The evolution of cognitive load theory and its application to medical education

Empirical findings supporting CLT principles come from four types of measures: (1) indirect measures of cognitive load through task performance accuracy [9‐11] or time needed for task performance [12, 13]; (2) dual-task performance measures [14, 15]; (3) bio-measures such as functional magnetic resonance imaging (fMRI) [16] or specific electroencephalographic (EEG) [17, 18] or eye-tracking variables [19]; and (4) subjective rating scales [4, 20, 21]. For a more detailed overview of findings of each of these types of measures, the reader is referred to Leppink, Van Gog, Paas, and Sweller [22]; the current review briefly highlights findings that have had a profound impact on CLT’s evolution.

Indirect measures have provided support for the previous assertion that less working memory capacity remains if more capacity is taken up by either extraneous or intrinsic cognitive load. Firstly, one found that students tend to demonstrate more accurate task performance [9, 10] and need less time for task performance [12, 13] if a strategy employed demands less problem-solving search. Especially among novice learners, engaging in problem-solving search contributes to extraneous cognitive load, leaving less room for dealing with intrinsic cognitive load, resulting, in turn, in less accurate task performance or an increase in time needed for task performance. Conversely, another study found that increased error rates in task performance could point to an elevated intrinsic complexity of information [11]. Excessive administration of intrinsic cognitive load, even with extraneous cognitive load being kept to a minimum, resulted in a very high overall cognitive load—and perhaps cognitive overload—with an increase in error rates as a logical consequence.

Further evidence in favour of a limited working memory resources model comes from dual-task studies, in which participants are instructed to simultaneously perform a primary task and a secondary task that is typically unrelated to the primary task. Learners who need to allocate fewer working memory resources to the primary task have more working memory resources available for more accurate or faster performance on the secondary task [14, 15].

Researchers using bio-measures have found, amongst others, increased activity in the dorsolateral prefrontal cortex during information processing [16], and in eye-tracking studies, negative correlations of saccade rate and amplitude with cognitive load [23, 24] and a positive correlation between fixation length and cognitive load [25].

One of the difficulties with particular bio-measures, for instance pupil dilation, is that its relation to cognitive load may vary with age [26]. Further, the approaches hitherto employed are not always practicable in the sense that they are heavily task-related, and mostly require special equipment and an even more careful study design and planning. In an educational setting this is not always feasible, which provides a logical explanation for their infrequent use so far. Subjective rating scales that measure cognitive load experienced by the learner are much easier to use and frequently encountered in the literature. The first scale made its appearance in the early 1990s in the form of a 9-point one-dimensional mental effort rating scale [4]. A more inclusive variation of this scale may be represented by the NASA task load index (TLX) [27], which seeks to capture five dimensions: mental, physical and temporal demands, own performance, and effort and frustration.

Apart from practical challenges, some conceptual challenges appeared. The two-factor framework one had relied on so far, that is, the division of cognitive load into intrinsic and extraneous cognitive load, did not appear to hold when considering that in some cases an increase in cognitive load could bolster learning. It therefore appeared plausible that a third type of cognitive load was involved, which in some sort of way was beneficial to learning. This concept came to be known as germane cognitive load [28]. Incorporating germane cognitive load into the framework, however, did not solve the riddle. On the contrary, it incited the desire to find a way to measure each type of cognitive load separately. So far, the existence of distinct types of cognitive load had been largely theoretical; although attempts to measure cognitive load had been plentiful, none of these had sought to measure each cognitive load separately.

At this point, CLT rested on the assumptions that (1) extraneous cognitive load should be kept to a minimum; and (2) germane cognitive load could arise only if intrinsic cognitive load had reached a specific level [28]. In view of these beliefs, the question arose as to how a third type of cognitive load could be assimilated into a limited working memory, and more importantly, how each of the specified cognitive loads could be quantified.

This unresolved issue sparked several efforts to devise an instrument with which the various types of cognitive load could be measured [29‐33]. Unfortunately, in these studies, single items instead of multiple items were used for one or more types of cognitive load. The use of multiple indicators of the separate types of cognitive load might yield a more precise measurement and might enable researchers to separate types of cognitive load more clearly than the use of a single indicator for each. Besides, none of the studies could address the so-called expertise reversal effect [34, 35]. Succinctly put, instructional support in a learning task that is beneficial for novice learners loses its effectiveness or even becomes detrimental as learners become more proficient in that type of task.

Unfortunately, these conceptual and methodological issues were left largely unrecognized and, instead, a return to former principles was deemed imminent [36‐38]. Germane cognitive load was reconceptualised as a subtype of intrinsic cognitive load, resulting in a plea for a move back to a two-factor intrinsic/extraneous cognitive load framework. To prevent any return to former principles from being rooted in methodological flaws in previous studies, one final attempt was made to develop a psychometric instrument that might distinguish between the three types of cognitive load [20] as defined in the late 1990s [28]. A set of four coherent studies [20] appeared to provide evidence in favour of the three-factor framework. Though being the first time since the inception of CLT that such an instrument measuring different types of cognitive load received empirical support, two follow-up studies [21] failed to provide further evidence for the germane cognitive load factor. It was therefore suggested that the three factors in the instrument be interpreted to represent intrinsic cognitive load, extraneous cognitive load, and a subjective judgment of learning [21, 22]. Table 1 presents the eight items of the questionnaire that reflect intrinsic cognitive load (items 1–4) and extraneous cognitive load (items 5–8) respectively; the questionnaire can also be used without items 4 and 8 [20‐22].

To rephrase the gist of CLT, learning in fact revolves around dealing with intrinsic cognitive load [22, 39]. Given the number of information elements in a task, a more proficient learner will experience a lower intrinsic cognitive load than a novice learner, because some of the information elements in the task are already part of the cognitive schema of the more proficient learner, leaving fewer new elements needing to be processed. Providing learners with a task that comprises more information elements that are not yet part of their schemas will impose a higher intrinsic cognitive load on their minds.

The contention that scaling up intrinsic cognitive load can bolster learning found support in a recent randomized controlled experiment using mixed methods [40]. In this study, Lafleur and colleagues contrasted a typical Objective Structured Clinical Examination (OSCE) with a so-called Hypothesis-Driven Physical Exam (HDPE [41]). Whereas the OSCE concerned a part-taskexamination, the latter required students to perform a whole-task physical examination. In this latter group, students first made a list of anticipated findings related to several competing diagnoses; they then selected and performed physical examination manoeuvres to elicit findings, interpreted these, corrected initial manoeuvres, to finally arrive at a motivated diagnosis. Throughout the experiment, extraneous cognitive load had been kept at a minimum in both groups. Intrinsic cognitive load, however, resulted somewhat higher in the whole-task HDPE condition. Curiously, this group also revealed a better performance with regard to backward and forward diagnostic reasoning, indicating that a higher administration of intrinsic cognitive load could actually be beneficial to learning.

Of course, this reasoning does not hold till cognitive overload and beyond. As we have seen before, excessive administration of intrinsic cognitive load can do more harm than good. More precisely, when the sum of intrinsic and extraneous cognitive load exhausts working memory capacity, it cannot reasonably be expected that any integration of new information elements into existing knowledge will occur. The same holds true for trivial administration of intrinsic cognitive load: when tasks are made very easy for learners in the light of the knowledge they already possess, boredom may prevail over learning [42]. By extension, we have seen that the assimilation of an additional germane cognitive load into a limited working memory was not so easy to envisage. A two-factor framework that meets all the conditions previously addressed seems much more plausible, especially so if one considers that the concept of germane cognitive load has never really found support in empirical research. The two-factor framework is therefore a logical starting point from which to proceed. How insights yielded so far can inform medical education design and research is the topic of the remainder of this paper.

One of the challenges medical education designers are faced with, is how to keep extraneous cognitive to a minimum. To achieve this, six strategies have been crystallized out which are particularly geared towards neophyte learners [28, 43] and which are briefly highlighted in the following.

With exception of the strategy aimed at avoiding redundancy, these strategies tend to be effective only for novice learners. The aforementioned expertise reversal effect [34, 35] indicates that instructional support through worked examples, completion tasks, non-specific goals, dual-mode auditory/visual presentation or avoidance of split attention may be redundant for more proficient learners and as such increase rather than decrease extraneous cognitive load. Thus, teachers and curriculum designers should monitor their learners’ proficiency in each phase of a curriculum and in a given course within a curriculum, and use this information to attune medical education design accordingly. This is necessary not only to keep extraneous cognitive load to a minimum, but also to raise intrinsic cognitive load to the most favourable level: the more the extraneous burden is eased, the more scope remains for intrinsic cognitive load to be processed. The following paragraphs outline how such desirable levels of intrinsic cognitive load can be attained.

Whether a learning task is regarded as complex depends on the individual learner’s cognitive schema of this type of task. The less knowledge learners can fall back on, the more complex a task will be, and, consequently the greater the amount of intrinsic cognitive load they will experience. The following three strategies can help to optimize the intrinsic cognitive load and stimulate learners to allocate their working memory resources to deal with that intrinsic cognitive load.

The aforementioned strategies come together in one coherent approach to medical education design: given end goals and assessment criteria of a course, a curriculum, one of the lines within a curriculum (e.g., skills training, research) or an individual learning trajectory, task complexity and fidelity should increase while instructional support should decrease as learners become more proficient. This approach can be visualized in a cube as in Fig. 1.

This model distinguishes three dimensions: task fidelity (all the way from literature to real patients), task complexity (the number of information elements in a task), and instructional support (all the way from worked examples to autonomous problem solving). These three dimensions together constitute three steps to proficient learning: (I) start with high support on low-fidelity low-complexity tasks and gradually fade that support as learners become more proficient; (II) repeat I for low-fidelity but higher-complexity tasks; and (III) repeat I and II in that order at subsequent levels of fidelity. This order of steps is based on the assumption that one needs a sufficient proficiency at lower-fidelity levels to avoid overload and disorientation at higher-fidelity levels, and sufficient proficiency at lower-level fidelity levels can be reached through an iterative process of gradually decreased instructional support and gradually increased learning task complexity.

For instance, if one cannot yet get beyond low-complexity textual descriptions of diagnoses, one is unlikely to be successful in dealing with simulated patients let alone real patients. It is like with learning a language. If one is still struggling with simple word completion tasks in a textbook, one cannot expect a more solid understanding of sentence patterns (low-fidelity task) that is needed to have a somewhat meaningful communication in that language with a teacher in class (medium-fidelity task) let alone for some meaningful communication in an everyday environment of native speakers (high-fidelity task). Likewise, if one does not understand what a standard deviation is, one cannot understand statistical tools including that concept (low-fidelity task), and consequently, one cannot expect to be able to appropriately apply these tools to simulated cases (medium-fidelity task) let alone to empirical research data or to real patient cases (high-fidelity task). One appears to need a sufficient proficiency at lower layers to continue the journey at a subsequent layer.

How many fidelity levels and complexity levels within fidelity levels are needed depends on the aims of the course, curriculum, line within a curriculum, or individual learning trajectory, and of course, the same holds for defining the relatively high-support low-complexity low-fidelity starting point; the numbers in Fig. 1 are included only to explain the basic rationale behind this holistic model. However, as the final section of this paper argues, the three steps of this model generate hypotheses that could be addressed in future research.

CLT could be extended further if three factors were taken into account, the first factor being emotion. More specifically, the question as to how emotions can influence our ability to process information in a variety of circumstances is one worthy of answering. To date, this factor has received minor thought. Recent findings such as that emotion could inhibit learning and decision-making if it stimulates bias [54] or when pondering over its cause [55, 56] should serve as a point of departure for more in-depth research, for instance how emotion affects learning and task performance across levels of task fidelity as envisioned in the model in Fig. 1.

Another factor the impact of which has been equally under-researched is assessment. While CLT’s emphasis has been essentially on learning and instruction, it has been less concerned with addressing how end terms and assessment criteria can affect the extent to which we actually engage in learning. The experiment by Lafleur and colleagues [40] demonstrates that slightly more challenging end terms and assessment criteria can stimulate learners to invest more effort in learning. In the context of the model in Fig. 1, more challenging end terms and assessment criteria may result in a larger cube that comprises more levels of complexity and/or fidelity. This may undermine learning when the combination of intrinsic and extraneous cognitive load is taken to the limits of working memory but may stimulate learning when extraneous cognitive load is kept to a minimum and intrinsic cognitive load is around an optimum. The aforementioned experiment should be replicated in a variety of settings, with different kinds of courses, at different stages of a curriculum, and in different contexts of learning at the workplace to further assess this statement.

Finally, the distinction between intrinsic and extraneous cognitive load—which is crucial for medical education design—cannot be made by a single-item measure, and it is unclear to what extent the NASA TLX could help us to make that distinction. However, the NASA TLX could provide inspiration to researchers when considering for instance to build forth on the instrument in Table 1, especially in the context of learning procedures and skills. Although the instrument in Table 1 already appears to take into account mental demands, performance, and effort, the other aspects addressed by the NASA TLX are not really visible here. Including the NASA TLX in the further development of measures for intrinsic and extraneous cognitive load could result in an instrument that still enables us to distinguish between intrinsic and extraneous cognitive load but also takes into account physical and temporal demands as well as frustration. In any case, the development of new measures should involve not only quantitative but also qualitative methods. A strength of research inspired by CLT carried out since its introduction is a certain rigour in study design and procedure. However, not that many studies have included qualitative methods, and this is perhaps one of the reasons why most attempts to develop measures for intrinsic and extraneous cognitive load have failed. One of the strengths of medical education research is that one can easily find examples of excellent use of quantitative as well as of qualitative methods. We will need both in the further development and refinement of measures, even if the combined use of subjective measures along with indirect measures, secondary tasks, and/or bio-measures is considered. In fact, when feasible, studies involving different types of measures would be recommended. This could shed light for instance on how specific bio-measures relate to intrinsic and/or extraneous cognitive load.