This paper gives an overview of my PhD research that aims to support the clinical decision making scenario by providing a tool that allows documenting and structuring clinical discussions in all the steps of the decision making process, by means of argumentation-based techniques. The goal is to highlight contradicting opinions, to recognize invalid reasoning steps, to discover hidden assumptions, or to identify missing evidences. Indeed, in a clinical discussion, participants make important pieces of knowledge explicit, by presenting different opinions, providing evidences that support hypotheses, and possibly arguing about clinical guidelines. However, this knowledge is often lost after the meeting is closed, since only the final decisions are written in patient records. The approach proposed in this paper aims at overcoming these limitations.
Multidisciplinary clinical discussions are becoming a routine activity in hospitals and
healthcare structures in general. Clinical discussions provide the forum for specialists
of various medical disciplines to focus on critical cases, debate about the diagnostic
hypotheses, therapeutic protocols or follow-up of patient conditions, and to devise the
most appropriate treatment. Unfortunately, as witnessed by many physicians we have
interviewed, most of this knowledge just exists during the discussion and is lost after
the meeting is closed. Only the final decisions that determine the specific actions to
perform - such as further examinations, surgical operations, or therapeutic treatments
are reported in medical records [
My research aims to investigate how we can support the decision making scenario by providing a tool that allows documenting and structuring a clinical discussion in all the steps of the decision process, and monitoring and revising the decisions. In this context, the exigency arises to design an interactive computer-based tool to help physi
cians manage clinical discussions effectively and efficiently and to support their
decision making activity. This tool could be useful for simplifying the work flow,
improving the quality of healthcare service and reducing costs ; since time costs money,
especially when several medical specialists are involved, it is important to keep the team
meetings as short as possible, without reducing the quality of the discussion. A clinical
discussion support tool should focus on two main objectives:
1. Documentation: it is fundamental to record and document a clinical discussion in an
informal yet well structured way. The representation language adopted should
account for the logical and temporal structure of a discussion, but at the same time, it
should be easily understandable by physicians. Discussion documentation will serve
as a memory support for the following meetings and as a justification record for the
decisions made.
2. Logical analysis: once a discussion has been properly documented, it can then be
interpreted from a logical point of view, on the basis of a set of reasoning patterns
(formalized as argumentation schemes [
On the basis of these objectives, we have decided to design a new interactive software tool, called ArgMED, structured in three components; the first component for the documentation of discussion; the second component for logical analysis of discussion; the third component for the creation of argumentation schemes. The paper is organized as follows: Section 2 presents background and related works; Section 3 introduces our research methodology; Section 4 and 5 describe the proposed approach; finally, Section 6 concludes the paper and outlines future research issues. 2
In order to investigate the research problem of documentation and visualization of the
discussion, and the problem of analyzing the discussion, we have analyzed many
existing formal and informal approaches, and tools for argument visualization and clinical
decision support. Most of interactive tools are based on argumentation theory.
Argumentation theory is a framework for practical and uncertain reasoning viewed as a
process of arguments production and evaluation. Arguments are entities including a
supported conclusion and a set of premises that represent not necessarily deductive reasons
to believe the conclusion itself [
To resolve the visualization problem several tools have been proposed in the
literature. For instance, Van Gelder [
known systems in this research area – like ArgVIS [
In medical domain, a variety of Clinical Decision Support Systems (CDSSs) based
on argumentation theory have been proposed to help medical personnel solve conflicts
and deepen the motivations underlying decisions, such as: REACT [
There are also several informal approaches (mainly based on ethnographic studies)
that have been proposed to document and visualize clinical discussions. Frykholm and
Groth [
The methodology applied in this research is composed of the following steps: 1. analysis of the current state of the art in medical contexts, with respect to: (i) existing approaches and tools for argument and visualization; (ii) existing approaches and tools for clinical decision support; (iii) informal approaches and ethnographic tools; (iv) argumentation theory and argumentation schemes; 2. analysis of real clinical discussions to understand the structure and how clinical dis
cussions take place, and to identify a minimal set of argumentation schemes; 3. proposal of solutions to overcome the limitations of existing proposals, e.g., models
and technology solutions; 4. definition of user requirements and functional requirements of the proposed system; 5. design, prototyping, and implementation of the proposed system; 6. validation of the system through user experiments; 7. refinement through iterations of the system. Proc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org). Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.
The first issue that we have investigated, is the clinical discussion structure. For this
issue, we have analyzed several real cases. A clinical discussion is usually constituted
by a set of sessions (or meetings), taking place in a strict temporal sequence. Each
session is based on a set of facts shared by all participants (for example, the general state
of the patient, the results of clinical tests, the effects of the on-going treatment, etc.
[
Each session evolves in general through three stages: 1. The session begins with the medical staff introducing the patient and providing detailed information about present and earlier diseases, subjective symptoms, and general health status. Objective observations and results of clinical tests are also reported. Such general facts and observations are assumed to be shared by all participants and are not disputable. 2. After the presentation, the participants in the meeting assert their proposals about possible diagnosis or treatment and support them through their own experience, literature cases, or clinical guidelines. During the session, participants may attack the assertions posted by colleagues or may support them. 3. At the end, participants identify a set of acceptable conclusions and then decide for one of them.
In particularly critical cases such as a difficult diagnosis, a rare pathology or a new treatment, several meetings are necessary before a valid and shared conclusion is reached, thus giving rise to an articulated clinical discussion.
The second issue about discussion documentation, is concerned with the acquisition and representation of a clinical discussion in a structured but at the same time intuitive way that can be understood and validated by the participants and serve as memory support for the following meetings. The analysis of real cases and the feedback we directly got from potential users on the represented discussions allowed us to identify some basic requirements to achieve this objective: Representation of a discussion according to the way the discussion has been carried out: physicians require a structured representation of a previous discussion in a way that somewhat adheres to the way the discussion has been carried out. In particular, they do not accept to structure the information according to a predefined scheme if this does not reflect the order in which information has been pointed out. For instance, if a hypothetical diagnosis has been proposed before looking for symptoms, they do not accept a discussion representation where this order is reversed, e.g. presenting the symptoms first and then a diagnosis as a possible cause. Free and no discussion protocol: physicians are not willing to follow any discussion protocol, but they want to feel free to participate in the discussion according to their usual habits. For instance, sometimes they want to point out all of a patient’s sympProc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org). Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.
toms, other times they want to focus on a subset of them to identify a diagnosis, other times they tentatively reason about a diagnosis and look for the corresponding symptoms. No specific informative need: observations, symptoms and results of diagnostic tests are directly collected by the practitioners in charge of the clinical case at hand and more information is generally not necessary. No abstract representation of propositions: physicians adopt a specific medical terminology with a shared meaning, and do not accept to characterize their propositions in abstract ways, e.g. identifying a major premise or distinguishing between data and general rules. No willingness to make argument structure explicit: even though physicians interact by pointing out arguments and counterarguments, they are not willing to make the relevant structure explicit during the discussion, let alone conform to a predefined scheme. Clear and easy language: the language used to document the clinical discussion must
be clear, easy to understand and specific to the medical domain. 4.2
To illustrate our approach, we focus on a literature example concerning a breast cancer
case [
MD1(A1): Given the type of cancer, I believe that we should undertaken a lumpectomy.
MD2 (A2): Agree.
MD1 (A3): I have undertaken a lumpectomy with clear margins and an auxiliary lymph node dissection, with no positive nodes, therefore I believe that there is no need of other treatment.
MD1(A4): No radiotherapy required. Given her age, tumour size, grade, margin status, the local recurrence risk is 1.3% with radiotherapy and 4.5% without radiotherapy, and no benefit in overall survival. The local control benefit is not large enough to justify the offer of treatment.
MD2 (A5): Agree.
MD1 (A6): No chemotherapy required, her 10 year death rate will only reduce from 5% to 4% with chemotherapy and has only a small effect on the local recurrence rate (now 3.2%).
MD3 (A7): I disagree. I think she should have chemotherapy, because while the death rate is similar, her 10year relapse rate will fall from 25% to 11% with chemotherapy, meaning there is more life without cancer.
MD1(A8): Yes. Good point, I had nFotigco.n1s.idTerhede tbhreeraelsatpcseannucmerbecras.sSeh.e should have chemotherapy. 4.3
The part of the tool to be used for tracking and managing clinical discussions has been developed through an iterative approach, including the design of paper-based and interactive prototypes, and various interviews with representative users (students in medicine and physicians). This activity has led to define the terminology to be used in the Proc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org). Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.
system and to understand how to support the creation and modification of a discussion.
The idea is structuring each discussion about a clinical case as a tree diagram,
somewhat resembling the IBIS-like notation of Rationale [
The documentation activity can be performed by a domain expert user e.g. a member
of the medical staff, in real-time during the discussion, or after the discussion, with the
help of a video recording and written notes. Figure 2 shows a screenshot of the
resulting system with an example (breast cancer case). Here, there are one diagnosis node
and 3 hypothesis nodes. Each hypothesis is supported by motivation pro or attacked by
motivation con node. For example the first hypothesis “Should surgery be performed?”
is supported by the motivation “Given the type of cancer, I believe that we should
undertaken a lumpectomy”. In [
After the documentation of the discussion it should then be possible to post-process this
first representation and produce a more logically sound formalization including, for
example, identification of missing data, need for verification activities, existence of
hidden conflicting opinions, etc. In order to investigate this issue, it is useful to regard
clinical discussions as argumentation processes. We have decided to rely on
argumentation schemes [
An argumentation scheme represents a reasoning patterns consisting in the statement of a presumption in favor of a given conclusion. All argumentation schemes share the same basic structure; namely, a name, a set of premises, a conclusion, and a set of critical questions. The conclusion is considered to be true if all premises are true and all critical questions receive a positive answer (are satisfied). Critical questions are indeed intended to challenge the validity of an argument and provide a sieve to make sure that the reasoning pattern is applied in the correct way. Premises, conclusion, and critical questions contain variables that allow the instantiation of the argumentation scheme to a particular case, thus yielding an argument. 5.1
Walton identified twenty-five argumentation schemes in the legal field [
Argumentation schemes must result from a deep knowledge acquisition activity, where knowledge engineers and domain experts work together to define the set of accepted reasoning patterns. To support this task an editor of argumentation schemes (see fig. 4) has been developed, which allows easy manipulation of the components of an argumentation scheme and supports the definition of an argument-relation graph used to define the relations among argumentation schemes.
Argument from Medical Expert Opinion (AMEO) Variables: Physicians <PH1, PH2, PH3, ... >, Domain <DOM>, Assertion <A>.
Premise 1: Physicians <PH1, PH2, PH3, ...> are specialists in domain <DOM>.
Premise 2: Physicians <PH1, PH2, PH3, ...> assert <A>.
Conclusion: <A>.
CQ1: Is <A> inconsistent with other experts’ assertions? CQ2: Is <A> inconsistent with recent studies? CQ3: Is there no evidence that substantiates assertion <A>? CQ4: Is the assertion <A> not in domain <DOM>?
Fig. 3. An example of argumentation scheme for the medical domain.
Proc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org).
Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes.
This volume is published and copyrighted by its editors.
Discussion analysis consists in the iteration of two interrelated activities:
A discussion statement is examined and a suitable argumentation scheme is selected
for its formal representation. The argumentation scheme is then instantiated into an
actual argument by substituting variable terms with the corresponding linguistic
items extracted from the statements being considered. Arguments are stored in a data
base compatible with the Argument Interchange Format (AIFdb) [
When the user creates a new argument by selecting the button “New Argument”, a popup window will be activated (see Fig. 6) asking him/her the argument name, to select the type of argument, the relation type, and related argument (if it exists). There are 3 types of relation: attack, support and neutral.
In our case, considering for example the statement A6 “No chemotherapy required, her 10 year death rate will only reduce from 5% to 4% with chemotherapy and has only a small effect on the local recurrence rate (now 3.2%)”, the user has selected the argumentation scheme “Argument from Medical Expert Opinion” (see Figure 3), with the purpose of instantiating a new argument. To this end, the user has selected the attack type. The new argument attacks the argument related to the hypothesis “Is chemotherapy required?”. After that, the user can instantiate premises by associating them with propositions appearing in the beginning of the discussion, so the variables <PHn> assumes the value “MD1”, <DOM> assumes the value “chemotherapy” and <A> assumes the value “No chemotherapy required, her 10 year death rate will only reduce from 5% to 4% with chemotherapy and has only a small effect on the local recurrence rate (now 3.2%)”. As a consequence, also the conclusion and critical questions are instantiated. Moreover, each premise can be connected by the user to the discussion statements that refer to the premise itself, and the same can be done for the conclusion and critical questions. In the case at hand, the user links the conclusion of the new argument to the statement A6, and the first critical question (“Is "No chemotherapy required, her 10 year death rate will only reduce from 5% to 4% with chemotherapy and has only a small effect on the local recurrence rate (now 3.2%)" inconsistent with other experts’ assertions?”) to the statement A7 (“I disagree. I think she should have chemoProc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org). Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.
therapy, because while the death rate is similar, her 10-year relapse rate will fall from 25% to 11% with chemotherapy, meaning there is more life without cancer”).
After the linkage of the critical question to a discussion statement, the user can answer it with a positive or negative answer. In the case at hand, the first critical question received a positive answer. Then the system, on the basis of the argument-relation graph, instantiates another argumentation scheme (type of “Argument from Medical Expert Opinion”), giving arise to a new argument that attacks the previous one. This means that the new argument (related to the statement A7) attacks the argument related to the statement A6. After that, the user can instantiate the variables of the new argument with suitable values. The variables <PHn> assumes the value “MD3”, <DOM> assumes the value “chemotherapy” and <A> assumes the value “she should have chemotherapy, because while the death rate is similar, her 10-year relapse rate will fall from 25% to 11% with chemotherapy, meaning there is more life without cancer”.
Considering the statement A8 “Yes. Good point, I had not considered the relapse numbers. She should have chemotherapy”, the user may create a new argument. To this end, the user selects the support type. The new argument supports the argument related to the statement A7.
Doing the same thing to all statements will give rise to a set of arguments (Arg1,
Arg2,...,Arg11). On the basis of the set of arguments created, the system is able to
create the argument graph. In case at hand, the resulting graph is shown in Fig. 7. This
graph can be processed by a standard conflict resolution algorithm to determine the set
of arguments (corresponding to physicians’ assertions) that, according to a specific
argumentation semantics [
Acknowledgments. The author wishes to thank his supervisor Giovanni Guida and cosupervisors Daniela Fogli and Massimiliano Giacomin for their support in carrying out this research. Proc. of CHItaly 2015 Doctoral Consortium, Rome (Italy), September 28th 2015 (published at http://ceur-ws.org). Copyright © 2015 for the individual papers by the papers' authors. Copying permitted for private and academic purposes. This volume is published and copyrighted by its editors.