=Paper=
{{Paper
|id=Vol-3180/paper-92
|storemode=property
|title=Explaining Artificial Intelligence Predictions of Disease Progression with Semantic
Similarity
|pdfUrl=https://ceur-ws.org/Vol-3180/paper-92.pdf
|volume=Vol-3180
|authors=Susana Nunes,Rita T. Sousa,Filipa Serrano,Ruben Branco,Diogo F. Soares,Andreia S. Martins,Eleonora Auletta,Eduardo N. Castanho,Sara C. Madeira,Helena Aidos,Catia Pesquita
|dblpUrl=https://dblp.org/rec/conf/clef/NunesSSBSMACMAP22
}}
==Explaining Artificial Intelligence Predictions of Disease Progression with Semantic
Similarity==
https://ceur-ws.org/Vol-3180/paper-92.pdf
Explaining Artificial Intelligence Predictions of
Disease Progression with Semantic Similarity
Susana Nunes† , Rita T. Sousa† , Filipa Serrano† , Ruben Branco, Diogo F. Soares,
Andreia S. Martins, Eleonora Auletta, Eduardo N. Castanho, Sara C. Madeira,
Helena Aidos and Catia Pesquita*
LASIGE, Faculdade de Ciências, Universidade de Lisboa, Portugal
Abstract
The complexity of neurodegenerative diseases has motivated the development of artificial intelligence
approaches to predicting risk of impairment and disease progression. However, and despite the success
of these approaches, their mostly black-box nature hinders their adoption for disease management.
Explainable artificial intelligence holds the promise to bridge this gap by producing explanations of
models and their predictions that promote understanding and trust by users. In the biomedical domain,
given its complexity, explainable artificial intelligence approaches have much to benefit from being able
to link models to representations of domain knowledge – ontologies. Ontologies afford more explainable
features because they are semantically enriched and contextualized and as such can be better understood
by end users; and they also model existing knowledge, and thus support inquiry into how a given
artificial intelligence model outcome fits with existing scientific knowledge. We propose an explainability
approach that leverages on the rich panorama of biomedical ontologies to build semantic similarity-based
explanations that contextualize patient data and artificial intelligence predictions. These explanations
mirror a fundamental human explanatory mechanism - similarity - while tackling the challenges of data
complexity, heterogeneity and size.
Keywords
Semantic Similarity, Ontology, Amyotrophic Lateral Sclerosis, Explainable AI
1. Introduction
Amyotrophic Lateral Sclerosis (ALS) is a neurodegenerative disease targeting the upper and
lower motor neurons, leading to progressive and diffuse paralysis and eventual respiratory
death. The underlying causes of ALS are not fully understood, and there is a heterogeneity of
clinical symptoms and a highly fluctuating life expectancy. Therefore, multiple approaches have
been proposed to predict disease progression to improve patient-personalized treatment [1, 2].
The iDPP CLEF 2022 [3, 4] is focused on the evaluation of Artificial Intelligence (AI)
CLEF 2022: Conference and Labs of the Evaluation Forum, September 5–8, 2022, Bologna, Italy
*
Corresponding author.
†
These authors contributed equally.
$ scnunes@ciencias.ulisboa.pt (S. Nunes); risousa@ciencias.ulisboa.pt (R. T. Sousa); fserrano@lasige.di.fc.ul.pt
(F. Serrano); rmbranco@ciencias.ulisboa.pt (R. Branco); dfsoares@ciencias.ulisboa.pt (D. F. Soares);
amartins@lasige.di.fc.ul.pt (A. S. Martins); eauletta@ciencias.ulisboa.pt (E. Auletta); ejcastanho@ciencias.ulisboa.pt
(E. N. Castanho); sacmadeira@ciencias.ulisboa.pt (S. C. Madeira); haidos@ciencias.ulisboa.pt (H. Aidos);
clpesquita@ciencias.ulisboa.pt (C. Pesquita)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073 CEUR Workshop Proceedings (CEUR-WS.org)
algorithms to predict ALS progression. Task 2, in particular, is dedicated to predicting time of
impairment targeting different events: non-invasive ventilation (NIV), percutaneous endoscopic
gastrostomy (PEG), and death. It includes highly-curated datasets based on real ALS patients,
followed at clinical institutions in Lisbon, Portugal, and Turin, Italy.
Machine learning (ML) approaches are increasingly successful in predicting the progression
of ALS, including the prognostic prediction of the need for non-invasive ventilation [5, 6] and
patient profiling [7]. The complexity of the underlying domain and data, coupled with the
black-box nature of many of the algorithms employed hinder their widespread adoption on the
management of these diseases. To tackle this, one of the challenge’s goals is also to propose
new approaches to make those prediction algorithms explainable.
Explainable artificial intelligence (XAI) is focused on developing approaches that ensure
algorithmic fairness, identify potential bias/problems in the training data, ensure that the
algorithms perform as expected, and bridge the gap between the ML community and other
scientific disciplines [8]. The lack of trust is the main barrier in the adoption of AI in clinical
practice [9]. XAI approaches can be widely classified into two categories: transparent models,
which are interpretable by design and include decision trees and linear models [10] and post-
hoc explainability[11, 12], where models are explained through external techniques. This last
category can also be divided into model-agnostic techniques, designed to be applied to any
ML model, and model-specific techniques tailored to explain a particular ML model. Recently,
post-hoc methods have been applied to explain recurrent neural networks that predict the need
for non-invasive ventilation [13].
In the biomedical domain, given its complexity, XAI approaches have much to benefit from
being able to link ML models to representations of domain knowledge [14, 15]. These Explainable
Knowledge-enabled Systems include a representation of the domain knowledge and support
explanation approaches that are context-aware and provenance enabled [16]. This knowledge
integration is crucial, since well-grounded explanations should be domain-dependent, which
means that they should be set within a context that depends on the task, background knowledge
and expectations of the user [17].
Semantic web technologies such as ontologies and knowledge graphs represent an unpar-
alleled solution to the problem of knowledge-enabled explanations since they provide the
needed semantic context [18]. Ontologies are semantic models that encode the knowledge of a
domain, in which each element is precisely defined, and the relationships between elements are
parameterized or constrained [19].
Ontologies add two very relevant dimensions to explainability: one the one hand, they afford
explainable features, i.e., features that are semantically enriched and contextualized and as such
can be better understood by end users; and on the other, they model existing knowledge, and
thus support inquiry into how a given AI model outcome fits with extant knowledge. These
two aspects are key for explainability and trust.
This work builds on the recently developed Brainteaser Ontology [20] and explores how
this ontology coupled with the rich panorama of more general biomedical ontologies can
support similarity-based explanations for patient end-stage event predictions that build upon
the contextualization of patient data and AI predictions.
2. Proposal
Similarity assessment is a natural explanatory mechanism [21] since the identification of similar
features to group similar objects is a basic cognitive ability. This is a frequent strategy in clinical
settings, where similarity between patients can be used by clinicians to help decide on the best
course of action through analogical reasoning [22]. Measuring similarity is also fundamental to
many ML algorithms, making similarity both a representation that can be explored by ML and
explanation approaches. The basis of a similarity-based explanation is that a prediction for a
given entity is formulated based on its similarity to other relevant entities with known outcome.
While comparing the numerical values assigned to patient features such as weight or blood
pressure can be achieved directly, the comparison of more complex concepts, such as symptoms,
is not straightforward. However, when ontologies are used to describe data, the semantic
annotations of data objects can be used to compute their semantic similarity[23]. A semantic
annotation assigns real-world entities in a domain to their semantic description (i.e. class of
an ontology) [24], and the ontology data model can be applied to a set of individual entities to
create a knowledge graph [25]. Ontologies thus provide the scaffolding for comparing entities
at a higher level of complexity by comparing their meaning. A semantic similarity measure is
then a function that, given two entities described in the knowledge graph, returns a numerical
score reflecting the closeness in meaning between them.
Figure 1 presents a subgraph of the Brainteaser Ontology (BO) enriched with the National
Cancer Institute Thesaurus (NCIt) ontology [26], where patients are described according to
the symptoms that characterize them. Symptoms are organized into a hierarchical structure
given by the is_a links. Patients B and C can be considered more similar because they share
the annotations "Spasticity" and "Feeding difficulties and mismanagement". Patient A is also
somewhat similar to Patient B since they are both annotated with "Muscle symptom". Most
semantic similarity measures are taxonomic in nature, they explore the hierarchical relations
between classes in an ontology. For instance, Patient B is directly annotated with "Muscle
twitching" whereas Patient A exhibits "Muscle symptom". By exploring the hierarchical relations
modeled in the ontology, both patients share an annotation to "Muscle symptom". Consequently,
ontologies with a rich taxonomic backbone typically allow more precise measures of semantic
similarity.
Semantic similarity-based explanations are based on understanding why particular objects, in
this case, patients, are considered similar or different and how that relates to the AI predictions.
Classical semantic similarity returns a single score, which on the one hand is easy to understand
and condenses the representation of a potential very large feature set into a single score, but on
the other affords little insight. To improve on the level of detail supported by similarity-based
explanations, we propose to compute several semantic similarity scores that describe similarity
in meaning under different semantic aspects encoded by the ontologies. A semantic aspect
(SA) corresponds to a subgraph of the ontology rooted at a class of interest. For example, an
explanation may be grounded on three different semantic similarity scores: co-morbidities
similarity, lifestyle similarity and symptom similarity.
These semantic similarity-based explanations are considered post-hoc, since they provide
explanations for the models’ predictions, as opposed to approaches that explain the models’
workings. They are also considered local, since they focus on explaining specific predictions
Figure 1: Semantic annotation of patients using a subgraph of Brainteaser Ontology enriched with the
NCIt ontology.
made by the model, instead of providing global explanations that apply to all predictions [27].
In particular, semantic similarity explanations focus on how a certain prediction fits with the
patient features and their interpretation according to existing scientific knowledge encoded in
the ontologies.
3. Methodology
In the context of the iDPP CLEF 2022, we propose a novel approach that generates semantic
similarity-based explanations for patient-level predictions. The underlying idea is that we can
explain the prediction for one patient by considering aspect-oriented semantic similarity with
other relevant patients based on the most important features used by ML approaches or selected
by users.
To build rich and easy to understand semantic-similarity based explanations, our approach
requires five steps (see Figure 2) : (1) the enrichment of the Brainteaser Ontology through
integration of other biomedical ontologies; (2) the semantic annotation of patients (if not already
available); (3) the similarity calculation between patients; (4) selection of the set of patients
to explain a specific prediction; and (5) the visualization of the generated similarity-based
explanations.
3.1. Ontology Integration
To ensure a rich and comprehensive semantic annotation of the data, we enrich the Brainteaser
Ontology with links to other ontologies. The BO models the data collected to describe disease
progression in ALS and MS. It reuses single classes from other biomedical ontologies listed in
Table 1. However it does not include import statements which hinders its ability to support a
more complete semantic similarity calculation since it is missing the context of the reused classes.
By explicitly linking BO to the other 9 biomedical ontologies and controlled vocabularies it
reuses through import statements we can establish a rich semantic landscape with approximately
770,680 classes in total and covering a variety of domains.
Table 1
Main ontologies and number of classes used to create the Brainteaser Ontology.
Ontology Classes
National Cancer Institute thesaurus (NCIt) 173,001
Ontology of Genes and Genomes (OGG) 69,689
Uberon 20,849
Medical Action Ontology (MAxO) 15,086
Ontology for MIRNA Target (OMIT) 90,916
SNOMED Clinical Terms (SNOMED CT) 358,483
Anatomical Therapeutic Chemical Classification (ATCC) 6,567
International Standard Classification of Occupations (ISCO) 619
Experimental Factor Ontology (EFO) 35,470
Figure 3 represents the impact of this enrichment via integration. On the left side the subgraph
of the neighborhood of the classes "Rheumatoid Arthritis", "Pulmonary Sarcoidosis" and "Heart
Disorder" as modeled in the BO is shown. Using the unenriched BO ontology, would result
in patients exhibiting theses diseases or disorders being all considered equally similar: the
shared meaning of these ontology classes is simply that they are all subclasses of "Disease or
Disorder". The right side illustrates the enriched subgraph including the full NCIt hierarchy,
which encodes considerably more contextual information, and supports semantic similarity
measures in correctly identifying that "Rheumatoid Arthritis" and "Pulmonary Sarcoidosis" are
more similar, since they are both Autoimmune Diseases.
3.2. Patient Semantic Annotation
Semantic annotation needs to be performed only when patient data is not annotated with the
ontology. Semantic annotations can be performed manually, but there are a number of existing
tools that facilitate this step such as the well known NCBO Annotator [28, 29]. An annotation
assigns an ontology class to describe an entity. In our case an annotation is a tuple a = that defines an ontology class to describe a patient at a given point
in time (clinical visit). A patient can then be represented by a vector of all their annotations,
i.e. the ontology classes that describe them and the time points where the assignment was
registered.
𝑃 = {𝑎1 , ..., 𝑎𝑛 } (1)
Figure 2: Overview of the semantic-similarity explanations methodology
3.3. Semantic Similarities between Patients
We take as input the most important features employed by ML methods to predict disease
progression and/or other features considered relevant by users, and use them to define the
different semantic aspects for which semantic similarity is calculated.
More formally, the similarity between two patients corresponds to a vector of semantic
similarity scores corresponding to each relevant aspect. This similarity may be computed
irrespective of time of annotation (considering all annotations for all time points) or by time
point.
𝑠𝑖𝑚(𝑃1 , 𝑃2 ) = {𝑠𝑖𝑚𝑆𝐴1 (𝑃1 , 𝑃2 ), ..., 𝑠𝑖𝑚𝑆𝐴𝑛 (𝑃1 , 𝑃2 )} (2)
Several well-established works explore the taxonomical (hierarchical) component of the
ontologies to measure the shared meaning between two entities in the ontology [30].
A popular semantic similarity measure is ResnikBMA based on the class-based measure
proposed by Resnik [31] in which the similarity between two classes corresponds to the In-
Figure 3: Impact of the Brainteaser Ontology enrichment through the integration of several relations
between domains.
formation Content (IC) of their most informative common ancestor. The IC is a measure the
reflects the specificity of a class in the ontology.
𝑠𝑖𝑚(𝑐1 , 𝑐2 ) = 𝑚𝑎𝑥{𝐼𝐶(𝑐)) : 𝑐 𝜖{𝐴(𝑐1 ) ∩ 𝐴(𝑐2 )}} (3)
where 𝑐 is a class in 𝐴(𝑐𝑖 ), the set of ancestors of 𝑐𝑖 .
ResnikBMA then uses the Best-Match Average to consider the best scoring pairs of classes
from each entity.
Σ𝑐1 𝜖𝐴(𝑒1 ) 𝑠𝑖𝑚(𝑐1 , 𝑐2 ) Σ𝑐2 𝜖𝐴(𝑒2 ) 𝑠𝑖𝑚(𝑐1 , 𝑐2 )
𝐵𝑀 𝐴(𝑒1 , 𝑒2 ) = + (4)
2|𝐴(𝑒1 )| 2|𝐴(𝑒2 )|
where 𝐴(𝑒𝑖 ) is the number of annotations for entity 𝑒𝑖 and 𝑠𝑖𝑚(𝑒1 , 𝑒2 ) is the semantic similarity
between the class 𝑐1 and class 𝑐2 .
Recently, more sophisticated approaches based on knowledge graph embeddings allow the
representation of each entity with a vector that approximates the similarity properties of the
graph and can then be used to compute similarity using operations such as cosine similarity [32].
3.4. Building Semantic Similarity Explanations
The models proposed by our team for the task of predicting the time of impairment for ALS of
the iDPP CLEF 2022 addressed the prediction of the event separately from predicting the time
window to the event. For event prediction, all time points were considered but the first three
proved to be the most relevant for the models. These were then also used in predicting time to
event. Likewise, to develop explanations better suited for this approach, we also developed two
types of explanations: event explanation and time window to event explanation.
A key aspect is selecting the most relevant patients to explain an event prediction. We
define three types of explanatory patients: (1) The N most similar patients exhibiting the same
outcome; (2) The N least similar patients with the same outcome; (3) The N most similar patients
exhibiting a different outcome. Since we compute several similarities in the previous step, to
select these explanatory patients, we propose an aggregated similarity where different features
can be assigned weights. Candidate explanatory patients are then ranked by their semantic
similarity to the target patient and the final set of explanatory patients is selected.
A semantic similarity explanation of a given target patient can then be defined as a two-
dimensional tensor where the first dimension represents patients and the second dimension the
semantic similarities according to different aspects:
𝑆𝑆𝐸(𝑃𝑡𝑎𝑟𝑔𝑒𝑡 ) = {𝑠𝑖𝑚(𝑃𝑡𝑎𝑟𝑔𝑒𝑡 , 𝑃𝑖 ), ..., 𝑠𝑖𝑚(𝑃𝑡𝑎𝑟𝑔𝑒𝑡 , 𝑃𝑛 )} (5)
where 𝑃𝑖..𝑛 correspond to patients sampled from the three types.
Regarding the prediction of a time window to a specific event, we follow a similar approach:
(1) The N most similar patients exhibiting the same event in the same time window; (2) The
N least similar patients exhibiting the same event in the same time window; (3) The N most
similar patients exhibiting the same event in a different time window. N can be defined by the
end-user, and may take a different value for each patient type.
The explanations for the time window to event predictions take the form of a three-dimensional
tensor, where the first dimension is the explanatory patients, the second dimension corresponds
to the semantic similarities according to different aspects and the third dimension represents
the time points.
3.5. Explanation Visualization
Finally, one fundamental aspect of explainability is communication to the end user. Our proposal
orchestrates semantic similarity explanations into a visualization that combines global and
aspect-oriented similarity for different sets of relevant patients.
We chose to use heatmaps to visualize the similarity explanations. For event prediction
explanations the representation of the two-dimensional tensor is straightforward, the y-axis
contains the patients and the x-axis the similarity scores according to the different aspects. Each
cell in the map is a colour-coded representation of the similarity of each patient to the patient
we want to explain for each feature, where a darker colour corresponds to higher similarity.
The visualization also includes the global similarity employed to select the different types of
patients. For time to event explanations, the third dimension is encapsulated within the x-axis
with three cells (one per time point).
4. Semantic Similarity Explanations
To showcase our proposal for semantic similarity explanations, we simulated patients with
annotations to the BO. One of them is the target patient for whom an event is predicted and an
explanation is built. The remaining patients are patients for whom the outcome is known and
they are used to explain the target patient.
To explain an event prediction, we set N=2 for the three types of explanatory patients.
To explain time window prediction we employed N=1,5,1 for each explanatory patient type
respectively. The heatmaps were built using the following selection of features: lifestyle, onset,
co-morbidities, symptoms and pharmacological substances. Figure 4 depicts the event prediction
explanation, while Figure 5 illustrates the time window to event prediction explanation. Patients
in green correspond to the most similar patients to the target patient (P1) and whose outcome
event or time to event is the same. Patients in white correspond to the most similar patients
similar to P1, but with a different prediction. Finally, patients in pink correspond to the least
similar patients to the target patient, but who share the same prediction.
Figure 4: Semantic similarity explanation for the event prediction.
Figure 5: Semantic similarity explanation for the time window to the event prediction.
5. Conclusions
Similarity is a natural explanatory approach, but computing similarity for complex data is
challenging. Semantic similarity provides an opportunity for explainability in the context of
the iDPP CLEF 2022 since it allows more complex patient comparisons supported by the
scientific context encoded in ontologies. Our proposal is based on first enriching the Brainteaser
Ontology with explicit imports of reused ontologies to support a more granular computation
of patient semantic similarity. The semantic similarity explanations are based on calculating
specific semantic similarity values according to relevant aspects either identified as relevant for
the ML predictions or selected by the end users. The end result is heatmap-based visualizations
that allow the comparison of the target patient whose predictions we want to understand with
other relevant patients with known outcomes. This method can be integrated with other types
of similarities, for instance, comparisons of Revised Amyotrophic Lateral Sclerosis Functional
Rating Scale (ALSFRS-R) scores.
A challenge for this explainability method is that it requires patient data to be sufficiently
detailed to ensure high quality annotations. The data currently available for the challenge can
be used to support these methods, but it does not yet afford a complete patient representation
under the BO ontology. As data becomes richer and more detailed, the potential value of
semantic similarity explanations increases. However, their true value must be measured in user
studies, where predictions and explanations are shown to experts and their ability to help users
understand a prediction is assessed.
Acknowledgements
The authors are funded by the FCT through LASIGE Research Unit (ref. UIDB/00408/2020 and
ref. UIDP/00408/2020),AIpALS (PTDC/CCI-CIF/4613/2020), PhD research scholarships to RTS
(SFRH/BD/145377/2019), DFS (2020.05100.BD), ENC (2021.07810.BD); and by the BRAINTEASER
project and KATY project that have received funding from the European Union’s Horizon
2020 research and innovation programme under the grants agreements No 101017598 and No
101017453, respectively.
References
[1] V. Grollemund, P.-F. Pradat, G. Querin, F. Delbot, G. Le Chat, J.-F. Pradat-Peyre, P. Bede,
Machine learning in amyotrophic lateral sclerosis: achievements, pitfalls, and future
directions, Frontiers in neuroscience 13 (2019) 135.
[2] S. Pires, M. Gromicho, S. Pinto, M. d. Carvalho, S. C. Madeira, Patient stratification
using clinical and patient profiles: Targeting personalized prognostic prediction in als, in:
International Work-Conference on Bioinformatics and Biomedical Engineering, Springer,
2020, pp. 529–541.
[3] A. Guazzo, I. Trescato, E. Longato, E. Hazizaj, D. Dosso, G. Faggioli, G. M. Di Nunzio,
G. Silvello, M. Vettoretti, E. Tavazzi, C. Roversi, P. Fariselli, S. C. Madeira, M. de Car-
valho, M. Gromicho, A. Chiò, U. Manera, A. Dagliati, G. Birolo, H. Aidos, B. Di Camillo,
N. Ferro, Intelligent Disease Progression Prediction: Overview of iDPP@CLEF 2022, in:
A. Barrón-Cedeño, G. Da San Martino, M. Degli Esposti, F. Sebastiani, C. Macdonald,
G. Pasi, A. Hanbury, M. Potthast, G. Faggioli, N. Ferro (Eds.), Experimental IR Meets Mul-
tilinguality, Multimodality, and Interaction. Proceedings of the Thirteenth International
Conference of the CLEF Association (CLEF 2022), Lecture Notes in Computer Science
(LNCS) 13390, Springer, Heidelberg, Germany, 2022.
[4] A. Guazzo, I. Trescato, E. Longato, E. Hazizaj, D. Dosso, G. Faggioli, G. M. Di Nunzio,
G. Silvello, M. Vettoretti, E. Tavazzi, C. Roversi, P. Fariselli, S. C. Madeira, M. de Carvalho,
M. Gromicho, A. Chiò, U. Manera, A. Dagliati, G. Birolo, H. Aidos, B. Di Camillo, N. Ferro,
Overview of iDPP@CLEF 2022: The Intelligent Disease Progression Prediction Challenge,
in: G. Faggioli, N. Ferro, A. Hanbury, M. Potthast (Eds.), CLEF 2022 Working Notes, CEUR
Workshop Proceedings (CEUR-WS.org), ISSN 1613-0073, 2022.
[5] D. Soares, R. Henriques, M. Gromicho, S. Pinto, M. d. Carvalho, S. C. Madeira, Towards
triclustering-based classification of three-way clinical data: A case study on predicting
non-invasive ventilation in als, in: International Conference on Practical Applications of
Computational Biology & Bioinformatics, Springer, 2020, pp. 112–122.
[6] A. S. Martins, M. Gromicho, S. Pinto, M. de Carvalho, S. C. Madeira, Learning prognostic
models using diseaseprogression patterns: Predicting the need fornon-invasive ventilation
in amyotrophic lateralsclerosis, IEEE/ACM Transactions on Computational Biology and
Bioinformatics (2021).
[7] T. Leão, S. C. Madeira, M. Gromicho, M. de Carvalho, A. M. Carvalho, Learning dynamic
bayesian networks from time-dependent and time-independent data: Unraveling disease
progression in amyotrophic lateral sclerosis, Journal of Biomedical Informatics 117 (2021)
103730.
[8] L. H. Gilpin, D. Bau, B. Z. Yuan, A. Bajwa, M. Specter, L. Kagal, Explaining explanations:
An overview of interpretability of machine learning, in: 2018 IEEE 5th International
Conference on data science and advanced analytics (DSAA), IEEE, 2018, pp. 80–89.
[9] E. Glikson, A. W. Woolley, Human trust in artificial intelligence: Review of empirical
research, Academy of Management Annals 14 (2020) 627–660.
[10] C. Rudin, Stop explaining black box machine learning models for high stakes decisions
and use interpretable models instead, Nature Machine Intelligence 1 (2019) 206–215.
[11] M. T. Ribeiro, S. Singh, C. Guestrin, "Why should i trust you?": explaining the predictions
of any classifier, in: Proceedings of the 22nd ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, KDD ’16, Association for Computing Machinery,
New York, NY, USA, 2016, p. 1135–1144. doi:10.1145/2939672.2939778.
[12] R. Guidotti, A. Monreale, S. Ruggieri, D. Pedreschi, F. Turini, F. Giannotti, Local rule-based
explanations of black box decision systems, arXiv preprint arXiv:1805.10820 (2018).
[13] M. Müller, M. Gromicho, M. de Carvalho, S. C. Madeira, Explainable models of disease
progression in als: Learning from longitudinal clinical data with recurrent neural networks
and deep model explanation, Computer Methods and Programs in Biomedicine Update 1
(2021) 100018.
[14] A. Holzinger, C. Biemann, C. S. Pattichis, D. B. Kell, What do we need to build explainable
ai systems for the medical domain?, arXiv preprint arXiv:1712.09923 (2017).
[15] B. Wollschlaeger, E. Eichenberg, K. Kabitzsch, Explain yourself: A semantic annotation
framework to facilitate tagging of semantic information in health smart homes., in:
HEALTHINF, 2020, pp. 133–144.
[16] I. Tiddi, et al., Directions for explainable knowledge-enabled systems, Knowledge Graphs
for eXplainable Artificial Intelligence: Foundations, Applications and Challenges 47 (2020)
245.
[17] D. Gunning, M. Stefik, J. Choi, T. Miller, S. Stumpf, G.-Z. Yang, XAI - explainable artificial
intelligence, Science Robotics 4 (2019) eaay7120. doi:10.1126/scirobotics.aay7120.
[18] F. Lécué, On the role of knowledge graphs in explainable ai, Semantic Web 11 (2019)
41–51.
[19] S. Staab, R. Studer, Handbook on ontologies, Springer-Verlag, Berlin Heidelberg, 2010.
doi:10.1007/978-3-540-92673-3.
[20] M. Bettin, G. M. Di Nunzio, D. Dosso, G. Faggioli, N. Ferro, N. Marchetti, G. Silvello,
Deliverable 9.1 – Project ontology and terminology, including data mapper and RDF graph
builder, BRAINTEASER, EU Horizon 2020, Contract N. GA101017598. https://brainteaser.
health/, 2021.
[21] D. Wang, Q. Yang, A. Abdul, B. Y. Lim, Designing theory-driven user-centric explainable
ai, in: Proceedings of the 2019 CHI conference on human factors in computing systems,
2019, pp. 1–15.
[22] D. Dumas, P. A. Alexander, L. M. Baker, S. Jablansky, K. N. Dunbar, Relational reasoning in
medical education: Patterns in discourse and diagnosis., Journal of Educational Psychology
106 (2014) 1021.
[23] C. Pesquita, Towards semantic integration for explainable artificial intelligence in the
biomedical domain., in: HEALTHINF, 2021, pp. 747–753.
[24] J. Jovanović, E. Bagheri, Semantic annotation in biomedicine: the current landscape,
Journal of biomedical semantics 8 (2017) 1–18.
[25] L. Ehrlinger, W. Wöß, Towards a definition of knowledge graphs., SEMANTiCS (Posters,
Demos, SuCCESS) 48 (2016) 1–4.
[26] S. de Coronado, L. W. Wright, G. Fragoso, M. W. Haber, E. A. Hahn-Dantona, F. W. Hartel,
S. L. Quan, T. Safran, N. Thomas, L. Whiteman, The nci thesaurus quality assurance life
cycle, Journal of biomedical informatics 42 (2009) 530–539.
[27] R. Guidotti, A. Monreale, S. Ruggieri, F. Turini, F. Giannotti, D. Pedreschi, A survey of
methods for explaining black box models, ACM computing surveys (CSUR) 51 (2018) 1–42.
[28] C. Jonquet, N. Shah, C. Youn, M. Musen, C. Callendar, M.-A. Storey, Ncbo annotator:
Semantic annotation of biomedical data, ISWC (2009).
[29] A. Tchechmedjiev, A. Abdaoui, V. Emonet, S. Melzi, J. Jonnagaddala, C. Jonquet, Enhanced
functionalities for annotating and indexing clinical text with the NCBO Annotator+,
Bioinformatics 34 (2018) 1962–1965. URL: https://doi.org/10.1093/bioinformatics/bty009.
doi:10.1093/bioinformatics/bty009.
[30] C. Pesquita, D. Faria, A. O. Falcão, P. Lord, F. M. Couto, Semantic similarity in biomedical
ontologies, PLOS Computational Biology 5 (2009) 1–12. doi:10.1371/journal.pcbi.
1000443.
[31] P. Resnik, Using information content to evaluate semantic similarity in a taxonomy, in:
Proceedings of the 14th International Joint Conference on Artificial Intelligence - Volume
1, IJCAI’95, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1995, p. 448–453.
[32] Q. Wang, Z. Mao, B. Wang, L. Guo, Knowledge graph embedding: A survey of approaches
and applications, IEEE Transactions on Knowledge and Data Engineering 29 (2017) 2724–
2743. doi:10.1109/TKDE.2017.2754499.