Privacy-Preserving Linkage of Distributed Biological, Clinical, and Imaging Data Supporting Artificial Intelligence in Pediatric Oncology Authors: Dieter Hayn, Martin Baumgartner, Karl Kreiner, Emanuel Sandner, Bernhard Jammerbund, Martin Schalling, Ulrike Poetschger, Vanessa Duester, Blanca Martinez de Las Heras, Adela Cañete Nieto, Ana Jiménez-Pastor, Luis Martí-Bonmatí, Ruth Ladenstein, Günter Schreier Published: 2026 ABSTRACT BackgroundCancer remains the leading cause of disease-related mortality in children over the age of one in Europe, with over 35,000 new pediatric cases and more than 6,000 deaths annually. Due to the rarity of pediatric cancers, clinical trial protocols often substitute for formal treatment guidelines, resulting in many children being enrolled in multiple trials, with biological samples and genomic data stored in various biobanks. Data collection in pediatric oncology is challenging, with sparse data acquired over extended periods, underscoring the need for optimal utilization of all available information through linked, privacy-preserving datasets.MethodsHere, we report the development of a distributed, privacy-preserving data infrastructure for the PRIMAGE project, a European initiative aimed at supporting artificial intelligence (AI)-driven image analysis for pediatric cancer prognostics. The infrastructure leverages the European Patient Identity (EUPID) Services for Privacy-Preserving Record Linkage, enabling pseudonymized data integration across clinical, biological, and imaging sources. The system incorporates EUPID's hashing and phonetic matching protocols to pseudonymize patient identifiers and link distributed datasets, facilitating secondary data use in compliance with the General Data Protection Regulation.ResultsData from over 700 neuroblastoma patients from European trials and hospitals were linked and uploaded to the PRIMAGE platform, where AI models predict clinical outcomes.ConclusionThis infrastructure successfully facilitated AI model development, advancing pediatric oncology research, and offering a scalable framework for future European health data initiatives, such as the European Health Data Space. Introduction Distributed Data in Pediatric Oncology Cancer is the most frequent cause of disease-related death among children over the age of 1 year in Europe. Each year, more than 35,000 children in Europe are diagnosed with cancer, leading to more than 6,000 pediatric cancer-related deaths in Europe per year. 1 In the field of rare diseases, such as pediatric oncology, even the routine patient treatments are closely related to ongoing trials, as formal clinical guidelines are often lacking and trial protocols are widely used instead. 2 As a result, the majority of children with cancer are enrolled in at least one clinical trial during their patient journey, have their samples and genomic profiles stored in one or more biobanks, and are registered in at least one data registry. Data-Related Open Issues in Pediatric Oncology Childhood cancer is a rare disease, making data acquisition time-consuming, costly, and limited in terms of up-to-date information compared with common diseases. Therefore, it is crucial to maximize the value of every data point, including through the linking of distributed data sources. Linked data are particularly valuable for long-term research that goes beyond the primary objectives of individual research projects (“secondary use”). In Europe, the General Data Protection Regulation (GDPR) 3 permits secondary use of data previously collected for other purposes, provided the patient has given appropriate consent. However, specifics of secondary use are often defined by GDPR derogations. The European Health Data Space (EHDS) 4 aims to support standardized primary and secondary use of health data across Europe in the future. The more data collected and the more records from diverse contexts linked together, the higher the potential to analyze and extract valuable insights. Various artificial intelligence (AI) techniques, including data mining, predictive modelling, and clustering, can assist researchers and clinicians in handling these large datasets. In its 2022 position paper, 5 the European Society for Pediatric Oncology (SIOPE) calls for increased funding to support innovations in pediatric cancer research, particularly to enhance research through big data and AI applications. 6 Data Infrastructures in European Pediatric Oncology SIOPEN-R-NET The SIOPEN-R-NET is a European infrastructure for neuroblastoma research, initially developed through an FP5 project that began in 2002. 7 Since then, the IT infrastructure has been continuously operational. The SIOPEN-R-NET supports patient registration, enrolment in clinical trials, collection of biological sample analysis results, image management, core data review, and various other functions. It is connected to multiple electronic data capture systems for clinical trials, including HR-NBL1, 8 VERITAS (NCT03165292), 9 OMS/DES (opsoclonus myoclonus syndrome/dancing eye syndrome), 10 and LINES (NCT01728155). 11 These systems collect clinical trial data in a pseudonymized format. Biological data of all SIOPEN patients can be collected in a central biology database, which has recently been updated to the SIOPEN BIOPORTAL (NCT05192980). 12 While core image-related outcomes and risk factors are typically collected within the electronic data capture systems, raw imaging data are stored on a central SIOPEN image server. 13 Privacy-Preserving Record Linkage in Pediatric Oncology Privacy-Preserving Record Linkage (PPRL) enables the linking of distributed datasets collected in different contexts without disclosing the participants' identities. Due to GDPR restrictions on recording identifiable information, such as patient names, the datasets to be linked often need to be pseudonymized. To protect patient privacy, each data source generates a unique pseudonym for each patient, with different pseudonyms used across contexts as mandated by the GDPR. The significance of PPRL in pediatric oncology was recently highlighted in a Lancet Oncology commentary, where SIOP Europe proposed six recommendations to enhance the current legislative framework, including recommendation 4: “support for privacy-preserving data linkage.” 14 A comprehensive overview of PPRL use cases in pediatric oncology is provided in the study by Hayn et al. 15 Multiple PPRL solutions have been published previously. In our previous paper 16 we listed 21 PPRL services that are available on the market. To facilitate the comparison of PPRL approaches, Vatsalan et al introduced a comprehensive PPRL taxonomy in 2013, defining 15 dimensions for characterizing existing techniques. 17 Building on this work, Gkoulalas-Divanis et al published a systematic review in 2021 that evaluates contemporary PPRL methods and systems using a taxonomy organized into four major categories: computational, utility, privacy, and practical aspects. 18 In addition, Christen et al provided a broad overview of the field, covering the fundamental concepts, methodologies, and challenges of PPRL. 19 Current state-of-the-art PPRL approaches employ a wide range of techniques, including phonetic encoding, 20 21 cryptographic hashing, 22 23 reference-value methods, 24 25 embedding-based representations, 26 27 differential privacy, 28 and secure multiparty computation, which can be applied either to quasi-identifiers (QIDs) or directly to clinical data. 29 30 31 32 Hampf et al described a PPRL solution based on federated trusted third parties. 33 While some methods are limited to linking records that match exactly, more advanced approaches support probabilistic record linkage by tolerating minor discrepancies, such as typing errors, spelling variations, or missing values. EUPID Services In pediatric oncology across Europe, the European Patient Identity (EUPID) Services are widely used to link pseudonymized and personalized records. The development of the EUPID Services began during the FP7 EU project European Network for Cancer Research in Children and Adolescents (ENCCA) . 34 35 As described in a recent paper, 36 the EUPID Services support various PPRL methodologies, including: Hashing of patient identifiers (IDs), such as social security numbers. Hashing of patient identifiers (IDs), such as social security numbers. Hashing of QIDs, such as names, date of birth, etc. Hashing of QIDs, such as names, date of birth, etc. Phonetic hashing of QIDs and hashing of identifiers with common typing errors (e.g., flipped day and month of birth, “Katherina” vs. “Katerina”). Phonetic hashing of QIDs and hashing of identifiers with common typing errors (e.g., flipped day and month of birth, “Katherina” vs. “Katerina”). When registering a patient in a given context, IDs and/or QIDs are hashed locally on the user's computer, and only hashed data are sent to the EUPID Services. The EUPID Services then compare the submitted hashes with those of all previously registered patients. If the patient has not yet been registered in the respective context, a new pseudonym is issued. If matching hashes are identified, the newly registered patient is linked to the existing unique identifier of the corresponding patient. If no match is found, a new unique identifier is generated. Additionally, for patients who already have a EUPID pseudonym in one context, EUPID enables the creation of a new pseudonym for another context (“replication”) without requiring reentry of IDs or QIDs. In this work, the EUPID Services were selected because they offer a high degree of customizability, support both ID- and QID-based PPRL, and have already been successfully applied in pediatric oncology, including several of the datasets analyzed in this study. State of the Art Currently, extensive research focuses on applying AI to multimodal oncology data (e.g., 37 38 39 40 ). However, the described infrastructures do not yet support privacy-preserving linkage of preexisting, distributed datasets. Meanwhile, some research infrastructures under development have the potential to enable AI tasks based on privacy-preserving linkage of distributed biological, clinical, and imaging data. These include the European Joint Program on Rare Diseases (EJP RD) virtual platform, 41 the European Rare Disease Registry Infrastructure (ERDRI), 42 and the EHDS. 4 Nonetheless, none of these infrastructures has progressed sufficiently and supports all required use cases to be applied in specific research projects as described below. Objectives The project PRedictive In-silico Multiscale Analytics to support cancer personalized diaGnosis and prognosis, Empowered by imaging biomarkers (PRIMAGE) 43 aimed at deriving image related risk factors based on AI approaches, considering the following types of data: Personalized data entered directly by physicians at the patients' hospitals. Personalized data entered directly by physicians at the patients' hospitals. Pseudonymized data describing biological samples as previously collected in SIOPEN studies and stored in the SIOPEN Biology database. Pseudonymized data describing biological samples as previously collected in SIOPEN studies and stored in the SIOPEN Biology database. Pseudonymized imaging data stored on an image server. Pseudonymized imaging data stored on an image server. Pseudonymized clinical trial data, stored in different electronic data capture systems (SIOPEN LINES trial, SIOPEN HRNBL1 trial, German Neuroblastoma Registry NB2016 of the GPOH, data from five Spanish hospitals and one Italian hospital, cooperating in PRIMAGE). Pseudonymized clinical trial data, stored in different electronic data capture systems (SIOPEN LINES trial, SIOPEN HRNBL1 trial, German Neuroblastoma Registry NB2016 of the GPOH, data from five Spanish hospitals and one Italian hospital, cooperating in PRIMAGE). Fig. 1 illustrates how this setting maps to the six dimensions related to PPRL use cases in pediatric oncology, as described in. 15 Mapping the PRIMAGE use case to the six dimensions relevant to privacy-preserving record linkage in pediatric oncology, as described in Hayn et al. 15 In previous publications, Martí-Bonmatí et al 43 described the setup and application of the PRIMAGE IT infrastructure for AI-based data analysis of neuroblastoma data. Veiga-Canuto et al 44 45 presented results of the developed AI models. This paper focuses on the PPRL processes used to link distributed data sources for AI model training and validation while ensuring privacy of involved patients. Methods Data Sources The data were collected in the course of the PRIMAGE project, 43 including patients with neuroblastic tumor diagnosis and pathological confirmation between 2002 and 2022. Clinical data of patients with neuroblastoma, ganglioneuroblastoma, or ganglioneuroma were included from the following data sources: SIOPEN/HR-NBL1, 8 SIOPEN/LINES, 11 GPOH/NB2016, 46 single cases from Spanish hospitals (La Fe University and Polytechnic Hospital, Hospital La Paz, Hospital Niño Jesús, Hospital Son Espases, Hospital Reina Sofía) and single cases from Italian hospitals (Pisa University Hospital, Istituto Gaslini). Clinical trial data and biological data had previously been collected in respective electronic data capture systems for each of the trials. Images of all patients included in the study were collected from the local sites in a pseudonymized way. Therefore, the local sites were asked to submit the respective data in DICOM format to the study center of the HR-NBL1 and LINES study, respectively. Prior data transmission, QIDs were manually replaced by SIOPEN-R-NET IDs from the DICOM data. The image management server (IMS) was based on the open-source, Java-based information management software Alfresco (Alfresco Software Inc). Various types of DICOM image datasets (magnetic resonance imaging [MRI], computed tomography [CT], positron emission tomography [PET] scans, MIBG [metaiodobenzylguanidine], etc.) were stored in a hierarchical folder structure, where access to folders and subfolders could be granted to specific user roles. The following folder hierarchy was applied: trial (LINES/HR-NBL1)—patient—time point in the trial (preinduction/postinduction/end of treatment/etc.)—modality (MRI/CT/PET, MIBG/etc.). After login, each user was given access to the folders, depending on their respective user roles. This platform had been developed and used previously for collecting and core reviewing of neuroblastoma imaging data. 13 The PRIMAGE study had the institutional ethics committee approvals from all involved institutions. EUPID Patient Pseudonymization For SIOPEN studies, pseudonymization in early years was done based on an identifier called “SIOPEN-R-NET ID,” which was unique for each patient and used as a PSN for all SIOPEN studies, including LINES and HR-NBL1. In more recent SIOPEN studies, starting with the infrastructure of the OMS/DES in children with and without neuroblastoma (NBpos and NBneg) trial, which went online in 2016, 47 the SIOPEN-R-NET ID was replaced by study-specific EUPID PSNs. Additionally, EUPID PSNs derived from the SIOPEN-R-NET ID were added (in addition to the SIOPEN-R-NET ID) for the earlier SIOPEN studies. Therefore, clinical data from the HR-NBL1 and LINES trials, along with biological sample data from SIOPEN studies, were pseudonymized with the EUPID Services in a dedicated SIOPEN Context years before the PRIMAGE project started. For patients not registered to these SIOPEN trials, new registrations were made directly to the PRIMAGE data platform via the EUPID Services based on their QIDs, ensuring duplicate prevention. The IMS was intended not only for data from HR-NBL-1 and LINES but also as a sustainable solution for future SIOPEN data. To prevent mixing PSNs from different trials within the IMS, a new EUPID Context was created, assigning unique IMS PSNs to all related patients, distinct from each trial's PSNs. EUPID Registration was supported by two processes: for retrospective patients, a unique ID (such as the SIOPEN-R-NET ID, used before EUPID development), was hashed for record linkage. For prospective patients, probabilistic linkage based on hashed QIDs (first name, last name, date of birth) was applied: Argon2 hashes were used to identify “full matches.” “Partial matches” were determined from phonetic hashes of first or last names, and from hashes with flipped month and day of birth. PSNs derived from unique IDs could not directly be linked to those from QIDs. Therefore, if a SIOPEN-R-NET ID-derived PSN existed for a patient, generating a new PSN from QIDs required entering the respective SIOPEN-R-NET ID in addition to the QIDs. EUPID Registration for prospective and retrospective patients and EUPID Replication are illustrated in Fig. 2 . EUPID Registration API and EUPID Replication API, supporting prospective registration based on quasi-identifiers (QIDs), retrospective registration based on unique IDs such as the SIOPEN-R-NET ID, and linkage of pro- and retrospectively generated pseudonyms (PSNs). EUPID, European Patient Identity. Pseudonymization and Privacy-Preserving Record Linkage Infrastructure Fig. 3 provides an overview of pseudonymization and PPRL process in PRIMAGE. All data had initially been collected in several hundred hospitals across Europe. Those hospitals have access to biological, clinical, and imaging data of their patients, and they know their patients' QIDs and PSNs. Overview of the pseudonymization and privacy-preserving record linkage process in PRIMAGE. The local hospitals (top) hold all data modalities (biological, clinical, image-related data) and quasi-identifiers (QIDs) of their own patients. Subsequent components hold pseudonymized data of a specific modality of patients from all hospitals. The PRIMAGE data platform combines selected data of all modalities and of all patients from these components and makes them accessible for AI models. Color-codes correspond to the six dimensions as introduced in Fig. 1 . For pseudonymization, the local hospitals were contacted, since only they had access to the patients' QIDs. PRIMAGE Platform The PRIMAGE data platform 48 was a cloud-based solution designed to support decision making in the clinical management of malignant solid tumors. It offered predictive tools that assist with diagnosis, prognosis, therapy selection, and treatment follow-up. These tools leverage novel imaging biomarkers, in silico tumor growth simulations, and advanced visualization techniques. Additionally, predictions with weighted confidence scores and machine-learning were employed to translate this knowledge into predictors for the most relevant disease-specific clinical endpoints. The PRIMAGE platform, based on Quibim Precision (v2.9.22), was a centralized platform deployed in the Microsoft Azure cloud. It included functionalities for data ingestion and pseudonymization. For patient's pseudonymization, an API-based integration with EUPID Services was performed. Clinical and molecular data was incorporated in the database through eCRFs that were designed together with the clinical partners of the project and implemented to guarantee consistency and completeness of the data. Data Upload and EUPID Data Transition to the PRIMAGE Platform The privacy-preserving infrastructure supported several ways of providing pseudonymized data to the PRIMAGE platform. Data that had not been pseudonymized via the EUPID Services were manually uploaded to the PRIMAGE platform. For all images from non-SIOPEN trials, QIDs were removed from the DICOM files and PRIMAGE EUPID PSNs were entered in the DICOM headers using the PRIMAGE platform. For clinical and biological data of patients with EUPID PSNs, new EUPID PSNs for the EUPID Context PRIMAGE platform were generated via the EUPID Replication API, the PSNs were manually exchanged within the datasets and the datasets with the PRIMAGE PSNs were uploaded to the PRIMAGE platform. Fig. 4 shows the different paths followed for data ingestion in the PRIMAGE project. Two pathways followed to transfer data to the PRIMAGE platform, depending on the data source. For transferring imaging data from the SIOPEN Image Server to the PRIMAGE data platform (path 2 in Fig. 4 ), a semi-automated data transition tool was used. EUPID Data Transition is defined as the transfer of pseudonymized data from one context to another, where PSNs for the source context are replaced by PSNs for the target context. Semi-automated data transition was done according to the process illustrated in Fig. 5 . Semi-automated data transition from the SIOPEN Image Management Server (IMS) to the PRIMAGE platform, replacing IMS pseudonyms (PSNs) by PRIMAGE PSNs. Manual tasks are colored in purple, automated tasks in green. Results We will first describe the characteristics of the source data which were linked (section “Source Data Characteristics”). Thereafter, in section “Record Linkage Accuracy,” results concerning privacy-preserving linkage of patient identities from different pseudonymized sources are described. Section “Data Transition and Target Data Characteristics” concerns the subsequent transfer of biological, clinical, and imaging data. Source Data Characteristics In the following, the amounts of data in the source systems of the described PPRL infrastructure at the time of writing are described. Please note that only selected data items of those data were used in the present study. The SIOPEN-R-NET contained clinical neuroblastoma data from six different trials with more than 250 centres in over 25 countries. Overall, 5,764 of the patients registered to the SIOPEN-R-NET portal were registered to a study linked to the SIOPEN-R-NET, based on a unique SIOPEN-R-NET ID, and 5,686 were also registered to EUPID. The SIOPEN Biology database contained data from 1,982 patients. All of those were pseudonymized based on the SIOPEN-R-NET ID. Additionally, 1,759 also had respective EUPID PSNs already, since they were enrolled to an EUPID-supporting SIOPEN study. The IMS contained 446.95 GB of imaging data, comprising 1,150,846 images across 8,822 series from 740 neuroblastoma patients. The GPOH/NB2016 trial incorporated 285 neuroblastoma patients with 737 imaging exams in the PRIMAGE platform. Finally, 205 neuroblastoma patients from additional hospitals were collected in the PRIMAGE platform. Record Linkage Accuracy Overall, data from two preexisting EUPID Contexts (SIOPEN-R-NET, SIOPEN IMS) and data without a preexisting EUPID Context were uploaded to the PRIMAGE platform that needed to be linked correctly with one another. In addition, pseudonymized datasets of already registered patients were updated several times, as new data of those patients became available. Despite the complex process of data upload to the PRIMAGE platform, the EUPID Pseudonymization Service and EUPID Replication Service ensured that 844 patients were uniquely added to the PRIMAGE platform, without any known double registration. Fig. 6 provides an overview of the number of patients registered to these three contexts. Venn diagram illustrating the number of patients registered in the three EUPID Contexts SIOPEN-R-NET, SIOPEN Image Management, and PRIMAGE, and patients from other trials not connected to the EUPID Services, including number of “overlapping” patients, i.e., patients with datasets in multiple contexts. EUPID, European Patient Identity. Record linkage was successfully conducted in 655 patients that were registered to the SIOPEN-R-NET, the IMS and the PRIMAGE platform. A total of 81 patients were linked between SIOPEN-R-NET and IMS but were not transferred to the PRIMAGE platform (e.g., due to insufficient image quality or completeness of imaging data). Four patients were registered to the SIOPEN IMS context only, which was unexpected. Detailed inspection of those patients revealed that three of those cases corresponded to faulty entries, most likely due to typing errors when entering the patients' SIOPEN-R-NET IDs (two represented empty datasets without any images on the server, one patient held a single image that was later reuploaded to another patient). In those three cases, the error had obviously been identified by the respective researcher, since for all three patients, correct datasets (without typing error) had been uploaded to the server—the incorrect datasets had, however, not been deleted. The fourth patient corresponded to a patient, for whom—most likely also due to a typing error—a second PSN was generated. Some of the patient's imaging data (MRI and CT images) were uploaded to the first PSN, some (mIBG images) to the second. Only the data of the PSN generated without typing error were uploaded to the PRIMAGE platform. The data uploaded to the faulty PSN had unintendedly not been considered for further analyses. Table 1 provides an overview of the number of patients registered based on the different types of QIDs, namely first name, last name and date of birth, SIOPEN-R-NET ID, or both. Table 1Number of patients which contain hashed data derived from first name, last name, and date of birth (prospective setting), hashed data derived from SIOPEN-R-NET IDs (retrospective setting) and those with both kinds of hashed data (retrospective patients that were updated prospectively), depending on the contexts to which the respective patient was registered for all patients present in more than one contextEUPID contextSIOPEN-R-NET and IMSPRIMAGE and other sourcesSIOPEN-R-NET and IMS and PRIMAGETotalFirst name, last name, date of birth(prospective)01890189SIOPEN-R-NET ID(retrospective)440652696First name, last name, date of birthand SIOPEN-R-NET ID(prospective and retrospective)0033Total44189655888Abbreviation: EUPID, European Patient Identity. Note: the overall number 888 corresponds to 44 + 189 + 655 patients in overlapping regions in the Venn diagram in Fig. 6 . Number of patients which contain hashed data derived from first name, last name, and date of birth (prospective setting), hashed data derived from SIOPEN-R-NET IDs (retrospective setting) and those with both kinds of hashed data (retrospective patients that were updated prospectively), depending on the contexts to which the respective patient was registered for all patients present in more than one context Abbreviation: EUPID, European Patient Identity. Note: the overall number 888 corresponds to 44 + 189 + 655 patients in overlapping regions in the Venn diagram in Fig. 6 . In addition to the 844 patients in the EUPID PRIMAGE context, due to privacy concerns of some participating centres, 301 patients were pseudonymized without the use of the EUPID services and registered directly to the PRIMAGE platform, without support of PPRL, leading to overall 1,145 patients in the PRIMAGE platform. Data Transition and Target Data Characteristics Manual data upload proofed to be feasible but time-consuming. While the data volumes for clinical and biological data were small (several Megabyte), large amounts of imaging data needed to be uploaded to the IMS and the PRIMAGE platform. At the start of the project, we faced several challenges during imaging data upload including pseudonymization to the IMS. Mainly, this concerned slow data upload rates leading to server timeouts and firewalls in some hospital infrastructures in combination with the original proxy-server-settings of the service. Due to these problems, in several cases, datasets needed to be uploaded more than once, with the need to not only add new patients and record, but to avoid duplicates and edit existing rather than add new datasets. Semi-automated data transition from the IMS to the PRIMAGE platform proofed to be stable and less time-consuming than manual upload. However, as shown in Fig. 4 , several manual steps were still required. We did not identify any privacy breaches or other issues that might affect data confidentiality in our study. However, certain privacy risk factors in relation to manual data management processes were identified, including the use of the wrong pseudonym during uploading, storage of data on insecure (temporary) infrastructure, incomplete deletion of temporary data, and sending credentials for access to temporary data to wrong accounts. The PRIMAGE platform that represented the target system of the record linkage process contained 1,08 TB of data. Detailed data characteristics are shown in Table 2 and Fig. 7 . Venn diagram illustrating the number of patients within the PRIMAGE platform, depending on the available data modality. Table 2Characteristics of the resulting neuroblastoma dataset in the PRIMAGE platformPropertyValueNumber of patients (female)1,145 (528)Age at diagnosis (mean ± standard deviation in months)34.46 ± 34.61Number of patients depending on tumor histology Neuroblastoma869 Ganglioneuroblastoma50 Ganglioneuroma18 No/insufficient material8 Not otherwise specified (NOS)1 Not available200 Characteristics of the resulting neuroblastoma dataset in the PRIMAGE platform The infrastructure has proven suitable for privacy-preserving linkage of distributed biological, clinical, and imaging data, supporting AI in pediatric oncology. The developed solution has been successfully utilized in various research activities, as summarized in. 43 44 45 Even as this paper was written—approximately 15 months after the end of the PRIMAGE project—except for the PRIMAGE platform, which was intended for use only during the project, the infrastructure remained operational. Discussion We have developed a privacy-preserving infrastructure that applies record linkage between multimodal data sources to support the development of AI models for imaging-based risk factors in pediatric cancer patients. The infrastructure was successfully implemented and utilized during the PRIMAGE project. The EUPID Services proved to be a valuable tool for linking prospective and retrospective pseudonymized data from various sources, supporting not only primary but also secondary use of research data. Additionally, this infrastructure can serve as a reference model for other AI projects with distributed multimodal data sources and for upcoming infrastructures, such as the EHDS. The PRIMAGE platform also contained 69 diffuse intrinsic pontine glioma patients, leading to 1,214 patients in the PRIMAGE platform. Since all these patients were directly entered to the PRIMAGE platform, without the need to link records from different source contexts before, in this paper we only focused on the 1,045 neuroblastoma patients. So far, privacy-preserving linkage in health research has largely focused on single domains or limited patient datasets, often not integrating clinical, biological, and imaging data within a cohesive infrastructure that also supports AI applications. Existing or upcoming European infrastructures like EJP RD/ERDERA virtual platform, 41 ERDRI, 42 and eventually the EHDS 49 aim to establish frameworks for privacy-preserving, cross-border health data sharing, yet they remain at conceptual or early development stages and do not support real-world applications at the scale achieved here. This study's implementation within the PRIMAGE project advances beyond previous methods, such as the SIOPEN-R-NET, 7 by introducing a comprehensive solution that integrates diverse data modalities using the EUPID Services' advanced pseudonymization capabilities. The resulting infrastructure provides an interoperable and scalable framework directly supporting artificial intelligence and predictive modelling, an emerging priority in pediatric oncology. 6 Furthermore, by operationalizing this privacy-preserving approach within the GDPR-regulated European landscape, this study responds directly to SIOPE's calls for improved data linkage capabilities and sets a benchmark for future data-driven health research infrastructures across Europe. 14 Limitations The EUPID Services supported all workflows required for the PRIMAGE project, although most data were used retrospectively, as they had already been recorded long before the project started. Record linkage could be achieved either by comparing hashes of QIDs (names, date of birth) or by hashing previously generated pseudonyms and using them as IDs for the EUPID Services. However, for patients where neither names nor pseudonyms were known across different contexts, linkage could only be established by contacting the patients' treating hospitals, which had access to both pseudonyms and QIDs. Due to privacy concerns, some centres participating in the study opted for not using the EUPID Services for patient registration, since a GDPR audit of the EUPID Services was ongoing. Therefore, 301 patients were pseudonymized based on another pseudonymization service that did not support PPRL. Record linkage for these patients needed to be done in a manual way, based on direct contact with the attending physicians. In the meantime, the security of the EUPID Services has been confirmed by an independent auditor, and the EUPID Services were certified as GDPR compliant by an external certification body (the certificate is publicly available at https://eupid.eu ). Therefore, these privacy concerns are considered obsolete for future applications. We found four patients that could not be linked correctly by the EUPID Services, most likely due to typing errors when entering the SIOPEN-R-NET IDs. While for three patients the error had been identified in time, for one patient we only found the missing linkage during the additional in-depth analysis for the present paper. Therefore, in our previous publications, 44 45 the data uploaded with typing error could not be linked to the right patient and were therefore not considered in the AI models. This error might have slightly reduced the accuracy of the AI models. However, no significant further consequences are expected. Record linkage could not be applied between patients registered exclusively based on first name, last name and date of birth (prospective setting) and patients registered based on SIOPEN-R-NET IDs (retrospective setting). If the researchers expected potential matches, they entered both kinds of QIDs. Therefore, we cannot rule out that additional linkages were missed. In addition to the EUPID Services, various other PPRL services are described in literature. Unfortunately, many currently used record linkage services do not preserve privacy as they require submitting IDs and/or QIDs to an external identity provider. While some PPRL methods have been successfully implemented and published within specific research projects, they are often not available for sustainable use. In parallel, several mature solutions have emerged, including commercial platforms such as Datavant 50 and HealthVerity's fedRAMP, 51 as well as open-source alternatives like Linkja 52 or E-PIX. 53 References to multiple further PPRL solutions can be found in. 16 Since our project required the linkage of both retrospectively and prospectively pseudonymized data, and since most of the retrospective data had already been pseudonymized with the EUPID Services, our first choice was using the EUPID Services. They were a well-established and sustainably maintained solution and had already been successfully applied to a subset of the data included in our study. As described here, they did fulfil all requirements of our study. However, for other projects, it is recommended to choose the best-fitting PPRL solution, depending on the individual project's needs. Since linkage of retrospective data comes with additional challenges, the PPRL service should be selected at an early stage. Data transition from the source context (IMS) to the target context (PRIMAGE platform) has been conducted in a semi-automated manner. All manual steps come with a certain privacy risk due to possible human errors. Currently, information security relies on manual tasks, such as granting access to uploaded data and deleting temporary files. In the future, this process will be further automated—to simplify the use of linked data for AI applications and to ensure end-to-end information security at a technical level. Linking pseudonymized data always carries a risk that patients may become identifiable due to the additional data available. Moreover, the privacy of linked data are constrained by the level of anonymity of the source data, which is particularly pertinent for imaging data. All images on the SIOPEN IMS are pseudonymized by removing identity tags from the DICOM files. However, pseudonymizing DICOM images presents challenges, as identifiable information may be contained in various (and even vendor-specific) DICOM tags, and the images themselves may allow for the reidentification of specific patients. Consequently, imaging data must always be handled with special care, both on the SIOPEN IMS and within the linked data in the PRIMAGE platform. Outlook As the next step, we are working toward fully automating the process of record linkage and data transition to target systems such as the PRIMAGE platform. These automation processes will be applied across various application scenarios within our research infrastructure, including the transition of data from distributed contexts to the SIOPEN BIOPORTAL, integration with the International Neuroblastoma Risk Group database, 54 and support for other research projects such as MONALISA. 55 The number of connected contexts is continually growing, with more IT systems being integrated into the SIOPEN BIOPORTAL and EUPID Services. Furthermore, the infrastructure is being linked to other European research initiatives, namely the EJP RD/ERDERA Virtual Platform and, conceptually, the EHDS. Conclusion To tackle the full potential of distributed data sources, PPRL is becoming increasingly important, especially in rare diseases such as pediatric oncology, where data are inherently fragmented across institutions and countries. We have demonstrated that the EUPID Services can support PPRL in prospective and—besides some limitations—even in retrospective settings. These findings provide evidence that privacy-preserving linkage can be integrated into existing research workflows without compromising patient confidentiality, thereby facilitating the creation of larger, more comprehensive datasets for collaborative research. Building on the results of this study, the existing European pediatric oncology research infrastructure is currently being further developed to incorporate PPRL as a core capability. This will enable more efficient longitudinal follow-up, cross-registry analyses, and federated studies while supporting compliance with data protection regulations. In the longer term, robust PPRL infrastructures will be a key enabler for trustworthy artificial intelligence applications, allowing high-quality, multicenter datasets to be assembled without centralizing identifiable patient data. Such solutions will also be required in federated analyses settings, to avoid duplicates in distributed records. Future work should focus on improving retrospective linkage methods, harmonizing data collection and identifier generation across participating centres, and evaluating the scalability and interoperability of PPRL services in larger international research networks. These developments will help establish a sustainable foundation for privacy-preserving, data-driven pediatric oncology research and ultimately contribute to improved outcomes for children and adolescents with cancer.