Procedural Learning in Mixed Reality: Assessing Cognitive Load and Performance

Abstract

Immersive technologies offer promising advancements in medical education, particularly in procedural skill acquisition. However, their implementation often lacks a foundation in learning theories. This study investigates the application of the split-attention principle, a multimedia learning guideline, in the design of knot-tying procedural content using a mixed reality (MR) technology, specifically Microsoft HoloLens 2. A total of 26 participants took part in a between-group design experiment comparing integrated and split-source formats for learning arthroscopic knots, with the performance and the cognitive load assessed. The initial hypotheses were not confirmed, as results did not show significant differences in performance during recall, nor in extraneous and germane cognitive load. However, the findings on intrinsic cognitive load highlight the complexity of participant engagement and the cognitive demands of procedural learning. To better capture the split-attention effect, future research should address the high element interactivity in MR representations. The study provides some foundation for designing procedural simulation training that considers both learners’ needs and cognitive processes in highly immersive environments. It contributes to the ongoing exploration of instructional design in MR-based medical education, emphasizing both the potential and challenges of multimedia learning principles in advanced technological contexts.

1. Introduction

Technological advancements have propelled immersive technologies into vital roles across various fields, including education (Chen, 2023). Extended reality (XR) encompasses (Milgram & Kishino, 1994) virtual reality (VR), augmented reality (AR), and mixed reality (MR), each offering distinct levels of digital integration into real-world environments. VR creates fully virtual spaces (Allcoat et al., 2021), AR overlays digital elements onto physical settings, and MR enables seamless interaction between the two (Park et al., 2021; Pimentel et al., 2022). Given the discussions on MR’s definition (Milgram & Kishino, 1994; Parong, 2021), we separated VR, AR, and MR. These technologies, particularly MR tools, are increasingly recognized for their potential to enhance healthcare systems (Masson, 2023) and may improve medical training (Silvero Isidre et al., 2023).

Advantages presented by MR are abounding. Some scholars assume that surgical training should no longer rely on patient-based learning (Mayo, 1927), since MR tools can enhance traditional training methods (de Sá et al., 2022) by enabling surgical technique practices (Grantcharov, 2008; Mergen et al., 2024) including high-risk procedures (Asoodar et al., 2024; Forgione & Guraya, 2017; Nagayo et al., 2022). In fact, procedural learning (such as suturing and knot-tying) may benefit from MR especially in countries where certain techniques like arthroscopic knot-tying are absent from curricula (Lacroix et al., 2021). Overall, these technologies offer some responses to ethical, economic, and time-related issues in medical training (Forgione & Guraya, 2017; Nagayo et al., 2022).

Nevertheless, their integration into the surgical education must tackle some challenges, sparking discussions on their effectiveness and appropriate implementation (Co et al., 2023). Some educators, it is argued, adopt these tools without effectively leveraging them to enhance and support meaningful learning experiences (Sandars et al., 2015). Others noted that despite the abundance of learning theories, few are designed to help educators integrate procedural simulations into their curricula (Rivière et al., 2018).

Thus, researchers have emphasized the importance of grounding the educational designs of MR techniques in empirical evidence, especially in medical training (Issa et al., 2011; Mayer, 2020). For example, studies have shown that learning was more effective when instructional materials included both words and images, compared to text-only formats (Mayer, 2002). Poorly designed educational content can hinder learning by increasing extraneous cognitive load (Parong, 2021). This underscores the significance of applying multimedia learning principles and the Cognitive Theory of Multimedia Learning (CTML) (Mayer, 2021) to minimize the unnecessary cognitive load and manage task complexity (Rivière et al., 2018). Additionally, the importance of monitoring the learners’ cognitive load during procedural simulations, such as suturing (McInnis et al., 2021), is highlighted, with cognitive load theory (CLT) emerging as a particularly relevant framework (Fraser et al., 2015).

Against this backdrop, this paper investigated the application of the split-attention principle in the design of knot-tying procedural content using a specific MR technology, HoloLens 2 (Palumbo, 2022). The principle is anchored in CLT (Ayres & Sweller, 2021) and CTML, and has shown potential for designing procedural simulations in a MR medium. In the following sections we succinctly present this principle, which is the focus of our attention.

1.1. Design of Learning Scenarios: The Case of the Split-Attention Principle

Multimedia learning principles guide the design of learning content to support cognitive processes. These principles can be categorized into three groups: (1) minimizing the extraneous cognitive load (ECL) by limiting unnecessary information, (2) focusing the learners’ attention on essential information, and (3) promoting generative processing to help the learners synthesize information into meaningful representations (Mayer & Fiorella, 2021).

For instance, the split-attention principle reduces ECL by presenting essential learning elements spatially close to each other, thereby minimizing the need for learners to switch between different information sources (Ayres & Sweller, 2021). Studies have shown that integrated content formats, like those used to learn the kidney structure (Cierniak et al., 2009) and orthopedic physical therapy (Pociask & Morrison, 2008), led to a better performance than split-source formats. Split-source formats increased the ECL (Cierniak et al., 2009), making tasks more difficult and requiring more concentration, but had no impact on task completion times (Pociask & Morrison, 2008).

The recent studies in immersive environments explored the application of this principle in physics lab experiments, and the results of these pioneering studies were mixed. Krüger and Bodemer (2022) found no significant effects of spatial contiguity (a form of split-attention principle) on the cognitive load and the declarative knowledge in an AR-based botanical garden course, although there was a reduction in temporal demand and an increase in perceived performance. In contrast, Thees et al. (2020) found a significant reduction in the ECL with an MR-adapted integrated condition, leading to better conceptual knowledge compared with a traditional setup. However, another study comparing a HoloLens (integrated condition) with an iPad (separate-display condition) did not show the expected benefits (Thees et al., 2022).

Despite these variations, Ginns (2006) reported a large effect size (d = 0.85) in a meta-analysis of spatial contiguity, which includes the split-attention effect, suggesting that integrated formats generally improve learning by reducing the ECL. The split-attention and spatial contiguity principles are essentially the same (Schroeder & Cenkci, 2018), with the latter being more specific, as emphasized by the CLT proponents (Fiorella & Mayer, 2021).

Moreover, there is evidence in the literature that the split-attention principle has yet to be explored in the context of MR-based medical, procedural training, especially as, to the best of our knowledge, no studies have been conducted regarding its implementation.

It is essential to briefly address two key theories that underpin the split-attention (or spatial contiguity) principle: CLT and its counterpart, CTML.

1.2. Cognitive Load Theory (CLT)

CLT (Sweller et al., 1998, 2011) aligns with classical cognitive architecture models, which propose that the cognitive system includes a limited-capacity working memory (WM) for learning and conscious thought, and an unlimited long-term memory (LTM) that stores automated schemas (Chanquoy et al., 2007; Tricot, 1998).

During learning, two key mechanisms are involved—the acquisition of schemas and the transfer of learned procedures from controlled to automated processing. Initially, knowledge is organized into schemas, which guide how the new information is processed. Over time, the handling of this information shifts to automatic processing, and without this automation performance remains slow, clumsy, and error-prone (Sweller, 1994). The CLT offers a theoretical framework to design surgical learning environments that account for the cognitive processes involved in learning (Tokuno et al., 2023).

CLT identifies three types of the cognitive load (CL) that affect working memory: (1) intrinsic cognitive load (ICL), related to the processed information, (2) extraneous cognitive load (ECL), caused by the information presentation, and (3) germane cognitive load (GCL), which relates to the effort needed to process and integrate relevant information into meaningful schemas (Chanquoy et al., 2007). These distinctions stress the importance of designing instructional scenarios that consider the learners’ CL, particularly in complex environments where an excessive load can hinder learning (Fraser et al., 2015).

In their systematic review, Çeken and Taşkın (2022) reported that most CL assessments relied on subjective measures, such as questionnaires, or indirect objective measurements. Thus, some subjective scales assessed the intensity of mental effort perceived by learners (F. G. Paas, 1992), while others attempted to capture all three types of CL (Leppink et al., 2013, 2014). However, those studies raised concerns about the psychometric and theoretical validity of these approaches in measuring and managing CL in educational contexts (Schroeder & Cenkci, 2020).

1.3. Cognitive Theory of Multimedia Learning (CTML)

We can argue that the CLT considers the relevant aspects of human cognitive architecture when designing good instructional materials (Paas & Sweller, 2021), shedding some light on how the different cognitive loads affect working memory capacity.

The CTML (Mayer, 2021), partly shaped by the CLT (Camp et al., 2021), focuses on how people can learn meaningfully from multimedia materials and asserts three types of demands on the limited cognitive capacity during learning (extraneous, essential, and generative processing), leading to the following three instructional goals: to reduce extraneous processing, to manage essential processing, and to foster generative processing (Mayer, 2024). Indeed, the CLT has helped provide the CTML with a solid foundation in instructional design research, and some CTML principles (such as the spatial contiguity principle) are subsets of the split-attention effect, originally developed by CLT’s tenets, which have evolved (Camp et al., 2021). Furthermore, the concept of multimedia has evolved (Camp et al., 2021), with recent works illustrating how the CTML is being applied in extended reality learning technologies (Altmeyer et al., 2020, 2021; Laumann et al., 2024; Strzys et al., 2018; Thees et al., 2020, 2022).

The deployment of immersive technologies in education underscores the need to address CL and optimize the learning content (Buchner et al., 2022; Poupard et al., 2024), yet many studies still overlook the established learning theories or frameworks (Radianti et al., 2020; Wu et al., 2020). Recently, Makransky and Petersen (2021) introduced the cognitive affective model of immersive learning (CAMIL), emphasizing the integration of cognitive and emotional factors to enhance the learning outcomes in virtual environments through thoughtfully designed immersive experiences.

1.4. Learning Surgical Procedures in Medical Education

Chiniara et al. classified simulations like suturing and knot-tying as procedural simulations, aimed at developing technical skills for real-life applications (Chiniara et al., 2013). For these critical tasks, researchers explore suitable learning technologies, ranging from traditional to innovative.

Previous studies on the procedural learning of knot-tying have compared 2D laptop-based presentations with 3D VR or MR environments (Nagayo et al., 2021; Yoganathan et al., 2018), but these often lacked of grounding in educational theories. Furthermore, the prevailing medium comparison approach has been criticized for the difficulty of setting up experimental conditions that are comparable in all aspects, and also focus on global effects rather than the nuanced features and interactions inherent in specific technologies (Buchner et al., 2022; Gonnermann-Müller & Krüger, 2024). Beyond these limitations, most past research has been conducted in less immersive 2D environments, highlighting the need for new investigations into the emerging educational contexts (Çeken & Taşkın, 2022).

Immersive technologies are widely used in medical education, particularly for anatomy and surgical training (Asoodar et al., 2024), with features like immersion and interaction (Makransky & Petersen, 2021). MR appears to be particularly well-suited for surgical training, as it offers a higher degree of interaction with 3D information within a real-world environment, compared with AR and VR (Sánchez-Margallo et al., 2021; Bautista et al., 2023). Recent studies have highlighted the growing use of MR in operating rooms, where it offers benefits such as real-time visualization of anatomical data, information overlay, and medical training applications (Magalhães et al., 2024).

The Microsoft HoloLens 2 is predominant in these contexts, and is used in various surgical procedures (Gregory et al., 2018; Liu et al., 2019). HoloLens 2, along with the Microsoft Dynamics Guide 365 software, enables the creation of personalized training guides using holograms, videos, images, and 3D models, making it valuable for medical education and simulation (Park et al., 2021; White, 2019).

A randomized controlled study conducted in Japan assessed the usability of an MR system for learning (Nagayo et al., 2022), in which 43 medical students with no prior suturing experience performed a subcuticular suture on synthetic skin using either a HoloLens 2 or a laptop. Performance scores showed no significant differences, but participants preferred the HoloLens 2 for its ability to merge virtual information with real objects, thereby aiding in surgical instrument manipulation. These findings suggest that the novelty effect (Parong, 2021) of technologies like the HoloLens 2 may hinder procedural learning (Nagayo et al., 2022).

Despite their growing popularity, MR technologies must overcome potential pitfalls (Pimentel et al., 2022). A too technocentric approach, focusing on technology over the learners’ cognitive needs (Mayer & Fiorella, 2021), often creates a gap between expectations and realities, sometimes leading to rejection.

1.5. Current Study

This research may be classified as a preliminary investigation aimed at addressing the various limitations raised, and functioning as a proof of concept to explore the split-attention effect within the context of knot-tying simulations utilizing MR technology. In alignment with the prior studies to which we alluded previously, we postulated that:

H1: 

CL scores of participants learning knot-tying in the non-split-attention condition will be significantly lower than those in the split-attention condition.

H2: 

Performance scores of participants learning knot-tying in the non-split-attention condition will be significantly higher than those in the split-attention condition.

H3: 

Execution time scores of participants learning knot-tying in the non-split-attention condition will be significantly lower than those in the split-attention condition.

H4: 

Average learning time of participants learning knot-tying in the non-split-attention condition will be lower than those in the split-attention condition.

2. Materials and Methods

2.1. Participants, Sampling, and Measures

The inclusion criteria of the participants were as follows: individuals aged 18 and more, a minimum B2 proficiency in French, low and no experience in knot-tying, and new users of mixed-reality headsets. The exclusions criteria were physical or mental health issues, motor or visual impairments, epilepsy, heart conditions, phobias, and cybersickness risks. To counter potential bias, we also excluded individuals pursuing medical education, as this could have exposed them to knot-tying and suturing practice. Indeed, the study plan was adapted to different issues which will be addressed in the limitations section.

A total of 26 participants (women = 14, men = 12; M_age = 23, SD_age = 2.735; 11 pursuing bachelor’s degrees, 15 pursuing master’s degrees; 24 advanced and 2 intermediate French speakers, with low expertise in knot-tying: M_expertise = 1.8, SD_expertise = 0.962) were recruited for the experiment, primarily through online advertisements, social media associated with the university, and direct contact. The results of one participant were excluded due to interruptions during the session. All participants were students, mainly in psychology, who received credits for their participation. Some students were in the science, technology and mathematics (STM) field. Experimental sessions were conducted from May to June 2023 in France, on the campus of a public university focusing on humanities and social sciences.

Participants were randomly assigned to either of the two conditions [independent variable (IV)]: 13 in the non-split-attention/integrated condition (C) and 13 in the split-attention/separated condition (NC). A counterbalancing method minimized the transfer effects in knot recall by using computer-generated sequences for up to 100 participants (Appendix I). Participants completed the informed consent and socio-demographic Limesurvey questionnaires using a laptop. The entire procedure is summarized (Figure 1); the experiment was planned to last 45 to 60 min per participant, depending on the time each participant chose to spent on learning knots.

The separated condition (Figure 2) required the participants to alternate their visual attention between a holographic learning screen positioned behind them (at 180°) and the knot-tying task in front of them, while the integrated condition (Figure 2) enabled simultaneous viewing of both elements in front of them.

The CL (measured using a 7-item Likert scale, with a 10-point scale for each item ranging from 1 (very low) to 10 (very high)) and performance (knot quality and execution time) were the dependent variables (DVs). The CL was assessed using a subjective Likert scale (Leppink et al., 2013, 2014; Ouwehand et al., 2021) adapted for the present study (Appendix F:

Table A5. Original, adapted, and translated items of the CLS.

#	Original Item (English)	# New	Adapted Item (English)	Adapted Selected Item (French)
1	The content of this activity was very complex.	1	-	-
2	The problem/s covered in this activity was/were very complex.	2	The procedures covered in these activities were very complex.	Les procédures abordées dans ces activités étaient très complexes.
3	In this activity, very complex terms were mentioned.	3	-	-
4	I invested a very high mental effort in the complexity of this activity.	4	I invested a very high mental effort in the complexity of these activities.	J’ai investi un effort mental très important dans la complexité de ces activités.
5	The explanations and instructions in this activity were very unclear.	5	The instructions for these activities were very unclear.	Les instructions de ces activités n’étaient pas très claires
6	The explanations and instructions in this activity were full of unclear language.	6		-
7	The explanations and instructions in this activity were, in terms of learning, very ineffective.	7	The instructions for these activities were, in terms of learning very ineffective.	Les instructions de ces activités étaient, en termes d’apprentissage, très inefficaces.
8	I invested a very high mental effort in unclear and ineffective explanations and instructions in this activity.	8	I invested a very high mental effort in unclear and ineffective instructions in these activities.	J’ai investi un effort mental très important dans les instructions peu claires et inefficaces de ces activités.
9	This activity really enhanced my understanding of the content that was covered.	9	-	-
10	This activity really enhanced my understanding of the problem/s that was/ were covered.	10	-	-
11	This activity really enhanced my knowledge of the terms that were mentioned.	11	-	-
12	This activity really enhanced my knowledge and understanding of how to deal with the problem/s covered.	12	These activities really enhanced my knowledge and understanding of how to realize the procedures covered.	Ces activités ont vraiment amélioré ma connaissance et ma compréhension de la manière de réaliser les procédures abordées.
13	I invested a very high mental effort during this activity in enhancing my knowledge and understanding.	13	I invested a very high mental effort during these activities in enhancing my knowledge and understanding of the procedures.	J’ai investi un effort mental très important au cours de ces activités pour améliorer ma connaissance et ma compréhension des procédures.

). Leppink et al. (2013) developed a 10-item scale to assess ICL, ECL, and GCL, initially linked to statistical knowledge but applicable across other domains, including language learning (Leppink et al., 2014). Overall, the modifications made, such as adding items and language translation, did not challenge its internal structure (Thees et al., 2020). For this study, the adaptations included retaining only the following 7 items from the original 10 items (previously 13 items) and their translation into French: ICL (items 2: Complexity and 4: Mental Effort-Complexity), ECL (items 5: Clarity, 7: Efficiency, and 8: Mental Effort-Clarity & Efficiency), and GCL (items 12: Knowledge & Understanding, 13: Mental Effort-Knowledge & Understanding). We conducted an exploratory factor analysis to assess the internal structure (Appendix D:

Table A4. KMO, reliability of adapted self-reported CLS.

CL Type/Item	KMO	Reliability α
ICL		0.826
Item 1	0.509
Item 2	0.560
ECL	0.393	0.735
Item 3	0.393
Item 4	0.749
Item 5	0.477
GCL		0.067
Item 6	0.550
Item 7	0.698

) of the instrument modified. For the three subscales, excellent reliability was obtained for the ICL and ECL (α_ICL = 0.826, α_ECL = 0.735) but not for the GCL (α_GCL = 0.0665). Overall, the instrument demonstrated acceptable reliability for this research (α_ICL,ECL,GCL = 0.744).

For the performance assessment, the quality of knot execution was evaluated using a rating grid created by authors, after the assessment of the video recordings of knot-tying (each knot was tied in 3 scenarios, divided into 3 sub-scenarios). Each correctly performed task in a sub-scenario earned one point (otherwise it earned zero). The execution time was recorded using a stopwatch.

The user satisfaction with the HoloLens 2 was assessed with an 11-item (rated on a 10-point Likert scale (1 = very low to 10 = very high) Smart Glass User Satisfaction (SGUS) questionnaire (Helin et al., 2018; Xue et al., 2019) (Appendix G:

Table A6. Translation of SGUS Questionnaire.

#	Original Item (English)	Translated Item (French)
1	GL1 With AR-glasses I could access information at the most appropriate place and moment.	GL1 Avec les lunettes AR, j’ai pu accéder à l’information à l’endroit et au moment les plus appropriés.
2	GL2 Content displayed on the AR-glasses made sense in the context I used it.	Gl2 Le contenu affiché sur les lunettes AR avait un sens dans le contexte où je l’ai utilisé.
3	GL3 AR-glasses provided me with the most suitable amount of information.	GL3 Les lunettes AR m’ont fourni la quantité d’informations la plus appropriée.
4	GL4 AR-glasses allowed a natural way to interact with information displayed.	GL4 Les lunettes AR permettent d’interagir de manière naturelle avec les informations affichées.
5	GL5 I had a good conception of what is real and what is augmented when using AR-glasses.	GL5 J’ai une bonne conception de ce qui est réel et de ce qui est augmenté lorsque j’utilise des lunettes AR.
6	GL6 The interaction with content on AR-glasses captivated my attention in a positive way.	GL 6 L’interaction avec le contenu des lunettes AR a captivé mon attention de manière positive.
7	GL7 The instructions given by AR-glasses helped me to accomplish the task.	GL7 Les instructions données par les lunettes AR m’ont aidé à accomplir la tâche.
8	GL8 I understood what is expected from me in each phase of the task with the help of AR-glasses.	GL8 J’ai compris ce que l’on attendait de moi à chaque phase de la tâche avec l’aide des lunettes AR.
9	GL9 Performing the task with the help of AR-glasses was natural to me.	GL9 L’exécution de la tâche avec l’aide des lunettes AR était naturelle pour moi
10	GL10 While using AR-glasses, I was aware of the phase of the task at all times during the execution of the task.	GL10 En utilisant les lunettes AR, j’étais conscient de la phase de la tâche à tout moment pendant l’exécution de la tâche.
11	GL11 While using AR-glasses, I was able to pay attention to the essential aspects of the task all the time.	GL11 En utilisant les lunettes AR, j’ai pu prêter attention aux aspects essentiels de la tâche à tout moment.

). It evaluated the user experience with smart glasses, focusing on MR perception, interaction, and usability. Grounded in web-based learning and mobile MR evaluation frameworks, it assesses interface design, MR-specific features, and learning effectiveness (Xue et al., 2019). An exploratory factor analysis was used to assess its internal structure (Appendix H:

Table A7. KMO Measure of Sampling Adequacy per Item of SGUS questionnaire.

Items	MSA
Overall	0.661
GL1	0.760
GL2	0.735
GL3	0.874
GL4	0.519
GL5	0.193
GL6	0.678
GL7	0.667
GL8	0.703
GL9	0.581
GL10	0.412
GL11	0.726

) and had an excellent reliability: α = 0.860.

2.2. Materials and Experimental Design

Participants viewed three knot-tying scenarios using a HoloLens 2. The knots chosen were of the following sliding types: Samsung Medical Center knot (SMC) (Kim & Ha, 2000), Double Twist knot (Rolla & Surace, 2002), and Duncan loop (Nottage & Lieurance, 1999). The knots were selected for their relevance in surgical training (Lacroix et al., 2021) and for their varying difficulty levels (Baumgarten & Wright, 2007). We based our design of the knot videos on traditional step-by-step instructions (Figure 3), dividing each video into three key moments before uploading them to the Dynamics 365 Guide. The application had 2 screens: one of the virtual screens displayed instructional reminders, reiterating the previously provided guidelines to help participants progress through the learning process (e.g., “Click the arrow to move to the next sub-step of the knot-tying procedure”). On the other screen, the video of the procedure was projected. Notably, audio guidance was deliberately omitted from the instructional videos.

To begin, each participant was required to scan a QR code related to a task by looking directly at it and staying about (12–20 inches) away from it. To navigate through different steps (stop, move forward, go back), the participant used eye-tracking interactions. Before starting the knot-tying task, the participants completed a familiarization task consisting of a Lego assembly procedure to learn how to navigate the HoloLens interface for about 15–20 min (Figure 3).

The Lego task and watching instructional videos independently helped the participants practice using the HoloLens commands. Knot-tying was then conducted using multicolored cords attached to adhesive plastic hooks. To prevent confusion, each type of knot was assigned a specific cord color: pink (Duncan), green (SMC), or blue (Double Twist). The videos were designed, recorded using an iPhone 13 Mini, and then edited for instructional purposes. The mixed-reality setup allowed for the real-time monitoring of the participant’s perspective, which also enabled capturing photos/videos (Figure 4).

During the learning stage of knot-tying, participants had up to three rounds for each knot, with the time recorded but no limits set. Specifically, after completing the first round, participants could perform two additional rounds—only then could they use the option to go back. However, during the knot-tying learning task (with each knot consisting of three sub-steps), they could not go back to view a sub-step that had already been completed. Statistical analyses were conducted using Jamovi and LimeSurvey for questionnaires.

Globally, the analyses included t-tests for independent samples, with the normality verified using the Shapiro–Wilk test and the homogeneity of variances using Levene’s test. If the assumptions were not met, the Mann–Whitney U test was used instead. A pilot study with 2 volunteers (1 Man, 1 Woman; M_age = 22; bachelor’s degree; no prior experience in knot-tying; novice with an MR tool; from psychology and STM fields) allowed the assessment of questionnaire clarity, image quality, and experiment duration.

3. Results

3.1. Cognitive Load

The descriptive results of each item are reported (Appendix A:

Table A1. Descriptive statistics of the variables relating to the CL perceived according to the conditions of presentation.

	95% CI
	Conditions	Mean	Lower	Upper	Median	SD	Minimum	Maximum
Complexity	C	4.15	2.83	5.48	5	2.19	1	7
	NC	5.77	4.33	7.21	6	2.39	1	10
Mental Effort-Complexity	C	5.54	4.39	6.69	6	1.90	2	8
	NC	6.77	5.21	8.33	8	2.59	2	10
Clarity	C	2.31	1.11	3.50	2	1.97	1	8
	NC	3.08	1.93	4.22	3	1.89	1	7
Efficiency	C	2.92	1.56	4.28	2	2.25	1	8
	NC	3.69	2.27	5.12	3	2.36	1	8
Mental Effort-Clarity & Efficiency	C	3.15	1.62	4.69	3	2.54	1	10
	NC	4.77	3.31	6.23	5	2.42	1	9
Knowledge & understanding	C	7.85	6.93	8.76	8	1.52	6	10
	NC	6.08	4.38	7.78	6	2.81	1	10
Mental Effort-Knowledge & Understanding	C	6.15	5.00	7.31	7	1.91	3	9
	NC	7.31	6.04	8.57	8	2.10	4	10

), for each category of CL (

Table 1. Descriptive statistics of ICL, ECL, GCL in different conditions.

	95% CI
	Conditions	Mean	Lower	Upper	Median	SD
ICL	C	4.846	3.775	5.917	5.500	1.772
	NC	6.269	4.843	7.695	6.500	2.360
ECL	C	2.795	1.663	3.927	2.333	1.873
	NC	3.846	2.804	4.888	4.000	1.725
GCL	C	7.000	6.240	7.760	7.500	1.258
	NC	6.692	5.585	7.799	7.500	1.832

). The results of the t-test or the Mann–Whitney test are provided, considering p < 0.05 as the threshold for statistical significance.

Overall, although the mean CL scores may suggest some benefit of the integrated condition for ECL, with the confidence interval CI 95% [−∞; 0.257], no statistically significant difference was detected, t (24) = −1.489, p = 0.075, Cohen’s d = −0.584. To obtain a clear estimate of CI, we applied bootstrap resampling (1000 iterations) (Appendix E) which provided CI [2.333; 4.167]. For the GCL, U = 77.5, p = 0.651, no statistically significant difference was noted (CI 95% [−∞; 1.5] effect size r = 0.083).

However, with 95% CI ([−∞; −0.023] after bootstrap resampling [4.5; 6.5]) for the ICL, the study found a statistically significant difference in favor of the integrated condition: t (24) = −1.423, p = 0.047, with a moderate effect size, Cohen’s d = −0.682.

3.2. Overall Performance Related to Knot Quality During Recall Phase

We began to assess the quality of knots in each sub-step of the recall phase. From this we calculated the overall recall scores for each participant, for each knot, which were as follows: Qt SMC (overall score for the SMC knot), Qt Double Twist (overall score for the Double Twist knot), and Qt Duncan (overall score for the Duncan). The descriptive results are provided below (

Table 2. Descriptive statistics of variables related to knot quality in different conditions of presentation.

	95% CI
	Conditions	Mean	Lower	Upper	Median	SD	Minimum	Maximum
Qt SMC	C	1.231	0.728	1.73	1	0.832	0	2
	NC	1.462	0.781	2.14	2	1.127	0	3
Qt Double Twist	C	1.231	0.966	1.50	1	0.439	1	2
	NC	1.154	0.927	1.38	1	0.376	1	2
Qt Duncan	C	2.538	2.069	3.01	3	0.776	1	3
	NC	2.000	1.396	2.60	2	1.000	1	3

). We detected no statistically significant differences with Qt SMC, U = 72, p = 0.762; Qt Double Twist, U = 78, p = 0.326; Qt Duncan, U = 60, p = 0.082, with, respectively, r = 0.1479, r = 0.652, and r = 0.0769.

The overall performance was assessed by summing the points each participant earned across all three sequences: Quality = Qt SMC + Qt Double Twist + Qt Duncan. The results indicate no significant difference: U = 75, p = 0.316.

3.3. Performance Related to Knot Execution Time During Recall Phase

We measured the time taken by each participant to complete each of the three knots. The descriptive statistics for these results are presented (

Table 3. Descriptive statistics of recall time variable according to presentation conditions.

	95% CI
	Conditions	Mean	Lower	Upper	Median	SD	Minimum	Maximum
Recall SMC	C	79.9	50.4	109.4	67	48.8	23	177
	NC	38.4	28.1	48.7	31	17.0	17	69
Recall Double Twist	C	144.1	97.4	190.8	125	77.3	32	270
	NC	140.3	105.4	175.2	125	57.8	79	288
Recall Duncan	C	48.2	36.5	59.8	39	19.2	23	81
	NC	33.5	24.2	42.8	32	15.4	17	71

). No statistically significant differences were detected for SMC (CI 95% [7; ∞] after bootstrap resampling (Appendix E) [31; 66]) U = 37, p = 0.993, with a moderate effect size (r = 0.562). Similarly, for, respectively, Double Twist (CI 95% [−42; ∞] after bootstrap resampling (1000 iterations) [102, 158] and Duncan (CI 95% [2.94; ∞] gave [30; 39]), no significant differences were observed: t (24) = 0.141, p = 0.555, d = 0.055 (small effect); and t (24) = 2.142, p = 0.979, d = 0.84 (large effect). In contrast, mean scores were in favor of the integrated condition.

3.4. Learning Time for Knot-Tying

Participants were allowed to complete up to three rounds of learning. All participants completed the 1st round (T1); however, most chose not to proceed with the 2nd and/or the 3rd round. Descriptive statistics for all rounds are elaborated in (Appendix B:

Table A2. Descriptive statistics of the variable relating to the learning time per turn, according to the conditions.

	Conditions	N	Mean	Median	SD	Min	Max
T1 SMC	C	13	5.85	4	4.413	2	18
	NC	13	8.92	8	4.591	3	16
T2 SMC	C	12	2.33	2.00	1.155	1	5
	NC	8	5.13	4.00	3.796	2	14
T3 SMC	C	4	1.50	1.50	0.577	1	2
	NC	4	3.00	2.00	2.828	1	7
T1 Double Twist	C	13	12.23	10	9.057	3	32
	NC	13	17.08	17	8.108	7	30
T2 Double Twist	C	5	4.40	4	1.140	3	6
	NC	4	5.25	4.00	3.304	3	10
T3 Double Twist	C	0	NaN	NaN	NaN	NaN	NaN
	NC	0	NaN	NaN	NaN	NaN	NaN
T1 Duncan	C	13	2.85	3	1.573	1	6
	NC	13	3.92	4	1.656	2	8
T2 Duncan	C	9	1.44	1	0.527	1	2
	NC	7	1.43	1	0.535	1	2
T3 Duncan	C	2	1.50	1.50	0.707	1	2
	NC	1	3.00	3	NaN	3	3

).

For T1, we conducted statistical analyses on the SMC, Double Twist, and Duncan knots by applying the Mann–Whitney U test (

Table 4. Inferential statistics relating to the first round learning time under different conditions.

		Statistic	p	Effect Size
T1 SMC	Mann–Whitney U	47.5	0.030	0.438
T1 Double Twist	Mann–Whitney U	51.0	0.044	0.396
T1 Duncan	Mann–Whitney U	47.0	0.025	0.444

). Given the one extreme unclear value of CI, we also applied bootstrap resampling (1000 iterations) to estimate 95% CIs (Appendix E). The participants in the integrated condition completed the tasks in less time than those in the separated condition. Specifically, the different times required to complete the knots were [95% CI: SMC [4, 8.5], Double Twist Knot [8, 17], and Duncan Knot [2.513, 4.0]]. Indeed, the results revealed a statistically significant difference between the conditions, with moderate effect sizes (Table 4).

3.5. User Satisfaction with the Use of HoloLens 2

The descriptive statistics for the satisfaction scores are presented (Appendix C:

Table A3. Descriptive statistics of user satisfaction.

	Conditions	N	Mean	Median	SD	Min	Max
GL1	C	13	6.92	7	2.47	1	10
	NC	13	7.23	8	2.98	2	10
GL2	C	13	7.77	8	2.62	2	10
	NC	13	7.77	8	2.42	3	10
GL3	C	13	7.69	8	2.53	3	10
	NC	13	7.77	8	1.96	3	10
GL4	C	13	6.69	7	2.66	1	10
	NC	13	7.69	8	2.32	2	10
GL5	C	13	9.08	9	1.04	7	10
	NC	13	8.54	9	1.81	5	10
GL6	C	13	7.54	8	1.98	2	10
	NC	13	7.85	8	2.08	4	10
GL7	C	13	7.46	8	2.18	3	10
	NC	13	8.15	8	1.57	6	10
GL8	C	13	7.77	8	2.39	3	10
	NC	13	8.08	8	1.38	5	10
GL9	C	13	6.00	6	1.83	3	9
	NC	13	6.46	7	2.63	1	10
GL10	C	13	8.54	9	1.39	7	10
	NC	13	7.69	8	2.32	1	10
GL11	C	13	6.85	8	3.08	1	10
	NC	13	7.46	8	2.54	3	10

), with the unexpected observation that mean scores for certain SGUS items (participants’ satisfaction) were slightly higher in favor of the separated condition.

Overall, the participants reported general satisfaction with the use of the HoloLens 2 for the task across the two conditions. There were no significant differences observed, and the results were as follows: GL1: U = 72.0, p = 0.750; GL2: U = 82.0, p = 0.458; GL3: U = 79.5, p = 0.405; GL4: U = 64.0, p = 0.862; GL5: U = 76.0, p = 0.332; GL6: U = 76.5, p = 0.675; GL7: t (24) = −0.927, p = 0.819; GL8: U = 77.5, p = 0.363; GL9: t (24) = −0.519, p = 0.696; GL10: U = 70.5, p = 0.235; GL11: U = 76.0, p = 0.681).

4. Discussion

Our study aimed to examine the impact of two presentation formats of knot-tying procedures using MR HoloLens 2. Grounded in CLT and CTML, expectations were that the integrated/non-split-attention format would prove benefits pertaining to cognitive load and performance. The findings did not reveal significant differences for the ECL, nor the GCL; notably, a significant effect was detected for the ICL. Although no significant differences were observed in terms of the quality of knot realization during recall and the time needed to perform the task, some significant differences were noted for the first learning round of all of the knots.

4.1. Empirical Contributions

These results contrasted with the previous studies conducted in traditional learning environments, which demonstrated benefits for integrated formats in terms of performance and CL (Ginns, 2006; Schroeder & Cenkci, 2018). Before proceeding further, it is necessary to address the surprisingly significant difference in intrinsic cognitive load (ICL) observed over the course of the experiment.

We noted a marked difference in how the participants engaged with the complexity of the procedures studied. This finding was unexpected, as most prior research has not reported such discrepancies, except for Laumann et al. (2024), who identified significant ICL differences attributed to the learners’ prior knowledge. In the context of this study, it is important to highlight that declarative knowledge appears to be the primary focus of the studies implementing the split-attention principle. Declarative knowledge is largely associated with declarative memory, which stores facts and events. One proposed explanation is that declarative knowledge requires directed attention (ten Berge & van Hezewijk, 1999). During the experiment, we observed that procedural knowledge, although reliant to some extent on declarative knowledge (e.g., user manuals, cooking recipes) in various forms (verbal, visual), is closely tied to skill acquisition and pattern recognition for goal completion. Repeatedly practicing a sequence of actions enhances performance (Wit et al., 2023), as over time the skills become automatic and require less attention (ten Berge & van Hezewijk, 1999). This observation raises three key considerations regarding the ICL and the ECL.

First, the differences in the ICL may be attributed to the number of training rounds participants chose to complete in order to master the procedures. Our analysis of participant engagement shows that although the number of repetitions decreased over time, the participants in the integrated condition were more likely to complete the second and third rounds of training. This may have contributed to perceiving the task’s complexity and the associated mental effort as lower.

Second, concerning ECL, the procedural learning characteristics facilitated by the HoloLens may have mitigated the split-attention effect, in contrast to other studies primarily focused on declarative knowledge and conducted in a medium-comparison format. Participants could complete the procedure without frequently shifting attention between the screen, their hands, and the cords used to tie the knots. This differs from situations requiring continuous attention shifts between multiple essential sources of information, such as manuals, texts, or virtual representations, to successfully complete a task, such as recalling knowledge (Pociask & Morrison, 2008) or responding to heat and temperature conduction (Thees et al., 2020). This considers the necessity of a high element interactivity (Leahy & Sweller, 2019) if any effect is to be detected. Certainly, even with a large simple size, if there are few connections between the sources of information, even when the distance between the elements is there (Ayres & Sweller, 2021), the research may come to null findings.

Regarding the results for ECL, a certain degree of split-attention was implemented; as we observe the mean scores of the participants, we may presume that for longer or more complex tasks, the split-attention effect might become more pronounced. However, the lack of significant differences in the performance and cognitive load suggests the possibility of a Type II error, where we fail to reject the null hypothesis despite there being a true effect. This occurs if the study is underpowered, resulting in insufficient data to detect a significant difference when one exists. The near-threshold p-value and effect size (p = 0.075; d = −0.584) could indicate that a larger sample size and improved design might have revealed significant effects.

Third, a challenge arises when comparing the prior research findings to our study, which focuses on procedural learning, whereas most prior studies measure declarative knowledge. Cierniak et al. (2009) observed a learning benefit for an integrated presentation, with the split-attention effect noted only when measuring declarative knowledge about nephron facts, but not when performing inference tests. This issue also applies to Krüger and Bodemer’s study (Krüger & Bodemer, 2022) on the declarative learning of botanical material, which showed no reduction in the cognitive load but an increase in perceived performance. We suggest that this reflects the broader trend in immersive environment studies, typically focused on medium comparisons and declarative learning tasks.

Meanwhile, the results of the study align with the recent studies on the application of multimedia learning principles in immersive environments, which also failed to find conclusive results on reducing cognitive load or improving performance (Strzys et al., 2018; Altmeyer et al., 2020; Thees et al., 2020; Altmeyer et al., 2021; Thees et al., 2022).

Having discussed how the study relates to the previous theoretical and empirical research, it seems necessary to note some other points that may account for the results obtained.

The variability in the amount of learning rounds selected by the participants could be an indirect indicator of perceived task difficulty or a reflection of intrinsic motivation. As noted, intrinsic motivation can be driven by interest or challenge (Ryan & Deci, 2000). When participants complete the first round successfully, they may perceive the task as sufficiently mastered or uninteresting, which reduces their motivation to continue. Conversely, those who find the task more challenging may engage in additional rounds to reinforce their skills. As highlighted in the motivation research, beliefs such as self-efficacy and the perceived value of a task significantly influence effort investment (Feldon et al., 2019). These factors, combined with the cognitive appraisal of task complexity and utility, appear to shape the participants’ decisions to repeat or abandon the subsequent learning stages.

Moreover, the self-regulation strategies of the participants likely influenced the results (Zimmerman, 2002). During the experiment, some participants mentally rehearsed the steps or mimicked them before deciding whether to proceed, demonstrating metacognitive skills. These strategies may have reduced the ECL by enabling the learners to focus on essential procedural aspects, thereby minimizing the complexity of the control condition. This raises questions about the interplay between self-regulation strategies and the cognitive load (de Bruin et al., 2020; Seufert, 2018).

Exploratory findings indicated that the participants were generally satisfied with the presentation and interaction with the content, aligning with (Xue et al., 2019). However, certain items, such as GL1 and GL7, showed higher mean scores in the control condition suggesting areas that warrant further exploration.

4.2. Practical Implications

Given the amount of research that promotes the value of mixed reality for surgical training (Barsom et al., 2016; Lu et al., 2020; Nagayo et al., 2021; Magalhães et al., 2024; Gonnermann-Müller & Krüger, 2024), the question pertains of whether our study calls into question its benefits. One should be balanced, and avoid extremes such as a technocentric approach and a non-informed learner-centered approach. Avoiding the latter extreme prevents the study from being naïve and ignoring various limitations, and then trying to fix them by using educational theories.

The cost and development of MR tools are crucial in surgical training, requiring institutions to assess the value of their financial investment (Parong, 2021). Effective MR development involves collaborations among educators, technologists, and healthcare professionals to develop relevant learning scenarios. While MR tools like the HoloLens offer much affordance, their battery time limits varying with the brightness levels can hinder practicality. A balanced approach is essential to integrate MR tools with traditional training methods.

The study explored the contribution of cognitive and educational psychology theories on surgical training (Sullivan, 2020), specifically by proposing a basic study to adapt the split-attention principle, which is currently underexplored in VR/AR/MR environments (Çeken & Taşkın, 2022), and in the design of a surgical course scenario. Using the Dynamics Guide 365 on the HoloLens 2, we developed procedural learning guides incorporating text and video demonstrations. While not utilized in this study, the possibility of integrating 3D holographic models offers a promising direction for the enhancement of knot-tying instructions.

4.3. Limitations and Future Directions

One of the main limitations of our study is the sample size and the lack of medical students among the participants, which can hinder the generalization of results.

During the study planning phase, we had to address the challenge of not being able to reach the intended sample size. Various constraints prevented us from achieving this goal, despite considering key works such as Ginns’ meta-analysis (Ginns, 2006). Barriers included the conditions of the controlled experiment, which required individual testing sessions, as well as material constraints (a single HoloLens device that required recharging, especially in high light conditions). Also, as it was necessary to control some environmental factors (such as lighting issues and noise disturbances) by managing the logistical limitations (availability of a testing room in the university laboratory) and the time required for each session (at least one hour), an adjustment of our initial approach was imperative. To cope with the fact that the study was not conducted in the context of a classroom or laboratory session (which is achieved in most studies, and provides the possibility to recruit more participants), the decision was to rely on somewhat of a convenience sample, to ensure relevance and feasibility.

As a pilot study, the current research provides food for thought. To address this, one could, for instance, rely on Ginns’ major meta-analysis (Ginns, 2006) and perform sample size calculations (effect size: 0.72, error threshold: 0.05, statistical power: 0.8), which indicate that a minimum of 50 participants are required. It provides perspective on how to assess the split-attention, not only by considering the high element interactivity but also the prior experiences of participants.

The participants’ prior experiences with XR tools were not accounted for, despite some reporting familiarity with such technologies. This potentially influenced the results through a novelty effect (Parong, 2021). Additionally, the influence of prior knowledge is crucial if any effect is to be observed, otherwise there is a risk of reverse results (Ayres & Sweller, 2021).

The absence of a maximum learning time likely introduced variability, affecting performance and cognitive load. Additionally, the absence of a standardized evaluation grid for knot-tying may have affected the reliability and comparability.

Future research should address these limitations and explore the following areas to optimize methodology and enrich findings. First, it is of utmost importance to conduct the research with a larger sample size, and to consider medical students as the population of interest, since they are the ones concerned with such education as future surgeons. Second, the use of standardized and task-relevant medical materials can enhance study comparability. Third, including a broader range of knots, both sliding and non-sliding, would better represent the diversity of surgical skills. Fourth, integrating 3D holograms via the Dynamics 365 Guide software could provide clearer and more interactive demonstrations and guidance in the learning knot-tying. Fifth, objective measures of knot quality, such as tension, could be incorporated. Sixth, a multimodal approach combining subjective and objective data (behavioral and physiological) could provide deeper insights into attention and mental effort, and potentially respond to the critics of the self-reported scales while assessing the cognitive load. Finally, examining the relationship between the cognitive load and self-regulated learning would clarify how the learners adjust their pace, repeat steps, or use metacognitive strategies to manage tasks. It seems that exploring the connection between the cognitive load theory and motivation is of great importance.

5. Conclusions

This study contributes valuable insights into the use of MR for knot-tying simulations. While the anticipated benefits of the integrated formats on the cognitive load were not fully realized concerning the ECL and the GCL, the significant findings related to the intrinsic cognitive load underscore the complexity of participant engagement and the learning peculiarity of procedures. To effectively observe the split-attention effect, future research should address high element interactivity in XR representations. Moreover, the integration of a MR in medical training should consider the cost and development of that technology, and balance it with some traditional methods. Ultimately, further exploration of the interplay between the cognitive load and self-regulated learning is necessary.