Lecture 8 Biomedical Decision Making: Reasoning and Decision Support

Holzinger, Andreas

doi:10.1007/978-3-319-04528-3_8

Andreas Holzinger²

2256 Accesses
2 Citations

Abstract

At the end of the eighth lecture, you:

Download chapter PDF

Keywords

These keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.

1 Learning Goals

At the end of the eighth lecture, you:

would be able to apply your knowledge gained in Lecture 7 to some example systems of decision support.
would have an overview about the core principles and architecture of decision support systems.
would be familiar with the certainty factors as for example used in MYCIN.
would be aware of some design principles of DSS.
would have seen the similarities between DSS and KDD on the example of computational methods in cancer detection.
would have seen CBR systems.

2 Advance Organizer

Case-based reasoning (CBR):: Process of solving new problems based on the solutions of similar past problems
Certainty factor model (CF):: A method for managing uncertainty in rule-based systems
CLARION:: Connectionist Learning with Adaptive Rule Induction ON-line (CLARION) is a cognitive architecture that incorporates the distinction between implicit and explicit processes and focuses on capturing the interaction between these two types of processes. By focusing on this distinction, CLARION has been used to simulate several tasks in cognitive psychology and social psychology. CLARION has also been used to implement intelligent systems in artificial intelligence applications
Clinical decision support (CDS):: Process for enhancing health-related decisions and actions with pertinent, organized clinical knowledge and patient information to improve health delivery
Clinical decision support system (CDSS):: Expert system that provides support to certain reasoning tasks, in the context of a clinical decision
Collective intelligence:: Shared group (symbolic) intelligence, emerging from cooperation/competition of many individuals, e.g., for consensus decision making
Crowd sourcing:: A combination of “crowd” and “outsourcing” coined by Jeff Howe (2006), and describes a distributed problem-solving model; example for crowd sourcing is a public software beta-test
Decision making:: Central cognitive process in every medical activity, resulting in the selection of a final choice of action out of several alternatives
Decision support system (DSS):: Is an IS including knowledge-based systems to interactively support decision-making activities, i.e., making data useful
DXplain:: A DSS from the Harvard Medical School, to assist making a diagnosis (clinical consultation), and also as an instructional instrument (education); provides a description of diseases, etiology, pathology, prognosis, and up to ten references for each disease
Expert system:: Emulates the decision-making processes of a human expert to solve complex problems
GAMUTS in radiology:: Computer-supported list of common/uncommon differential diagnoses
ILIAD:: Medical expert system, developed by the University of Utah, used as a teaching and testing tool for medical students in problem solving. Fields include pediatrics, internal medicine, oncology, infectious diseases, gynecology, pulmonology, etc.
MYCIN:: One of the early medical expert systems (Shortliffe 1970, Stanford) to identify bacteria causing severe infections, such as bacteremia and meningitis, and to recommend antibiotics, with the dosage adjusted for patient’s body weight
Reasoning:: Cognitive (thought) processes involved in making medical decisions (clinical reasoning, medical problem solving, diagnostic reasoning)

3 Acronyms

DSS:: Decision support system
CBR:: Case-based reasoning
CDS:: Clinical decision support
CF:: Certainty factor
KDD:: Knowledge discovery from data
LISP:: LISt processing (programming language)
KE:: Knowledge engineering
mRNA:: Messenger ribonucleic acid

4 Key Problems

Slide 8-1: Key Challenges

Health care Information Networks (HINs) help professionals and patients access the right information at the right time and invite a new design and integration of decision support systems within these collaborative workflow processes. The need to share information and knowledge is increasing (e.g., shared records, professional guidelines, prescriptions, care protocols, public health information, health care networks, etc.). The well-established “Evidence-Based Medicine” (EBM) and “Patient-centered medicine” paradigms representing different visions of medicine are suggesting behaviors, so different that they are also raising dilemmas. Attempts made to standardize care are potentially ignoring the heterogeneity of the patients (Fieschi et al. 2003).

Challenges in the development of DSS

The development of medical expert systems is very difficult—as medicine is an extremely complex application domain—dealing most of the time with weakly structured data and probable information (Holzinger 2012).

Some challenges (Majumder and Bhattacharya 2000) are as follows:

(a) Defining general system architectures in terms of generic tasks such as diagnosis, therapy planning and monitoring to be executed for (b) medical reasoning in (a); (c) patient management with (d) minimum uncertainty. Other challenges include (e) knowledge acquisition and encoding; (f) human–computer interface and interaction (HCI); and (g) system integration into existing clinical environments, e.g., the enterprise hospital information system; to mention only a few.

In the previous lecture we have got an overview about some fundamentals of decision making from the human factors perspective; now we will have a closer look on technological solutions. We follow the definition of Shortliffe (2011) and define a medical DSS as any computer program designed to support health professionals in their daily decision making processes (Shortliffe 2011).

In a funny cartoon a physician offers a second opinion from his computer. The patient looks horrified: How absurd to think that a computer could have better judgment than a human doctor! But computer tools can already provide valuable information to help human doctors make better decisions. And there is good reason to wish such tools were broadly available.

5 Decision Support Systems

Slide 8-2: Two Types of Decision: Diagnosis Versus Therapy

In the previous lecture we have got an overview about some fundamentals of decision making from the human factors perspective; now we will have a closer look on technological solutions. We follow the definition of Shortliffe (2011) and define a medical DSS as any computer program designed to support health professionals in their daily decision making processes. Dealing with data in the health care process is often accompanied by making decisions. According to Bemmel and Musen (1997) we may determine two types of decision:

Type 1: Decisions related to the diagnosis, i.e., computers are used to assist in diagnosing a disease on the basis of the individual patient data. They include the following questions:

(a)
What is the probability that this patient has a myocardial infarction on the basis of given data (patient history, ECG)?
(b)
What is the probability that this patient has acute appendicitis, given the signs and symptoms concerning abdominal pain?

Type 2: Decisions related to therapy, i.e., computers are used to select the best therapy on the basis of clinical evidence, e.g.,

(c)
What is the best therapy for patients of age x and risks y, if an obstruction of more than z % is seen in the left coronary artery?
(d)
What amount of insulin should be prescribed for a patient during the next 5 days, given the blood sugar levels and the amount of insulin taken during the recent weeks?

For both types we need medical knowledge. On the basis of the available knowledge we can develop decision models on the basis of the available patient data.

5.1 Decision Models

Slide 8-3: Taxonomy of Decision Support Models

In this slide we see that decision models can be grouped into two main categories:

1.
Quantitative: based on formal statistical methods to test the probability of the occurrence of an event, e.g., to test that the probability for “healthy” is higher than that for a certain disease as we have seen in differential diagnostics.
2.
Qualitative: relying on symbolic methods, rather than following a strictly formal mathematical basis. Such models are inspired by insights on human reasoning, thus often called heuristics, and perform deductions on symbolic models using logical operations to conclude a diagnosis based on a case model. According to Van Bemmel we should avoid the word heuristics and use the term symbolic, because such methods may be composed of elementary two-class, single-feature decision units from the first category, e.g., E = “ x > L ”_.

5.2 Evolution of DSS

Slide 8-4: History of DSS is a History of Artificial Intelligence

In the early 1950s decision trees and truth tables were used, followed by systems based on statistical methods, finally followed by expert systems. The history of DSS is very closely related to artificial intelligence (AI), the roots can be traced back to attempts to automate chess play. A famous sample was a fake: the Mechanical Turk (see slide, below, left). Built in 1770 by Wolfgang von Kempelen (1734–1804), the device appeared to be able to play against a human, as well as perform the knight’s tour, which requires moving a knight to visit every square of a chessboard only once. The “real” start of AI research was in 1955, when John McCarthy coined the term AI and defined it as the science and engineering of making intelligent machines. Edward Feigenbaum was one of first to construct an artificial expert and while looking for an appropriate field of expertise, he met Joshua Lederberg, the Nobel laureate biochemist, who suggested that organic chemists need assistance in determining the molecular structure of chemical compounds (Rheingold 1985).

Slide 8-5: Evolution of Decision Support Systems

In 1965 Feigenbaum, Lederberg and Buchanan began work on DENDRAL (see top root in the slide), a procedure for non-redundantly enumerating topologically distinct arrangements of any given set of atoms, consistent with the rules of chemical valence (Lindsay et al. 1993). Conventional systems had failed to support organic chemists in forecasting molecular structures. Human chemists know that the possible structure of any chemical compound depends on a number of rules about how different atoms can be bound to one another; and many facts about different atoms in known compounds. By discovering a previously unknown compound, they can gather evidence about the compound by analyzing it with a mass spectroscope, which provides a lot of data, but no clues to what it all means. Look at the slide (Shortliffe and Buchanan 1984): DENDRAL was followed by MYCIN; and actually MYCIN was the inspiration for many other systems.

Slide 8-6: Early Knowledge-Based System Architecture

DENDRAL was well known to computational chemists who have incorporated many parts of it in their own software. Although it does no longer exist today, it had a major impact on a newly developed field:

Knowledge engineering (KE), which is both science and engineering of Knowledge-based systems (KBS) and applies methods from artificial intelligence, data mining, expert systems, decision support systems, and mathematical logic, as well as cognitive science. A great amount of work is spent in observing human experts and the design of models of their expertise.

One of the first spinoffs from DENDRAL was Meta-DENDRAL, an expert system for people whose expertise lies in building expert systems. By separating the inference engine from the body of factual knowledge, Buchanan was able to produce a tool for expert-system builders. In this slide we see the first architecture of the basic principle of any expert systems, consisting of a knowledge base, an inference engine, and a dedicated user interface to support the HCI process (Shortliffe and Davis 1975).

Slide 8-7: Static Knowledge Versus Dynamic Knowledge

MYCIN was programmed in Lisp and used judgmental rules with associated elements of uncertainty. It was designed to identify bacteria causing severe infections (bacteremia, meningitis), and to recommend antibiotics, with the dosage adjusted for the patient’s body weight. Edward Shortliffe, both a physician and computer scientist was confronted with problems associated with diagnosing a certain class of brain infections that was an appropriate area for expert system research and an area of particularly importance, because the first 24 h are most critical for the patients. In the slide we see the idea of the separation of static knowledge (the rules and facts) and dynamic knowledge (the entries made by the human user and deductions made by the system). This is the principle of rule-based systems (Shortliffe and Buchanan 1984).

Slide 8-8: Dealing with Uncertainty in the Real World

We are already well aware about the notion of probable information. The problem is that classical logic permits only exact reasoning: IF A is true THEN A is non-false and IF B is false THEN B is non-true—however, most of our real-world problems do not provide this exact information, mostly is inexact, incomplete, uncertain, noisy, and/or unmeasurable. This is a big problem in the biomedical area.

Slide 8-9: MYCIN: Rule-Based System: Certainty Factors

Shortliffe was aware of the problems involved with classic logic and introduced the certainty factor (CF) which is a number between −1 and +1 that reflects the degree of belief in a hypothesis.

Positive CF’s indicate evidence that the hypothesis is valid. If CF = 1, the hypothesis is known to be correct (and contrary for CF = −1). If CF = 0, there is either no evidence regarding the hypothesis or the supporting evidence is equally balanced, suggesting that the hypothesis is not true. MYCIN’s hypotheses are statements regarding values of clinical parameters for the various nodes in the context tree.

Let us look on an original example in the next slide.

Slide 8-10: Original Example from MYCIN

This MYCIN example makes the Certainty Factor CF clear (Shortliffe and Buchanan 1984).

Slide 8-11: MYCIN Was Not a Success in the Clinical Practice

MYCIN was not a success in the clinical practice; however, it was a pioneering work for practically each following system; for example ONCOCIN evolved from this work and assisted physicians in managing complex drug regimens for treating cancer patients. It has been built on the results of the MYCIN experiments while gaining experience with regular clinical use of an advice system for use by physicians. The work has also been influenced by data regarding features that may be mandatory if decision support tools are to be accepted by clinicians. Clinical oncology was selected due to the fact that this medical domain meets many of the criteria that have been identified for building an effective consultation tool using AI techniques (Shortliffe 1986). Up to date, the main architecture of a DSS is the same as that developed in the 1970s.

5.3 Design Principles of DSS

Slide 8-12: Basic Design Principles of a DSS

As we have already heard at the very beginning of this lecture, the development of medical expert systems is very difficult—as medicine is a complex application domain—dealing most of the time with weakly structured data (Holzinger 2012). Problems include (Majumder and Bhattacharya 2000):

(a) defining general system architectures in terms of generic tasks such as diagnosis, therapy planning and monitoring to be executed for (b) medical reasoning in (a); (c) patient management with (d) minimum uncertainty. Other challenges include (e) knowledge acquisition and encoding, (f) human–computer interface and HCI; and (g) system integration into existing clinical environments, e.g., the enterprise hospital information system.

Slide 8-13: Cybernetic Approach to Medical Diagnostics

This slide shows the typical workflow of a medical reasoning system: Abduction, deduction, and induction represent the basic elements of the inference model of medical reasoning. Clinical patient data is used to generate plausible hypotheses, and these are used as start conditions to forecast expected consequences for matching with the state of the patient in order to confirm or reject these hypotheses (Majumder and Bhattacharya 2000).

Slide 8-14: State-of-the-Art Architecture of DSS

Present-day DSS consist of three main components:

1.
Knowledge base, the heart of the system, contains the expert facts, heuristics, judgments, predictions, algorithms, etc., and the relationships—derived from human experts.
2.
Inference engine, examines the status of the knowledge base, and determines the order the inferences are made; it also includes the capability of reasoning in the presence of uncertainty (compare with MYCIN).
3.
User interface enables effective HCI—additionally there are external interfaces providing access to other databases and data sources (Metaxiotis and Psarras 2003).

Slide 8-15: On the Design and Development of DSS

DSS deal with problems based on available knowledge. Some of this knowledge can be extracted using a decision support tool (data mining) which is in fact part of a KDD process (Lecture 6). Data mining tools are usually difficult to exploit because most of the end users are neither experts in computing nor in statistics. It is difficult to develop a KDD system that fits exactly to the end users’ needs. Those difficulties can only be tackled by including end users into DSS development. It is necessary to combine methods from Software Engineering (SE) and HCI. Abed et al. (1991) proposed an approach to combine (1) the Unified Process (UP) from SE and (2) the U model from HCI.

The U model (see next slide) considers those steps which do not exist in traditional SE models.

Slide 8-16: Example: Development Following the U Model 1/2

For the effective use in it is necessary to combine methods from Software Engineering (SE) and HCI.

In this U-model we determine two phases:

1.
A descending phase for specification and human–computer systems design and development.
2.
An ascending phase for the evaluation of the system.

The validation consists of comparing the model of the theoretical tasks specified in the descending phase with the model of the real tasks highlighted in the ascending phase, according to the original principles suggested by Abed et al. (1991). The result of the comparison either validates the system or highlights its deficiencies.

Slide 8-17: Improved U Model 2/2

The final model resulting from the assessment allows a generalization of the end users specific behavior under particular work conditions and context—what traditional often ignore. Ayed et al. (2010) proposed a modified version of the U-model, specifically adapted to DSS and knowledge discovery (KDD):

1.
The analysis of the domain, including the definition of the system objectives, which allows the first functional and structural description of the system to be developed.
2.
The development of the first interface prototypes (models) for the DSS in question, which, by giving future users an idea of the possible solutions, allows them to be implicated as early as possible in the project life cycle.

Slide 8-18: Remember the Similarities Between DSS and KDD

If you look at this slide and compare the DSS process with the KDD (data mining) process, then you will recognize the similarity between decision-making processes and data mining processes.

Slide 8-19: The Design Phases

In this slide we can see the various phases (A to E) of the U-Model approach, which is based on the principle of iterative and incremental development, which allows each task accomplished to be evaluated as soon as the first iterations of the development process have been completed:

A = Requirements analysis (needs capture)

B = Analysis and specification

C = Design and prototyping

D = Implementation

E = Test and evaluation

Be aware of the user-centered design process, which will be discussed in Lecture 12!

5.4 Clinical Guidelines

Slide 8-20: Clinical Guidelines as Decision Support and Quality Measure

Guidelines have to be formalized (transformed from natural language to a logical algorithm) and implemented (using the algorithm to program decision support software which is used in practice). Work on formalization has focused on narrative guidelines, which describe a process of care with branching decisions unfolding over time (Medlock et al. 2011). Systematic guidelines have potential to improve the quality of patient care.

Quality. The demand for increased quality assurance has led to increased interest in performance indicators and other quality metrics. In order for the quality of care to improve as a result of these measures, they must be linked to a process of care. For example, a rule such as “80 % of diabetic patients should have an HbA1c below 7.0” could be linked to processes such as: “All diabetic patients should have an annual HbA1c test” and “Patients with values over 7.0 should be rechecked within 2 months.” These measure quality and performance at the population level, but in order to improve the quality of care, action is required at the patient level.

Condition-action rules specify one or a few conditions which are linked to a specific action, in contrast to narrative guidelines which describe a series of branching or iterative decisions unfolding over time. Narrative guidelines and clinical rules are two ends of a continuum of clinical care standards.

Clinical rules represent elementary, isolated care recommendations, while narrative guidelines describe a coherent, unified care process.

Slide 8-21: Clinical Guidelines

Most work in developing computer-interpretable guidelines has focused on the difficult problem of formalizing the time-oriented structure of guidelines.

Medlock et al. (2011) propose the Logical Elements Rule Method (LERM), although presented linearly in the text, in practice some steps may be done in parallel, as shown in this slide. Some steps, such as extracting data elements or checking for conflicts between rules, may need to be repeated with the results of later steps as input.

Slide 8-22: Example Exon Arrays

Progress in genomics has increased the data available for conducting expression analysis, used in transcriptomics. This can be very helpful for decision support. It deals with the study of mRNA and the extraction of information contained in the genes. This is reflected in the exon arrays requiring techniques to extract information. This slide shows the correlation of two probe intensities—among 11 tissues (breast, cerebellum, heart, kidney, liver, muscle, pancreas, prostate, spleen, testes, and thyroid): The black boxes represent exons; grey boxes represent introns; (b) Probe design of Exon arrays. Four probes target each putative exon; below: The top color bar indicates the probe annotation type, core probes (red), extended probes (blue), full probes (yellow). The signal intensities of core probes tend to have high correlation (top right corner of the heatmap) (Kapur et al. 2007).

Corchado et al. (2009) provided a tool based on a mixture of experts model which allows the analysis of the information contained in the exon arrays, from which automatic classifications for decision support in diagnoses of leukemia patients can be made. The proposed model integrates several cooperative algorithms characterized for their efficiency for data processing, filtering, classification and knowledge extraction. This is a mixture of expert tools that integrates different cognitive and statistical approaches to deal with the analysis of exon arrays.

Slide 8-23: Computational Leukemia Cancer Detection 1/6

Exon arrays as seen in Slide 8-22 are chips which allow for a large number of data to be analyzed and classified for each patient (six million features per array). The high dimensionality of data makes it impossible to use standard techniques for expression array analysis (which contain approximately 50,000 probes).

High dimensionality of data from each exon array implies problems in handling and processing, thus making it necessary to improve each of the steps of expression array analysis in order to obtain an efficient method of classification. An expression analysis basically consists of three steps:

1.
Normalization and filtering
2.
Clustering and classification
3.
Extraction of knowledge

These steps can be automated and included within an expert system. Since the problem at hand deals with high dimensional arrays, it is important to have a very good preprocessing technique that can facilitate automatic decision making with regard to selecting the most vitally important variables for the classification process. In light of these decisions, it will be possible to reduce the set of original data. After the organization of groups, patients can be classified and assigned into the group with which they share the most similarities. Finally, an extraction of knowledge system facilitates the interpretation of the results obtained after the preprocessing and classification steps, thus making it possible to learn from the information acquired from the results. The process of extracting knowledge shapes the knowledge obtained into a set of rules that can be used for improving new classifications.

In this slide we see such an exon array structure: (1) Exon–intron structure of a gene. Grey boxes represent introns, rest represent exons. Introns are not drawn to scale. (2) Probe design of exon arrays. Four probes target each putative exon. (3) Probe design of 3′ expression arrays. Probe target the 3′ end of mRNA sequence (Corchado et al. 2009).

Slide 8-24: Computational Leukemia Cancer Detection 2/6

The proposed model by Corchado et al. (2009) incorporates the mixture of three experts in sequential form, having the advantage of integrating different techniques, considered to be optimal for using in the stages of the expression analysis for the problem of classifying leukemia patients. Techniques that offer good results in each phase are combined and the model considers the characteristics of each expert in order to achieve an appropriate integration. The structure of the modules can be seen in Fig. 8-15, the steps include:

1.
Preprocessing and filtering
2.
Clustering
3.
Extraction of knowledge
4.
Information representation

The different modules work independently, to facilitate the modification of any of the proposed experts, or to incorporate new techniques (including new experts). This affects the expert of a single module, while the others remain unchanged. This allows a generalization and making it possible to select the expert best suited to apply in each particular problem.

The initial problem description is composed of all the individuals D = {d₁, … d_t} together with the n probes. The first expert preprocesses and filters the probes, reducing the set of probes to s elements but maintaining the t individuals. The second expert executes the clustering, creates r groups and assigns the new individual (t + 1) to one of these groups. The third expert explains how the individual elements have been classified into groups by means of a knowledge extraction technique, and by obtaining a graphical representation (a tree). The final module represents the probability of assigning individuals to each of the groups depending on the probes selected, taking into account the knowledge extracted (Corchado et al. 2009).

Slide 8-25: Computational Leukemia Cancer Detection 3/6

This slide shows the classification performed for patients from groups CLL and ALL. The X axis represents the probes used in the classification and the Y axis represents the individuals. Above we can see, represented in black, most of the people of the CLL group are together, coinciding with the previous classification given by the experts. Only a small portion of the individuals departed from the initial classification. Below we see the classification obtained for the ALL patients. It can be seen that, although the ranking is not bad, the proportion of individuals misclassified is higher. Groups that have fewer individuals have a high classification error.

Classification obtained for (a) ALL patients and (b) CLL patients. Each of the values obtained correspond to the fluorescence intensity for an individual. At the bottom of the image is shown the fluorescence scale of values; the lowest level is 2 (blue), while the highest is 12 (red) (for interpretation of this images in color please refer to the original article (Corchdo et al. Bajo 2009)).

Slide 8-26: Computational Leukemia Cancer Detection 4/6

Following the decision tree shown in this slide, the patients were assigned to the expected groups. Only one of the patients was assigned to a different group by both methods. The healthy patients were eliminated in order to proceed with the classification.

The values of the leaf nodes represent the predicted group and the number of elements assigned to each of the groups following the order (ALL, AML, CLL, CML, NOL, MDS). The rest of the nodes represent the probe and the fuzzy value to compare the individual to classify. If the condition is true, then the branch on the left is selected, otherwise, the branch on the right is selected. The tree helps to obtain an explanation of the reason why an individual has been assigned to a group.

Slide 8-27: Computational Leukemia Cancer Detection 5/6

The work of Corchado et al. (2009) demonstrates a model of experts that uses exon arrays to perform an automatic diagnosis of cancer patients. The system incorporates experts at each phase of the microarray analysis, a process that is capable of extracting knowledge from diagnoses that have already been performed, and that has been used to increase the efficiency of new diagnoses. The model combines:

1.
Methods to reduce the dimensionality of the original set of data.
2.
Preprocessing and data filtering techniques.
3.
A clustering method to classify patients.
4.
Modern extraction of knowledge techniques.

Slide 8-28: Computational Leukemia Cancer Detection 6/6

The system of Corchado et al. (2009) works in a way that is similar to how human specialist teams work in a lab, is also capable of working with big data and making decisions automatically and reduces the time needed for making predictions. The main advantage of this model is the ability to work with exon array data 0; very few tools are capable of working with this type of data because of the high dimensionality. The proposed model resolves this problem by using a technique that detects the importance of the genes for the classification of the diseases by analyzing the available data. For the time being, three experts have been designed, one for each phase of the model.

6 Case-Based Reasoning

Slide 8-29: Thinking–Reasoning–Deciding–Acting

Note: Always remember that Thinking–Reasoning–Decision–Action is intrinsically tied together. A good primer for clinical thinking is Alfaro-LeFevre (2013).

Slide 8-30: Case-Based Reasoning (CBR) Basic principle

CBR is a problem-solving paradigm, different from other AI approaches. Instead of relying solely on general knowledge of a problem domain, or making associations along generalized relationships between problem descriptors and conclusions, CBR is able to utilize the specific knowledge of previously experienced, concrete problem situations (cases). A new problem is solved by finding a similar past case, and reusing it in the new problem situation. A second important difference is that CBR also is an approach to incremental, sustained learning, since a new experience is retained each time a problem has been solved, making it immediately available for future problems. The description of a problem defines a new case. This new case is used to RETRIEVE a case from the collection of previous cases. The retrieved case is combined with the new case—through REUSE—into a solved case, i.e., a proposed solution to the initial problem. Through the REVISE process this solution is tested for success, e.g., by being applied to the real world environment or evaluated by a teacher, and repaired if failed. During RETAIN, useful experience is retained for future reuse, and the case base is updated by a new learned case, or by modification of some existing cases (Aamodt and Plaza 1994).

Slide 8-31: The Task-Method Decomposition of CBR

In this slide we see the task-method structure: Tasks have node names in bold letters, while methods are written in italics. The links between task nodes (plain lines) are task decompositions, i.e., part-of relations, where the direction of the relationship is downwards. The top-level task is problem solving and learning from experience and the method to accomplish the task is CBR (indicated in a special way by a stippled arrow). This splits the top-level task into the four major CBR tasks corresponding to the four processes: retrieve, reuse, revise, and retain. All four tasks are necessary in order to perform the top-level task. The relation between tasks and methods (stippled lines) identify alternative methods applicable for solving a task. A method specifies the algorithm that identifies and controls the execution of subtasks, and accesses and utilizes the knowledge and information needed to do this (Aamodt and Plaza 1994).

Slide 8-32: CBR Example: Radiotherapy Planning 1/6

Example: Radiotherapy planning for cancer treatment is a computationally complex problem. An example from Petrovic et al. (2011) shall demonstrate it: Prostate cancer is generally treated in two phases. In phase I, both the prostate and the surrounding area, where the cancer has spread to, will be irradiated, while in phase II only the prostate will be irradiated. The total dose prescribed by the oncologist is usually in the range of 70–76 Gy, while the dose ranges in phases I and II of the treatment are 46–64 Gy and 16–24 Gy, respectively. The dose is delivered in fractions, each fraction being usually 2 Gy.

Slide 8-33: CBR Example: Radiotherapy Planning 2/6

In this slide we see the workflow of radiotherapy: (1). CT scanning, (2) Tumor localization, (3) Skin reference marks, (4) Treatment planning, (5) Virtual simulation, (6) Radiotherapy treatment.

Slide 8-34: CBR Example: Radiotherapy Planning 3/6

The patient is first examined and then CT scans or MRI is carried out. Thereafter, the generated scans are passed onto the planning department. In the planning department, first, the tumor volume and the organs at risk are outlined by the medical physicist so that the region that contains the tumor can be distinguished from other parts that are likely to contain microscopic (tiny) tumor cells. Afterwards, the medical physicist in consultation with the oncologist defines the planning parameters including the number of beams to be used in the radiation, the angle between beams, the number of wedges,^{Footnote 1} the wedge angles and generates a Distribution Volume Histogram (DVH) diagram for both phases I and II of the treatment. DVH presents the simulated radiation distribution within a volume of interest which would result from a proposed radiation treatment plan. The next task is to decide the dose in phases I and II of the treatment so that the tumor cells can be killed without impairing the remaining body, particularly the organs lying close to the tumor cells, i.e., rectum and bladder. The organs lying close by should preferably not be impaired at all by the treatment. However, the oncologist usually looks for a compromise of distributing the inevitable dose among the organs. Rectum is a more sensitive organ compared to the bladder and is the primary concern of oncologists while deciding the dose plan. There is a maximum dose limit for different volume percentages of the rectum, and it has to be respected by oncologists when prescribing a dose plan. In certain cases, this condition may be sacrificed to some extent so that an adequate dose can be imparted to the cancer cells. Oncologists generally use three groups of parameters to generate a good plan for each patient. The first group of parameters is related to the stage of cancer. It includes Clinical Stage (a labelling system), Gleason Score evaluates the grade of prostate cancer and is a integer between 1 and 10), and Prostate Specific Antigen (PSA) value between 1 and 40. The second group of parameters is related to the potential risk to the rectum (degree of radiation received by different volume percentages of the rectum. It includes the DVH of the rectum for Phases I and II at 66, 50, 25, and 10 % of the rectum volume. Example: the DVH states that 66 % of the rectum will receive 50 % of radiation. It means that if the dose prescribed by the oncologist in the phase I of the treatment is 60 Gy, then the amount of radiation received by 66 % of the rectum is 30 Gy. The final PSA value is a parameter related to the success rate of the patient after the treatment.

Slide 8-35: CBR System Architecture 4/6

In the system developed by Petrovic et al. (2011), the cases which are similar to the new case are retrieved using a fuzzy similarity measure. A modified Dempster–Shafer theory is applied to fuse the information from the retrieved cases and generate a solution as shown in this slide.

1.
The clinical stage of the cancer is of ordinal type and can be divided in seven different categories T1a, T1b, T1c, T2a, T2b, T3a, T3b.
2.
The value of the Gleason Score is an integer number from [1, 10] interval.
3.
PSA is a real numbers from [1, 40].
4.
DVH is a real number between [0, 1].

In order to use features of different data type, measurement units and scale together in the similarity measure we need to normalize them. However, it would not be easy to define a preferably linear mapping in the [0, 1] interval. Instead, we define fuzzy sets low, medium, and high for each feature. They are normalized fuzzy sets whose membership functions take value from [0, 1] interval. In addition, fuzzy sets enable expression of preference of the oncologist. An example of membership functions of fuzzy sets low, medium and high Gleason score is given in Figure Slide 8-36.

Slide 8-36: Membership Function of Fuzzy Sets in Gleason Score 5/6

The parameters of these membership functions are set in collaboration with the oncologist. Each attribute l (Gleason score (l = 1), PSA (l = 2)) of case cp is represented by a triplet (vpl1, vpl2, vpl3), where vplm, m = 1, 2, 3 are membership degrees of attribute l in the corresponding fuzzy sets low (m = 1), medium (m = 2), and high (m = 3).

Slide 8-37: Case-Based Reasoning 6/6

This final slide demonstrates the adaptation mechanism. In this example, the final outcome of the Dempster–Shafer theory is a dose plan having 62 Gy and 10 Gy of radiation in phases I and II of treatment, respectively. This is not a feasible dose plan because the dose received by 10 % of the rectum is 56.2 Gy which is larger than the prescribed maximum dose limit (55 Gy). Hence, in order to generate a feasible dose plan, the repair mechanism is performed. The dose corresponding to the phase II of the treatment is decreased by 2 Gy, which leads to the new dose plan 62 and 8 Gy, which is a feasible dose plan.

Note: The Dempster–Shafer theory (DST) is a mathematical theory of evidence and allows the combination of evidence from different sources resulting in a degree of belief (represented by a belief function) that takes into account all the available evidence (Zadeh 1986).

7 Future Outlook

Slide 8-38: Future Outlook

Modern approaches, as for example the IBM Watson Technologies (Holzinger et al. 2013) may considerably assist future medical doctors. Following the hypothesis that medicine is a data problem; the answer to your problem is “out there” in the masses of data, but literally, a human has little chance to find it. Hence, we need machine intelligence to help to harness these large amounts of data and to provide decision support for the clinician. A long debate is ongoing whether such technological approaches will replace medical doctors. To answer such a question we can remember that the same question was asked in the 1960s, when computer was first used for teaching and learning purposes—and up to now, still human teachers are in the classroom …

8 Exam Questions

8.1 Yes/No Decision Questions

Please check the following sentences and decide whether the sentence is true = YES; or false = NO; for each correct answer you will be awarded 2 credit points.

8.2 Multiple Choice Questions (MCQ)

The following questions are composed of two parts: the stem, which identifies the question or problem and a set of alternatives which can contain 0, 1, 2, 3, or 4 correct answers, along with a number of distractors that might be plausible—but are incorrect. Please select the correct answers by ticking ☒ - and do not forget that it can be none. Each question will be awarded 4 points only if everything is correct.

8.3 Free Recall Block

Please follow the instructions below. At each question you will be assigned the credit points indicated if your option is correct (partial points may be given).

9 Answers

9.1 Answers to the Yes/No Questions