EXpectation Propagation LOgistic REgRession on permissioned blockCHAIN (ExplorerChain): decentralized online healthcare/genomics predictive model learning.

OBJECTIVE: Predicting patient outcomes using healthcare/genomics data is an increasingly popular/important area. However, some diseases are rare and require data from multiple institutions to construct generalizable models. To address institutional data protection policies, many distributed methods keep the data locally but rely on a central server for coordination, which introduces risks such as a single point of failure. We focus on providing an alternative based on a decentralized approach. We introduce the idea using blockchain technology for this purpose, with a brief description of its own potential advantages/disadvantages.
MATERIALS AND METHODS: We explain how our proposed EXpectation Propagation LOgistic REgRession on Permissioned blockCHAIN (ExplorerChain) can achieve the same results when compared to a distributed model that uses a central server on 3 healthcare/genomic datasets, and what trade-offs need to be considered when using centralized/decentralized methods. We explain how the use of blockchain technology can help decrease some of the problems encountered in decentralized methods.
RESULTS: We showed that the discrimination power of ExplorerChain can be statistically similar to its counterpart central server-based algorithm. While ExplorerChain inherited some benefits of blockchain, it had a small increased running time. DISCUSSION: ExplorerChain has the same prerequisites as a distributed model with a centralized server for coordination. In a manner similar to secure multi-party computation strategies, it assumes that participating institutions are honest, but "curious."
CONCLUSION: When evaluated on relatively small datasets, results suggest that ExplorerChain, which combines artificial intelligence and blockchain technologies, performs as well as a central server-based method, and may avoid some risks at the cost of efficiency.

INTRODUCTION

Predictive modeling and cross-institutional collaboration

Applying machine learning methods to predict patient outcomes using data sources such as electronic health records and genomics data or healthcare predictive modeling, is an increasingly popular and important area and can facilitate research studies and advance quality improvement initiatives. Specifically, such modeling methods based on artificial intelligence (eg, logistic regression or deep learning algorithms) can help generate scientific evidence for comparative effectiveness research, accelerate biomedical discoveries, and improve patient care. For example, a medical center may use predictive analytics to identify and take measures that can potentially prevent 30-day readmissions, thus improving care while avoiding penalties from regulatory agencies., Although the concept of predictive modeling is appealing, some diseases and conditions are rare and therefore require data from multiple institutions. To obtain enough data for model construction, sharing healthcare/genomic data across institutions is an intuitive solution to improving models. While a single institution may only have few patient records, taken together, a set of institutions may provide adequate numbers for a generalizable model. Aggregating all cases in a registry or data coordinating center database has been a common strategy to manage these databases. However, this leads to a single point of failure or breach, inaccessibility of data, or improper data disclosure that may place sensitive protected health information at risk, respectively. Once data are passed on to other entities, it is difficult to ascertain the chain of custody, leading to potential misuse.

Distributed predictive modeling and potential blockchain solution

As a result, combining cross-institution data into a single repository for model learning has been challenged as the ideal model to protect privacy of individuals, and/or institutions, while ensuring that no institution has full responsibility for all the data. Multiple algorithms have been proposed to conduct predictive modeling in a distributed way without the need to aggregate data centrally. That is, by disseminating machine learning models instead of transferring observation-level patient data, the privacy of individuals can be further protected. However, many of these privacy-preserving modeling algorithms still need a central server to intermediate the process of modeling and combine the global model. Such an architecture carries its own risks (Figure 1a), such as failure to build, tune, or evaluate a model in case the central coordinating server is down, and the potential for this central coordinating server to try to breach the privacy of individuals or institutions by examining aggregate statistics.

Figure 1.

The network topology for a 4-site (s, s, s and s) setting, with each site holding protected health information data (d, d, d and d). (a): Centralized network topology adopted by state-of-the-art learning methods. Such an architecture carries several risks: (R1) the sites may not be allowed to transmit data outside specific computer environments due to their institutional policies, (R2) the data being disseminated and the transfer records could be altered without a clear way to determine immutability, (R3) trained models could also be tampered within the central server without being noticed by other participating sites and thus undermines provenance,, (R4) the server represents a single point of failure,, (R5) additional security to protect data is not offered, (R6) the client-server architecture may present synchronization problems,, and (R7) the sites cannot join/leave the network at any time. Furthermore, long-term sustainability of the whole network becomes dependent on the institution that serves as the central coordinating node (R8). Typically, once a coordinating institution is chosen, the network architecture is built around the coordinating center, with nodes serving as data providers that are unable to assume the role when needed. (b): Decentralized network topology (blockchain) adopted by ExplorerChain. Five desirable features make blockchain suitable to mitigate the problems faced by centralized architectures: (R1) Blockchain is, by design, decentralized; the verification of transactions is achieved by majority voting., Each institution can control the use of computational resources. (R2) A blockchain provides an immutable audit trail. That is, changing the data or records is very difficult., (R3) The traceable origins certify data provenance., In our case, each trained model is recorded in a collaborative and distributed ledger, which cannot be updated silently by any of the sites without being noticed. (R4) The peer-to-peer architecture of blockchain ensures that there is no risk of single point of failure,, and thus improves security and robustness. Also, by removing the dependency on a central node, blockchain increases the availability of the models at all sites at all times. (R5) The enhanced security/privacy features of blockchain further protect data and models. Additionally, (R6) The blockchain mechanism can remove synchronization conflicts.,,, (R7) Each site can join/leave the network freely without imposing overhead on a central server or disrupting the machine learning process.,, Finally, network long-term sustainability (R8) is increased because its architecture is fully transparent and each participating site can collaborate with low operation/maintenance costs.

To address these challenges, a plausible solution is to adapt distributed databases to build and evaluate predictive models in a peer-to-peer architecture. These peer-to-peer distributed models cannot work with all the data at once. However, partial batch training (ie, batch training at each site) is possible, followed by an online or “transfer learning” strategy in which the next site or peer gets a partially trained model and improves it using its own data. In this sequential peer-to-peer process, it is important that no site constitutes a single point of failure and that all sites can easily verify how the model is evolving.

One of the main problems with online machine learning models (centralized or distributed) is the potential for a limited number of initial inputs force gradient descent methods to be trapped into limited solutions. Although the proposed strategy does not eliminate or necessarily mitigate this problem per se, enabling the use of data from multiple sites for model training in a way that does not centralize the data allows more institutions to participate, so the model parameters can evolve in different directions. Also, a major challenge with distributed databases that rely on a central coordinating server is the vulnerability to a single point of failure and potential tampering with logs at the central server side.,, On the other hand, for distributed databases that rely on a peer-to-peer network, one of the main problems relates to the coordination of access. If the path from peer-to-peer is easily traceable and broadcasted to the whole network, this problem is solved. Transparency of usage in the form of an immutable ledger that can be easily verified by the whole network is therefore a desirable feature. When asking queries and building machine learning models using sensitive data from several institutions, maintaining a robust immutable ledger of access is critical. Additionally, resilience to failure at any node and disintermediation (ie, lack of a central authority) are desirable in clinical data research networks. Blockchain technology,,,,,,,, has many of these desirable features, in addition to being developed by a community of developers as opposed to being a custom-tailored software solution that is hard to document and maintain. We propose to treat each institution in a network as a node in a blockchain network in a way that allows batch training at a site, but online training across sites, in a sequential fashion that is resilient to node failure and immutably recorded. We thus take advantage of known characteristics of blockchain technology (resilience to single point of failure, ability to create an immutable ledger, transparency of use, and development by the community) to build a multi-center predictive model.

Blockchain is a decentralized peer-to-peer technology that has evolved significantly since the development of Bitcoin blockchain, which is the underlying technology of Bitcoin crypto currency. Blockchain nodes (ie, computers running the blockchain software) compose a distributed peer-to-peer network (Figure 1b) that follow an agreed-upon time stamp and consensus protocol to create a chain of transaction blocks (ie, “blockchain”). Such a blockchain allows nodes to collaboratively generate a consensus ledger without the need of a central intermediary., The chain can store arbitrary data in either space reserved for the “metadata” space of the transactions,, or as part of “smart properties” managed by “smart contracts.” A blockchain network can be either permissionless (ie, a public network in which any node can participate, such as Bitcoin) or permissioned (ie, a private network in which only authorized nodes can participate, such as most networks that focus on healthcare).,

Advantages and disadvantages of blockchain-based modeling

A recent review paper summarized the key benefits and potential challenges of blockchain technology and compared blockchain to traditional distributed database management systems. The use of one of the main blockchain platforms ensures that code is maintained, enhanced and documented by the community at large, and not limited to the site that created the infrastructure’s code. Among many other decentralized systems, including less-connected systems based on gossip algorithms, blockchain brings benefits and addresses many current challenges in distributed networks. On the other hand, some challenges related to how certain types of cryptography can be incorporated, scalability, and threat of a 51% attack still exist.,, Also, there might be potential network security concerns such as leaving particular Internet ports open to allow remote blockchain access, although in permissioned networks this can be protected. A robust method to integrate this distributed ledger technology with the distributed learning algorithms is yet to be investigated in depth. Several methods have been proposed to solve the distributed learning problem with or without blockchain, as shown in Table 1. GLORE, EXPLORER, WebGLORE and SMAC-GLORE are based on a client-server architecture, which possess potential risks such as single point of failure, while ExplorerChain is based on peer-to-peer architecture without such weaknesses. GloreChain and HierarchicalChain are based on blockchain, however, it is not resilient in situations in which a site leaves or joins the network. ModelChain, was blockchain-based and proposed online learning, but it was not actually evaluated in practice. Finally, LearningChain is an online learning method on blockchain, but it focuses on improving privacy (ie, adopting differential privacy scheme) and security (ie, defending against the Byzantine attack) on stochastic gradient descent algorithm, and not on healthcare institutions’ needs for immutable ledgers, and transparency of use.

Table 1.

Comparison of the state-of-the-art distributed learning methods

Method	Author	Reference	Architecture	Learning	Focus	Status
GLORE	Wu et al	10	Client-Server	Batch	Healthcare	Evaluated
EXPLORER	Wang et al	11	Client-Server	Online	Healthcare	Evaluated
WebGLORE	Jiang et al	12	Client-Server	Batch	Healthcare	Evaluated
SMAC-GLORE	Shi et al	13	Client-Server	Batch	Healthcare	Evaluated
GloreChain	Kuo et al	14	Peer-to-Peer	Batch	Healthcare	Evaluated
HierarchicalChain	Kuo et al	15	Peer-to-Peer	Batch	Healthcare	Evaluated
ModelChain	Kuo et al	16 , 17	Peer-to-Peer	Online	Healthcare	Proposed
LearningChain	Chen et al	18	Peer-to-Peer	Online	Privacy/Security	Evaluated
ExplorerChain	Kuo et al	–	Peer-to-Peer	Online	Healthcare	Evaluated

OBJECTIVE

We developed and evaluated a decentralized blockchain-based predictive modeling method, EXpectation Propagation LOgistic REgRession on Permissioned BlockCHAIN (ExplorerChain), to support cross-institution healthcare/genomic research and quality improvements initiatives. This approach should, theoretically, remove certain risks introduced by the client-server architecture nature of distributed networks that rely on a central node. We aimed at empirically showing results of such a design on healthcare and genomic datasets.

MATERIALS AND METHODS

ExplorerChain

In developing the consensus protocols for predictive modeling blockchain networks, we adopt online distributed learning,, to train predictive models on the decentralized network architecture. Our working hypothesis was that online algorithms on a distributed blockchain-based network that does not rely on a coordinating node can provide better security/robustness than models that use a central node, without sacrificing predictive power.,, We chose the online logistic regression algorithm of EXPLORER as our predictive modeling method. In ExplorerChain, we utilize the metadata space of the transactions to disseminate the online machine learning models and integrate the blockchain network with distributed online learning (Figure 2)., Intuitively, permissioned/private blockchain platforms, are feasible for ExplorerChain, because the sites need to be permitted to participate in the predictive model learning process. We selected MultiChain, as the underlying permissioned blockchain platform based on our prior review.

Figure 2.

A simplified example of ExplorerChain. Only the aggregated data (ie, the machine learning model) and the meta information are stored on-chain, while the protected health information (PHI, the observation-level patient data) are stored off-chain. This design ensures that the institutions can share information to improve the predictive model without transmitting PHI. It should be noted that the amount of transactions is set to be zero, indicating that the blockchain serves purely as a nonfinancial distributed ledger.

Datasets

In our experiments, 3 clinical/genomic datasets were utilized to test ExplorerChain. The first dataset (“Edin”) is the Edinburg Myocardial Infarction (MI) data, which contains predictor features and one binary MI outcome (ie, presence of disease). The clinical question pursued by this data set is how accurate a prediction model for MI can be when only signs, symptoms and certain electrocardiographic data are available at presentation. The second dataset (“CA”) has cancer biomarkers (ie, CA-19 and CA-125) and a binary cancer outcome (ie, presence of cancer). The clinical question at hand is whether combining the level values of the 2 biomarkers helps identify pancreatic cancer. The third dataset, “THA,” was used to predict hospital length of stay following unilateral primary total hip arthroplasty (THA) surgery. Following the practice used in other anesthesia publications, the outcome to be predicted was a dichotomized prolonged hospital length-of-stay (LOS) outcome (ie, whether the actual LOS is greater than the expected LOS (3 days) for THA at our institution)., The Institutional Review Board at University of California San Diego (UCSD) approved this study with Project Number 171344X on February 9, 2018, and the informed consent requirement was exempted by our institution’s Human Research Protections Program because the dataset was defined as Health Insurance Portability and Accountability Act (HIPAA) “deidentified.” The statistics/features of the test datasets are summarized in Table 2. Given the age and limited size of the Edin and CA datasets, they are presented for illustration purposes only. The THA dataset was recently used in a quality improvement project at an academic medical center.

Table 2.

Statistics and features of the datasets tested in our experiments. The class distribution (ie, the percentage of the positive/negative classes) is also included. The numerical covariates are labeled with an asterisk symbol (“*”), while the categorical ones are converted into binary through dummy coding. The values for the myocardial infarction and cancer biomarker datasets are adapted from

Dataset	MyocardialInfarction(Edin)	CancerBiomarker(CA)	Length of Hospitalization(THA)
# of Covariates	9	2	34
# of Samples	1,253	141	960
Class Distribution	0.219 / 0.781	0.638 / 0.362	0.278 / 0.722
Outcome	Presence of Disease	Presence of Cancer	Hospital Length of Stay is greater than 3 days
Covariates	Pain in Right Arm	CA-19*	Male Sex	SA - Posterior
	Nausea	CA-125*	Age ≥ 65 years old	SA - Anterolateral
	Hypo Perfusion	–	Preoperative METs < 4	SA - Anterior
	ST Elevation	–	General Anesthesia (versus Neuraxial Anesthesia)	CM - Chronic Kidney Disease
	New Q Waves	–	Non-English Speaker	CM - Chronic Obstructive Pulmonary Disease
	ST Depression	–	OG - Mild	CM - Congestive Heart Failure
	T Wave Inversion	–	OG - Moderate	CM - Coronary Artery Disease
	Sweating	–	OG - Severe	CM - Hypertension
	Pain in Left Arm	–	OG - Avascular Necrosis	CM - Diabetes Mellitus
	–	–	CHD - No osteoarthritis	CM - Obstructive Sleep Apnea
	–	–	CHD - Mild osteoarthritis	CM - Dialysis
	–	–	CHD - Moderate osteoarthritis	CM - Psychiatric history (depression, anxiety, or bipolar disease)
	–	–	CHD - Severe osteoarthritis	CM - Active Smoker
	–	–	CHD - Previous Surgery (ie, hip replacement)	CM - Asthma
	–	–	CHD - Avascular Necrosis	CM - Thrombocytopenia (platelets < 150 000/uL)
	–	–	Obesity (BMI > 30kg/m2)	CM - Anemia
	–	–	Preoperative Opioid Use	CM - Dementia

Abbreviations: BMI, body mass index; CHD, Contralateral Hip Description; CM, comorbidities; METS, metabolic equivalents; OG, osteoarthritis grade (operative side); SA, surgical approach; THA, total hip arthroplasty.

Experiment settings

We aimed at determining whether ExplorerChain could reach equivalent predictive performance as state-of-the-art central server-based learning approaches, while taking advantage of some key benefits of blockchain technology. Specifically, we compared this approach with a central server-based method using an online machine learning method that we adapted for ExplorerChain. We selected the centralized version of EXPLORER as our baseline comparison method, and tested ExplorerChain on the iDASH private HIPAA-compliant computing environment network., Each dataset was evenly and randomly split into 2-, 4-, and 8-site combinations. Then, for each site, a training dataset was generated by randomly selecting 80% of records, while the test dataset was created using the remaining 20%. We adopted the area under the receiver operating characteristic curve (AUC), to evaluate whether the discrimination performance of ExplorerChain was comparable to that of EXPLORER. The averaged AUC among all N sites was used as the result for one trial. We repeated the above dataset splitting process (including both multiple-site and training/test splitting) over 30 trials. Also, we evaluated the Pearson Correlation Coefficient (PCC) between the AUC results of the 2 methods, to verify if their results were linearly correlated (the closer the PCC is to 1, the higher positive linear correlation between 2 methods). Furthermore, we used 2-sample t-test of AUCs between 2 methods with alpha = 0.05, and obtained a mean difference (delta) between the AUC results of the 2 methods that was not statistically significant (ie, P value > .05). We also computed iterations of ExplorerChain and compared the execution times with those of EXPLORER.

RESULTS

Correctness

The predictive performance results are shown in Table 3. The differences of the means (≤ 0.016) and the standard deviations (≤ 0.011) of the AUC values were relatively small. The PCC values (> 0.7) indicated high linear correlation between 2 methods. The P values for the 2-sample t-tests were above .05 for all datasets, demonstrating that the discrimination performance of ExplorerChain was very similar to that of EXPLORER (mean AUC difference ≤ 0.002). ExplorerChain approximates EXPLORER, and thus the latter may outperform the former; however, the difference was small and not statistically significant. Also, for both methods, the mean AUC values of the Edin data set are the highest and the ones of the THA are the lowest among all 3 datasets. The standard deviations of AUC of the CA data set are the highest and the ones of the Edin data set are the lowest. This pattern appears repeatedly for simulations involving N = 2, 4, and 8 sites, suggesting that the performance of ExplorerChain is consistent with that of EXPLORER.

Table 3.

The experiment results for the myocardial infarction (Edin), the cancer biomarker (CA), and the length of hospitalization (THA) datasets. The evaluation metric is the averaged full area under the receiver operating characteristic curve (AUC) among N sites, for 30 trials. The Pearson Correlation Coefficient (PCC) was computed to evaluate the linear correlation between 2 methods. Finally, the alpha in the 2-sample t-test was 0.05, and the p-values larger than 0.05 (shown in bold italic) indicate no statistically significant difference between the AUC results of EXPLORER and ExplorerChain

		EXPLORER		ExplorerChain		Correlation	Two-Sample t-Test
Dataset	N	Mean AUC	Standard Deviation	Mean AUC	Standard Deviation	PCC	Delta	Test Statistics	P-value
Edin	2	0.965	0.013	0.965	0.013	0.999	0.000	−1.559	0.130
	4	0.962	0.010	0.960	0.011	0.867	0.000	1.868	0.072
	8	0.957	0.014	0.954	0.015	0.906	0.002	1.371	0.181
CA	2	0.893	0.054	0.891	0.055	0.977	0.000	1.106	0.278
	4	0.862	0.075	0.853	0.078	0.932	0.000	1.694	0.101
	8	0.892	0.060	0.876	0.071	0.746	0.000	1.811	0.080
THA	2	0.734	0.035	0.733	0.036	0.995	0.000	1.622	0.116
	4	0.738	0.047	0.735	0.047	0.975	0.000	1.529	0.137
	8	0.718	0.040	0.712	0.040	0.909	0.000	1.878	0.070

Abbreviations: AUC, area under the receiver operating characteristic curve; CA, cancer biomarker; PCC, Pearson correlation coefficient; THA, total hip arthroplasty.

Iteration

The iteration results (ie, rate of convergence) are shown in Table 4 and Figure 3. In general, ExplorerChain required a low number of iterations (2 or 3), indicating a fast rate of convergence. This held for N = 2, 4, 8 sites on the Edin and the CA data sets, as well as for N = 2 for the THA. For N = 4 and 8 sites on the THA, higher mean iterations were required (8 or 9). For the THA and N = 8, although the maximum number of iterations (10) was reached in almost every one of the 30 trials (Figure 3), the AUC was still comparable to that of EXPLORER (PCC = 0.909 and P value of the 2-sample t-test = .070 in Table 3), indicating that an increase in the limit of iterations may not be necessary. These speedy convergence results are consistent with EXPLORER (ie, < 10 iterations).

Table 4.

Number of iterations and time results of ExplorerChain among N sites for 30 trials of the 3 datasets. All time measurements are averaged over N sites, and the total time includes both running time and synchronization time. For ExplorerChain, the per-iteration time is computed by dividing the time by the mean number of iterations, and the additional pausing time (240 seconds per trial) between trials for result collection was deducted

		# of Iterations		Time (Seconds)
		ExplorerChain		EXPLORER		ExplorerChain
Dataset	N	Mean	Standard Deviation	Total	Running	Total	Running	Total/ Iteration	Running/ Iteration
Edin	2	2.433	1.455	2.477	2.426	144.519	7.609	59.391	3.127
	4	3.033	1.542	2.451	2.399	165.890	9.939	54.689	3.277
	8	3.633	2.157	2.432	2.383	184.086	12.145	50.666	3.343
CA	2	2.000	0.000	2.000	1.945	129.618	6.125	64.809	3.062
	4	2.700	1.088	2.011	1.949	154.175	8.829	57.102	3.270
	8	3.233	1.547	1.996	1.947	172.654	10.938	53.398	3.383
THA	2	2.533	1.814	2.399	2.348	147.259	8.045	58.128	3.176
	4	8.000	2.533	2.366	2.315	317.266	23.865	39.658	2.983
	8	9.833	0.913	2.364	2.314	365.355	32.713	37.155	3.327

Figure 3.

Number of iterations of ExplorerChain on 3 datasets and 2-, 4- and 8-site settings for each of the 30 trials. The number of iterations increases with the number of sites, with an upper limit to the maximum number of required iterations (10 in our experiment) to reach consensus. However, the relatively large standard deviation, especially when the number of sites is small (eg, N = 2), suggests that the number of required iterations is highly dependent on the dataset.

Time

The time results are also shown in Table 4. Compared to EXPLORER, our proposed ExplorerChain required significantly longer total time because of the synchronization time required to collect errors from all other sites, and also longer total times as the number of sites increased. Such a synchronization time is related to hyper-parameters of ExplorerChain and blockchain/network speed. Regarding running time, ExplorerChain was slower than EXPLORER because of the additional waiting time to find the consensus model. On the other hand, the per-iteration time results showed that, in each iteration, the running time of ExplorerChain was similar to that of EXPLORER. ExplorerChain required about 60 seconds of synchronization time on average. Specifically, the 30 seconds of waiting time period contributed to about half of the 60-second average synchronization time per iteration. The results of the per-iteration total time show variations (30–70 seconds) among the 3 datasets, while the results of the per-iteration running time remained stable (around 3 seconds).

DISCUSSION

Findings

Our study suggests that the proposed blockchain-based ExplorerChain method can learn cross-institution predictive models without transferring observation-level patient data and without relying on a central coordination node. Because of the distributed architecture that is not based on a central node, this approach may avoid concerns encountered in architectures that rely on a central node. The discrimination of the model learned using ExplorerChain was comparable to that of EXPLORER, a central server-based online learning method. ExplorerChain added a layer of protection, provided by the blockchain technology, that helped improve the security/robustness of the learning process. The cost for such improvement was paid mainly in terms of synchronization time, since, without a central server, it took a larger number of iterations for all sites to find a final consensus model (a similar trade-off noted for databases,,,). As the number of sites increased, the number of iterations and time grew linearly. We used relatively small but real-world data for our experiments. The THA data set is contemporary and motivated by medical center needs that address quality of care from the clinical and administrative perspectives. Although surgical outcomes registries such as American Joint Replacement Registry (AJRR) are helpful to develop predictive models for orthopedic surgery outcomes, they suffer from the following problems: registry fees, required data entry and maintenance, and institutional data privacy restrictions which may limit the inter-institution data comparisons. With machine learning and block chain technology, there exists the possibility to process sensitive data, such as that related to complications and cost, without human intervention in a decentralized, more private fashion.

Potential clinical applications

Gabriel et al showed that it is possible to build accurate models for patient LOS after total hip arthroplasty. Their model was based on a single institution and performed well, but if more institutions could participate in a way that preserves individual and institutional privacy, the model would be more robust. This is particularly valuable in the era of value-based care where the focus is on incentivizing higher quality care at lower costs. Given the sensitive nature of administrative processes and the fact that many institutions are competing with each other, institutions may want to keep their data hidden from other institutions and insurance companies, while still contributing to the development of a better predictive model. The method presented here enables this participation, while preserving institutional privacy, using a novel strategy based on blockchain technology. Future applications of this technology may be seen in development of a risk-stratification model for reimbursement. At present, in the bundled care model, there is a single reimbursement fee for all lower extremity arthroplasties, for the entire episode of care, regardless of patient factors that can affect complication rates. This has led to the practice of “cherry-picking” younger, healthy patients who are less likely to have complications. Such a decentralized model with protection of institutional privacy through the blockchain technology presented could encourage participation from a diverse group of practices to build an accurate model. The results of the THA data set show that the differences in AUC from the blockchain-based model and a benchmark centralized model do not have clinical significance: the models’ performances are essentially the same (AUC difference of 1%–2%).

Limitations

Two major concerns for the heavy use of blockchain includes scalability of the approach and acceptability by users. The scalability issue consists of several aspects, such as (S1) the number of users, (S2) the number of records in the datasets, and (S3) the number of the covariates in the datasets. For (S1), we expect a small number of users in most of the research networks, and therefore the impact to the blockchain network should be negligible. For (S2), the increment of the number of the records will not affect the size of transactions. This is because the size of the transaction is only related to the number of covariates. Additionally, the blockchain transaction speed will still be fast enough. Thus, this issue may not be a serious barrier for a well-designed blockchain network. For (S3), the transaction metadata size for the largest dataset (THA with 34 covariates) was small when compared to the default upper limit transaction metadata size for MultiChain.

Regarding the acceptability by users, one feasible solution is to start from a small number of pilot institutions within a system, and build the blockchain network in parallel with current operating platforms. Users should not be concerned about the technology underlying the network as long as it is reliable. Misconceptions about blockchain technology need to be overcome at the design, development, and dissemination levels. Given the hype surrounding blockchain and some unwarranted claims that it is an ideal solution for various problems, it is not surprising that the healthcare sector is exercising caution before adopting the technology. At the same time, it would be unwise not to consider the technology for situations in which its built-in advantages overcome its limitations.

We presented here a specific situation in which blockchain technology features such as disintermediation, immutable logging capability, resilience to node failure, and transparency of transactions match the needs of machine learning models being built across healthcare and research institutions. We also presented the practical drawbacks in terms of additional time for synchronization and commented on potential vulnerabilities. We anticipate that this information, when coupled with direct comparisons of platforms and use cases that are now emerging in the healthcare industry will help decision-makers choose among various technologies available for distributed machine learning.

CONCLUSION

ExplorerChain integrates 2 important technologies, artificial intelligence (online machine learning) and a distributed ledger (blockchain), without a central authority, to learn a predictive model across institutions in a distributed architecture (ie, without the need for sharing the patient-level data nor the need for a central coordinating node). We discussed the advantages/disadvantages of adopting blockchain, and showed empirically on 3 datasets that ExplorerChain produces results that are equivalent to the level of the state-of-the-art central server-based method. Additionally, through a traceable and immutable “path,” and as a peer-to-peer blockchain network without a central intermediary, ExplorerChain can improve security and robustness. These are useful for healthcare and genomics applications, as bottlenecks associated with local servers and network downtimes are eliminated. While ExplorerChain requires more iterations and time to reach a consensus without a central intermediary, the extra time may prove irrelevant for the development of machine learning models. ExplorerChain’s promising results warrant further evaluation and refinement, while suggesting that it is feasible to adopt blockchain-based artificial intelligence methods to build decentralized learning models across multiple sites. There are still mixed opinions and profound misunderstandings with corresponding questions about blockchain technology and its potential applications in healthcare/genomics. Examples, to name a few, include whether blockchain technology will prove to be practical and useful for healthcare; whether blockchain will only become practical as networks get faster and computer speeds increase; and whether the turnkey solution of blockchain will become available to all biomedical informatics researchers. ExplorerChain, as an early stage feasibility study, may serve as a step towards answering these open yet critical questions about the future of blockchain technology for applications in healthcare and genomics.

FUNDING

The authors were funded by the US National Institutes of Health (NIH) (OT3OD025462, R00HG009680, R01HL136835, R01GM118609, and U01EB023685) and a UCSD Academic Senate Research Grant (RG084150). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

AUTHOR CONTRIBUTIONS

T-TK contributed in conceptualization, data curation, formal analysis, funding acquisition, investigation, methodology, project administration, resources, software, validation, visualization, and writing (original draft). RAG contributed in data curation and writing (review and editing). KRC contributed in critical discussion and writing (review and editing). LO-M contributed in conceptualization, funding acquisition, project administration, resources, supervision, and writing (review and editing).