable for practical applications with large database containing millions of records as they are typical for administrative or medical databases. One proven technique for PPRL for large scale applications is PPRL based on Bloom filters.MethodUsing appropriate parameter settings, Bloom filter approaches show linkage results comparable to linkage based on unencrypted identifiers. Furthermore, this approach has been used in real-world settings with data sets containing up to 100 Million records. By the application of suitable blocking strategies, linking can be done in reasonable time.ResultHowever, Bloom filters have been subject of cryptographic attacks. Previous research has shown that the straight application of Bloom filters has a nonzero re-identification risk. We will present new results on recently developed techniques to defy all known attacks on PPRL Bloom filters. These computationally simple algorithms modify the identifiers by different cryptographic diffusion techniques. The presentation will demonstrate these new algorithms and show their performance concerning precision, recall and re-identification risk on large databases.

IntroductionBusinesses worldwide are increasingly adopting the storage, compute and analytical services provided by cloud computing. Yet, few operational linkage units are keeping pace with this world of technological change - most use legacy systems approaching their limits with the rapidly increasing size and range of datasets now required for linkage.
 Objectives and ApproachTo meet the demands of linkage for the near future, it is important that new solutions for linkage consider the services provided by public cloud infrastructure for compute, storage and analytics. We examined Platform as a Service (PaaS) offerings for use in the development of a cost-effective cloud model for scalable, privacy-preserving record linkage (PPRL). PPRL techniques were adapted to maximise the quality of linkage and to automate as much of the process as possible. Finally, a prototype was created to demonstrate the capabilities and potential of the model.
 ResultsWe present our cloud model for PPRL, a platform for record linkage that provides rapid scaling of resources to meet demand, and the results of how our prototype performed on massive datasets.
 Conclusion/ImplicationsThe application of record linkage using relatively inexpensive cloud infrastructure represents a significant step towards providing an efficient and scalable record linkage service to researchers and government. Larger datasets can be linked efficiently, including national or cross-jurisdictional datasets, with little investment in private infrastructure, and improved turnaround times for researchers.

Record linkage (RL) is the process of identifying pairs of records that correspond to the same entity, for example the same patient. The basic approach assigns to each pair of records a similarity weight, and then determines a certain threshold, above which the two records are considered to be a match. Three different RL methods were applied under privacy-preserving conditions on hospital admission data: deterministic RL (DRL), probabilistic RL (PRL), and Bloom filters. The patient characteristics like names were one-way encrypted (DRL, PRL) or transformed to a cryptographic longterm key (Bloom filters). Based on one year of hospital admissions, the data set was split randomly in 30 thousand new and 1,5 million known patients. With the combination of the three RL-methods, a positive predictive value of 83 % (95 %-confidence interval 65 %-94 %) was attained. Thus, the application of the presented combination of RL-methods seem to be suited for other applications of population-based research.

Infinity Laplacian on graphs with gradient terms for image and data clustering

SummaryAccurate prediction of the solubility of chemical substances in solvents remains a challenge. The sparsity of high-quality solubility data is recognized as the biggest hurdle in the development of robust data-driven methods for practical use. Nonetheless, the effects of the quality and quantity of data on aqueous solubility predictions have not yet been scrutinized. In this study, the roles of the size and the quality of data sets on the performances of the solubility prediction models are unraveled, and the concepts of actual and observed performances are introduced. In an effort to curtail the gap between actual and observed performances, a quality-oriented data selection method, which evaluates the quality of data and extracts the most accurate part of it through statistical validation, is designed. Applying this method on the largest publicly available solubility database and using a consensus machine learning approach, a top-performing solubility prediction model is achieved.

ObjectiveTo introduce Linkja, an open-source, certified, Privacy-Preserving Record Linkage (PPRL) tool with the primary objective of facilitating secure and efficient record linkage across diverse datasets, while preserving individual privacy, providing a comprehensive and privacy-centric approach to program assessments. ApproachLinkja has been used to evaluate Chicago’s Flexible Housing Pool program by securely linking various datasets, including Homeless Management Information Systems (HMIS), Electronic Health Records, Jail records, Medical Examiner reports, and Public Health data. Data contributors used separate honest brokers for record linkage, and generation of salt and cryptography keys. ResultsLinkja enabled this multi-sector linkage by facilitating legal and data sharing agreements, avoided licensing fees and procurement processes for all parties, and provided a de-identified for analysis. The insights gleaned from the analysis motivated the FHP program to plan a follow-up evaluation that will add emergency serviced records from a separate City of Chicago department. The successful linkage provided a nuanced understanding of the association between stable housing and reduced healthcare utilization. ConclusionsLinkja's cryptographic certification ensures the highest standards of security during data linkage, safeguarding sensitive information. The matching rules, validated on hard-to-match populations, prove essential in maintaining accuracy and reliability in the linkage process. ImplicationsThe open-source nature of Linkja promotes transparency and collaboration in the development and enhancement of privacy-preserving record linkage techniques. This tool has broad implications for sectors dealing with sensitive data, fostering responsible data-sharing practices, and enabling secure collaborative research initiatives without compromising individual privacy.

Examine the accuracy of privacy preserving record linkage (PPRL) matches in real world data (RWD). We conducted a systematic literature review to identify articles evaluating PPRL methods from January 1, 2013 to June 15, 2023. Eligible studies included original research reporting quantitative metrics such as precision and recall in health-related data sources. Covidence software was used to manage the review process. Five studies met our inclusion criteria. Tokenization and hash functions were used to hash and encrypt personally identifiable information (PII) including first and last names, dates of birth (DOB), and Social Security Numbers (SSNs) in a variety of RWD. All identified studies utilized deterministic matching. Combinations of tokenized or hashed PII that included "quasi-identifiers" like names and DOBs had consistently high precision (>95%) but lower recall, likely due to misspelled or inconsistently spelled names and name changes. SSN-based combinations demonstrated high precision but variable recall due to incomplete SSN data in RWD. Studies that employed algorithms in which at least one match was identified from a specified set of PII combinations provided high precision and high recall. The systematic review indicates that PPRL methods generally provide highly accurate patient data linkage while maintaining privacy. Researchers should carefully consider the completeness and stability of each PII element selected for PPRL and may want to employ a strategy that allows for patient records to be matched if they meet at least one of several combinations of PII.

Record linkage is an emerging research area which is required by various real-world applications to identify which records in different data sources refer to the same real-world entities. Often privacy concerns and restrictions prevent the use of traditional record linkage applications across different organizations. Linking records in situations where no private or confidential information can be revealed is known as privacy-preserving record linkage (PPRL). As with traditional record linkage applications, scalability is a main challenge in PPRL. This challenge is generally addressed by employing a blocking technique that aims to reduce the number of candidate record pairs by removing record pairs that likely refer to non-matches without comparing them in detail. This paper presents an efficient private blocking technique based on a sorted neighborhood approach that combines k-anonymous clustering and the use of public reference values. An empirical study conducted on real-world databases shows that this approach is scalable to large databases, and that it can provide effective blocking while preserving k-anonymous characteristics. The proposed approach can be up-to two orders of magnitude faster than two state-of-the-art private blocking techniques, k-nearest neighbor clustering and Hamming based locality sensitive hashing.

Semi-supervised learning is drawing increasing attention in the era of big data, as the gap between the abundance of cheap, automatically collected unlabeled data and the scarcity of labeled data that are laborious and expensive to obtain is dramatically increasing. In this thesis, we first introduce a unified view of density-based clustering algorithms. Then, we build upon this view and bridge the areas of semi-supervised clustering and classification under a common umbrella of density-based techniques. We show that there are close relations between density-based clustering algorithms and the graph-based approach for transductive classification. These relations are then used as a basis for a new framework for semi-supervised classification based on building-blocks from density-based clustering. This framework is not only efficient and effective, but it is also statistically sound. We also generalize the core algorithm of the framework HDBSCAN* so that it can also perform semi-supervised clustering by directly taking advantage of any fraction of labeled data that may be available, rather than instance-level pairwise constraints. Experimental results on a large collection of datasets show the advantages of the proposed approach both for semi-supervised classification, as well as for semi-supervised clustering. In addition, we evaluate the semi-supervised learning algorithms to determine relationships between chemical structure and biological activity in datasets from Medicinal Chemistry. The datasets evaluated in this area are characterized by a low number of labeled examples, a high dimensionality, and in some cases, do not have a clear relationship between chemical structure and biological activity, which makes it difficult to use classification techniques and analyze chemical phenomena. We implement and validate semi-supervised classification approaches that are appropriate for data analysis in Medicinal Chemistry.