Towards Semantic Integration for Explainable Artiﬁcial Intelligence in

the Biomedical Domain

Catia Pesquita

LASIGE, Faculdade de Ci

encias da Universidade de Lisboa, Portugal

Keywords:

Ontology Alignment, Knowledge Graph Alignment, Ontology Matching, Knowledge Graphs, Ontologies,

Semantic Web, Explainable Artiﬁcial Intelligence, Machine Learning, Healthcare, Clinical Research, Health

Informatics.

Abstract:

Explainable artiﬁcial intelligence typically focuses on data-based explanations, lacking the semantic context

needed to produce human-centric explanations. This is especially relevant in healthcare and life sciences

where the heterogeneity in both data sources and user expertise, and the underlying complexity of the domain

and applications poses serious challenges. The Semantic Web represents an unparalleled opportunity in this

area: it provides large amounts of freely available data in the form of Knowledge Graphs, which link data to

ontologies, and can thus act as background knowledge for building explanations closer to human conceptu-

alizations. In particular, knowledge graphs support the computation of semantic similarity between objects,

providing an understanding of why certain objects are considered similar or different. This is a basic aspect of

explainability and is at the core of many machine learning applications. However, when data covers multiple

domains, it may be necessary to integrate different ontologies to cover the full semantic landscape of the un-

derlying data. We propose a methodology for semantic explanations in the biomedical domain that is based on

the semantic annotation and integration of heterogenous data into a common semantic landscape that supports

semantic similarity assessments. This methodology builds upon state of the art semantic web technologies and

produces post-hoc explanations that are independent of the machine learning method employed.

1 INTRODUCTION

Recent successes in black-box models, such as deep

neural networks, are revolutionizing artiﬁcial intelli-

gence (AI) applications, but despite their impressive

successes, their effectiveness and integration in real-

world applications are still limited by their inability

to explain their decisions in a human-understandable

way. These limitations stem from ethical concerns,

but also accountability, safety and liability (Guidotti

et al., 2018). In critical use cases, for instance, in

clinical decision making, there is reluctance in the

deployment of such models because the cost of mis-

classiﬁcation is potentially very high, endangering pa-

tients’ health and lives (Miotto et al., 2018). More-

over, models that predict natural phenomena may bet-

ter contribute to scientiﬁc advancements when re-

searchers are able to understand them. This is ev-

ident in the application of black-box models in ge-

nomics, drug-discovery and pathology, among others

(Min et al., 2017).

This is one of the historical challenges of AI:

the ability of a model to afford explanations of how

and why it arrived at a particular outcome. How-

ever, the deﬁnition of explainable AI (XAI) is still

not agreed upon by the community and is often used

interchangeably with interpretable or comprehensible

AI (Guidotti et al., 2018). While interpretation re-

quires transparency in the underlying mechanisms of

a system, a comprehensible one can be opaque while

emitting symbols a user can reason over. Both en-

able explanations of decisions, but they do not yield

explanations themselves, leaving explanation genera-

tion to human analysts who may deduce different ex-

planations depending on their background knowledge

about the data and its domain. Relatively few works

address these issues and they are typically based on

researchers’ intuitions of what constitutes a ‘good’

explanation, without taking into account how humans

explain decisions and behaviour to each other, ar-

guably a strong starting point to improve human in-

teractions with explanatory AI (Miller, 2019).

Pesquita, C.

Towards Semantic Integration for Explainable Artiﬁcial Intelligence in the Biomedical Domain.

DOI: 10.5220/0010389707470753

In Proceedings of the 14th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2021) - Volume 5: HEALTHINF, pages 747-753

ISBN: 978-989-758-490-9

747

2 SEMANTIC TECHNOLOGIES

FOR XAI IN BIOMEDICINE

One can argue that what is needed for humans to

understand each other can be transferred to what is

needed to make AI outcomes understandable for hu-

mans. We need to fulﬁl the properties of human un-

derstanding, namely, that human explanations imply

social interaction which is grounded in a shared con-

text, and that users select explanations from a large

space of possible explanations based on their under-

standing of the context (Miller, 2019). The vast ma-

jority of works in XAI lack the ability to integrate

background knowledge into the process to create a

shared context, rendering them inadequate to build

explanations for common users without AI exper-

tise. Providing this contextualization is an even bigger

challenge in areas such as systems medicine where

data comes from different domains and with differ-

ent levels of granularity or personalized medicine,

which often relies on highly diverse data, ranging

from molecules, organelles, cells, tissues, organs, all

the way up to individuals, environmental factors, pop-

ulations, and ecosystems (Holzinger et al., 2019).

2.1 Explainable Knowledge-enabled

Systems

In the scientiﬁc and healthcare domains, where ma-

chine learning (ML) methods methods and particu-

larly black-box methods such as deep learning have

been gaining traction, it has been proposed that suc-

cessful explainable-AI systems need to be able to link

ML models to representations of domain knowledge

(Holzinger et al., 2017; Wollschlaeger et al., 2020).

Recently (Chari et al., 2020) deﬁned Explainable

Knowledge-enabled systems as ”AI systems that in-

clude a representation of the domain knowledge in the

ﬁeld of application, have mechanisms to incorporate

the users’ context, are interpretable, and host expla-

nation facilities that generate user-comprehensible,

context-aware, and provenance enabled explanations

of the mechanistic functioning of the AI system and

the knowledge used.”

However, most existing approaches that tackle

XAI in knowledge-enabled systems focus on inter-

pretability and not on building explanations. A re-

cent survey (Chari et al., 2020) in this area presents

neuro-symbolic approaches (Hitzler et al., 2020) as

a potential solution, and although some preliminary

works tackle explanation of deep learning for image

recognition (e.g., (Zhou et al., 2018; Sarker et al.,

2017)) or transfer learning (Chen et al., 2018) most

simply allow for the inclusion of knowledge in the

machine learning approaches but do not yield expla-

nations themselves. Few XAI approaches exist in

clinical application areas and most are still focused in

statistical explanations (e.g.,(Lundberg et al., 2018)).

(Phan et al., 2017) employs neuro-symbolic learning

to predict human behaviour, producing simple expla-

nations based on the identiﬁcation of key features.

However, explanations are ﬂat, limited in expressiv-

ity.

2.2 Ontologies and Knowledge Graphs

We have established that for AI outcomes to be truly

useful, they need to support interpretation, and this

requires semantic context. By semantic context, we

mean the situation in which a term or entity appears.

In relational databases and spreadsheets, semantic

context is sometimes lacking because important in-

formation about what the various data ﬁelds mean and

how they relate to one another is often implicit in the

names of database tables and column headers. What

is needed is a way to express the semantic connec-

tions between data items in a way that is expressive

enough to capture nuanced relationships while at the

same time formalized and restrictive enough to allow

software as well as humans to make inferences based

on the links.

Semantic web (SW) technologies and artefacts

(such as ontologies and knowledge graphs) are a

potential solution to the problem of human-centric,

knowledge-enabled explanation since they provide

this semantic context (Lecue, 2019; d’Amato, 2020).

Ontologies establish a conceptual model that repre-

sents the concepts of a domain and their relationships

with one another (Munn and Smith, 2013), in a way

that can be understood both by humans and machines.

In general, an ontology encoding a domain of knowl-

edge reﬂects the consensus of specialists or communi-

ties dealing with that domain. Domains are expressed

mainly through classes referencing real-world entities

(e.g. “Head”, “Fever”), and the relationships found

between them (e.g. “Face is part of Head”, “Fever is

a Vital Sign Finding”). The classes are frequently ac-

companied by lexical properties, such as preferred la-

bels, synonyms, textual descriptions, etc., which pro-

vide human-readable deﬁnitions. In contrast, relation-

ships between classes are represented not in text but

through formal axioms, or statements, which make

their intended meaning more amenable to automatic

manipulation and reasoning, thus allowing the use of

modern computational power to operate on the mean-

ing, rather than the structure, of the real-world enti-

ties. A knowledge graph (KG) is frequently taken

to mean a collection of data items with relations es-

HEALTHINF 2021 - 14th International Conference on Health Informatics

748

tablished between them and described according to

an ontology. The ontological layer of a KG thus de-

scribes and imposes some order on the data of a do-

main of interest.

In the biomedical and healthcare domains, the Se-

mantic Web represents an unparalleled opportunity

since it provides large amounts of freely available

data and a set of technologies dedicated to data shar-

ing, integration, management, and reasoning (Fer-

reira et al., 2020). The availability of over 1,000

open biomedical ontologies in BioPortal and more

than 2 billion data items publicly available as a KG

(i.e., Linked Open Data) represents a unique oppor-

tunity to integrate clinical and biomedical data. In

the biomedical domain, these datasets range from se-

mantic annotations for the functions of gene prod-

ucts, abnormal phenotypes related to diseases and

genes, or drug adverse events. In electronic health

records, patient-level data is commonly described us-

ing standardized coding schemes and ontologies, such

as UMLS, SNOMED-CT and ICD-9/10. This anno-

tation is mostly conﬁned to ﬁnal diagnosis and proce-

dures, which are frequently used for billing purposes,

whereas ﬁner-grained clinical information is typically

found in free text format, making its linking to on-

tologies a greater challenge. However, once data and

AI outcomes are integrated with ontologies and KGs,

they can serve as background knowledge to XAI ap-

plications and in this way afford the semantic context

that is essential for explanations closer to human con-

ceptualizations and thus more useful in real-world ap-

plications.

3 A METHODOLOGY FOR

SEMANTIC EXPLANATIONS

XAI techniques can be categorized according to how

they support human reasoning into inductive reason-

ing, querying and similarity modelling (Wang et al.,

2019). Semantic explanations can actually support

the three kinds of explanations, since KGs naturally

support reasoning, querying and semantic similarity

computation. To produce semantic explanations we

need to address three challenges: (1) how to link in-

put data and AI outcomes to their meaning; (2) how to

link this meaningful data with what is already known

and (3) how to use this contextual information to build

effective explanations.

Figure 1 depicts the proposed methodology for

semantic explanations in biomedical AI applications.

It builds upon our well-established experience in se-

mantic technologies within the biomedical domain

and addresses the main challenges faced when build-

ing semantic explanations.

The core of the methodology is an integrated KG

that supports the XAI approaches. The integrated KG

is built by connecting heterogeneous data (scientiﬁc,

clinical, etc.) to existing domain ontologies to provide

a rich semantic layer to the data. This is achieved by

performing (1) Ontology selection to determine the

optimal set of ontologies to adequately describe the

data, (2) Semantic annotation, to link the data to

the ontologies, and (3) Semantic integration to es-

tablish links between the ontologies. The ﬁnal step

is to build (4) Semantic explanation approaches that

explore background knowledge afforded by the KG.

This methodology focuses on model agnostic expla-

nations which are able to work regardless of the ma-

chine learning model employed and can be integrated

into already existing approaches.

Let us consider a simple example of semantic ex-

planations in healthcare, where we have trained a ma-

chine learning model that identiﬁes patients with res-

piratory tract infections based on EHR data. Three

patients arrive at the hospital with similar complaints.

Patient A is described as having ”fever”, B as having

”fever w/ infection/cough”, and C as having a ”respi-

ratory tract infection”. Figure 2 describes the seman-

tic annotation of these cases using a subgraph of the

SNOMED-CT.

Our machine learning model has classiﬁed both B

and C as positive examples. While for patient C the

classiﬁcation as having a respiratory tract infection

is straightforward, the same is not true for patient B.

However, by having the patients described within the

same semantic landscape which now includes back-

ground knowledge about symptoms and diseases, we

are able to understand why patient B was also classi-

ﬁed as having a respiratory tract infection. Notice that

although the semantic annotation links each record to

different concepts in the ontology, it is possible to rea-

son that both patients B and C are more similar to each

other than to patient A since both B and C suffer from

an ”Infectious process” associated with a ”Respira-

tory ﬁnding”. Of course in this simpliﬁed example,

patients are only described by easily comprehensible

textual features, but the methodology extends to cases

where objects are described my several annotations,

across multiple domains.

The following describes each step of the method-

ology, highlights the main challenges that need to be

addressed and provides a brief overview of state of

the art semantic web technologies and tools that can

be employed to tackle them.

Towards Semantic Integration for Explainable Artiﬁcial Intelligence in the Biomedical Domain

749

Figure 1: A methodology for semantic explanations for heterogeneous biomedical data.

3.1 Ontology Selection

Regarding the ﬁrst challenge, it is expectable that

multiple ontologies are needed to achieve a good cov-

erage of semantic annotations, especially in multi-

domain applications. However, to improve reason-

ing support, the selection should be focused on the

minimum set of ontologies that still provides ade-

quate granularity and scope to ensure the best possi-

ble coverage. Automated ontology recommendation

services, such as BioPortal Recommender are able

to recommend one or more ontologies that provide

adequate coverage for input textual data (Mart

ınez-

Romero et al., 2017) considering aspects such as

complementarity of resources and semantic richness.

When multiple ontologies are selected, they should as

much as possible be aligned and integrated to build a

single uniﬁed semantic landscape. In previous work,

we have developed automated approaches to select

appropriate ontologies for data integration follwoing

these principles (Faria et al., 2014).

3.2 Semantic Annotation

Once suitable ontologies are selected, the next step is

semantic annotation, i.e. connecting data to its mean-

ing that is encoded in an ontology creating a knowl-

edge graph. To ensure a high-quality semantic de-

scription of the data explored by the machine learning

models, we need to not only annotate feature values

but also metadata (e.g. feature labels) and classiﬁca-

tion targets in the case of supervised learning. The se-

mantic annotation of biomedical text needs to address

several challenges (Jovanovi

c and Bagheri, 2017): in

the case of clinical notes, the use of abbreviation

and acronyms as well as the prevalence of spelling

mistakes and of meaningless notes (e.g., ﬁlling in a

mandatory ﬁeld with a period); in the case of biomed-

ical terms, there is a high degree of ambiguity both

in terms of polysemy and homonymy, which is fur-

ther compounded with the use of acronyms that corre-

spond to words (e.g. the CAT gene). These challenges

are addressed by semantic annotation tools speciﬁ-

cally designed for the semantic annotation of biomed-

ical and clinical text (Tchechmedjiev et al., 2018)

and recent advances in word embeddings speciﬁcally

trained in the biological (Lee et al., 2020) and clin-

ical domains (Alsentzer et al., 2019) are improving

the performance in biomedical semantic annotation

(Gonc¸alves et al., 2019).

3.3 Semantic Integration

It is not uncommon that a speciﬁc application re-

quires multiple ontologies to describe the underlying

data. One one hand, because complex applications

such as clinical care and research require the integra-

tion of multiple domains of knowledge, and on the

other because uncoordinated development often re-

sults in the adoption of multiple ontologies and con-

trolled vocabularies that cover the same or similar do-

mains. In these situations, where more than one on-

tology is used to annotate the data and to establish

a ”shared context”, we need to identify the connec-

tions and relations between different ontologies and

knowledge graphs. Discovering the semantic links

or alignments between ontologies and the data sets

that they organize can be very difﬁcult, particularly

if the datasets are large and complex, as is routinely

the case in the biomedical domain. Biomedical and

clinical datasets are particularly challenging to align

for several reasons. Massive amounts of multimodal

HEALTHINF 2021 - 14th International Conference on Health Informatics

750

Figure 2: Example of the semantic annotation of patient records and machine learning target.

and diverse data are currently being generated by re-

searchers, hospitals and mobile devices around the

world, and their combined analysis presents unique

opportunities for healthcare, science, and society. The

data can range from molecular to phenotypic, be-

havioural to clinical, individual to population, genetic

to environmental. Biomedical Big Data goes well be-

yond the recognized challenges in handling large vol-

umes of data or large numbers of data sources, and

presents speciﬁc challenges pertaining to the hetero-

geneity and complexity of data as well as to the com-

plexity of its subsequent analysis.

Recent advances in semantic technologies sup-

port the rapid integration of datasets automatically

or semi-automatically by encoding machine-readable

representations of the meanings of data and metadata

items; in particular, the success of the Linked Data

movement demonstrates both the possibility of, and

need for, semantic integration of data from diverse

sources on a massive scale. This challenge can be ad-

dressed by employing ontology matching and linked

data matching techniques which are able to identify

meaningful links between entities described with dif-

ferent ontologies or vocabularies, in effect building a

KG that connects all relevant entities through contex-

tualized relations.

However, ﬁnding these relations is challenging,

because biomedical vocabulary is rich and complex,

different ontologies may model related concepts in

a different way, and the relations between concepts

may be themselves complex and semantically rich. In

previous work, we have developed the Agreement-

MakerLight ontology matching system which em-

ploys a set of diverse computational techniques rang-

ing from lexical matching, to machine learning, rea-

soning, graph visualization and user interaction to

perform the alignment of ontologies and knowledge

graphs. It is particularly suited to handle the chal-

lenges in biomedical ontology matching (Faria et al.,

2018), including ﬁnding complex correspondences

(Oliveira and Pesquita, 2018).

3.4 Semantic Explanations

When the outcomes of the ML models are semanti-

cally integrated with the input data, a shared semantic

landscape can be explored by explanations. Although

explanations based on reasoning and querying can

be employed with the proposed methodology, since

KGs naturally support both activities, here we focus

on semantic similarity-based explanations. A natu-

ral process in human learning is to identify similar

and distinguishing features to group similar objects

and discriminate different ones. At their core, many

types of AI approaches take into account similarity

modelling, including distance-based methods, such

as clustering; classiﬁcation into different kinds, such

as supervised learning; and dimensionality reduction,

such as matrix factorization or autoencoders. Expla-

nations for these approaches can be based on under-

standing why certain objects are considered similar or

different (Wang et al., 2019).

The semantic similarity between two objects can

be measured by comparing the ontology entities that

describe them(Pesquita, 2017). A simple semantic

similarity measure based on the ratio of shared ontol-

ogy classes, would score the similarity between pa-

tient B and patient C as 3/6. Computing the simi-

larity between the classiﬁcation target and each pa-

tient is also possible, with patient A having a score

of 1/6, B having a score of 1, and patient C having

a score of 2/5. Both types of similarities, between

instances, and between an instance and a classiﬁca-

tion target can be presented as explanations. There is

Towards Semantic Integration for Explainable Artiﬁcial Intelligence in the Biomedical Domain

751

a large number of semantic similarity measures that

take into account different ontology properties and

object properties, providing more sophisticated sim-

ilarity measures(). There are several challenges in

measuring semantic similarity in the biomedical do-

main, namely how to address the multiple aspects

that a KG can represent in the context of a speciﬁc

application (Sousa et al., 2020), how to adequately

consider the speciﬁcity of ontology classes (Aouicha

and Taieb, 2016) and how to employ multiple ontolo-

gies(Ferreira and Couto, 2019). We have extensive

experience in biomedical semantic similarity, having

developed methods for its computation and evalua-

tion, e.g. (Pesquita, 2017; Cardoso et al., 2020).

4 CONCLUSIONS

This work proposes a methodology to enable seman-

tic explanations of machine learning applications in

the biomedical domain. The methodology tackles

the main challenges in providing human-centric ex-

planations based on a contextualized understanding

of the data and AI outcomes. It leverages the large

amounts of freely available biomedical data and meta-

data in the form of Knowledge Graphs, and builds

upon state of the art solutions for semantic annota-

tion and integration to embed the data and AI out-

comes with already established knowledge within the

domain. It then explores semantic similarity between

instances and between instances and outcomes to sup-

port similarity-based explanations. This methodology

affords post-hoc explanations that are built indepen-

dently of the machine learning algorithms employed,

and can thus be integrated into any application for

which data can be semantically annotated with exist-

ing biomedical ontologies.

In future work, we will employ this methodol-

ogy to build a semantic explanation system integrat-

ing our existing contributions in semantic annotation,

integration and similarity and apply it to the explana-

tion of biomedical machine learning applications, in-

cluding protein-protein interaction prediction, gene-

disease association and disease progression predic-

tion.

ACKNOWLEDGEMENTS

This work was funded by the Portuguese FCT through

the LASIGE Research Unit (UIDB/00408/2020 and

UIDP/00408/2020), and also by the SMILAX project

(PTDC/EEI-ESS/4633/2014).

REFERENCES

Alsentzer, E., Murphy, J., Boag, W., Weng, W.-H., Jindi,

D., Naumann, T., and McDermott, M. (2019). Pub-

licly available clinical bert embeddings. In Proceed-

ings of the 2nd Clinical Natural Language Processing

Workshop, pages 72–78.

Aouicha, M. B. and Taieb, M. A. H. (2016). Computing

semantic similarity between biomedical concepts us-

ing new information content approach. Journal of

biomedical informatics, 59:258–275.

Cardoso, C., Sousa, R. T., K

ohler, S., and Pesquita, C.

(2020). A collection of benchmark data sets for

knowledge graph-based similarity in the biomedical

domain. Database, 2020.

Chari, S., Gruen, D. M., Seneviratne, O., and McGuinness,

D. L. (2020). Directions for explainable knowledge-

enabled systems.

Chen, J., Lecue, F., Pan, J., Horrocks, I., and Chen, H.

(2018). Knowledge-based transfer learning explana-

tion. In 16th International Conference on Principles

of Knowledge Representation and Reasoning, pages

349–358. AAAI Press.

d’Amato, C. (2020). Machine learning for the semantic

web: Lessons learnt and next research directions. Se-

mantic Web, (Preprint):1–9.

Faria, D., Pesquita, C., Mott, I., Martins, C., Couto, F. M.,

and Cruz, I. F. (2018). Tackling the challenges of

matching biomedical ontologies. Journal of biomedi-

cal semantics, 9(1):4.

Faria, D., Pesquita, C., Santos, E., Cruz, I. F., and Couto,

F. M. (2014). Automatic background knowledge se-

lection for matching biomedical ontologies. PloS one,

9(11):e111226.

Ferreira, J. D. and Couto, F. M. (2019). Multi-domain se-

mantic similarity in biomedical research. BMC bioin-

formatics, 20(10):23–31.

Ferreira, J. D., Teixeira, D. C., and Pesquita, C. (2020).

Biomedical ontologies: Coverage, access and use.

In Wolkenhauer, O., editor, Systems Medicine Inte-

grative, Qualitative and Computational Approaches,

pages 382 – 395. Academic Press, Elsevier.

Gonc¸alves, R. S., Kamdar, M. R., and Musen, M. A. (2019).

Aligning biomedical metadata with ontologies using

clustering and embeddings. In European Semantic

Web Conference, pages 146–161. Springer.

Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Gian-

notti, F., and Pedreschi, D. (2018). A survey of meth-

ods for explaining black box models. ACM computing

surveys (CSUR), 51(5):1–42.

Hitzler, P., Bianchi, F., Ebrahimi, M., and Sarker, M. K.

(2020). Neural-symbolic integration and the semantic

web. Semantic Web, 11(1):3–11.

Holzinger, A., Biemann, C., Pattichis, C. S., and Kell,

D. B. (2017). What do we need to build explainable

ai systems for the medical domain? arXiv preprint

arXiv:1712.09923.

Holzinger, A., Langs, G., Denk, H., Zatloukal, K., and

uller, H. (2019). Causability and explainability of

HEALTHINF 2021 - 14th International Conference on Health Informatics

752

artiﬁcial intelligence in medicine. Wiley Interdisci-

plinary Reviews: Data Mining and Knowledge Dis-

covery, 9(4):e1312.

Jovanovi

c, J. and Bagheri, E. (2017). Semantic annota-

tion in biomedicine: the current landscape. Journal

of biomedical semantics, 8(1):44.

Lecue, F. (2019). On the role of knowledge graphs in ex-

plainable ai. Semantic Web, (Preprint):1–11.

Lee, J., Yoon, W., Kim, S., Kim, D., Kim, S., So, C. H.,

and Kang, J. (2020). Biobert: a pre-trained biomedi-

cal language representation model for biomedical text

mining. Bioinformatics, 36(4):1234–1240.

Lundberg, S. M., Nair, B., Vavilala, M. S., Horibe, M.,

Eisses, M. J., Adams, T., Liston, D. E., Low, D. K.-

W., Newman, S.-F., Kim, J., et al. (2018). Explain-

able machine-learning predictions for the prevention

of hypoxaemia during surgery. Nature biomedical en-

gineering, 2(10):749–760.

Mart

ınez-Romero, M., Jonquet, C., O’connor, M. J., Gray-

beal, J., Pazos, A., and Musen, M. A. (2017). Ncbo

ontology recommender 2.0: an enhanced approach

for biomedical ontology recommendation. Journal of

biomedical semantics, 8(1):21.

Miller, T. (2019). Explanation in artiﬁcial intelligence: In-

sights from the social sciences. Artiﬁcial Intelligence,

267:1–38.

Min, S., Lee, B., and Yoon, S. (2017). Deep learn-

ing in bioinformatics. Brieﬁngs in bioinformatics,

18(5):851–869.

Miotto, R., Wang, F., Wang, S., Jiang, X., and Dudley, J. T.

(2018). Deep learning for healthcare: review, oppor-

tunities and challenges. Brieﬁngs in bioinformatics,

19(6):1236–1246.

Munn, K. and Smith, B. (2013). Applied ontology: An in-

troduction, volume 9. Walter de Gruyter.

Oliveira, D. and Pesquita, C. (2018). Improving the inter-

operability of biomedical ontologies with compound

alignments. Journal of biomedical semantics, 9(1):1.

Pesquita, C. (2017). Semantic similarity in the gene ontol-

ogy. In The gene ontology handbook, pages 161–173.

Humana Press, New York, NY.

Phan, N., Dou, D., Wang, H., Kil, D., and Piniewski, B.

(2017). Ontology-based deep learning for human be-

havior prediction with explanations in health social

networks. Information sciences, 384:298–313.

Sarker, M. K., Xie, N., Doran, D., Raymer, M., and Hitzler,

P. (2017). Explaining trained neural networks with se-

mantic web technologies: First steps. In Twelfth Inter-

national Workshop on Neural-Symbolic Learning and

Reasoning 2017, London, UK, July 17-18, 2017.

Sousa, R. T., Silva, S., and Pesquita, C. (2020). Evolving

knowledge graph similarity for supervised learning in

complex biomedical domains. BMC bioinformatics,

21(1):6.

Tchechmedjiev, A., Abdaoui, A., Emonet, V., Melzi, S.,

Jonnagaddala, J., and Jonquet, C. (2018). En-

hanced functionalities for annotating and indexing

clinical text with the ncbo annotator+. Bioinformat-

ics, 34(11):1962–1965.

Wang, D., Yang, Q., Abdul, A., and Lim, B. Y. (2019). De-

signing theory-driven user-centric explainable ai. In

Proceedings of the 2019 CHI conference on human

factors in computing systems, pages 1–15.

Wollschlaeger, B., Eichenberg, E., and Kabitzsch, K.

(2020). Explain yourself: A semantic annotation

framework to facilitate tagging of semantic informa-

tion in health smart homes. In HEALTHINF, pages

133–144.

Zhou, B., Bau, D., Oliva, A., and Torralba, A. (2018). In-

terpreting deep visual representations via network dis-

section. IEEE transactions on pattern analysis and

machine intelligence, 41(9):2131–2145.

Towards Semantic Integration for Explainable Artiﬁcial Intelligence in the Biomedical Domain

753