Federated Learning in Precision Oncology: Bridging Multi-Institutional Artificial Intelligence and Clinical Implementation

Introduction

Precision oncology is one of the major achievements of modern medicine that allows making specific diagnostics, prognoses, and treatment recommendations based on unique patient data at the molecular, pathological, radiological, and clinical levels. Meanwhile, the use of AI technologies in oncology has made it possible to improve the processes of cancer screening, diagnostics, molecular analysis, treatment selection, trials matching, prognosis estimation, and even patient surveillance. Yet the creation of advanced and generalizable oncology algorithms is hampered by lack of accessible data due to privacy and institutional regulations, which do not allow for gathering large-scale databases of patient data. Federated learning (FL) is one such method that offers hope in addressing this issue. Through the provision of mechanisms through which multiple institutions can learn from the same model without having to share patient-level data, FL enables effective and private collaboration to build on the strength of distributed data sources. Besides its application as an approach to data privacy, FL can act as the underlying infrastructure in creating multi-institutional clinical-grade artificial intelligence ecosystems within the scope of precision oncology. This, however, would require advancements in various areas.^1,2 Despite remarkable progress in the field of oncology AI, the failure to generate sufficiently diverse datasets through collaborations across multiple institutions remains a significant bottleneck in translating the research into a clinical setting. We believe that FL is an essential technological backbone capable of addressing this problem, allowing collaborative creation of models without jeopardizing patients’ privacy. However, the future success of FL depends on prospective validation, governance, interpretability, and clinical workflow integration to a much larger extent than on technology itself. Therefore, FL should not be seen merely as a privacy-protected tool but as a platform that empowers clinically validated, multi-center-enabled precision oncology.

Why Precision Oncology Requires Federated Learning

Precision oncology is dependent on the combination of various data types, including imaging, digital pathology, genomics, transcriptomics, proteomics, EHR, and clinical outcomes. All of these data points contribute to creating an exhaustive dataset that helps understand tumor biology and develop customized treatment regimens for patients. ³ The main issue with the current state of oncology data is that most health care institutions still struggle with data fragmentation, and even big cancer centers cannot boast diverse enough sets of data to build clinically reliable AI models for rare cancers, molecular variants of tumors, pediatric tumors, and marginalized populations. In this regard, FL presents a viable alternative. It enables model building through collaboration among health care facilities without the need to share data. Using FL, health care organizations will be able to collaborate on building a model using their datasets while maintaining patient data on premises, thus gaining access to larger distributed datasets. Overall, FL allows building more diverse and representative AI models.^2,4

Federated Learning Across the Precision Oncology Continuum

The applications of FL in the different phases of cancer can be summarized as follows. Within imaging, FL has been shown to offer potential benefits in many radiomics applications, including tumor detection, segmentation, response evaluation, recurrence assessment, and survival analysis. Multi-hospital FL models have already demonstrated performance comparable to centralized models while offering increased generalizability and reduced institutional bias. Another rapidly evolving field where FL has been successfully implemented is that of digital pathology. Through federated network architectures capable of analyzing whole-slide images without sharing any patient information, FL approaches can assist in tumor classification, prognostic assessment, biomarker discovery, and disease progression prediction. This method comes in handy considering the diversity in staining techniques, scanners used, and even annotation methods in different hospitals. The application of genomics and multi-omics analysis has been widely applied in precision medicine, which helps in identifying actionable mutations, classifying molecular types, and treating patients through targeted therapy. It is possible to analyze genomics, transcriptomics, epigenomics, and proteomics data using FL methodology without violating patient privacy.

FL could also improve molecular tumor board functionality by allowing institutions to jointly analyze the efficacy of treatments linked to rare molecular markers, which will increase evidence-based decision-making for better precision treatments. In addition to disease diagnosis and treatment plan formulation, FL is expected to play an integral role in the process of matching patients with clinical trials and the creation of real-world evidence using federated NLP and EHR models. Finally, federated longitudinal outcome modeling could enable continuous learning based on treatment reactions, side effects, and patient survival.^5,6

Current Evidence and Emerging Trends

FL applications in oncology have gained extensive evidence in the recent years. Multicenter imaging studies have shown promising results in terms of FL application for tumor segmentation, lesion detection, response prediction, and radiomics-based prognosis. Large-scale federated efforts for glioblastoma segmentation across numerous institutions and patients have provided evidence of FL feasibility for privacy-preserving collaboration. Federated deep learning approaches in digital pathology have proven efficient in tumor diagnosis, prognostics, and prediction of disease progression. Recently developed explainable federated pathology models have achieved enhanced interpretability and have shown potential for clinical applications. There have been numerous research works on federated genome-wide association studies in order to discover biomarkers, classify molecular subtypes of cancers, and predict therapeutic response while ensuring genomic privacy. In addition, FL models for electronic health records and natural language processing have also shown promise in terms of clinical data analysis, toxicity prediction, treatment suggestion, and clinical trial recruitment. The most promising avenue of FL in oncology appears to be multimodal FL encompassing imaging, pathology, genomics, laboratory values, and clinical outcome data.

Nonetheless, the majority of researches published till date are mostly retrospective and experimental in nature. Performance measures are quite impressive, but the number of findings in terms of practical application to clinical practice is rather small. Hence, FL may be regarded as a promising but yet emerging part of precision oncology infrastructure.^5–7

From Technical Performance to Clinical Readiness

It is important to emphasize the difference between technical success and clinical success. FL studies often present impressive results such as high accuracy, sensitivity, specificity, Dice similarity coefficient, and area under receiver operating characteristics curve. Nevertheless, good prediction may not lead to better patient outcomes. The majority of oncology models based on FL have been validated retrospectively using carefully selected data, while the clinical reality is characterized by high variability related to differences in patients, imaging, pathology, sequencing, and treatments. Thus, models demonstrating high predictability in experimental studies may fail in the same conditions during clinical use. Furthermore, successful implementation cannot be ensured by predicting alone. There is a need for AI to be able to communicate effectively with doctors, to provide data that can be interpreted easily, to affect the decision-making process, and to demonstrate its effectiveness through results. The use of prospective studies in multiple centers is not common, and papers on better patient outcomes are few. Future research in the field should focus on prospective validation, implementation science, integration into routine workflow, monitoring, and outcome assessment. Therefore, in the case of FL, apart from the predictive power, it will be essential to measure its effect on clinical care.^1,8

Governance, Trust, and Implementation Barriers

However, even with its potential benefits, FL poses some obstacles that need to be overcome before its implementation in clinical practice. The first issue that could prevent FL from becoming commonplace is the issue of heterogeneity in institutional data. Datasets used for oncology studies tend to be heterogeneous within institutions because of different imaging systems, pathology processes, genomics, and patients’ characteristics. This is likely to lead to the existence of non-identically and independently distributed (non-IID) data. Another problem related to data is that of quality and annotation consistency. There could be differences in tumor segmentation, labeling of outcomes, and other annotations that would bias results. Clinicians will be unwilling to use the “black box” nature of any artificial intelligence model. Explainable models that give clinically significant explanations of the results of such predictions might increase transparency and allow better decision-making and even compliance with regulations. While FL decreases the necessity of exchanging raw data, the possibility of cybersecurity threats still exists. Such attacks might include the model inversion attack, membership inference attack, and malicious parameter manipulation. There are also privacy-enhancing technologies that could assist in dealing with this problem.

However, several significant regulatory and governance issues are yet to be answered. As more institutions collaborate on creating and implementing AI tools, complexities related to auditability, accountability, responsibility, and liability increase. These issues will necessitate cross-sector collaboration between health care institutions, developers, and regulators.^9,10

Practical Priorities for Clinical Translation

There are some key considerations that need to inform the next stage of FL research within precision oncology. One is the establishment of prospective multicenter studies for validating clinical utility and ensuring patient benefits. Another is the standardization of processes for data annotation, reporting, benchmarking, and model evaluation for improved consistency across federated research projects. Explainable federated AI systems will have a high priority for increased transparency, credibility, and acceptance among clinicians and regulators. Federated infrastructures should consider security and privacy from the start and not as an afterthought. Governance structures for issues such as accountability, auditing, and liability need to develop along with technology itself. Last, FL should integrate into the evolving world of precision oncology with its molecular tumor boards, rare cancer consortia, multi-modal clinical decision support, and continuously learning health care ecosystems that incorporate real-world evidence. ⁴

Conclusion

FL is advancing from being a machine learning technique that protects privacy through computation to becoming a promising infrastructure for precision oncology. With FL, institutions can build and validate artificial intelligence algorithms without sharing patient data, offering an exciting method to overcome a significant barrier to multicenter and precision cancer research and treatment. Recent research reveals that FL could be applied to a range of tasks in precision oncology, including: radiomics, digital pathology, genomics, multi-omics, molecular tumor boards, clinical trial matching, prediction of longitudinal outcome, and multimodal AI.

Despite recent technological developments, FL is still at a crucial stage. Most studies have been retrospective in design, and numerous issues regarding prospective validation, workflow integration, interpretability, cybersecurity, governance, auditability, and regulation need to be solved for clinical implementation to become possible. As a result, should not only be seen as a machine learning technique that preserves privacy but also as a fundamental element of developing trustworthy multi-center artificial intelligence infrastructure for precision oncology.

Authors’ Contributions

A.H. and A.H.: Conceptualization, investigation, data curation, writing—original draft preparation, writing—review and editing, validation. All authors critically reviewed and approved the final version of the article.