A Grand Unified Theory of Rehabilitation (We Wish!). The 57th John Stanley Coulter Memorial Lecture

Definitions 

A theory can be defined as “a plausible or scientifically acceptable general principle or body of principles offered to explain phenomena.”3 A theory systematizes a set of underlying laws or processes into a unifying causal structure, just as a theorem about triangles systematizes laws about parallel lines, or complementary angles, to explain the general properties of triangles.4 Theories can exist on a spectrum from “macro” theories, which seek to explain a wide range of phenomena in the world (eg, the laws of thermodynamics), to more “micro” theories, which seek to provide a mechanistic explanation for very specific phenomena (eg, the relationship between stimulation of cellular receptors and the rise in production of cyclic adenosine monophosphate).

Many theories are formally stated as such, but whenever a scientist or clinician makes a prediction of some outcome, based on a set of observations, they are engaging, at least implicitly, in theorizing. For example, if a clinician plans a set of strengthening exercises for a patient with limited ambulation, he or she is at least implicitly theorizing that (1) there is a lawful relationship between strength and ambulation ability and (2) the set of prescribed exercises will cause a change in strength. When pressed, the clinician may be able to further elaborate these theories by specifying the attributes of the specific patient that suggest that weakness is at the heart of the ambulation deficit and the nature of the exercise program that is key to improving strength.

A hypothesis can be defined as “a tentative assumption made in order to draw out and test its logical or empirical consequences.”5 That is, a hypothesis is a statement of the implications of a particular theory with respect to specific predictions about a phenomenon of interest. The hypothesis may be quite tentative, stated primarily as a thought experiment to allow the investigator to consider the specific implications of a theory and to spot its shortcomings and potential theoretical alternatives, or it may be stated as a clear prediction of an outcome that is ready for an empirical test.

A theoretical model (or model for short) is a “description or analogy used to help visualize something (as an atom) that cannot be directly observed.”6 As such, a model deliberately simplifies and ignores causal factors that are small compared with those that are large, in a tradeoff favoring ease of use and conceptual clarity over comprehensiveness. Consider, for example, Newtonian mechanical models that leave out friction in making predictions about the interactions of force, mass, and acceleration. The model from the outset is known to be “false” yet to be useful in providing approximations to many real-world phenomena.4 A model really is a theory, but one that is “put to work” by being specified in a form in which actual computations about outcomes can be made. Thus, we could have a general theory of Newtonian mechanics that says that “as the force applied to a body increases, so will its acceleration” or we could have a theoretical model, F=M×A, that allows us to compute the expected acceleration given the force and the mass. Because of their inherent and intentional simplification, it makes no sense to ask whether a model is “true.” Models are always “false” in the absolute sense, in overlooking real-world details. Rather, the test of a theoretical model is how well it approximates the phenomena of interest in the real world.

Taxonomy is the study of the general principles of scientific classification.7 Thus, the main job of a taxonomic system is to group phenomena or observations into categories that are objective, mutually exclusive, and useful in scientific inquiry. There is no specific requirement that a taxonomic system be based in theory. However, it is often the case that the categories that are useful in a taxonomic system are carved along lines that have theoretical relevance. Consider 2 familiar taxonomies from other areas of science. The periodic table of elements seeks to define elements in terms of their atomic number (number of protons) and atomic weight (protons plus neutrons) (fig 1). To do its job, it need only specify a one-to-one correspondence between the atomic number and the identity of the element (ie, it specifies that atomic number is the basis of chemical classification). Yet the row and column structure of the periodic table shows that elements in the same column also have strikingly similar chemical characteristics (eg, 1 column contains all of the inert gases). This in turn implies that if a new element of a given atomic number is discovered, some of its chemical characteristics can be predicted based on its location in the periodic table.

View Large Image Download to PowerPoint

Fig 1. The periodic table of the elements. Reprinted with permission from the Lawrence Berkeley National Laboratory.18

The Linnean phylogenetic tree is another familiar scientific taxonomy that identifies the relationships among plant and animal species (fig 2). Again, the requirement is that this taxonomy specifies the boundaries between distinct species and organizes the evolutionary relationships among species appropriately. This classification system was originally based on phenotypic characteristics of member species—number and types of leaves for plants, presence or absence of a spine for animals, etc. However, with the development of modern molecular genetics, these classifications have also assumed theoretical relevance by predicting that species whose “branches” are close together on the tree should have greater overlap in their genetic codes than those that are more distantly related. Conversely, analyses of organisms’ genetic sequences have largely supported the original structure of the phylogenetic tree while also pointing to occasional branches that need to be relocated based on deoxyribonucleic acid analyses as opposed to structural similarities.8

View Large Image Download to PowerPoint

Fig 2. The phylogenetic tree. Image by CSS Inc, for the National Biological Information Infrastructure.

The Impact of a Theoretical Orientation 

Having examined definitions of some of the key terms in theory development, it is worth considering the contributions that a theoretical orientation can make to scientific progress in general and to progress in rehabilitation more specifically.

I have argued previously that treatment theories can be very powerful tools for shaping treatment research.9 A treatment theory specifies the (hypothesized) active ingredients of a treatment and, related to this, the mechanism(s) by which the treatment is expected to produce its effect. Clearly, treatment theory helps very directly to define a treatment, in that it suggests that all activities that carry those active ingredients should be members of a particular class of treatments, and that activities that lack those ingredients are members of a different class. But treatment theory also supports other aspects of treatment research. It helps to shape appropriate inclusion and exclusion criteria for a treatment efficacy study by suggesting subgroups of people who would and would not be expected to respond to the treatment. For example, if the mechanism of a particular rehabilitation intervention is the conscious learning of a new compensatory skill, then it stands to reason that people with very severe memory impairments that interfere with conscious learning may not be appropriate treatment candidates. Treatment theory also suggests outcome measures that are appropriate for measuring the effects of treatment by pointing out areas where change is and is not expected based on the treatment mechanism. In a related example, if the mechanism of a treatment for moderate memory impairment is, again, a learned compensatory strategy, such as making lists of to-be-remembered items, then it would be illogical to use a standardized memory test (where making lists is considered “cheating”!) as a measure of treatment success.

But theory has a much larger role in scientific research than simply constraining choices in clinical trials. According to Isaacson, “Einstein had a good feel for experimental findings, and he used this knowledge to find certain fixed points upon which he could construct a theory. But his emphasis was primarily on the deductive approach.”2(p117) Thus, Einstein required a few key empirical facts as starting points for his theories, but most of his work consisted of logically connecting these few “fixed points” with thought experiments. His process of theory development never involved a “systematic review” of all the empirical data in a given domain in the hopes of sifting through it for meaningful patterns. Indeed, there were others who were frequently more knowledgeable about the details of a domain than he was. One might even argue that his ability to develop far-reaching theories rested in part on not being encumbered by too many empirical results. Einstein himself said that “The deeper we penetrate and the more extensive our theories become, the less empirical knowledge is needed to determine those theories.”2(p118) Isaacson summarized this ability by noting that “Einstein’s great strength as a theorist was that he had a keener ability than other scientists to come up with what he called ‘the general postulates and principles which serve as the starting point.’”2(p351) By their nature, theories are anchored by a set of fixed points but have implications far beyond those anchoring points. That is, theories, once framed, have logical extensions that invite new directions in empirical research that would not have been apparent from the anchoring facts alone.

Two Case Examples 

Let us consider 2 theoretical frameworks, one more micro and one more macro, that have implications for rehabilitation research. First is a connectionist model of object naming, promulgated by Schwartz et al.10, 11 Connectionist models are a general class of computational models that have emerged from the intersection of artificial intelligence and neuroscience. These models loosely simulate neural processes by positing a layer of input nodes, analogous to sensory and perceptual processes, that register and code incoming information; a layer of output nodes that represent the possible responses of the system; and an intermediate or “hidden” layer that performs computations that map patterns of input to patterns of output. Each input node is connected to each node in the hidden layer, and each node in the hidden layer is connected to each node in the output layer (fig 3). These connections are analogous to synapses. When nodes in the input layer are stimulated, that stimulation spreads to all of the intermediate nodes to which they are connected (and on to the response nodes in a similar fashion) through a process referred to as “spreading activation,” and the degree of downstream activation is controlled by the strength of connections between nodes. Although connections among nodes all start at the same connection strength, or “weight,” if the network is shown a set of inputs and given feedback on the correct outputs, it “learns” to alter the connection strengths to improve its performance. Brain lesions are simulated by weakening the connection strengths so that the network is less efficient and makes more errors.

View Large Image Download to PowerPoint

Fig 3. A computational model of object naming.10, 19 Reprinted by permission of Myrna Schwartz.

In the object-naming task model, when learning is assumed to be complete, the input nodes consist of detectors of particular semantic features; the intermediate layer corresponds to the “lemma” level—that is, the conceptual basis of words but without their phonological (sound) attributes; and the output layer consists of specific phonemes that can be joined together to utter the word. When the network is “shown” pictures of animals, the semantic features of those animals (4 legs, fur, size, etc) activate corresponding input nodes for those features. The activation of these nodes spreads to the lemma layer in proportion to the connection strengths between the activated semantic features and corresponding lemmas. At this stage there is also a mechanism to resolve competition among similarly activated lemmas, such that 1 lemma is ultimately “selected.” The activation of this lemma, in turn, propagates activation to corresponding phonemes, and a phonological output string is given as a “response,” with a second mechanism to resolve competition among similarly activated phonemes. Over its learning history, the network has learned that 1 set of semantic features should result in the response “c_a_t,” whereas another overlapping set of features should result in the response “d_o_g.”

The performance of the network, in terms of error rate and type, depends on the connection strength between the layers, but weakening the connections between semantic and lemma nodes has somewhat different effects (predominantly semantic errors) than weakening the connections between lemma and phoneme nodes (predominantly phonologic errors). Overall, when the network is severely lesioned (fig 4), it tends to produce a large number of errors, most of which are nonwords—that is, a set of unrelated phonemes. As the connection weights increase, the number of errors decreases, but the character of the remaining errors changes so that an increasingly large proportion are semantic errors—that is, real words that share semantic features with the target picture. With further increases in connection weights, errors are reduced to near zero. However, this overall severity pattern is moderated by the relative impairment of the semantic versus phonological connections.

View Large Image Download to PowerPoint

Fig 4. The relationship between connection strength and error frequency and type. Reprinted by permission of Myrna Schwartz.

The performance of the model can be assessed initially with data from actual object-naming performance. Indeed, data regarding aphasic subjects’ overall error rates, as well as the proportion of errors of different types, show a very good fit to the predictions of the model (fig 5). Although this theoretical model is anchored by only a few key assumptions from object naming, it makes predictions that go far beyond the starting assumptions. For example, it predicts that if healthy speakers perform object naming under speeded conditions that induce errors, these errors should be predominantly semantic in nature. It also predicts that patients recovering from Wernicke’s aphasia should evolve into an anomic profile—a recovery profile not well accounted for by “syndrome” models of aphasia. Both predictions of the model have been confirmed with empirical data. The model also has some interesting potential implications for treatment research. That is, it allows one to assess the relative severity of semantic and phonological “lesions” (ie, connection strength parameters) across patients whose current performance profiles may not be obviously semantic or phonological. Thus, to the extent that treatments for these 2 types of deficits differ, modeling the relative contributions of the 2 deficits may help guide rational treatment— a prediction of the model that can be subjected to empirical testing.

View Large Image Download to PowerPoint

Fig 5. The accuracy of the computational model as assessed by comparing the predicted and observed error patterns in a group of aphasic subjects.10 Abbreviations: Mn, mean; RMSD, root mean square deviation (the difference between observed and predicted error rates averaged across patients); VAF, total variance accounted for. Reprinted by permission of Myrna Schwartz.

Goal-directed behavior change offers another view into the utility of theory in rehabilitation. Across many domains of behavior, from psychotherapy to industrial production, it has been shown that explicitly setting performance goals has a powerful impact on behavior change.12 More specifically, goals that are neither currently achievable nor “way out of research” appear to have the biggest positive impact on performance. This can be operationalized, for example, by setting performance goals that are 20% higher than prior performance.12 Furthermore, the process of goal setting matters. Typically, goals chosen by the individual are more powerful than goals provided by another. In addition, relatively simple tasks seem to benefit from explicit performance goals, whereas more complex tasks benefit more from “learning goals” (eg, to discover the best strategy for accomplishing the task). Goals appear to have a directive function (promoting attention to goal-relevant stimuli over irrelevant stimuli), an energizing and persistence function (promoting ongoing effort toward performance improvement), and a discovery function (leading to the identification of goal-relevant knowledge and skills).

Like the connectionist theoretical models, this theoretical framework suggests research directions that might not have been apparent in its absence. For example, people with executive deficits that reduce self-awareness, verbal fluency, and the ability to retrieve information in open-ended formats may have difficulty identifying and articulating personally relevant goals. Similarly, once a goal is articulated, people with working memory deficits or “goal neglect” may be able to articulate a goal but not to hold it in mind over time and use it to drive ongoing behavior.13 Thus, we might predict that people with these deficits would benefit less from a goal-setting strategy. We might also predict that interventions that can enhance performance in these particular areas would, in turn, allow such people to benefit more from goal setting. In line with this, subjects with goal neglect who are trained to associate a randomly delivered tone with asking themselves, “What am I supposed to be doing?” appear to improve performance on self-directed tasks even though the tone contains no task-relevant information.14 Whether combining this type of intervention with a more self-selected goal-setting strategy would further enhance its impact remains to be seen.

A Unified Theory of Rehabilitation? 

Einstein sought a unified field theory that would account for such disparate phenomena as gravity, acceleration, and electromagnetism. As Isaacson noted, “He was disquieted when there were two seemingly unrelated theories for the same observable phenomenon.”2(p147) Indeed, in his lifetime, Einstein succeeded in unifying theories of gravity, acceleration, and the passage of time. He never did, however, find a unified theory to incorporate electromagnetism, and we have not succeeded in doing so to this day. Thus, there is no a priori way to know how many theories we need to account for the phenomena we observe. We can seek to combine and reduce, and if we succeed we know it, but if we fail, we know only that we have not yet succeeded.

Rehabilitation is such a broad discipline that it is worth asking how many theories we need to guide our research and practice. Is there a single overarching theory to guide rehabilitation? The place to start our search is the International Classification of Functioning (ICF), promulgated by the World Health Organization (WHO).15 The ICF is fundamentally a taxonomic system of human function. As such, it attempts to define unambiguously the domains of body structure and function, activities, and participation, as well as to define personal factors and environmental factors as additional influences on function (fig 6). It also hypothesizes interrelationships among these domains, such that changes in body structure and function may be predicted to impact activities, and activities to impact participation, etc. Like the other taxonomic systems we have considered, it has hints of theory—in this case a theory of the enablement and disablement process. That is, it does not simply tell us how to define activities, participation, and environment; it also says that the relationship between activities and participation can be modified by the environment, in the sense that wheelchair users (who lack the activity of ambulation) will participate differently in society as a function of the physical environment.

View Large Image Download to PowerPoint

Fig 6. The ICF model of human functioning. Adapted from WHO.20

Although the ICF contains the seeds of a unified theory, it is far from a theoretical model in the sense of being able to compute people’s level of societal participation from data about their body structure and function and their environment. What would it take to achieve a theoretical model of the enablement and disablement process? First, instead of a single box representing each concept, we would need a very large number of boxes to account individually for the structure and function of the leg extensor muscles, the shoulder joint, the frontal lobes, etc, not to mention a myriad of specific activities and a large set of different types of societal participation. That is, computing a person’s ambulation ability must surely be done from specific body functions (such as strength, balance, proprioception) that are particularly pertinent to ambulation. Second, it appears unlikely that quantitative predictions of function will be derived from simple linear combinations of these variables. For one thing, there are probably many nonlinear effects. For example, it is unlikely that leg strength has a linear relationship with ambulation ability, with some strength threshold required before any ambulation is possible, improvement in ambulation with increases in strength in the midrange, and no further improvements at strength levels above some ceiling.16 There are also almost certainly interactions among variables. For example, a person with a memory impairment but good problem-solving skills and self-awareness may function well in most activities but a similar person who also has difficulties in these other areas may fail to compensate for the memory difficulties. Finally, measuring the environmental inputs to participation is particularly challenging because it requires measuring a wide range of physical and social attributes in all of the locations that a person may encounter. It must also take into consideration that the environment does not have uniform facilitating or hindering effects across impairments: curb cuts are helpful to people in wheelchairs but not to those with visual impairments.

As daunting as these challenges are, they are achievable in principle, through concentrated programs of research. The most realistic route to progress is likely to be for different research groups to build specific pieces of the overall model (eg, a model of the body structures and functions that affect ambulation and the aspects of the environment that interact with them) and then to stitch them together over time into a larger framework. Doing this is an important and worthwhile task, but it will result in a unified theory of enablement and disablement, not a unified theory of rehabilitation.

A Theory of Rehabilitation 

The ICF, if fully elaborated into a theoretical model, will tell us what will happen to a person’s ambulation if we change his/her level of strength, but it will not tell us how to change that person’s level of strength. That is, a theory of enablement/disablement tells us how changes in 1 part of the model will be felt in other parts of the model, but it gives us no tools to make those changes. A theory of rehabilitation, in contrast, must address change because it is an applied science that seeks to enhance human function and participation, not merely to predict it.

One might assume that understanding how a system works automatically tells us how to fix it if it is broken, but this is by no means universal. Consider the domain of cognitive neuropsychology. We have relatively sophisticated theories of psycholinguistics, attention control, memory function, etc, that are able to predict many aspects of normative performance in those domains. In some cases, we also have well-developed theories of the nature of dysfunction in those domains in the context of brain damage. But none of these theories speaks directly to how to fix the broken system. Indeed, an examination of many of the attempts at neuropsychologic rehabilitation suggests that the “repair tools” were predominantly imported from behavioral psychology in the form of repeated trials with positive reinforcement, shaping, chaining, and successive approximations. Because these learning tools themselves were built around normal animal and human learning, it remains somewhat uncertain even at present whether the rules of “relearning” in a damaged system are similar to the rules of learning in an intact system. Thus, a science of rehabilitation needs a theory of how change is achieved, and then the elaborated ICF model can tell us how far-reaching that change will be.

Can we hope for a unified theory of rehabilitation (ie, change)? Einstein might be disappointed in us, but it is self-evident that no single theoretical framework can account for changes in organ structure and function, changes in activity performance, and changes in the social and physical environment. Rather, we will need to seek different tools for change in these different domains. Although a myriad of theories may ultimately prove relevant to rehabilitation, several classes of theory appear particularly relevant.

Relevant Theoretical Perspectives 

At the level of body structure and function, demand-based adaptation is a pervasive phenomenon. Tissues stretch and lengthen when subjected to prolonged and graded force; muscles get stronger when subjected to graded work demands; cardiovascular fitness improves in response to graded endurance demands, etc. Although the biochemical signals underlying these different adaptations differ, it is worth asking whether the nature of the demands that are optimal for promoting adaptation have any commonalities. For example, what is the therapeutically optimal rate of increase in demand across these differing domains?

Although demand-based adaptation is common, it is not universal. For example, current evidence suggests that placing increasing demands on the memory systems of amnesic patients does not “strengthen” their memory capacity.17 If there were no underlying capacity limits affecting demand-based change, then the only difference between talented and untalented musicians, athletes, and scholars would be the amount of demand they were subjected to in the past. Thus, theories of demand-based change are likely to have many useful applications, but we should not assume that they can be applied equally well in all arenas.

A range of learning theories can provide useful tools for change at the activity level, because resumption of many activities in the context of impairments involves developing adaptive skills. Thus, theories of classical and operant conditioning, declarative and procedural memory, and practice-based skill learning have broad applications for rehabilitation treatment. In the case of purely physical disabilities, these theories can probably be applied in a rather direct way. For people with cognitive impairment, however, we need to examine how damage to the “learning system” may alter the applicability of these learning theories. As discussed earlier, neural network models that are lesioned and then retaught may provide some useful insights.

A related and important set of theories pertains to motivation. Some of these theories address the body structures and functions related to motivation (eg, models of the control of arousal and of the interaction of frontolimbic structures in developing and maintaining intentions that drive behavior), whereas others address the issue at a more psychologic level (eg, theories of goal setting, stages of change theories, and theories related to self-management). Collectively these frameworks can help us explore the factors that determine the degree to which people with disabilities undertake the often-challenging work required for optimal adaptation. Thus, some of these theories will be aimed more at what the ICF currently calls “personal factors.”

Change at the person-environment interface requires still different theoretical frameworks, because these changes require collective action of social groups or political systems. Theories of the diffusion of innovation have important implications for how environmental changes that are supportive of people with disabilities are or are not widely adopted.15 These theories attempt to account for the common observation that advances as varied as improved farming techniques, evidence-based medical practice, and predictions about climate change are often available in the “expert community” long before they are adopted in general practice, and the importance of a charismatic and respected champion (consider Al Gore’s effect on the visibility of climate change, a phenomenon that was already widely known) in inducing widespread adoption of novel practices. A variety of political, social, and economic theories can also be applied to understand the forces that support and hinder needed environmental changes.

Conclusions 

I have argued that theories, theoretical models, and taxonomies play crucial roles in the development of all scientific domains. They do this not just by organizing current knowledge into useable forms but also by driving new research to explore unanticipated implications of these theories, once conceived. Critically, this new research does not simply add more facts to the domain but supports, refutes, or revises the theory that spawned it.

At present, many areas of rehabilitation are underdeveloped from a theoretical perspective. Thus, the field should invest energy in developing a body of well-articulated theories and examining the applicability of theories from other domains, in parallel with energy directed toward empirical research. We need not only to articulate relevant theories but also to refine them into theoretical models that can make specific predictions that can be empirically validated or refuted. As discussed, much work will be required to develop a theoretical model of the enablement and disablement process, based on the principles of the ICF. But we must also develop and test the applicability of theories addressing change at each of the ICF levels. Taken together, these theoretical developments can advance scientifically based treatment research, as well as help us predict the impact of those treatments on the larger enablement and disablement process. We cannot expect a unified theory to account for all of these effects, but we can seek, as Einstein did, to apply the minimal set of theoretical frameworks required to encompass the phenomena of importance to the rehabilitation process.

How should this work proceed? Given the size and varied nature of the task, it will need to proceed in parallel on many fronts, with effort made to link these separate efforts as progress is made. We will need work on macro theories related to the enablement and disablement process, and micro theories applied to change in specific domains. Can this large and complex task be directed by a central master plan? Probably no more effectively than China’s “Great Leap Forward” was able to centrally plan an efficient system of agriculture. At present, we may have to accept a somewhat chaotic work plan that is bolstered by frequent reexamination and efforts made to promote cross-talk among researchers exploring different pieces of the puzzle.

Most importantly, we need to acknowledge that empirical work alone will not develop the science of rehabilitation. Treatment trials are complex, challenging, and expensive. Surely it is critical that the investment of such resources produce more than a yes-or-no answer about the benefit of a particular treatment. Einstein had an enormous impact on theoretical physics from his armchair. We need to support the theoretical development of our field as energetically as we support the gathering of evidence, and we need to encourage ongoing conversation between theorists and empiricists in the development of our scientific foundations.