Articles :: More Healthcare IT Articles

Email Article
Email a friend

Formal Methods of Decision Making in the Clinical Domain

By Obi Igbokwe
A look at how formal techniques in clinical decision can be used to varying degrees of complexity and help the decision maker to be more systemic, explicit and rational in their decision making.

Clinical decision making is a complex skill that is learned gradually as clinicians deal with patients, observe more experienced clinicians and have discussions about their decisions and the rationale behind the decisions. Medical students, residents, postdoctoral fellows and practitioners are provided with an opportunity through clinical training to make their decisions and review them with others.

The underlying principles of decision making are ordinarily not formalized, which is why clinicians view judgement more art than science and why clinicians often characterized it as an informal, intuitive process.

This formal decision making can be seen as a framework that is explicit and around which can be built what was conceived as intuition. Throughout the development of the principles of a prescriptive formalization of the process of clinical decision (clinical decision analysis), the process of intuitive clinical reasoning has been compared with the procedures of formal making.

There is increasing evidence that formal decision techniques applied sensibly can improve poor decision making. In a study done to evaluate the usefulness and effectiveness of a decision-support system for preoperative staging of prostate cancer, it was found that not only was it more accurate that two of the attending physicians, and all the residents involved in the study, it also improved the residents’ staging accuracy to approximately that of the physicians.

Clinical decision analysis can be referred to as a cluster of formal techniques for the modelling, measurement and evaluation of clinical inputs, processes and outcomes. It can also be simply defined as the modelling of the components of a decision formally into a tree.

Requirements of designing of the decision tree include noting the relevant uncertainties and possible actions. Some of which might have been otherwise overlooked.
The tree not only identifies pieces of information that will be available before a decision must be faced, it also identifies information that need not be collected as it does not affect a decision.

The tree also allows one to concentrate on one part of the problem without losing the total picture and provides a means of integrating in a meaningful way one’s thinking concerning all parts of the problem.
Not only is the process is the process of structuring a clinical decision problem with a decision tree one of the most valuable part of decision analysis, there are also particularly helpful in dealing with typical situations for which routines have not been established by the clinicians.

The tree uses conventional notation, where chances are represented by a square node, chance events by circular nodes and outcomes by rectangles. Thus the tree is made of choice and chance nodes and branches with outcomes states.

In the clinical domain, the problem is structured as a decision tree, where choices, information and preferences are separated. Drawing of the tree can be aided by a computer using the TreeAge software package. The probability of every chance branch is assessed from literature, experienced colleagues and personal experience in that order.

In many cases medical literature might not contain anything relevant to the needed figure and thus the decision maker has to rely on his own as well as that of his colleagues’ subjective assessment.

To make assessments, people tend to rely on several discrete, often unconscious, mental processes known as heuristics. These heuristics have been identified and they include representativesness heuristics, availability heuristics and lastly anchoring and adjustments.

Represenativeness heuristics is where the process, of which the probability of an event was got, is ascertained by how closely its essential features resemble the essential features of the parent population.

For example if the clinician was to judge the probability of malaria by the degree to which the patient’s signs and symptoms closely resembles the clinician’s mental image of patients with malaria. If the patient has all the classical signs and symptoms of malaria, the clinician judges the patient to have malaria.

Availability heuristic is when the probability of the event is judged by the case with which similar events are remembered.

More easily remembered events are judged likely. A clinician who has cared for a patient who had a fever and later died would vividly remember malaria as a cause of fever more easily. The problem here is that the physician would overestimate the occurrence of malaria (a rare disease in temperature regions) in all patients who have a fever.

Anchoring is when the clinician has an initial estimate and then adjusts the estimate further based on further patient information (adjustment). Here the physician makes an initial estimate of the probability being 0.3. After finding that the patient has just come back from Africa, increases the probability because of the high incidence of malaria in Africa, a fact that is easily ascertained.

However probability ascertained in whatever manner can then be adjusted to take account of the new information gained from using diagnostics tests by the calculation of the posttest probability using Bayes’ Theorem. This method makes the implicit use of probability by the clinician explicit. These probabilities can now be used for obtaining the best possible outcome for the patient.

Next the utility of each outcome state is assed. Utilities should preferably elicit the patient’s own individual preferences. There are three techniques for doing this and they rate scale, the standard gamble and time trade off methods. Each of these methods has its own disadvantages but they all allow for the arrival at judgements which produce a better decision than that achievable in any attainable mode.

Rating scale consists of a line on a page with clearly defined end-points, with the most preferred health state placed at one end, the least preferred at another end and the remaining health states lying between them. The intervals or spacing between these states corresponds to the differences in the preferences as perceived by the patient.

For instance, the patient might be asked to choose the best health state of a batch of descriptions concerning things such as restrictions of activity and pain and be asked to locate other states on the scale, such that the distances between the locations are proportional to subject’s differences in preference.

In the time trade-off method, the patient is asked to choose the relative amount of time he would be willing to consider spending in various time states. For instance, in evaluation of a chronic health state such as diabetes, the patient would be offered a choice of remaining in that state for the rest of his life against returning to full health for a shorter period of time.

The amount of time the individual is willing to trade off can be used to obtain a preference value for the chronic health state. The same method can be used to evaluate temporary states.

Thirdly the standard gamble, which is the classical method of measuring preferences, can be used to measure preferences for chronic states preferred to death. Here the patient is offered two alternatives, either of which is the gamble, of a treatment with two possible outcomes (death or return to normal health for the remainder of his life), or the certain outcome of remaining in the chronic state for the remaining of his life.

For example, a patient with a slipped disc is offered the choice of not having an operation and suffering from the slipped disc for the rest of his life (certainty) or having an operation of which there are two possible outcomes (gamble).

He could either make a full recovery or he could be worst off than before he had the operation. The probability of a full recovery is now varied until the patient is indifferent between the gamble and the certainty and the probability got is used to calculate the preference value for the health state which in this case is full recovery.

Possible outcome could be described using Quality-Adjusted Life Years (QALYs). One of the approaches would be to convert poor health years into good health years by asking the patient to indicate the shortest period in good health that he would accept in return for a life in a state of ill health. The patient is asked the question for each outcome.

The final tree decision is now complete with the probability estimates and utility values for each outcome. The next step is to identify the optimal strategy by calculating the weight of each possible outcome and prune off all but the heaviest branch at each node. This process is called ‘folding back’ of the tree. This process is possible to do on computers using the TreeAge software package.

Each utility is multiplied by the probability of the branch and producing the ‘weighted utility’. Similar figures around each outcome for the option are then added to give the expected outcome. The option with the highest expected utility is then known as optimal strategy.

For instance, if a physician is faced with a hospitalized 68 year old with diabetes mellitus, hypertension and cardiovascular disease who suddenly develops a weakness, with unusual sensations and difficulty speaking but all these symptoms resolve within 15 minutes.

The physician believes that the patient has just had a transient ischemic attack and has a 09% risk of having a stroke and if it does happen, it is equally as likely to be major (paralysis and prolonged disability) as it is to be minor (less surely disabling).

One of the options would be treat the patient with anticoagulant medication. This has a probability of causing bleeding and has some chance of death in the patient. If the patient develops no further complication, the risk of having a stroke reduces and the patient is less likely to suffer from a major stroke than a minor stroke, if a stroke does eventually happen.

Another option, which carries greater risk of dying, is surgery but if the patient survives he is less likely to have a stroke than if he was on medication and he is even less likely to suffer a major stroke. Thus both the risk of having a stroke and the stroke being major are less likely with surgery than with medication.

Now a decision has been drawn and the probabilities and utilities (using patient’s preferences) have been worked out. The tree is now folded back and the expected utilities obtained. If the expected utility value for the medication is higher than surgery then medication is chosen as the optimal strategy or vice versa.

After choosing the optimal strategy one would now have to consider if varying either probability or utility would result in another option becoming the optimal strategy.

This is called a sensitivity analysis. For instance, in the above example, if the patient’s preference is changed for a certain outcome of surgery, it might have such an effect that the optimal strategy might be changed to surgery instead of medication or if the probability of dying from surgery is reduced, the same effect might be noticed. Variation of the judgement helps to test the robustness of the conclusions to changes in the date entered.

There is a variant of sensitivity analysis called threshold analysis which tells at what level of each of the probabilities, taken one at time, the decision maker would consider the currently favoured strategy no better than its nearest competitor.

The decision tree works well for problems involving chance events whose one occurrence is over a short period of time horizon. However when the natural history of disease involves repeated events (episodic haemorrhage with anti-coagulation) or over prolonged periods (such as myocardial infarction after coronary artery bypass surgery), the decision becomes “bushy” and the approach becomes cumbersome. Because the utility must be defined on when each good and each deleterious clinical event occurs, the utility structure becomes unmanageable.

The Markov model of prognosis can be used as an alternative to standard formal decision analysis. Thus a decision tree can be replaced outright by Markov model, which can be also be grafted onto standard decision analysis as an equivalent to the utility structure. The Markov model is noted for its simplicity, ease of use in calculating prognosis and its representation of many clinical problems and still be able to remain faithful.

Formal techniques in clinical decision can be used to varying degrees of complexity and help the decision maker to be more systemic, explicit and rational in their decision making.

1. Llewelyn, L.; Hopkins, A. Analysing how we reach clinical decisions. Royal College of Physicians London. 1993
2. Chang, P. L.; Li, Y. C.; Wang, T. M.; Huang, S. T.; Hsieh, M. L.; Tsui, K. H. Evaluation of a decision-support system for preoperative staging of prostate cancer. Med Decis Making. 1999; Oct – Dec 19(4):149-27
3. Weinstein, M. C.; Fienberg H. V.; Elstiein, A. S.; et al. Clinical Decision Analysis. W.B. Saunders Company 1980.
4. Eddy, D. M. Anatomy of a decision. JAMIA. 1990;263: 441-443
5. Lilford, R. J.; Parker, S. G.; Brainholtz, D.A., et al. Decision analysis and the implementation of research findings. BMJ. 1998; 317: 405-09
6. Logan, S.; Gilbert, R. Use of Decision analysis in clinical practice and policy. Evidence-Based Health Policy and Management. 1997
7. Owens, D. K.; Sox H, C. Medical decision making: Probabilistic medical reasoning. Chap. 3 in: Medical Informatics Computer Applications in Health Care. Shortliffe, Edward H. and Perreault, Leslie E. eds. Addison-Wesley Publishing Company, Reading, MA, 1990.
8. Beck, J. R.; Pauker, S.G. The Markov process in medical prognosis. Med Decis Making. 1983; 3(4): 419-458

Copyright © 2018 All Rights Reserved. Contact Us - About Us - Privacy Policy - Terms & Conditions - Resources

Can't find what you are looking for? View our Site Map