Explainability of intelligent systems in health-care domain is still in its initial state. Recently, more e orts are made to leverage machine learning in solving causal inference problems of disease diagnosis, prediction and treatments. This research work presents an ontology based causal inference model for hypothyroid disease diagnosis using case-based reasoning. The e ectiveness of the proposed method is demonstrated with an example from hypothyroid disease domain. Here, the domain knowledge is mapped into an expert de ned ontology and causal inference is performed based on this domain-speci c ontology. The goal is to incorporate this causal inference model in traditional case-based reasoning cycle enabling explanation for each solved problem. Finally, a mechanism is de ned to deduce explanation for a solution to a problem case from the combined causal statements of similar cases. The initial result shows that case-based reasoning can retrieve relevant cases with 95% accuracy.
With the outburst of AI applications, expectations have been increased for intelligent systems in all domains including the health-care domain. The growing capabilities of AI, especially the applications of Machine Learning (ML) leverage new requirements to be ful lled such as human-level intelligence. According to Pearl, at present three prime obstacles in AI and/or ML applications are there to achieve human-level intelligence, which are; i) adaptability or robustness, ii) explainability and iii) understanding of cause-e ect connections. Explainability is one of the major characteristics of human-level intelligence. In recent years, a good number of works have been done to equip systems with causal models in disease diagnosis. Incorporating causal model would facilitate explainability of the prevailing systems [18]. State-of-the-art techniques to develop AI applications are gaining more accuracy gradually but lack of proper explanation with the solution inhibits the reliability of those techniques. Another signi cant issue is to
include causality in the explanation to a solution. If these issues can be solved, most of the challenges to produce AI/ML applications with human-level intelligence will be conquered. Causality in intelligent health-care applications is still in an immature state and there are rooms for further research and improvement. According to Sene, medical information is doubling every 5 years but only 20% of the evidence based knowledge is used by medical practitioners [22]. Several research works have been carried out to make evident based knowledge useful to clinicians as well as intelligent health-care systems. Case-based reasoning (CBR) has been being used for health systems due to its close characteristics to human behaviour, intimidating previous experience as the medical practitioners do. In CBR, only the solution is produced for a new problem case with responsible features only. In this paper, we have adapted the traditional structure of the CBR cycle and introduce causal inference using domain ontology. This research work is aimed at making CBR systems more explainable for health-care services with the support of domain knowledge from an ontology. The objective of this work is further elaborated by considering the following example scenario.
Problem scenario: A patient is reported to be diagnosed for having thyroid diseases provided he/she has taken several tests and recorded his/her history of medication, treatment and symptoms. For this scenario, a physician with access to machines that are capable of classifying based on historical data, can diagnose the patient primarily. From state-of-the-art techniques, for example rough sets learning [19], the physician could get the nal verdict of the diagnosis. The verdict may provide that due to having Thyroid Stimulating Hormone (TSH) test value less than or equal to 6 mU/L, the patient may have negative hypothyroid. But for T SH > 6 mU/L, the patient might be diagnosed for negative, primary hypothyroid and compensated hypothyroid with 40%, 30% and 30% possibilities respectively. In this uncertain scenario, some causal statements in support of each verdict along with the contribution of other features from the diagnosis would facilitate the decision making for a physician. To overcome these uncertainties and obstacle for explainable AI applications, this research work is destined to produce human understandable explanation in natural language with the label or verdict generated by a classi cation mechanism from prede ned causal statements.
The remaining parts of this paper is divided into several sections. Section 2 describes the background of the concerned topics with several related works. Detailed description of the proposed causal model with formal de nitions and brief descriptions of developed components are discussed in Section 3. Section 4 presents a brief evaluation on case retrieval model and the description of extracting explanation for extracted solution. Finally, Section 5 states some of the future possibilities of this research work and conclusive statements respectively. 2
This section contains a short description of the methods used to achieve the goal of this work followed by a brief discussion on recent research works.
CBR is an AI approach that uses previous experiences to solve a current
problem which perfectly aligns with the characteristics of physicians in case
of patient for diseases. According to Kolodner [13], CBR is a reasoner that
solves a new problem by remembering and using past situations similar to the
current. CBR is an instance-based lazy learning method, i.e., it does not try to
reason until it has to [15]. The term \case" represents an experience achieved
from a previously solved problem. The term \based" means in CBR cases are
the source of reasoning. Finally, the term \reasoning" means the approach of
problem-solving, i.e., solving a problem by concluding using previously solved
cases [21]. Based on implementation techniques, CBR can be distinguished into
four main types: i) CBR using nearest neighbour, ii) CBR using induction, iii)
CBR using fuzzy logic and iv) CBR using database technologies [22]. In addition
to the basic implementation methodologies, Barua et al. have shown a distributed
architecture of CBR using XML les containing individual cases [
Ontology structures the concepts with de nition and relations. The term
ontology has been evolved from philosophy which means a systematic account of
existence [
Various forms of causal models are being investigated for long to be used in
disease diagnosis. Mostly causal models are comprised of mathematical models
that represent the causal relationships in a system [10]. Pearl introduced causal
models based on the Bayesian Network (BN) [17], a commonly used
methodology for prediction and classi cation tasks in di erent domains. Basically, a BN
consists of a directed acyclic graph (DAG) and a set of conditional probability
tables such that each node of the graph represents a variable and it is associated
with a conditional probability table that contains probability of each form of the
variable with every possible state of its parent states. Causal Bayesian Network
(CBN) [16] is upgraded from BN with autonomous causal relations with the do
operator by Pearl [25]. Several works have been done to learn and adopt the
CBN for developing causal models. A theoretical study has been carried out by
Eberhardth et al. on lower bound of worst case for the number of experiments
to be done to recover causal structures [
Recent works have been done to incorporate ontologies in systems to achieve
explainability. Besnard et al. have proposed ontology-based inference rules for
causal explanation [
The proposed model is an accumulation of several components from di erent concepts i.e., CBR, ontology and causal inference model. The formal de nitions followed by the description of each of the components and their overall collaboration are stated in the following subsections. 3.1
An ontology is a representation of a domain with respect to its entities,
relationship among entities and their attributes which is known as DERA knowledge
representation framework [
To facilitate causation in the system, we propose to de ne causal statement C comprised of two atomic statements, and as ABox connected with the keyword \causes" in the form of \C : causes ". Causal statements will be associated with cases as individual or in a set. For example, a possible causal statement is \C : TT4(Abnormal) causes Hypothyroid(Primary)" which summarises \Primary hypothyroid is caused by abnormal free thyroxine level (TT4)"
Explanation for the solution case will be generated by translating causal statements associated with the similar cases { because of , where is the solution and is the set of probable explanations to the solution. \because of " is used to make statements shorter which can be translated in natural language as \the probable solution is given because of the facts".
owl: Thing
An ontology was built to facilitate explanations of solutions produced using
CBR. Thyroid disease dataset [19] was used to build the concepts of the
ontology. This dataset was created by the Garvan Institute, Sydney, Australia which
is now available at UCI Machine Learning Repository2. The thyroid disease
dataset contains 3772 instances with 26 attributes each. The attributes
represent information of diagnosed patients for hypothyroid i.e. age, sex, ongoing
medications, disease history, values for di erent diagnosis tests for determining
levels of hormones { TSH, triiodothyronine (T3), thyroxine (TT4), thyroxine
utilization rates (T4U) and free thyroxine index (FTI). Instances of the dataset
are labelled with four classes: primary hypothyroid, compensated hypothyroid,
secondary hypothyroid and negative. Concepts and relations in the ontology
were developed and validated in respect of two expert curated medical
ontologies built by Shen et al. [23] and Can et al. [
Hormone class. { hasValueLevel: Represents the relationship between HormoneTest class and TestValueLevel class.
Data properties of the developed ontology were de ned afterwards which represent values for various entities in the ontology. Set A holds these data properties in equation de ned for DERA in section 3.1. The data properties of the hypothyroid ontology are described brie y below.
{ hasAge: Represents the age of a patient. { hasGender: Represents the gender of a patient. { hasTestDefinition: Represents the information of a diagnostic test. { hasTestName: Represents the name of a diagnostic test. { hasTestValue: Represents the value of a diagnostic test. { hasValue: Represents the reference values of a diagnostic test. { lowerLimit: Represents the regular lower limit of a diagnostic test. { upperLimit: Represents the regular upper limit of a diagnostic test.
Finally, the individuals were added to the ontology based on the experimental data from Thyroid Disease dataset. Most prominent way to store an ontology is to use web ontology language (OWL) or resource description format (RDF). In this work, the built ontology is stored using RDF since it facilitates to hold the inferred axioms of the ontology whereas OWL holds the de ned axioms only. 3.3
Cases in the case base are stored in the form of XML les. This method is
adopted from the work of Barua et al. [
: </solution> <causations> <causation id=1> <cause>Instance_1</cause> <effect>Instance_2</effect> </causation> : </causations> <similar_cases> <similar_case>case_id_1</similar_case> <similar_case>case_id_2</similar_case> : </similar_cases> The proposed causal model is incorporated within the naive CBR architecture. Figure 3 illustrates the working mechanism of the model. This model is described based on the basic four steps of a CBR system: i) retrieve, ii) reuse, iii) revise and iv) retain [21]. The rst step of the mechanism is to represent the new problem into de ned format of the old cases. After the new case is built, following steps are followed to generate solution with probable explanations.
Retrieve: Similar cases were retrieved from the case base using similarity function developed based on k-nearest neighbour (k-NN) algorithm associated with voting from the most similar cases. Detailed evaluation of the accuracy of retrieval model is discussed in section 4.
Reuse: In this step, the new solution was formed from the similar cases. To represent the causal statement associated with the solution, the inference engine combined all the relevant statements to the new case and prepared for the translation into explanations. For example, if three similar cases were found with two di erent solutions and . According to the de nition in section 3.1, the representation of the causal statements { C ;1 : causes , C ;2 : causes and C ;1 : causes .
Causal inference engine combines these causal statements into a single statement with union operator { C : because of f ; g t because of f g. Revise: The solved case from the previous step containing a solution with explanation translated from the associated causal statements by the explanation
New Case Learned Case n it a e
R Tested Case
Similarity Function
fs Retrieve Previous Cases Case Base Ontology
Revise Confirmed Solution
Explanation Translator
New
Case Retrieved Cases Causal Inference Engine Solved Case Suggested Solution e s u e R translator. Brief discussion on the explanation translator is discussed in the section 4. Finally, the suggested solution was manually curated by an expert on the basis of domain expertise to produce tested case.
Retain: By further generating explanation from modi ed suggested solution, con rmed solution was developed. This solution was inserted in the case base for future reference. In some cases, there were requirements to modify facts in the ontology, this was also done in this step. 4
To retrieve similar cases from the case base, k-NN was used. The value of k was determined after tuning the model parameters using grid search over di erent values of k. Moreover, 5-fold cross validation for all the values were done and k = 3 produced the highest mean cross-validation accuracy. Table 1 shows the accuracy for 5-fold cross validation with di erent values of k.
After retrieving similar cases from case base, explanation for the solution case was deduced by inferring associated causal statements of the solved cases. For better understanding of the mechanism, consider the following example demonstrating a translation mechanism to generate explanation from causal statements.
Figure 4 illustrates the hierarchy and causation of the concerned entities in this example. The path from the nodes representing individual causes to the verdict node leads to an explanation for the predicted verdict produced from CBR. From the generic statements C1;1, C2;1 and C2;2 it is found that, Pregnancy causes low TSH, Low TSH causes secondary hypothyroidism and Low TT4 causes hypothyroidism respectively. Finally, the probable explanation is extracted with prede ned natural language for each of the object properties as, \Pregnancy causes low TSH. Low TSH with low TT4 causes secondary hypothyroidism. Therefore, Pregnancy can be an explanation of secondary hypothyroidism. " (1) (2) (3)
The rapidly growing nature of AI applications in the aspects of capabilities and performance leverages the need of upgrading the intelligence. People seek more humanly intelligence from the machines. To be speci c, explanations are expected more often in addition to the result of any task. This work represents a framework for adding causation to CBR architecture using domain speci c ontology on hypothyroid disease diagnosis. Thyroid disease dataset was used to develop the ontology and causal model. The outcome of this work would have been more credible if there were instances of all types of hypothyroidism in the selected dataset though some evaluation method was applied to justify the use of k-NN in case retrieval step. Progressive works are still on to make this system more robust and test with more feasible datasets. However, this causal model was developed in a generalised fashion that can be adopted to any domain by using a domain speci c ontology and tuning several parameters in explanation translator which enables the model with adaptability. Moreover, representing cases in XML les can make provision to use this type of case-bases in distributed architecture which would contribute to overcoming challenges of being scalable. 10. Hitchcock, C.: Causal models. The Stanford encyclopedia of philosophy (2008) 11. Huang, B., Zhang, K., Sanchez-Romero, R., Ramsey, J., Glymour, M., Glymour, C.: Diagnosis of autism spectrum disorder by causal in uence strength learned from resting-state fmri data. arXiv preprint arXiv:1902.10073 (2019) 12. Ishak, M.B., Leray, P., Amor, N.B.: Ontology-based generation of object oriented bayesian networks. In: CEUR Workshop Proceedings. vol. 818, pp. 9{17 (2011) 13. Kolodner, J.L.: An introduction to case-based reasoning. Arti cial Intelligence Review 6(1), 3{34 (Mar 1992). https://doi.org/10.1007/BF00155578, https://doi. org/10.1007/BF00155578 14. Lamy, J.B., Sekar, B., Guezennec, G., Bouaud, J., Seroussi, B.: Explainable arti cial intelligence for breast cancer: A visual case-based reasoning approach. Arti cial intelligence in medicine 94, 42{53 (2019) 15. Mitchell, T.M.: Machine Learning. McGraw-Hill, Inc., New York, NY, USA, 1 edn.
(1997) 16. Pearl, J.: Causality. Cambridge University Press, New York, 2nd edn. (2009) 17. Pearl, J.: Probabilistic Reasoning in Intelligent Systems: Networks of Plausible
Inference. Elsevier (2014) 18. Pearl, J.: The Seven Tools of Causal Inference with Re ections on Machine Learning. Forthcoming, Communications of Association for Computing Machinery 1(1) (2018) 19. Radwan, E., Assiri, A.M.A.: Thyroid Diagnosis based Technique on Rough Sets with Modi ed Similarity Relation. International Journal of Advanced Computer Science and Applications 4(10), 120{126 (2013) 20. Raghu, V.K., Zhao, W., Pu, J., Leader, J.K., Wang, R., Herman, J., Yuan, J.M., Benos, P.V., Wilson, D.O.: Feasibility of lung cancer prediction from low-dose ct scan and smoking factors using causal models. Thorax pp. thoraxjnl{2018 (2019) 21. Richter, M.M., Weber, R.O.: Basic CBR Elements, pp. 17{40. Springer Berlin
Heidelberg, Berlin, Heidelberg (2013) 22. Sene, A., Kamsu-Foguem, B., Rumeau, P.: Telemedicine framework using casebased reasoning with evidences. Computer Methods and Programs in Biomedicine 121(1), 21{35 (2015). https://doi.org/10.1016/j.cmpb.2015.04.012, http://dx. doi.org/10.1016/j.cmpb.2015.04.012 23. Shen, Y., Colloc, J., Jacquet-Andrieu, A., Lei, K.: Emerging medical informatics with case-based reasoning for aiding clinical decision in multiagent system. Journal of Biomedical Informatics 56, 307{317 (2015). https://doi.org/10.1016/j.jbi.2015.06.012, http://dx.doi.org/10.1016/j. jbi.2015.06.012 24. Tong, S., Koller, D.: Active learning for structure in bayesian networks. In: International Joint Conference on Arti cial Intelligence. vol. 17, pp. 863{869. Citeseer (2001) 25. Tucci, R.R.: Introduction to Judea Pearl's Do-Calculus. arXiv preprint arXiv:1305.5506 (2013), http://arxiv.org/abs/1305.5506 26. Wang, H.T., Tansel, A.U.: Composite ontology-based medical diagnosis decision support system framework. Communications of the IIMA 13(2), 4 (2013)