Page 145 - AIH-2-3

Artificial Intelligence in Health Explainable solutions from AI for HSSs

care. To support health professionals as consultants, such learning algorithms have not been able to generate an
1,2
tools (e-consultants) must be based on solid, substantial explanation of their decisions.
knowledge. Machine learning algorithms, including deep Time-consuming and complex systems are designed
learning techniques, provide consultations by being trained for long-term operation and use. However, the terms of
on massive datasets. However, the specificity of medical use change over time, as do the user’s requirements for
information means no datasets are adequate for solving the functionality, interface, and even expertise. In medicine,
problem. An exception exists in simplified prediction or clinical guidelines, which reflect clinical knowledge and
risk assessment tasks, where datasets consist of thousands practice, are subject to change. We refer to the software
of uniform tuples of predictors. For pre-diagnostic tasks, a system for physicians that can update its knowledge as
large number of symptom checkers are provided.
medicine evolves as a “viable” system.
This study showed that for 10 randomly selected
clinical cases with a reliable diagnosis (five from PubMed Integrating advances in AI, software engineering, and
publications and five from the city hospital) and the five knowledge engineering can offer clinicians comprehensive,
most promising symptom checkers, the correct diagnosis medically intelligent systems that can provide credible
was in the top five in 40% of cases, and became the most answers, understand patients’ histories and medical
likely in only two cases. This confirmed that artificial records, and evolve with the knowledge of specialists.
intelligence (AI) services trained on data from one This paper aims to present an architectural principle and
institution work well on simple cases or data from the same a new technology for creating explanatory clinical decision
place. Fortunately, methods based on explicitly presented support systems (ExClDSS) and a tooling environment to
and manageable information are being developed. ensure their evolution.
In medicine, collections of text-based clinical The ontology-oriented production method is proposed,
guidelines for physicians, doctors, and clinicians are the contribution of which can be expressed in five aspects:
constantly being expanded and modified. Formalized (i) Independent (but coordinated) creation and support
clinical guidelines, which a program can interpret, make of each type of architectural component: knowledge
it possible to provide decision support tools with better bases, clinical dictionaries, software reasoners with
chances of influencing clinician behavior than narrative explanations (ontological interpreters), and user
guidelines. The methodology of creating interpretable interface. A single ontology ensures their consistency
clinical guidelines is rightly associated with knowledge- (ii) Cognitive scientists are engaged in the ontology of
based systems, a software system with a particularly knowledge, laying the foundation for forming various
3
important information component, the knowledge base. types of cause-and-effect, spatial, and temporal
This additional architectural component is created with relationships of medical concepts. Therefore, the
the participation of specialists and experts. The methods of knowledge base is close to the real knowledge that
automated recognition of such medical texts are still being doctors are taught at universities
developed and improved. Experts input these texts into (iii) Explanation is formed based on the applied knowledge
electronic databases in a conscious and meaningful way. base (in the format of the “ontology of explanation”
However, this process requires appropriate formalization approved by doctors)
tools. (iv) Evaluation of the created knowledge base with
The integration of explicitly written knowledge with various methods and tools. We evaluate the formal
machine learning allows the vast amount of human completeness and integrity of the knowledge base and
knowledge and the capabilities of machine learning to conduct a monitoring assessment of the current state
be used to achieve previously unattainable performance. and correctness of the replenished collection of cases
Integration can increase the reliability and robustness of (precedents), ensuring a monotonous improvement of
machine learning, facilitate interaction between humans the knowledge base
and machine learning systems, and make system decisions (v) Software reasoners (together with a graphical user
understandable to humans. Knowledge-enabled systems interface [GUI]) do not depend on knowledge and
4
(KES) are applicable to mediate between machine learning reference books. Therefore, they are repeatedly used
algorithms and human users. 5 to lay out many systems with knowledge bases.
To support health professionals as consultants, not only The research’s novelty is as follows. Unlike most existing
is high-quality knowledge important in such systems, but methods of knowledge-based product constructing –
these systems must also rely on approaches that explicitly that assume either the presence of a sufficient set of data
explain their recommendations. However, machine to extract the necessary knowledge from them or the

Volume 2 Issue 3 (2025) 139 doi: 10.36922/aih.5736