Keywords
Data sharing - federated learning - encryption - common data models - cancer
1 Introduction
Clinical informatics has transformed cancer diagnosis, treatment, and surveillance
over the past decade. Informatics efforts have allowed oncologists to better understand
rare tumors by creating larger cohorts of similar patients. Additionally, with improvements
in machine learning techniques, our ability to decode the complexity of cancer has
dramatically improved across different data streams, spanning genomics, diagnostic
imaging, and the electronic health record [[1]].
Despite increasing enthusiasm regarding the role of informatics and machine learning
in improving our understanding of cancer, the availability of clinical data remains
a persistent bottleneck to progress. This problem has become more apparent because
modern machine learning techniques often rely on larger amounts of high-dimensional
clinical data. Restrictions on sharing protected health information have made it increasingly
difficult to aggregate larger amounts of cancer data for study. Moreover, when data
sharing agreements are put in place, heterogeneity across institutions makes the process
of obtaining usable harmonized data for analysis a labor-intensive process. The process
of data harmonization itself serves as an additional barrier, as institutions must
devote resources to either adjusting models to account for differences in data or
harmonizing data before collaboration.
In response to these challenges, the cancer informatics community has attempted to
build solutions to ease the burden of sharing patient-usable data safely between institutions.
Here, we present a review of current difficulties in sharing clinical cancer data
and key literature on informatics solutions that have shown success. Specifically,
we focus on decentralized machine learning approaches, homomorphic encryption technology,
and common data models.
2 Methods
For this narrative review, we performed a search of MEDLINE with a focus on prominent
clinical informatics journals, including JCO Clinical Cancer Informatics, the Journal of the American Informatics Association, Applied Clinical Informatics, BMJ Health & Care Informatics, Informatics
for Health and Social Care, International Journal of Medical Informatics, Methods
of Information in Medicine, and the IMIA Yearbook of Medical Informatics. We reviewed articles published from
2018 to 2022 that were relevant to our discussion, with an emphasis on articles from
the past two years. We limited our review to studies in three domains: 1) decentralized
analytics/machine learning, 2) homomorphic encryption methods, and 3) common data
models. Although these domains do not cover all methods for data sharing, we chose
them because of their application in cancer informatics and their usefulness across
different types of data seen in oncology research. Finally, we limited our review
to studies that examined cancer across genomic, imaging, and clinical data streams.
3 Barriers to Sharing Patient Data
3 Barriers to Sharing Patient Data
In contrast to sharing non-clinical data streams, the sharing of patient-linked healthcare
data poses considerably higher risk. Both the United States Health Insurance Portability
and Accountability Act (HIPAA) and the European General Data Protection Regulation
(GDPR) impose strict rules regarding the exchange of personally identifiable healthcare
data for non-care purposes, including research. Due to such restrictions, multi-institutional
efforts to aggregate large patient cohorts are often burdensome, time-consuming, and
require significant resources [[2], [3]]. Successful examples of large collaborative datasets include the AACR-sponsored
Project GENIE, which provides genomic data with limited outcomes data across 18 institutions
representing 160,000 cancer patients in 6 countries [[4]]. Given such difficulties, large data collaborations are often limited to a small
number of like-minded institutions with adequate resources and enthusiasm. Although
such small efforts have value, they risk potential bias because they lack data representing
the entire spectrum of cancer patients. This is of particular importance given that
institutions and health systems without resources to participate in such studies may
represent different patient populations. Kaushal et al. [[5]] found that a minority of clinical deep learning research studies (24%) analyzed
multiple institutional data. Moreover, they found that many of the studies used data
from only three states within the US (California, Massachusetts, and New York), with
a paucity of studies from the other 47 states.
Patients have expressed concerns regarding the privacy of their healthcare data. Khullar
et al. [[6]] conducted a survey of patient perspectives on artificial intelligence within healthcare
found that over 70% of surveyed patients expressed concerns about privacy breaches
associated with healthcare data. This proportion was higher among non-white (74% vs.
68%) and older respondents (72% vs. 69%).
Another barrier to sharing patient data is a lack of interoperability between institutional
datasets. Despite increasing digitization of healthcare data, there is a lack of standardization
across many clinical data streams, most notably the clinical data within the electronic
health record [[7]]. Although the Digital Imaging and Communications in Medicine (DICOM) provide a
data standard in diagnostic imaging, differences in imaging parameters, equipment,
and protocols among different institutions make models trained on multi-institutional
datasets less likely to generalize across different patient populations and regions
[[7]]. Additionally, although data standards such as FHIR/HL7 have the potential to increase
interoperability, adoption has been lacking [[8]].
4 Anonymization
Anonymization remains the most common way of sharing patient level data that has shown
some success. Removing identifiable features from a dataset mitigates the risks associated
with sharing data across institutions. A main theoretical advantage of data anonymization
is that it allows aggregated data to be shared within the public domain for further
research studies. Successful examples of de-identified public cancer data repositories
include the Cancer Genome Atlas (TCGA) [[9]] and Cancer Imaging Archive (TCIA) [[10]] which have both required considerable investment by the US National Institute of
Health (NIH). The TCGA tiers its data into different layers of sensitivity, with only
the most general layer being openly available to users. However, anonymization still
presents challenges that make data sharing difficult. One of the challenges is that
true anonymization is difficult to achieve, and there is always a risk that patient
data can be re-identified despite best efforts to remove information. Second, effectively
anonymizing various data streams requires significant resources, expertise, and time
which can be difficult for resource limited institutions [[11]]. Third, anonymization may remove important information needed for analysis [[12]]. Lastly, there are differing views of what constitutes truly anonymized data, often
placing different institutional policies at odds with one another when trying to share
data [[13], [14]].
5 Decentralized Analytics
5 Decentralized Analytics
One emerging solution to help mitigate risks and resources required to anonymize data
for shared informatics projects is decentralized analysis. Decentralized analysis
allows institutions to keep healthcare data locally but conduct informatics analysis
collaboratively. These solutions have become popular in the setting of machine learning
techniques which have large data requirements. Typical decentralized machine learning
techniques involve sharing some parameters of a model during the training process
without sharing actual protected health information.
Chang et al. [[15]] demonstrated the efficacy of cyclical weight transfers to train deep learning models
to classify breast mammograms. Cyclical weight transfer involves multiple institutions
transferring parameters of a model during the training process. Because the model
is training on data from both institutions intermittently, there is less likelihood
of overfitting to one institutions dataset. The authors found that cyclical weight
transfer created a model with similar accuracy to a model created with all data aggregated
centrally. The method was scalable to 20 hypothetical institutions and maintained
relatively strong performance when one institution had lower quality data or imbalanced
datasets. One potential disadvantage of this method is that performance is strongest
when all institutions are training at the same time. Also, the cycle is dependent
on the computational speed of each institution meaning training may be limited by
the resources of the slowest institution.
Deist et al. [[16]] demonstrated one of the largest successful algorithms trained in a decentralized
manner. The authors trained a logistic regression model to predict 2-year survival
based on TNM staging on over 20,000 non-small cell lung cancer patients across five
countries. The model which used TMN stage as the inputs was trained in a distributed
manner using the Alternative Direction Method of Multipliers technique. It is unclear
whether this method is applicable to more complex data streams with larger numbers
of parameters, but nevertheless demonstrates the potential of decentralized machine
learning techniques to aggregate large international cohorts of patients.
Federated machine learning techniques were initially developed in 2015 and represent
a different method for decentralized machine learning [[17]]. Federated learning algorithms rely on distributing copies of a machine learning
algorithm to institutions which house their own protected data [[18]]. Training iterations are completed locally and return results to a central repository
for aggregation. The central repository then provides a new global model to re-distribute
to devices for further training. The major benefits of federated learning compared
to alternative decentralized machine learning techniques is the flexibility to operate
when some devices are off-line. The disadvantage of federated learning is that performance
potentially decreases if the incorrect aggregation strategy is chosen when configuring
the global model.
Federated learning has been shown to be effective for cancer image analysis. Recently
the German Cancer Consortium released a Joint Imaging Platform for federated clinical
image analysis [[19]]. The developers successfully created a federated learning platform which was implemented
across 10 institutions within Germany. The platform is currently being implemented
to house data for six multi-center clinical trials investigating a variety of different
cancer types. The platform highlights the potential of federated learning to improve
the efficiency of large clinical trials in oncology which are frequently multi-institutional
in nature.
Sarma et al. [[20]] demonstrated the utility of federated learning for biomedical image segmentation.
The authors trained a deep learning algorithm which successfully segmented prostates
on MRI across three institutions. The federated learning algorithm was more generalizable
and accurate when compared to models trained locally (Dice 0.812 vs. 0.889, p<.001).
Federated has also been applied to histopathological image analysis. Agbley et al. [[21]] demonstrated the ability to identify invasive carcinoma on breast pathology specimens
across three separate institutions. The authors noted that the federated learning
underperformed when there was significant class imbalance between institutions. This
suggests that federated learning systems may not perform well when combining very
disparate patient populations.
Lu et al. [[22]] similarly demonstrated federated learning to be an effective method for classification
of histopathological data. The authors were able to train image classification models
to predict survival from whole slide images across four different institutions. Although
the authors found federated models did show strong performance, the models did underperform
when compared to centrally trained models. It is unclear from the study if the performance
gap between central and federated models would be mitigated if they had a larger number
of institutions.
Although federated approaches have most often been focused on supervised machine learning
tasks, it has been shown to also be applicable to unsupervised learning algorithms.
Bercea et al. [[23]] developed a framework to train an unsupervised autoencoder to identify high grade
gliomas on brain MRIs across four institutions. The authors found their federated
approach improved glioma identification by 80% compared to locally trained models.
Specifically, the authors noted that the dramatic increase in performance was because
each individual institution did not possess enough cases to adequately train their
model. Only through a federated approach across multiple institutions were the authors
able to have a sample size large enough to successfully complete their task.
Federated learning using clinical data has been shown to be effective in small settings.
Rajendran et al. [[24]] demonstrated both neural network and logistic regression models predicting risk
of developing lung cancer from electronic health data could successfully be trained
in a cloud based federated environment. Notably the authors found that the logistic
regression model did not show significant improvements in performance when trained
using larger federated data source. In contrast the neural network showed significant
improvements in performance when trained on a larger dataset via federated learning.
Hansen et al. [[25]] similarly used clinical data across three countries to build a cox regression model
to identify factors associated with larynx cancer outcomes. The federated model showed
strong discriminatory ability with AUCs ranging from 0.67 to 0.77 but did not significantly
outperform non-federated localized cox models. The authors do note that the federated
model shows slightly better separation of risk groups compared to localized models.
These findings further highlight that certain data intensive machine learning methods
may benefit the most from federated environments.
Federated learning does have known challenges. Scalability of federated learning infrastructure
on cancer problems remains understudied. A systematic review of studies using federated
databases completed by Zerka et al. [[26]] found that published studies on federated learning involved less than 10 institutions
and often only analyzed a few hundred patients.
One of the largest demonstrations of federated learning at scale was published by
Pati et al. [[27]] With a group of 71 institutions across six continents the authors trained a successful
auto-segmentation model for glioblastomas on brain MRIs. As expected, the authors
found increased data improved overall model performance and made federated models
more robust to potential data quality issues at individual institutions.
Although federated learning does allow individual devices to retain protected data,
they may still be sensitive to privacy threats. These threats include attempted extraction
of training data information from intermediate/final models and corrupting models
to produce inaccurate results [[28]
[29]
[30]]. Evidence suggests that nefarious actors maybe able to reconstruct individual data
located on devices if given access to model parameters while training a federated
model [[31]].
6 Homomorphic Encryption
Encryption is alternative technique to facility patient data sharing which is thought
to be less sensitive to privacy threats than decentralized analytics. Based on the
fundamentals of number theory, encryption techniques transform original data into
an encoded format [[12]]. Among the most popular encryption techniques within healthcare is homomorphic
encryption. Homomorphic encryption is a specific type of encryption which enables
primitive mathematical operations (for example addition, multiplication) directly
on encoded data. The advantage of homomorphic encryption techniques is that they offer
more certain privacy of data. The disadvantage is the significant computational resources
to successfully encrypt medical data. There appears to be an efficiency security trade-off
where the most efficient scalable encryption methods are likely less secure [[18]].
Improving the efficiency of homomorphic encryption methods remains an area of continued
research. In 2018, the iDASH Privacy and Security Workshop organized a special competition
track to create secure parallel genome wide association studies using homomorphic
encryption [[32]]. The winning homomorphic encryption solutions from Duality Technologies [[33]] and UCSD [[32]] successfully completed full GWAS for 1,000 individuals in approximately 4 and 2
minutes respectively. Both solutions chose a similar common encryption framework.
CKKS/HEAAN appears to be amenable to numerical optimization for problems that involve
machine learning and statistical learning. Recently the group from Duality Technologies
improved upon their iDASH winning solution resulting in improved computational efficiency
and reduce computer memory usage [[34]]. They also demonstrated the scalability of their method on a larger dataset of
25,000 individuals.
Homomorphic encryption on images is particularly challenging given difficulty encoding
visual images. Khilji et al. [[35]] successfully demonstrated the ability to train a deep learning classification model
using homomorphic encryption. The model which attempted to diagnose the present of
Acute Lymphoblastic Leukemia from pathologic images had an accuracy of 77.9%. Their
classification model although somewhat accurate did underperform compared to non-encrypted
models (77.9% vs. 80.0%).
Homomorphic encryption on clinical data has been less studied. Son et al. [[36]] demonstrated the ability homomorphic encryption to securely train model for breast
cancer recurrence using clinical data from 13,000 patients. Interestingly the authors
found performance of the model trained using their homomorphic encryption method performed
equivalently to models trained on un-encrypted data.
Paddock et al. [[37]] demonstrated the feasibility of homomorphic encryption to identify exceptional
tumor responders within a real-word dataset. The authors were able to identify all
exceptional responders in 21 months over the course of the simulated study. Although
the encryption and decryption process was computationally expensive (often requiring
hours of computation), the authors argue the rate limiting step compares favorably
to cohort identification and aggregation using traditional methods.
Combining federated learning and homomorphic encryption is arguably the safest solution
to protect health information. Froelicher et al. [[29]] described a novel federated learning platform which leverages multiparty homomorphic
encryption to enable privacy preservation on distributed datasets. The authors showed
their platform reproduced two published centrally trained models: 1) predicting survival
after the receipt of immunotherapy and 2) predicting HIV viral load from genetic data.
These findings suggest that combining data sharing approaches may best ensure patient
data security.
7 Common Data Models
Both decentralized analytics and encryption techniques help address privacy concerns
which limit data sharing, but do not address interoperability issues that plague cancer
informatics projects. Common data models represent a potential solution to help improve
data sharing while also addressing interoperability issues among institutions with
different informatics infrastructure. Given there exists several different common
data models which are already in production across healthcare, these common data model
solutions may be the most likely to impact clinical practice in the short-term.
Anonymization is a potential advantage of common data models. Common data models often
are designed with privacy in mind and can easily eliminated elements with protected
health information when sharing across institutions. Additionally, common data models
can identify sensitive data elements and collectively build data sharing policies
which allow for tiered access based on privacy risk. The use of common data models
is not mutually exclusive to decentralized analytics or encryption techniques. Combining
techniques may potentially provide superior methods for data sharing. Combining common
data models with homomorphic encryption has been proposed by Rosario et al. [[38]] The authors demonstrated the feasibility to use homomorphic encryption and apply
it to the i2b2 common data model. This allows the potential to share common data model
elements in a secure fashion or more safely store them within a cloud environment.
One emerging disadvantage to common data models is difficulty harmonizing between
popular common data models [[39]]. Some common data models employ traditional relational database design in which
each table corresponds to a clinical domain (e.g., PCORnet CDM). In contrast, other
common data models use alternative structures. The commonly employed i2b2 common data
model employs a star-schema format which leverages one large ‘fact’ table to connect
various concepts. Similarly, the Observational Medical Outcomes Partnership (OMOP)
common data model uses a hybrid approach which blends domain tables and ‘fact’ tables
[[39]]. Recognizing the difficulty of harmonizing common data models, the ONC launched
a common data model harmonization initiative in collaboration with the FDA, NCI, NIH
and NLM. The three year project which was completed in 2020 developed a common data
architecture to facilitate interoperability between four common data models commonly
used in clinical medicine (Sentinel, PCORNet, i2b2, and OMOP) [[40]]. Additionally, common data models are limited to certain data elements and organizational
paradigms which may not align well with all clinical specialties. For example, various
common data models record radiation therapy for cancer treatment differently with
different degrees of detail [[41]]. Specifically within oncology, significant efforts have been made to attempt to
align current common data models to represent cancer data in a way which is most relevant
to cancer research [[42], [43]]. Although considerable literature has been devoted to common data models in healthcare,
we will review advances in common data models specifically within the domain of oncology.
The OMOP common data model is among the most popular common data models used in medicine.
Developed by Observational Health Data Science and Informatics (OHDSI) program, OMOP
enables analysis of disparate observational databases through common terminologies,
vocabularies, and coding schemes [[44], [43]]. There exist a number of extensions created by the OHDSI Oncology Subgroup and
a number of independent groups have also attempted to create their own ways to extend
OMOP for cancer informatics research.
Warner et al. [[44]] in collaboration with OHDSI have made OMOP more useful for cancer informatics by
extending OMOP to better capture the structure of chemotherapy regimens. Leveraging
a HemOnc.org a curated website of chemotherapy regimens, the authors were able to
successfully map content from HemOnc.org to a relational data model that is compatible
with the OMOP common data model. In addition to increasing the chemotherapy information
available to OMOP users, the authors also created an extension to the OMOP CDM to
handle episodes of care allowing for the capture cancer treatment information with
temporal information [[45]].
In addition to chemotherapy information OHDSI recently released an OMOP Oncology Module
[[43]] which extends the OMOP CDM and standardized vocabularies to better represent cancer
diagnoses, treatments, and episodes. The module incorporates information from seven
existing standards including the WHO International Classification of Diseases for
Oncology, HemOnc.org, North American Association for Central Cancer Registries, College
of American Pathologists Electronic Cancer Checklists, Nebraska Medical Clinical Ontology
Application, National Cancer Institute Thesaurus, and the Anatomical Therapeutic Chemical
Drug classification system. The module was successfully pilot tested at six institutions.
The developers noted that the integration of the electronic health record to institutional
tumor registry information was necessary to successfully fill the OMOP module.
Yu et al. [[46]] demonstrated the potential utility of common data models to collect adverse events
in patients receiving immunotherapy. Using the OMOP common data model, the authors
were able to identify ipilimumab induced hypopituitarism four months earlier than
the FDA Adverse Event Reporting System.
The OMOP common data model has also been used to conduct epidemiological studies regarding
cancer incidence. Lee et al. [[47]] used an OMOP created dataset across three hospitals in South Korea to test the
association between thiazide usage and non-melanomatous skin cancer prevalence. The
investigators were able to leverage OMOP common data elements to create a cohort of
over 600,000 patients over the course of years. Such large-scale studies would be
difficult to complete without the use of a common data model to efficiently aggregate
data elements of interest.
Despite the popularity of the OMOP common data model across a variety of clinical
disciplines [[48], [49]], one disadvantage for cancer informatics research is the lack of genetic information
included in the standard OMOP tables. To address this Shin et al. [[50]] developed a genomic common data model which allows genomic information to be integrated
within standard OMOP data tables. The authors showed that successful implementation
of their proposed common data model allowed for successful comparison of genetic data
between the Cancer Genome Atlas and Ajou University Hospital in South Korea.
Like the OMOP, PCORnet is an alternative common data model which was proposed by the
Patient Center Outcomes Research Institute to facilitate large-scale patient centered
research. Carnahan et al. [[51]] examined the utility of PCORnet to evaluate the utilization of molecular-guided
cancer treatment and testing across nine clinical research networks and two health
plan research networks. The authors found traditional billing codes to be effective
at identifying molecular testing, but unable to adequately capture cancer specific
details regarding the analyte being tested. In a sub-analysis, the authors found that
PCORnet failed to capture all patients who received molecular guided therapy and recommended
linkage of PCORnet with tumor registries to improve data capture.
One consistent challenge across proposed common data models remains the lack of specificity
to clinical oncology. Most common data models lack important cancer-specific variables
(staging, molecular testing, adverse events) that are important to cancer informatics
researchers. To address growing concerns, the Minimal Common Oncology Data Elements
(mCODE) initiative was started in 2018 [[52]]. The mCODE initiative which is led by the American Society of Clinical Oncology
(ASCO) attempts to providing infrastructure regarding data elements which can be used
across electronic health records. Some critiques of the mCODE initiative are the lack
of variables capturing smoking and drinking status and the ability for it to be implemented
outside of the United States [[53]]. Although relatively recent, mCODE has been created partly with the hope of guiding
future common data models to include more cancer specific data elements [[54]]. There is increasing enthusiasm regarding the use of mCODE to help better capture
and standardize oncology data. Specifically, CodeX the HL7 FHIR accelerator focused
on interoperability is implementing and testing the use of mCODE within specific use
cases [[52]]. Use cases include cancer registry reporting, EHR derived endpoints for cancer
clinical trials, cancer clinical trial matching, prior authorization, and capturing
radiation therapy treatment data for cancer patients.
An alternative to mCODE named OSIRIS was recently proposed by the French Institute
National du Cancer (INCa) [[55]]. OSIRIS is a minimum data set framework composed of 67 clinical and 65 omic items
which was validated on 300 patients across six clinical trials across different cancer
types. OSIRIS is compatible with the HL7 Fast Healthcare Interoperability Resources
(FHIR) format. Features of the OSIRIS common data include temporal structure to capture
longitudinal cancer events, a blend of omics and clinical concepts, ability to integrate
future data streams (e.g., proteomic), and the presence international terminologies.
8 Future Directions and Conclusions
8 Future Directions and Conclusions
Enabling data sharing, whether through use of common data models, federated learning,
or other approaches is key to reaching the vision of learning from every cancer patient.
There have been considerable advances to help facilitate data sharing within cancer
informatics. Decentralized machine learning, encryption, and common data models represent
three promising solutions which can help ease the burden of aggregating large cancer
datasets. Each solution poses advantages and limitations which are being investigated
in parallel. It is likely that no one informatics approach will be appropriate for
all types of data and clinical settings. More likely tailored approaches will be personalized
based on a specific informatics task, resources, and research expertise. Additionally,
combined solutions may offer the safest and most applicable method to facilitate collaboration.
Recent regulatory changes have implications for promoting the sharing of healthcare
data and advancing research. The Office of National Coordinator of Health Information
Technology (ONC) finalized the 21stCentury Cures Act and its companion Cures Act Final Rule on April 5th, 2021, which aim to facilitate data sharing with patients and may provide momentum
for institutions to create infrastructure for safe and effective sharing of PHI with
researchers [[56]]. In addition, the newly enacted NIH Data Management and Sharing Policy requires
prospective planning for managing and sharing scientific data during the application
process for NIH-funded projects, promoting greater transparency and data sharing [[57]]. Such initiatives are expected to encourage institutions to invest in data sharing
technology and infrastructure at scale. Furthermore, the NIH Cloud Platform Interoperability
effort represents a significant investment by the NIH in promoting infrastructure
for democratizing data for researchers and is likely to play a larger role in promoting
data sharing in the future.
Although significant progress has been made in prototyping these solutions, there
are few examples of wide adoption at scale within oncology. Current progress was frequently
limited to organizations with significant informatics expertise and resources. The
next iteration of research in this area will likely focus on developing scalable solutions
which are user-friendly for clinicians with varying levels of expertise. Given the
current prototyped solutions across different data types, oncology is a field well
positioned to develop such scalable solutions and become among the leading clinical
fields in this arena.
No conflict of interest has been declared by the author(s).
