Federated Health Recommender System

Sarah Pinon

1 a

, Simon Jacquet

, Colin Vanden Bulcke

2 b

, Edouard Chatzopoulos

3 c

Xavier Lessage

4 d

and Rapha

el Michel

4 e

NaDI, Namur Digital Institute, University of Namur, Belgium

IONS, Institute of NeuroScience, UCLouvain, Belgium

ICTEAM, Institute of Information and Communication Technologies, Electronics and Applied Mathematics, UCLouvain,

Belgium

CETIC, DSIDE, Charleroi, Belgium

Keywords:

Health, Precision Medicine, Patient-Oriented Decision Support System, Recommender System, Federated

Learning.

Abstract:

Precision Medicine is a new and growing approach to health care. This initiative includes different patient-

oriented Decision Support Systems (DSS), such as Health Recommender Systems (HRS). These patient-

oriented DSS aim to increase the accuracy and personalization of health care. However, the development

of these systems faces a major obstacle related to the conﬁdential and private nature of medical data. These

systems require, indeed, a large volume of data to run effectively. But medical data are dispersed among sev-

eral institutions and cannot be centralized for strict conﬁdentiality reasons. To address this issue, this position

paper proposes a system’s architecture in which Federated Learning is exploited to build a HRS. Federated

Learning allows exploiting the data maintained by different institutions to build the system without requiring

their sharing. To demonstrate the feasibility of our proposition, we build a Federated Drug Recommender Sys-

tem. The goal of the system is to assist doctors in their administration of drugs by using historical disease-drug

interactions and drug data. As a position paper, the objective of this use case is limited to a proof of concept

realized on non-sensitive open-source data. Our ambition is then to use the architecture proposed in this paper

to develop a Federated HRS on real medical data.

1 INTRODUCTION

Precision Medicine is currently increasing in popu-

larity with its ambitious promises of achieving an

incomparable quality of healthcare (Koenig et al.,

2017; Kosorok and Laber, 2019). To achieve its

promises, this healthcare approach aims to offer per-

sonalized treatment strategies to patients by consid-

ering a large amount of data. This data includes, for

example, the patient’s clinical condition, lifestyle, ge-

netics, biomarker information. By including all this

information, this medicine is able to offer more pre-

cise therapeutic strategies (Koenig et al., 2017).

The aspiration of Precision Medicine is perfectly

https://orcid.org/0000-0001-5392-1020

https://orcid.org/0000-0002-2210-1625

https://orcid.org/0000-0001-5848-7999

https://orcid.org/0000-0003-0861-0068

https://orcid.org/0000-0001-5505-9171

aligned with the mission of Health Recommendation

Systems (HRS). HRS are, indeed, intended to help

healthcare professionals to deal with medical infor-

mation overload (e.g. medical results, drug informa-

tion, pathology information) in their patient decision-

making (Tran et al., 2021). These patient-oriented de-

cision support systems consider these large volumes

of data to recommend, for example, the most appro-

priate medication, treatment or lifestyle (De Croon

et al., 2021).

The performance of HRS, like any RS, depends

crucially and positively on the volume of data ex-

ploited (Tan et al., 2020). In the era of Big Data,

the volume of medical data available is increasingly

important, coming from clinical institutions, individ-

ual patients, pharmaceutical industries,...(Xu et al.,

2021). This emergence represents a great opportunity

for HRS.

However, the accessibility of this data is challeng-

Pinon, S., Jacquet, S., Bulcke, C., Chatzopoulos, E., Lessage, X. and Michel, R.

Federated Health Recommender System.

DOI: 10.5220/0011722700003414

In Proceedings of the 16th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2023) - Volume 5: HEALTHINF, pages 439-444

ISBN: 978-989-758-631-6; ISSN: 2184-4305

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

439

ing. Indeed, medical data is, in the hospital context

for example, fragmented across several hospitals. For

conﬁdentiality and privacy reasons, these data cannot

be easily accessed and centralized on a server outside

the institutions to build an effective RS. This limits the

creation of a robust, generalizable and high-quality

system. Conversely, by developing small, local sys-

tems within each hospital, several biases will impact

the quality of the results, such as considering only a

subset of conditions or treatments (Xu et al., 2021;

Rieke et al., 2020).

To exploit medical data from several hospitals to

build a generalizable model, we propose to use Feder-

ated learning (FL). The privacy and security concerns

very important in the medical sector are, by this way,

addressed. Indeed, the FL allows institutions to train

the RS on their own infrastructure without exchang-

ing raw (e.i. sensitive) data. The only information

shared between the different parties for the construc-

tion of the global model corresponds to the model pa-

rameters (Kaissis et al., 2020; Li et al., 2020; Rieke

et al., 2020; Xu et al., 2021).

In this paper, we propose the architecture of a

HRS exploiting the FL, named Federated Health Rec-

ommender System (F-HRS). The particularity of the

proposed architecture lies in the integration of FL to

build the recommendation model. In this way, our

system overcomes the medical data conﬁdentiality is-

sues stated above. This position paper demonstrates

then the feasibility of our proposition by using non-

sensitive and open-source data to develop a Feder-

ated Drug Recommender System. In future work, we

will exploit real data from hospitals to demonstrate

the usefulness and effectiveness of our system in real

situations.

This paper is organized as follows: Section 2 re-

views the state of the literature about HRS and FRS.

Section 3 develops the architecture of the proposed

system, the selected model used in our system and

the federated learning method exploited. Section 4

presents our proof of feasibility. Section 4 discusses

our future works. Section 5 concludes the paper.

2 LITERATURE REVIEW

Recommendation systems in the health ﬁeld are in-

tended for two types of users: (1) patients or healthy

users and (2) health professionals. In the case of pa-

tients, these RS allow to involve patients in the co-

creation of their own health. HRS for healthcare

professionals aim at assisting these professionals to

provide precise patient-oriented decision (De Croon

et al., 2021; Tran et al., 2021). In this project, we

focus on HRS for healthcare professionals. The ma-

jority of these HRS concern, in the literature, the rec-

ommendation of drugs for different pathologies (e.g.

diabetes, migraine, infectious diseases).

A major challenge faced by existing HRS is the

accessibility of medical data that is sensitive and con-

ﬁdential(Tran et al., 2021). To deal with this prob-

lem, the commonly used approach is data encryption.

This method allows to keep the conﬁdentiality of the

data while exploiting them in the system of recom-

mendations (Tran et al., 2021). Indeed, data encryp-

tion, such as homomorphic encryption, involves en-

cryption of the data by the users before sending it to

another party. It is then the encrypted data, and not the

raw and complete data, which is used to compute the

interest function (Hoens et al., 2013). However, this

approach of data encryption raises problems due to its

high computational and communication costs, which

considerably reduces the performance of the recom-

mendation system(Tran et al., 2021).

A technique increasingly used by RS to exploit

larger and more widely distributed data while pre-

serving its conﬁdentiality and security is Federated

Learning (Alamgir et al., 2022). This type of learning

has the particularity, as opposed to data encryption, to

train the algorithm where the data are located (e.g. in

each hospital). After this local training, the model pa-

rameters are, after an encryption, returned to the cen-

tral model. By this way, the data conﬁdentiality and

privacy are protected which is especially important in

the medical ﬁeld (Kaissis et al., 2020).

To the best of our knowledge, despite the interest

of Federated Learning for Recommendation Systems

and medical data accessibility issues, no HRS has yet

integrated the Federated Approach to exploit all med-

ical data while preserving their privacy.

3 F-HRS ARCHITECTURE

F-HRS architecture proposed has three main compo-

nents, namely: the Input Data, the Recommender Al-

gorithm and the Federated Method. This section de-

velops each of these components. Figure 1 illustrates

the combination of these components in an architec-

tural way. This illustration is further explained in the

following subsections.

3.1 Input Data

F-HRS exploits different types of databases, as shown

in Figure 1. On the one hand, our system uses hospital

data related to patients such as their medical history,

their proﬁle, the results of their medical examinations.

HEALTHINF 2023 - 16th International Conference on Health Informatics

440

Figure 1: F-HRS Architecture.

These data are in the database of the individual hos-

pitals and cannot be shared due to their private status.

On the other hand, our system can be fed by an ex-

ternal database containing information on the recom-

mendation items. This database, unlike the others, is

open-source and can therefore be used by the different

hospitals.

3.2 Recommender Algorithm

The Recommender Algorithm used by our F-HRS is

the Neural Collaborative Filtering (NCF) Algorithm

(He et al., 2017a). This algorithm works on the princi-

ple of collaborative ﬁltering. This ﬁltering is famous

in RS and consists in ﬁnding users similar to the user

of interest and recommending to this user items liked

or used by these similar users (Schafer et al., 2007).

We have chosen this algorithm because the integration

of a neural network within the Collaborative Filter-

ing Algorithm allows to learn the interaction function

between users and items automatically from the data

rather than manually as is the case in the basic algo-

rithm. Moreover, the NCF Algorithm allows to easily

integrate content features representing users and/or

items and therefore offer more accurate and relevant

recommendations(He et al., 2017b).

Concretely, the NCF algorithm used is illustrated

in Figure 2. To process, the algorithm requires, as in-

puts, positive and negative examples of interactions

between users and items. In our case, we applied the

general approach which consists of randomly gener-

ate four negative examples for each user or patient by

using items for which the user has not yet interacted.

Positive and negative interactions between users and

items are stored into binary matrix. The users or items

matrix can, eventually, be enriched with additional

data, by concatenating the binary vectors, as shown

in Figure 2 for the items case. Each complete user’s

vector and item’s vector from the matrix are then em-

Figure 2: Neural Collaborative Filtering Architecture.

bedded in d dimensions, eight by default. These vec-

tors are concatenated and introduced in the neural net-

work. Inputs are passed to the hidden layers of the

multilayer neural network to learn these non-linear re-

lationships between users and items. Most NCF im-

plementations use 4-5 dense layers with less than 100

neurons per layer. The last layer of the network is a

sigmoid function which maps its inputs to a score be-

tween zero and one to represent the probability that

the user interacts with the item. The algorithm is

trained on ﬁve epochs, which a standard. For each

epoch, a new set of negative examples is composed

for each user.

3.3 Federated Learning

To integrate the Federated approach to the HRS, we

follow the ﬁve general steps, namely (Nishio and Yo-

netani, 2019):

1. Initialization: A global model is initialized ran-

domly or by pre-training with public data. For our

experiments, we randomly initialized the weights,

which is a common practice in federated learning.

2. Client Selection: The different parties (e.i. hospi-

tals) which will contribute to the federated train-

ing are selected. By default, our framework se-

lects clients (e.i. hospital servers) in their start or-

der without taking into account a possible client

to be discarded. If the need arises, it is quite pos-

sible to add additional functionality to this frame-

work to select clients based on various criteria (e.

i. such as performance or dataset quality).

3. Distribution: The weights from the global model

are shared with the parties. This step is criti-

cal and must incorporate the best approach to se-

Federated Health Recommender System

441

cure weight distribution (P. Treleaven and Pitha-

dia, 2022) based on, for example, homomorphic

encryption (HE) classical encryption (TLS), dif-

ferential privacy (DP) or Secure Multi-party Com-

putation (SMPC).

4. Update and Upload: Each party (e.i. hospital)

trains the model (e.i. Collaborative Filtering Al-

gorithm) on their local data, and subsequently up-

loads the updates model weights to the server.

5. Aggregation: The weights collected are aggre-

gated and the global model is replaced with the

model resulting from the aggregation. The aggre-

gation can be carried out by means of a weighted

sum for example, but there are many methods de-

pending on the type of data to be trained. In our

system, we use the Federated Averaging method.

This aggregation method is the most prevalent

one, because of both its simplicity and its perfor-

mance. Federated averaging consists in comput-

ing the weight of the global model as the average

of the weights of the clients models.

All steps except the Initialization are iterated until

the global model achieves a desired level of perfor-

mance.

Various frameworks are available for the deploy-

ment of this Federated approach. In the context of

this project, we opted for Flower (Beutel et al., 2021).

Flower is an open source framework dedicated to

facilitate the deployment of machine learning algo-

rithms in federated environments. The following ﬁg-

ure 3 shows its architecture based on an aggregation

server and several clients running with different envi-

ronments. The main particularity of Flower compared

to other Federated Learning frameworks is its support

for a wide variety of machine learning back-end li-

braries such as Pytorch, Tensorﬂow or Jax thanks to

its ML framework-agnostic implementation. Flower

is also ﬂexible, as it allows both single host simulation

and multi-host deployment of the federated procedure

(Liu et al., 2022).

Figure 3: Flower core framework architecture.

4 PROOF OF FEASIBILITY

To demonstrate the feasibility of the proposed archi-

tecture, we develop a simple Federated Drug Recom-

mender System based on non-sensitive open-source

data. To goal of this Recommender System is to as-

sist the doctors in the drugs administration.

To achieve this, our system exploits two types of

databases. The ﬁrst concerns interactions between

drugs and diseases (Dat, 2018). The second is re-

lated to drugs’ features (e.g. pharmaceutical class,

contraindications) (Dru, 2022).

To build our system, we divide the data set with

the drugs pathology’s interactions in a training and

test sets. The test set contains for each pathology the

last drug related in the database. First, we used the

training set to build our system following the archi-

tecture developed in the previous sections. In a sec-

ond step, we evaluated the performance of our system

by checking if the last drug associated to the pathol-

ogy for which we search for a drug is in the ten drugs

recommendations made by our system.

The performance of our Federated Drug Recom-

mender System is not presented in this paper. These

ﬁgures are, indeed, at the moment irrelevant since our

system is still being improved and requires better and

real data to illustrate the relevance of the approach.

Nevertheless, thanks to our use case, the technical fea-

sibility of the architecture could be validated.

5 POTENTIAL ISSUES AND

CONCERNS

During the development of our system architecture,

we considered potential problems related to our ap-

plication domain (i.e. the medical sector) and to the

applied approach.

First, the medical ﬁeld is facing a lack of stan-

dardization of data formats and protocols which can

negatively affect the integration of data from multiple

sources. To face this problem, we have two solutions.

In a ﬁrst step, it is possible to focus on data collected

by widely used software and therefore store it accord-

ing to a certain standardized structure. Secondly, this

problem of standardization is known within the liter-

ature and the medical section. It generates, indeed,

the development of standardized data formats such as

OMOP (Klann et al., 2019; Makadia and Ryan, 2014).

These new data formats would solve this heterogene-

ity problem.

Second, implementing the federated approach is

more resource intensive than a traditional centralized

system. Nevertheless, the potential beneﬁts of a fed-

HEALTHINF 2023 - 16th International Conference on Health Informatics

442

erated approach in the medical ﬁeld can far outweigh

the challenges of implementing it.

Third, the risk of data bias and incorrect recom-

mendations to the incomplete data in any machine

learning system is a reality. With the federated learn-

ing, we try to reduce that risk by training a module

with more data coming from multiple sources. By this

way, more comprehensive data can be available but

also more diverse datasets. Non-IID data across data

providers can, indeed, be a major issue in FL. How-

ever, a lot of research (Li et al., 2019; Zhao et al.,

2018) is carried out and techniques are being devel-

oped to improve the quality of federated systems in

such settings.

6 FUTURE WORK

Having developed the architecture of our F-HRS and

demonstrated its technical feasibility, our next goal

is to operationalize this system on real data. To

achieve this, we are in discussion with different hos-

pitals to get our local models developed on their real

database. This step will obviously require some para-

metric adaptations of the system.

Moreover, after several discussions with health

care professionals, it is clear that hospital data is

rarely in a structured format. Generally, medical re-

ports are available as PDF documents written in nat-

ural language. To be able to integrate this impor-

tant information into our recommender system, an im-

portant step related to the data preparation must be

considered. To do so, different techniques of Natu-

ral Language Processing will have to be mobilized in

order to transform these unstructured data into struc-

tured data.

7 CONCLUSIONS

In this position paper, we propose to integrate the Fed-

erated Learning to build a health recommender model.

By this way, we want to overcome the lack of data

that medical institutions face in developing patient-

oriented decision support systems. Indeed, federated

approach allows to train a general and robust recom-

mender model on data from several institutions with-

out the need to share the raw data. The technical feasi-

bility of our solution has been demonstrated via open-

source data in the context of a drug-recommender sys-

tem.

ACKNOWLEDGEMENTS

We would like to thank the Walloon region for the

funding of the ARIAC project, of which our F-

HRS project was born and is part. Thanks also to

the TRAIL organization which organized the TRAIL

Summer Workshop in September 2022 during which

our F-HRS project was developed.

REFERENCES

(2018). Recommendation medicines by using a re-

view. https://www.kaggle.com/code/chocozzz/

recommendation-medicines-by-using-a-review/data.

Accessed on 2022-09-08.

(2022). Drugcentral. https://drugcentral.org/. Accessed on

2022-09-08.

Alamgir, Z., Khan, F. K., and Karim, S. (2022). Federated

recommenders: methods, challenges and future. Clus-

ter Computing, pages 1–22.

Beutel, Qiu, T. M., Parcollet, de Gusmao, and Lane (2021).

Flower: A friendly federated learning framework. On-

device Intelligence Workshop at the Fourth Confer-

ence on Machine Learning and Systems (MLSys).

De Croon, R., Van Houdt, L., Htun, N. N.,

Stiglic, G.,

Abeele, V. V., Verbert, K., et al. (2021). Health recom-

mender systems: systematic review. Journal of Medi-

cal Internet Research, 23(6):e18035.

He, X., Liao, L., Zhang, H., Nie, L., Hu, X., and

Chua, T.-S. (2017a). Neural collaborative ﬁltering.

https://arxiv.org/.

He, X., Liao, L., Zhang, H., Nie, L., Hu, X., and Chua, T.-S.

(2017b). Neural collaborative ﬁltering. In Proceed-

ings of the 26th international conference on world

wide web, pages 173–182.

Hoens, T. R., Blanton, M., Steele, A., and Chawla, N. V.

(2013). Reliable medical recommendation systems

with patient privacy. ACM Transactions on Intelligent

Systems and Technology (TIST), 4(4):1–31.

Kaissis, G. A., Makowski, M. R., R

uckert, D., and Braren,

R. F. (2020). Secure, privacy-preserving and feder-

ated machine learning in medical imaging. Nature

Machine Intelligence, 2(6):305–311.

Klann, J. G., Joss, M. A., Embree, K., and Murphy,

S. N. (2019). Data model harmonization for the

all of us research program: Transforming i2b2 data

into the omop common data model. PloS one,

14(2):e0212463.

Koenig, I. R., Fuchs, O., Hansen, G., von Mutius, E., and

Kopp, M. V. (2017). What is precision medicine? Eu-

ropean respiratory journal, 50(4).

Kosorok, M. R. and Laber, E. B. (2019). Precision

medicine. Annual review of statistics and its appli-

cation, 6:263.

Li, T., Sahu, A. K., Talwalkar, A., and Smith, V. (2020).

Federated learning: Challenges, methods, and fu-

ture directions. IEEE Signal Processing Magazine,

37(3):50–60.

Federated Health Recommender System

443

Li, X., Huang, K., Yang, W., Wang, S., and Zhang, Z.

(2019). On the convergence of fedavg on non-iid data.

arXiv preprint arXiv:1907.02189.

Liu, X., Shi, T., Xie, C., Li, Q., Hu, K., Kim, H., Xu,

X., Li, B., and Song, D. X. (2022). Unifed: A

benchmark for federated learning frameworks. ArXiv,

abs/2207.10308.

Makadia, R. and Ryan, P. B. (2014). Transforming the pre-

mier perspective

 hospital database into the observa-

tional medical outcomes partnership (omop) common

data model. Egems, 2(1).

Nishio, T. and Yonetani, R. (2019). Client selection for fed-

erated learning with heterogeneous resources in mo-

bile edge. IEEE International Conference on Com-

munications, 2019-May.

P. Treleaven, M. S. and Pithadia, H. (2022). Unifed: Fed-

erated learning: The pioneering distributed machine

learning and privacy-preserving data technology. ieee.

Rieke, N., Hancox, J., Li, W., Milletari, F., Roth, H. R.,

Albarqouni, S., Bakas, S., Galtier, M. N., Landman,

B. A., Maier-Hein, K., et al. (2020). The future of

digital health with federated learning. NPJ digital

medicine, 3(1):1–7.

Schafer, J. B., Frankowski, D., Herlocker, J., and Sen, S.

(2007). Collaborative ﬁltering recommender systems.

In The adaptive web, pages 291–324. Springer.

Tan, B., Liu, B., Zheng, V., and Yang, Q. (2020). A fed-

erated recommender system for online services. In

Fourteenth ACM Conference on Recommender Sys-

tems, pages 579–581.

Tran, T. N. T., Felfernig, A., Trattner, C., and Holzinger,

A. (2021). Recommender systems in the healthcare

domain: state-of-the-art and research issues. Journal

of Intelligent Information Systems, 57(1):171–201.

Xu, J., Glicksberg, B. S., Su, C., Walker, P., Bian, J., and

Wang, F. (2021). Federated learning for healthcare

informatics. Journal of Healthcare Informatics Re-

search, 5(1):1–19.

Zhao, Y., Li, M., Lai, L., Suda, N., Civin, D., and Chandra,

V. (2018). Federated learning with non-iid data. arXiv

preprint arXiv:1806.00582.

HEALTHINF 2023 - 16th International Conference on Health Informatics

444