Back to List of Publications
To date, most advances in patient monitoring have been associated with the development of new clinical measurements, or improvements in the processing of existing ones. With the development of computational techniques drawn from artificial intelligence, prototype systems have been developed that advance monitoring in a quite different direction. Rather than simply displaying measurements for clinicians to interpret, the goal is now to develop intelligent patient monitors that assist clinicians in the task of interpretation itself [3][23].
It is also possible to envisage devices that close the loop between measurement and treatment. Empowered with the ability to automatically interpret clinical signals, one could construct therapeutic devices that automatically assess a patient's clinical state and then alter the treatment they deliver [10]. Experiments in closed loop control of drug delivery systems [2][21], or ventilators that are capable of automatically altering their settings [9][16] point the way. Ultimately, such technologies may contribute to the development of more robust artificial organs [11] with the ability to sense, intelligently interpret and respond to the varying contexts of the physiologic milieu they are placed in.
The motivations for developing intelligent patient monitors are numerous. The most pressing arise from the difficulties individuals face when they need to monitor and react to continuous data, and are not unique to clinical medicine. They are also an issue, for example, in the design of systems used by airline pilots and nuclear power plant operators. These human factors include the problems of data overload, vigilance, varying expertise, and human error.
It comes as no surprise that clinicians may have difficulty in using the vast amount of information that can be presented to them on current monitoring systems [29]. Not only is the amount of information available greater than can reasonably be assimilated or displayed, but the clinical environment provides sufficient distraction to reduce the effort that can be devoted to signal interpretation. Worse still, current alarm technology floods clinicians with false alarms, providing further unnecessary distraction [19].
The level of expertise that individuals bring to a task like the interpretation of clinical signals varies enormously, and it is not always possible to access more skilled colleagues to remedy such deficits. This frequently leads to errors in diagnosis and selection of treatment. Indeed, the majority of complications associated with anaesthesia, for example, result from inadequate training or insufficient experience of the anaesthetist [6][26].
We can conceptualise several distinct layers in the construction of an intelligent monitoring and control system, spanning the range from raw patient signals to control decisions. Each layer requires a number of different technologies, and for the present there remain research questions associated both with the technologies, and in the ways they are to be combined.
At the signal level, the development of novel transducers and signal processing techniques will continue to provide better ways of measuring physiological state. As important as the development of new sensors is the identification of those sensors that provide the most clinically relevant information. Fewer, more appropriate sensors are preferable to a proliferation of physiological signals that have to be correlated and interpreted before they give value [13].
After physiological signals are generated they need to be validated [15]. Automating the process of signal validation has probably attracted less attention than it deserves, and this has contributed to the proliferation of false alarms from monitoring systems. At present it is often up to the clinician to ascertain whether a measurement accurately reflects a patient's status, or is in error. However, validation is not necessarily a straightforward task. While in many situations, signal error is clear from the clinical context, it can also manifest itself as subtle changes in the shape of a waveform. The development of intelligent alarm systems which automatically validate signals prior to generating alarms remains an important focus for further work [15].
Once validated signals are available, and noise and artefact have been filtered [22][25], pattern recognition technologies come into play, detecting regularities within the signal to suggest particular clinical conditions. There are various approaches to this problem, ranging from traditional blackboard models which offer a syntactic approach to recognition [20], statistical approaches like hidden Markov models, Kalman filtering [25], to neural networks. Each of these techniques provides a method of matching signal to symbol.
At the next level we move beyond the signal to symbol layer, and perform inferences with the symbols extracted from clinical data. Central to this activity is the notion of model-based reasoning [5][27], where pathophysiological knowledge is encoded as some form of computational model. The inferences that we derive from these models may be diagnoses, explanations of observed behaviour, predictions about future pathophysiological states, or control actions. The automated diagnosis of clinical conditions can assist clinicians by identifying rare diseases or complex cases. For example, intelligent monitors can track several diagnostic hypotheses over time. By comparing the dynamic behaviour of measured patient variables with the behaviours expected of each hypothesis, they can assist clinicians in narrowing down the range of possible causes [3][24][31].
Finally, there is the task layer, in which the intelligent monitoring system models the activities of the clinician. These models capture knowledge about the needs of and constraints operating upon different individuals, the knowledge they bring to a particular task, and the type of dialogue the system is likely to have with them [4]. By explicitly capturing notions of clinical task we enhance the likelihood that intelligent systems meet clinical need.
At the heart of the challenge to medical AI is its ability to identify and satisfy fundamental clinical needs. There is little advantage in developing a complex system that mimics skills that most clinicians possess and use effectively. Rather, we should attempt to provide support for cognitive functions that clinicians perform poorly. The long lag in the introduction of computerised decision support into medicine is probably as much due to failure on this point as it is to the limitations of current technology. It is because the work in intelligent monitoring and control aims to satisfy perceived needs of working clinicians that it has a good chance of making a significant and positive contribution to health care. Yet, as we start to refine and apply this set of technologies, there are still important questions about the best directions in which to proceed.
In many ways, work to date in the design of patient monitoring systems has been fragmented. On the one hand, there has been a steady, if small, effort concentrated on the technical aspects of diagnosis, monitoring and control of physiological signals on the part of the medical AI community. On the other hand, there is a large, if diverse, literature on human factors in the clinical workplace, identifying causes of suboptimal performance and suggesting areas in which decision support might be appropriate [8]. Further removed from clinical medicine, there is a large literature on the cognitive aspects of process control [30]. These latter two bodies of work help us to characterise the task layer, and should provide the motivational impetus and focus for the development of intelligent patient monitoring systems. Unfortunately, this does not seem to have happened to any great extent, with work focusing (often necessarily) solely on technical issues. As the technical aspects of the domain start to become clearer, it may now be an appropriate time to undertake a closer examination of the clinical needs that originally motivated the field.
Diagnosis is one of the core abilities of an intelligent monitoring system. If it is to interact autonomously with a physiological system, then the correct identification of patient state must precede any control action. Most diagnostic techniques have been designed to deal with diagnosis in the largely static context of the traditional medical consultation. The complications of interpreting continuously varying physiological systems in real time stretches such existing diagnostic methods considerably.
In this issue, two papers are presented that deal with the challenging problems inherent in the diagnosis of time varying clinical state. Downing [7] presents work on diagnosing abnormalities from continuously varying physiological signals. In particular, he applies formal techniques of consistency based diagnosis to the task. Working in the domain of ventilator management, Uckun et al. [28] focus on the value that qualitative models of pathophysiologic processes bring to diagnosis. Both papers make important contributions in this area.
Independent of the design of autonomous physiological control systems, diagnosis has also been traditionally seen as a key task with which clinicians require assistance. This view was possibly motivated by early research in decision analysis which suggested that human judgements were prone to decision biases [18], and the corollary which stated that such flawed judgements could be normalised by relying on sounder, more formally based diagnostic systems. There is now growing evidence that questions such an assessment. Based upon studies of individuals in the field rather than in controlled laboratory situations, evidence now suggests that the primary effort for decision makers is not at the moment of choice, but rather in situation assessment [12]. In other words, it may well be the case that the majority of clinicians do not have difficulty in making a diagnosis, but rather in establishing a clear picture of the state of the world, and clarifying their goals and assumptions, prior to attempting to make a diagnosis. If it is the case that situation assessment is the key bottleneck in decision making, this has implications for the design of decision support tools for monitoring and control. Systems which assist clinicians in making an assessment of monitored data may be of more utility than systems that attempt to manufacture a diagnosis. Both the diagnosis papers in this issue [7][28] can be seen as attempts to build systems which establish the clinical context within which therapeutic decisions can be made, rather than simply providing diagnoses to clinicians.
Even if clinicians do not routinely need support with patient diagnosis, there is still a case for assistance with machine diagnosis. Here the problem seems to lie in the lack of specialised mechanical knowledge needed to work with complex machines, rather than any inherent difficulty in the logic of the task. For example, close to half of all anaesthetic mishaps associated with human error are related to operating the anaesthetic machine and ancillary equipment. Another smaller but significant portion is related to machine failures [6]. It may be that systems that can self-diagnose faults, or suggest solutions to technical problems during surgery will have real clinical utility.
While much medical AI work has focused on supporting clinicians with diagnosis, less has been said about how we might support their control tasks. This is surprising, since much of the patient care activity in real time domains like intensive care or anaesthesia is associated with control. Clinicians seek to maintain their patients in acceptable physiological states by augmenting delivery rates of drugs, gases and fluids, titrating dose against clinical response. Clinical context dictates such elements as choice of individual therapy or therapy combinations, or the decisions to switch therapy.
The state of research into intelligent control systems is somewhat analogous to that in diagnosis. The focus has largely been in the development of automatic controllers, which relieve clinicians of their control activities by taking them out of the control loop. Such systems remain largely in the proof of concept stage, their introduction being hampered by medico-legal and ethical considerations.
There has been much less interest in developing systems which assist clinicians when they need to remain within the control loop. It may be that supporting such activities has been seen as too risky given the state of decision support technology. However, such a perspective is still clouded by the diagnostic paradigm. It is not at all obvious that we have to make accurate therapy recommendations to build clinically useful control support systems. It may be quite acceptable to assist with the situational assessment aspects of control, assisting clinicians in visualising relevant clinical information, and leaving the final control choices to them. Indeed, there are some decision support paradigms in the medical AI literature that could support such control tasks. For example, critiquing systems that assist clinicians in refining treatment strategies may be useful in optimising patient specific global control strategies [23]. But more could be done.
The remaining two papers in this issue focus on the design of systems to support control. Ash et al. [1] examine one critical constraint for all monitoring and control systems - the need to produce a result in real time. They explore how one might balance the trade-offs between choosing a clinical action that is an optimal solution, and one that is less satisfactory but takes less time to construct. Working in the domain of ventilator management, Rutledge et al. [24] report on the implementation of a system which is designed to assist clinicians in choosing appropriate control settings. Choices are based on the evaluation of the predicted effects of control changes against preferred treatment plans.
Despite the excellent technical work in the mechanics of diagnosis and control, little work to date has explored the modes of interaction possible between intelligent monitoring or control systems and their users. Thus, while much work goes into the design of the computational components for example, little if anything is usually said about the principled design of user interfaces. This needs to be redressed, since the manner in which clinicians interact with intelligent systems will have a great impact on the systems' eventual utility [4].
User interfaces are an integral part of device design. They are integral because interfacing is essentially a task matching exercise, explicitly specifying user tasks at a conceptual level to permit optimal interaction. Without an explicit model of task layer knowledge, the likelihood is that in a complex medical environment, intelligent decision support systems will poorly match the needs of clinical users.
More often than not, this task knowledge has been captured only as an implicit set of design constraints, rather than being given any explicit attention. A detailed examination of the nature of clinical tasks can reveal important information about the type of support most appropriate for them [8]. It should thus be a priority for system designers to explicitly identify the clinical tasks they are servicing, and the manner in which clinicians will successfully use their systems. This by definition requires close work with and understanding of the clinical workplace, and the detailed design of user interfaces and the interactions they permit.
It is very easy to build complex technological artefacts that get in the way of, rather than assist with, the purpose for which they were designed. Paying close attention to user needs and work practices minimises this to a large extent. However, we also need to avoid the introduction of excessive complexity, or oversimplification in the conceptual models that intelligent systems present to clinicians.
Complex decision support systems introduce their own models of operation for the user to absorb, and if these are counterintuitive, imprecise or idiosyncratic they can make system use difficult. Further, any mismatch between the cognitive models used by a clinician and those designed into a decision support system means that the clinician has to alter the way he or she thinks before obtaining the benefit of the system. Although such exhortations for rational design may seem obvious in the light of the considerable work in human-computer interaction, adhering to them is not a trivial task.
In contrast to the introduction of unwanted complexity, oversimplification of conceptual models may also have its own price to pay. Reducing the need for the clinician to carry detailed models of patient state can reduce cognitive load, but may leave them in a worse position if they are caught with an unexpected problem. While a clinician may initially expend effort to build up an accurate picture of a patient's state, the accuracy of all models decays with time [17]. Summary displays, for example, may allow clinicians to keep their existing models superficially updated, but do not necessarily allow them to revise details if a major reconstruction of mental models was needed.
The same applies to autopilot or closed loop control systems where the loss of immediate real time information or out of the loop unfamiliarity is well documented [30]. Faced with a sudden new clinical problem, the amount of effort spent in defining and treating the cause may be greater than if the clinician had initially worked harder in keeping a detailed mental model of patient state up-to-date.
Stated in the most general terms, the use of intelligent monitoring and control systems should not result in their users developing impoverished mental models of the systems they seek to control. These lessons, and no doubt many others, come from the study of people interacting with systems in their natural working environment. Their implications for system design are often profound.
The four articles collected in this special issue of Artificial Intelligence in Medicine address several key issues in the intelligent monitoring and control of dynamic physiological systems. These papers make it abundantly clear that the technical challenges in diagnosis, monitoring and control of dynamic physiological systems are important ones that deserve ongoing effort.
As always, the underlying motivation must be to improve clinical outcome, whether it be through the reduction in error rates, or the optimisation of treatment delivery. Challenges still exist in the definition of the clinical role that such systems should fulfil, and the way in which they will be used by clinicians. It is only through a constant referral back to clinical motivation that one can hope to affect clinical outcome and meet genuine clinical needs. Most importantly of all, addressing technical issues, and defining the role and impact of intelligent systems should not be seen as parallel exercises. The clearer our understanding of the place of intelligent monitoring and control systems in the clinical workplace, the more ably will we set the technical research agenda.