Acceptance of an Adaptive Robotic Nursing Assistant for Ambulation Tasks

Abstract

The effective use of nursing assistant robots requires an understanding of key acceptance factors. The study examined the differences in attitudes among 58 nursing students while performing ambulation tasks with and without an Adaptive Robotic Nursing Assistant (ARNA) robot. An ARNA is driven by tactile cues from the patient through a force–torque-measuring handlebar, whose signals are fed into a neuro-adaptive controller to achieve a specific admittance behavior regardless of patient strength, weight, or floor incline. Ambulation tasks used two fall-prevention devices: a gait belt and a full-body harness. The attitude toward the robot included perceived satisfaction, usefulness, and assistance, replacing the perceived ease-of-use construct found in the standard technology acceptance model. The effects of external demographic variables on those constructs were also analyzed. The modified technology acceptance model was validated with the simultaneous estimation of the effects of perceived usefulness and assistance on satisfaction. Our analysis employed an integrated hierarchical linear mixed-effects regression model to analyze the complex relationships between model variables. Our results suggest that nursing students rated the ARNA’s performance higher across all model constructs compared to a human assistant. Furthermore, male subjects rated the perceived usefulness of the robot higher than female subjects.

1. Introduction

Robotics in nursing is a new development that has emerged in the last decade, coinciding with the gradual rise of service robotics [1,2,3]. Physical services rendered by commercial nursing robots found in hospitals can include transport, supply fetching, patient lifting, and ambulation tasks [4,5,6].

Patient fall prevention is an important public health issue. In 2024, medical costs for nonfatal fall injuries reached more than USD 50 billion, and those for fatal fall injuries exceeded USD 754 million in the United States [7]. Various strategies have been implemented to minimize the probability of inpatient falls during ambulation tasks, including patient fall risk assessments, nurse and patient education, and the use of different sensors [8,9,10]. However, these strategies focus solely on precautionary measures, whereas robots offer a valuable technological alternative by offering walking assistance, avoiding obstacles, and enhancing user balance and stability [4,8,11]. Using these devices reduces the need for human supervision and encourages user independence and safety, making them effective in mitigating fall-related injuries. Thus, it is crucial to consider how healthcare professionals perceive this new technology and what factors affect their perception. Research on nurses’ perceptions of such robots, especially in fall-prevention tasks, is still limited to a few pilot studies [4,5,12], and more work in this area is needed.

Our Adaptive Robotic Nursing Assistant (ARNA) was designed and built as a general-purpose nursing platform to handle physical tasks in hospital environments [4,5,13]. One of the primary functions of the ARNA is powered ambulation for recovering patients, with fall-prevention capabilities that enhance user safety [13,14]. In traditional transfer or ambulation tasks, fall-prevention devices such as gait belts are often used. To address fall prevention, the ARNA employs full-body fall-protection harnesses, which include straps secured across different body parts and connected to the robot. These safety features are further supported by a robust control system, known as model-free neuro-adaptive control (NAC), ensuring enhanced reliability and adaptability. A primary challenge in physical human–robot interaction (pHRI) is developing adaptive control strategies for robots that effectively respond to external forces from humans while ensuring physical safety and interaction performance. Conventional approaches like impedance and admittance control have been commonly implemented to manage mechanical compliance and enhance user interaction. However, safety and performance guarantees in pHRI remain unresolved issues [15,16]. Novel control techniques, such as those based on neural networks, including Parallel Neural Network User Interfaces (PNNUIs) [17] and neuro-adaptive control (NAC) [18], as well as advanced methods like Model Predictive Control (MPC) [19], offer potential improvements in robot adaptability. In recent years, NAC has been predominantly utilized as the controller for many of our robotic manipulators [20], with enhancements such as the addition of a human-intent estimator (HIE-NAC) [21], which integrates human-intent modeling into the neuro-adaptive controller. In this study, we used the NAC controller for the ARNA, aiming to achieve smooth, personalized, and responsive pHRI [20]. Further details of this control method are comprehensively described in the Methods section.

The present study assesses nursing students’ attitudes, such as perceived usefulness, perceived assistance, and perceived satisfaction, toward using the ARNA in walking tasks with fall-prevention devices across three scenarios: a gait belt with human assistance, a gait belt with the robot, and a harness with the robot. The primary goal is to examine the efficacy of the ARNA for aiding nurses in ambulating patients in a simulated laboratory environment.

A commonly used framework for evaluating behavioral intention to use technology is based on the technology acceptance models (TAMs) developed by Davis [22,23]. In these models, the psychological constructs of perceived usefulness and perceived ease of use are regarded as key explanatory variables. Later, TAMs were modified to include more external variables (demographic variables, system characteristics, and subjective norms) and constructs (satisfaction, enjoyment, and performance), which had a moderating effect on the relationships among TAM variables [24,25,26]. Several studies that use these models offer important additional insights for our current research. For example, differences in age have been found to influence abilities, attitudes, behavior, and openness to adopt new technologies, since older people have a lower intention to adopt them and are less willing to use them [27,28]. Gender has been suggested to play an important role in human–robot interaction due to different perceptions of robots by males and females. It has been found that men tend to view robots as more useful, show a greater intention to use them in the future, and are more willing to integrate them into their daily lives compared to women [27]. Empirical evidence on differences in perceived usefulness of robots based on race is mixed. Although some studies have found that non-Whites and ethnic minorities tend to exhibit lower usage of health-related technology compared to Whites [28], the results of other studies suggest that it is cultural background that mostly affects attitudes and expectations in terms of robots [29,30].

Structural equation modeling techniques and multiple regression methods are commonly used to obtain parameter estimates in the TAM framework [4,31]. In this study, we propose a novel approach geared toward repeated-measures experimental designs. It includes a hierarchical linear mixed-effects regression model with two random-effects terms to account for correlations both within subjects and within constructs for a given subject [32]. This approach allows for the estimation of parameters by capturing the nested structure of the data and addressing potential intra-individual and intra-construct variability, thus enhancing the robustness of the analysis in small-sample experimental settings.

The paper is structured as follows. Section 2 outlines the robot, hypothesis, experimental design, and methodology. Section 3 reports the experimental results. Section 4 and Section 5 summarize the main findings and conclusions.

2. Methods

2.1. ARNA Description

The ARNA was designed as a general-purpose robotic nursing assistant aimed at alleviating the physical demands on nurses by providing safe and effective support during patient ambulation tasks. Its primary objective is to assist nurses in hospital settings by safely supporting recovering or mobility-impaired patients during ambulation, thereby reducing the physical strain and risk of injury among healthcare providers. Target tasks specifically include patient walking assistance, fall prevention, and supportive interactions during transfers. The robot’s intended user population includes patients who require physical support during walking tasks and nursing personnel who are directly involved in patient ambulation and caregiving tasks.

2.1.1. Interfaces

To assist nurses and patients, the ARNA can be operated in both physical and nonphysical modes. The nonphysical mode of interaction, also called Patient Sitter mode [13], is controlled through remote telemanipulation using a joystick or tablet. For the physical operation mode, a handlebar (Figure 1) is fitted on the ARNA base to assist patients during ambulatory physiotherapy exercises, also called Patient Walker mode [13]. The handlebar sits on an ATI Axia 80 multi-axis force–torque sensor, and the user can guide the robot in different directions by pushing, pulling, or twisting the handlebar. Hence, this interface acts as a support device for walking and, at the same time, collects interaction forces to power the robotic walker through an adaptive admittance controller.

2.1.2. Electronics and Software

The ARNA integrates a network of sensors and processors for simultaneous localization and mapping (SLAM), navigation, and emergency stops (Figure 2). Twelve sonar sensors are positioned around the base to detect obstacles, while six bump sensors provide emergency-stop capability. When sensor readings exceed a defined threshold, an electronic relay disconnects the motor signals, causing an immediate halt. For SLAM and navigation, the ARNA uses a Velodyne Puck LiDAR with four IMU transducers for sensor fusion. An ASUS Xtion camera aids in tasks using a 7-DOF manipulator. This manipulator is a Kinova Gen3 robotic arm, which is integrated into the ARNA platform primarily for object manipulation tasks such as fetching supplies. While this arm is a key feature of the general-purpose ARNA platform, it was not used for the ambulation tasks in this specific experiment. The ARNA’s data processing, SLAM, intent estimation, and navigation are managed by an NVIDIA Jetson TX2 and VersaLogic EPU 4562 Blackbird, with a microcontroller handling data filtering. A Netgear Nighthawk AC1900 router supports networking, interfacing, and remote operation. The ARNA operates within a Robot Operating System (ROS) C++/Python environment, enabling modular and open-source software development. It is important to note that while the ARNA platform is equipped with LiDAR and processors for SLAM and autonomous navigation, this functionality was not utilized in the current study.

2.2. Model-Free Control Architecture

A model-free NAC controller was developed for the ARNA to facilitate smooth, personalized, and natural pHRI. Depending on the user and operating environment, this framework can adapt its control gains online. This ability to personalize and adapt is crucial, as nursing students, as observers, can witness seamless experiences and natural interactions between patients and robots.

The controller assumes that the user will try to guide the robot in a particular direction and continuously interact with the handlebar to maintain or correct the desired path. The outer intent estimation loop takes the force/torque measurements from the handlebar as its input to estimate the intended reference trajectory from the user (Figure 3). These estimated/approximate reference commands are fed to the inner loop of the NAC, which generates the required wheel motor torques to fulfill the goal. The outer intent estimation loop is supported by two neural networks, with one estimating the human gains that vary from user to user and the other exploiting these gains to generate the approximate reference commands. The inner-layer NAC approximates the ARNA’s motion dynamics using another neural network. The weights in the inner-layer neural network are adapted to generate a suitable feedback-linearizing control law, which enforces the tracking error to zero. A summary of the neural network-based model-free controller is presented here for completeness (see [6,20,21] for more details).

The user operates the ARNA through the sensorized handlebar, which generates three torque outputs ($f_h$ = ${\tau }_x,{\tau }_y,{\tau }_z$) corresponding to the interaction around the three axes. These torque signals are filtered for noise elimination and fed to the outer intent estimation loop along with the current pose and velocity vector: $P={\left[\begin{array}{c}x,y,\theta \end{array}\right]}^T$, $\dot{P}={\left[\begin{array}{c}{v}_x,v_y,{\omega }_z\end{array}\right]}^T$. Assuming the user wants to move the ARNA to a desired pose $P_r$ (unknown to the robot and controller), a set of tracking error vectors is defined as $e_d=P_r-P$ and $\dot{e_d}=\dot{P_r}-\dot{P}$ relative to a reference robot position $P_r$. The outer intent estimation loop uses two neural networks (NN I, NN II) to approximate the human model and the intended desired trajectory, respectively, thereby completely circumventing the need to tune the controller for different users. These neural networks are defined as

$$\left[\begin{array}{c}{\widehat{P}}_r\\ {\dot{\widehat{P}}}_r\end{array}\right]={\widehat{S}}^T\varphi \left(\xi \right)$$

(1)

$${\widehat{E}}_h={\widehat{U}}^T\varphi \left(\xi \right)$$

(2)

where ${\widehat{P}}_r$, ${\dot{\widehat{P}}}_r$, and ${\widehat{E}}_h=[{\widehat{K}}_h,{\widehat{D}}_h]$ are the estimated desired position, velocity, and user-specific gains, respectively. The variable $\xi$ denotes the combination of the interaction torques, desired position, and velocity: $\xi ={\left[\begin{array}{c}{f}_h^T,P^T,\dot{P^T}\end{array}\right]}^T$. The unknown weight matrices of NN I and NN II are denoted by $\widehat{S}$ and $\widehat{U}$, respectively. The symbol $\varphi (.)$ describes the sigmoid activation function used to build the neural networks. The estimated error and the discrepancy between the actual and estimated errors are defined as

$$\left[\begin{array}{c}{\widehat{e}}_d\\ {\dot{\widehat{e}}}_d\end{array}\right]=\left[\begin{array}{c}{\widehat{P}}_r-P\\ {\dot{\widehat{P}}}_r-\dot{P}\end{array}\right]$$

(3)

$${\overline{\tilde{e}}}_d=\left[\begin{array}{c}{\tilde{e}}_d\\ {\dot{\tilde{e}}}_d\end{array}\right]=\left[\begin{array}{c}{e}_d\\ e_d\end{array}\right]-\left[\begin{array}{c}{\widehat{e}}_d\\ {\dot{\widehat{e}}}_d\end{array}\right]=\left[\begin{array}{c}{P}_r\\ {\dot{P}}_r\end{array}\right]-\left[\begin{array}{c}{\widehat{P}}_r\\ {\dot{\widehat{P}}}_r\end{array}\right]={\tilde{S}}^T\varphi \left(\xi \right)+{ϵ}_1$$

(4)

where $\tilde{S}$ is the error in the weight estimation. The user-specific filtered error dynamics are given by

$$f_h={\widehat{D}}_h{\dot{e}}_a+{\widehat{K}}_h{e}_a=J({\widehat{E}}_h\odot {\overline{e}}_a)$$

(5)

where $J=[I_6,I_6]$, ⊙ is the Hadamard product and ${\overline{e}}_a=\left[\begin{array}{c}{e}_a\\ {\dot{e}}_a\end{array}\right]$. Defining a sliding-mode error variable $s={\widehat{e}}_d-e_a$, the weight update laws for the two NNs are chosen as

$$\begin{array}{cc}\hfill & \dot{\widehat{S}}=-A\varphi \left(\xi \right)s^{T}Jdiag\left({\widehat{E}}_h\right)+\kappa \left|\right|s\left|\right|A\widehat{S}\hfill \\ \\ \hfill & \dot{\widehat{U}}=-B\varphi \left(\xi \right)s^{T}Jdiag\left({\dot{e}}_a\right)+\kappa \left|\right|s\left|\right|B\widehat{U}\hfill \end{array}$$

(6)

where $\kappa \in R^{+}$ is a design constant, and A and B are design-specific positive-definite matrices. It can be proved that the weight update equations mentioned above make the estimated desired trajectory converge to the actual user-intended desired trajectory [6,20], and the error between them can be reduced by tuning the design scalar $\kappa$.

As the outer loop ensures that the estimated desired trajectory vector ${\widehat{P}}_r$ converges to the true user-intended trajectory $P_r$ (which is unknown), the inner NAC loop uses ${\widehat{P}}_r$ to compute the driving input torques for the motors actuating the wheels of the ARNA. Hence, the objective of the NAC loop is to make sure that a filtered error variable $r={\dot{\widehat{e}}}_d+\rho {\widehat{e}}_d,\rho \in R^{+}$ converges to as small a value as possible. Note that this will lead to tracking error convergence as the filtered error defines first-order stable dynamics.

The motion dynamics of the ARNA can be captured using an Euler–Lagrange model, which is expressed as

$$M\left(P\right)\dot{r}=C(P,\dot{P})r+f(r,\dot{r},P,\dot{P})+{\tau }_d+\tau$$

(7)

where $M(.)$ and $C(.)$ represent the inertia matrix and the Coriolis matrix, respectively. The term $\tau$ denotes the input torques to the motors, ${\tau }_d$ is the torque due to external disturbances, and the function $f(.)$ represents the cumulative modeling uncertainty, which has to be approximated by a neural network, i.e.,

$$\widehat{f}(.)={\widehat{W}}^T\varphi \left({\widehat{V}}^{T}x\right)$$

(8)

where $x=[{\widehat{e}}_d^T,{\dot{\widehat{e}}}_d^T,P^T,{\dot{P}}^T,{\ddot{P}}^T]$, and $\widehat{W}$ and $\widehat{V}$ are the weight matrices corresponding to the output and hidden layers, respectively. It was proved in [6] that a control law of the form

$$\begin{array}{c}\hfill \tau =K_{v}r+K_r\left(\right||\left[\begin{array}{cc}\widehat{W}& 0\\ 0& \widehat{V}\end{array}\right]\left|\right|+c_b)r+\widehat{f}(.)\end{array}$$

(9)

where $c_b$ is the upper bound for the weight matrices, ensures bounded tracking performance for the ARNA. The term $\widehat{f}(.)$ refers to the output of a neural network whose weights are updated in the following way:

$$\begin{array}{cc}\hfill & \dot{\widehat{W}}=M\widehat{\varphi }{r}^T-M{\widehat{\varphi }}^T{\widehat{V}}^{T}x{r}^T-{\kappa }_{1}M\left|\right|r\left|\right|\widehat{W}\hfill \\ \hfill & \dot{\widehat{V}}=Nx{r}^T{\widehat{W}}^T{\widehat{\varphi }}^T-{\kappa }_{1}N\left|\right|r\left|\right|\widehat{V}\hfill \\ \hfill & {\widehat{\varphi }}^T=diag\left(\varphi \left({\widehat{V}}^{T}x\right)\right)(I-diag{\left(\varphi \left({\widehat{V}}^{T}x\right)\right)}^T)\hfill \end{array}$$

(10)

where ${\kappa }_1$ is a design constant, and M and N are user-defined positive-definite matrices.

2.3. Participants

The test users of the ARNA included fifty-eight pre-licensed nursing students (BSN and MEPN) recruited via email and verbal communication from the School of Nursing, University of Louisville. Students were chosen based on their availability, willingness to participate in repeated sessions, and homogeneity in training background, which allowed us to control for variability in prior clinical experience. This population also represents a critical target group for early-stage testing, where the ARNA can be meaningfully assessed before testing in more complex real-world scenarios. Participation of all subjects was both voluntary and confidential. Students received compensation in the form of three hours of clinical credit. The inclusion criteria required that participants were from the School of Nursing and self-reported as being physically and psychologically fit to participate in simulated clinical interactions. The exclusion criteria included any self-reported physical limitations that prevented participants from completing the tasks, as well as any self-reported intellectual, cognitive, or developmental disorders that could interfere with task performance. Participants with prior experience in using robots in clinical settings were excluded to avoid bias due to familiarity with robotic systems.

An external participant, who was not a nursing student, assumed the role of the simulated patient throughout the experiment. Specifically, the simulated patient was trained prior to data collection to ensure a consistent interaction style with the ARNA across all trials. While we acknowledge that experience over time can influence the results of the experiment, the setup was designed to minimize variability on the patient’s level, since the focus of the study was on the ability to detect effort-related signals from future nurse practitioners, rather than the simulated patient’s behavior per se.

The University of Louisville Institutional Review Board granted approval to conduct the experiments under IRB no. 18.0659. We acknowledge that a portion of the sample data was previously used in an earlier publication [4]. In the present study, we analyze the full dataset (58 subjects vs. 38 in [4]), employ a different study design, and extend the analysis to additional variables within a unified framework that accounts for both within-subject and within-construct correlations.

2.4. TAM Hypotheses

In the standard technology acceptance model (TAM) framework, the two primary variables of perceived usefulness (PU) and perceived ease of use (PEOU) have a mediating role in the inter-relationships between the system characteristics (external variables) and the behavioral intentions and eventual usage of a specific technology [23,33]. However, since our nursing student subjects only assisted the simulated patient in walking tasks (holding the gait belt or harness back strap firmly behind the patient), they did not directly control the robot for safety reasons. We replaced the PEOU construct with perceived assistance (PA), which refers to how much support and help users believe the technology provides. In other words, we wanted to measure the extent to which participants viewed the ARNA as assistive [34], consistent with the intelligent systems application of the technology acceptance model [35]. While perceived assistance (PA) is not traditionally a standalone variable in classic TAMs, studies show that users are more likely to adopt technology that provides clear assistance aligned with their needs and values, as this perception supports both perceived usefulness and perceived ease of use [11,12,36]. Furthermore, since the subjects under study were not yet nursing professionals who could provide data on their actual intention to use the technology, the behavioral intention construct of the TAM was replaced with perceived satisfaction (PS) as an objective proxy measure, an approach consistent with that used in similar studies [37,38,39,40]. Figure 4 illustrates the proposed TAM relationship framework, with perceived satisfaction (PS) as the response variable influenced by external factors: perceived usefulness (PU), perceived assistance (PA), tasks, and demographic variables.

During our study, we tested the following hypotheses:

H1. 

Perceived usefulness and perceived assistance are significantly higher when using a robot in ambulation tasks compared to manual controls.

H2. 

Perceived usefulness impactsperceived satisfaction, as users who find the system useful tend to be more satisfied with it.

H3. 

Perceived assistance impacts perceived satisfaction by enhancing the user experience that supports task accomplishment.

We also tested whether demographic factors such as age, race, or gender had significant effects on perceived usefulness and perceived assistance.

2.5. Intervention Scenarios

Three intervention scenarios were designed to assist participants in walking a simulated patient using fall-prevention devices: “Gait Belt+Human” (gait belt with human assistance only, no robot), “GB+ARNA” (gait belt with the ARNA), and “Harness+ARNA” (a full-body fall-protection harness connected to the ARNA). The robotic arm, which is used to fetch objects, was not employed in this experiment, and no operator was involved in controlling it. As shown in Figure 5A, the interventions took place in LARRI’s lab with a simulated hospital room and a clearly marked path that resembled a 46.5-foot-long hospital corridor with one right-hand turn (placed 15 feet away from the starting point). Figure 5B depicts the process of placing a gait belt on a simulated patient. Figure 5C demonstrates the ambulation task with a harness and a robot.

2.6. Instruments and Variables

The perceived usefulness (PU) variable (measured using 4 items on a Likert-type scale ranging from 1 to 5) evaluated in this study was determined based on the core constructs of the technology acceptance model (TAM) [22,23]. The perceived usefulness measure was taken directly from the TAM questionnaire [23], but the referent for the technology was changed to the ARNA. Additional subjective measures, such as perceived assistance (PA) (measured using 4 items on a Likert-type scale ranging from 1 to 5) and perceived satisfaction (PS) (measured using 5 items on a Likert-type scale ranging from 1 to 5), were constructed based on best practices for corresponding measures of perceived assistance [34,35] and perceived satisfaction [38,39,40] by changing the referent for those items to the ARNA. The types of tasks and demographic variables were included based on the proposed experimental design.

A detailed per-item data dictionary for each construct is provided in

Table A1. Itemized questionnaire.

Construct	Item Statement	1	2	3	4	5
Perceived Usefulness (PU)
PU1	ARNA/gait belt enables nurses to complete patient care more quickly
PU2	ARNA/gait belt improves patient care and management
PU3	ARNA/gait belt increases nurses’ productivity in patient care
PU4	ARNA/gait belt makes nurses’ patient care and management easier
Perceived Assistance (PA)
PA1	ARNA/gait belt provides accurate assistant services
PA2	ARNA/gait belt provides reliable assistant services
PA3	ARNA/gait belt provides safe assistant services
PA4	ARNA/gait belt provides convenient assistant services
Perceived Satisfaction (PS)
PS1	I am completely satisfied with using the ARNA/gait belt in patient care
PS2	I feel very confident in using the ARNA/gait belt in patient care
PS3	I found it easy to use the ARNA/gait belt in patient care
PS4	I can accomplish tasks quickly using the ARNA/gait belt
PS5	I believe that using the ARNA/gait belt in patient care will increase the quality of nursing tasks

in Appendix A. Other independent variables included a dummy-coded experimental task variable (Task 1: “GB+Human”; Task 2: “GB+ARNA”; Task 3: “Harness+ARNA” with Task 1 as the reference), gender (reference: “female”), race (reference: “White”), and age.

2.7. Study Design

This study used a block-randomized Latin square design, in which each participant completed all three tasks. In this design, the order of tasks was systematically varied across blocks so that each task appeared exactly once in each position (first, second, third), thereby controlling for order and carryover effects. The first block started with a “Gait Belt+Human” task, the second with a “GB+ARNA” task, and the third with a “Harness+ARNA” task. Recruitment was carried out in separate time periods, and because each period was subject to different practical constraints (e.g., availability of eligible participants, scheduling), the number of participants enrolled in each block was not equal. There were 14 subjects randomized to the first block, 28 to the second, and 16 to the third. Nevertheless, a linear mixed-effects modeling technique to obtain parameter estimates effectively handles asymmetrical or unbalanced sample sizes [41].

2.8. Data Collection Procedure

During the experimental session, each participant was introduced to the ARNA and was provided with a short description of its capabilities in walking tasks. In all three scenarios, participants only ambulated the simulated patient, who actually walked while holding the robot’s handlebar as it tugged the patient to move forward, backward, sideways, or to turn. We provided no direct training for robot operation, except on how to stop it using an emergency button. The robot was also equipped with safety sensors to avoid unexpected collisions with the environment.

In all scenarios, the simulated patient was seated on a bed, and the ARNA was stationed at the room’s entrance. The participants, following scripted study procedures, entered the room and applied the test devices, such as a walking belt (gait belt) or a full-body fall-protection harness, to the simulated patient while standing beside them. The gait belt was fastened at the simulated patient’s waist. The harness was snugly fitted over the hips and waist of the patient, attached to two hooks on the robot’s frame, and secured by a ring located in the middle of the back between the patient’s shoulder blades. Next, the simulated patient walked with the robot along a predefined path from and back to the hospital bed while the participants held either the gait belt or the waistband of the harness behind the simulated patient.

The duration of the walking portion of the experiment in each task lasted 5–10 min, depending on whether the robot was used (10 min) or not used (5 min). The experiment was conducted under the same conditions for all participants. Following each of these three tasks, participants were asked to evaluate their experiences when performing tasks by marking their responses in a questionnaire administered via an iPad using Qualtrics. Data from the Qualtrics questionnaire were stored in a secure online Qualtrics cloud, and only approved research personnel had access to those data.

The questionnaire was adapted from the original TAM paper [22], as well as its extensions and usability studies [5,13,14,42]. The questionnaire’s content validity was established by a panel of this paper’s co-authors, who are faculty members with extensive expertise in assistive technologies in robotics, nursing education, and psychology. The panel included one expert from the Louisville Automation and Robotics Research Institute (LARRI), J.B. Speed School of Engineering, University of Louisville, with over 20 years of experience in human–robot interaction and system design (Dr. Popa); one expert in nursing education from the School of Nursing, University of Louisville, with 30 years of experience in geriatric care and clinical skills training (Dr. Logsdon); and three experts in psychology from Oklahoma State University, with backgrounds in behavioral assessment of human factors in healthcare-related research (Dr. Edwards, Dr. Erdmann, and Dr. Yu). Their diverse expertise ensured a rigorous content validation process. This approach aligns with commonly accepted practices in content validation literature, where a panel of five to ten subject matter experts is considered sufficient for ensuring content validity [43,44]. For internal consistency, Cronbach’s $\alpha$ for all constructs across each task was calculated.

2.9. Statistical Analysis

All statistical analyses were conducted with R software (version 4.4.1) using data exported in CSV format. After the quality checks, de-identification, and data cleaning were performed, Cronbach’s $\alpha$ and descriptive statistics were calculated using R packages (ltm, tidyverse).

An integrated hierarchical linear mixed-effects regression model was used to analyze the relationship between perceived satisfaction (the response variable) and other external variables. Two types of random intercepts were introduced: (i) within subjects and (ii) within constructs for each subject, to account for correlation between observations. This modeling technique explicitly accounts for within-subject correlation by including both fixed and random effects, thereby borrowing strength across repeated observations and accounting for variability more precisely. It was demonstrated that such models provide greater power to detect true effects, making them more robust and generalizable [32,45].

The parameter estimates for the model were obtained using Wald t-tests with the Satterthwaite approximation for degrees of freedom. The most parsimonious model was chosen based on a likelihood ratio test using a chi-square test statistic. The usual assumptions (normality, homoscedasticity, and independence of residuals) and the normality of random effects were tested after model fitting using R packages.

3. Results

3.1. Reliability of TAM Questionnaire

The adapted TAM questionnaire consisted of 13 items that included the PU, PA, and PS domains. The generally accepted threshold for Cronbach’s alpha is between 0.70 and 0.80, which is considered acceptable, while a value above 0.80 is regarded as good [46]. The Cronbach’s $\alpha$ values for the TAM construct scores across each task were calculated and are provided in

Table 1. Cronbach’s $\alpha$ values for TAM construct scores.

Task	PU	PA	PS
Task 1	0.912	0.879	0.913
Task 2	0.868	0.838	0.866
Task 3	0.884	0.828	0.857

PU = perceived usefulness; PA = perceived assistance; PS = perceived satisfaction.

. The results indicate acceptable reliability.

3.2. Characteristics of Subjects and Variables

There were 58 nursing students included in the analysis.

Table 2. Characteristics of subjects and variables $(n=58)$.

Variable		n	%	M	SD
Age				22.57	3.74
	18–20	13	22.41
	21–30	42	72.41
	31–40	3	5.17
Gender
	Male	5	8.62
	Female	53	91.38
Race
	White	34	58.62
	Asian	10	17.24
	Black/African American	14	24.14
PU
	Task 1			3.66	0.96
	Task 2			4.21	0.66
	Task 3			4.33	0.73
PA
	Task 1			4.03	0.80
	Task 2			4.42	0.63
	Task 3			4.53	0.52
PS
	Task 1			3.88	0.94
	Task 2			4.16	0.69
	Task 3			4.21	0.68

PU = perceived usefulness; PA = perceived assistance; PS = perceived satisfaction.

presents a summary of the descriptive statistics for the subjects’ demographic characteristics and other variables. There were 14 subjects assigned to the first sequence, 28 subjects to the second, and 16 to the third. The mean age of the subjects was 22.57 (SD = 3.74). Most subjects were female (91.38%, n = 53). Those identifying as White constituted more than half of the sample (58.62%, n = 34). The second-largest race in the sample was represented by Blacks/African Americans (24.14%, n = 14), followed by Asians (17.24%, n = 10). A comparison of the means of the TAM construct scores across all tasks demonstrated that participants generally regarded the tasks with the ARNA as more useful and providing a higher assistance level than those with only a human holding a gait belt. The mean scores of perceived satisfaction for both tasks, including those with the robot, were also higher than those without the robot. Formal tests of the differences in means between tasks in a repeated-measures experimental design with and without the ARNA are provided in the next section.

3.3. Hierarchical Mixed-Effects Model

Let $Y_{ik}^{PU}$, $Y_{ik}^{PA}$, and $Y_{ik}^{PS}$ represent the mean scores across corresponding survey items for perceived usefulness, perceived assistance, and perceived satisfaction, respectively, for the $i^{th}$ individual in the $k^{th}=1,2,3$ repeated measurement. Let $T_{ijk}$ be the indicator variables corresponding to the $j^{th}=1,2,3$ task of the $i^{th}$ individual for the $k^{th}$ repeated measurement. Consider $T_{ijk}$ equal to 1 if the individual i in the $k^{th}$ repeated measurement performs the task $j=2,3$; otherwise, it is set to 0. Let $c=1,2,3$ label the corresponding construct such as “PU”, “PA”, or“PS”. For the race dummy variable, let $R_{ir}$ equal 1 if $r=1,2$ (“Asian”, “Black/African American”) and 0 otherwise. For the gender dummy variable, let $G_i=1$ if male and 0 otherwise.

Based on these notations, we present the following hierarchical model that analyzes these constructs, tasks, and other external variables and their interrelationships in an integrated manner:

$$\begin{array}{cc}\hfill & Y_{ik}^c={\beta }_0+\sum _{j=2}^3{\beta }_j^c{T}_{ijk}+\sum _{r=1}^2{\beta }_r^c{R}_{ir}+{\beta }_g^c{G}_i+{\beta }_a^{c}Ag{e}_i+e_i+e_{ic}+{ϵ}_{ick}\hfill \end{array}$$

(11)

for $c=1,2$, and

$$\begin{array}{c}\hfill Y_{ik}^3=\sum _{c=1}^2{\gamma }_{3c}{Y}_{ik}^c+e_i+e_{i3}+{ϵ}_{i3k},\end{array}$$

(12)

where $e_i\sim N(0,{\sigma }_v^2)$ and $e_{ic}\sim N(0,{\sigma }_u^2)$ are the corresponding random effects for the $i^{th}$ individual and for the construct within the same individual, and ${ϵ}_{ick}\sim N(0,{\sigma }^2)$ is the random noise uncorrelated with any fixed or random effects for $c=1,2,3$. The nested structure of the random effects implies two levels of covariance: ${\sigma }_u^2+{\sigma }_v^2$ (the covariance within measurements of the same construct for the same individual) and ${\sigma }_v^2$ (the weaker covariance between measurements of different constructs within the same individual). As per these notations, ${\beta }_2^c$ and ${\beta }_3^c$ represent the differences in the effect of tasks 2 and 3 from the baseline task 1 for a specific construct c. These parameter effects in Equation (11) correspond to the first two arrows from the external variables to perceived usefulness and perceived assistance in Figure 4. Similarly, the effects of the perceived usefulness and perceived assistance constructs on perceived satisfaction are given by the ${\gamma }_{31}$ and ${\gamma }_{32}$ coefficients. These relationships are represented by arrows directed toward the perceived satisfaction node.

The relationship between PU and PA shown in the TAM diagram with an arrow connecting two boxes is not explicitly specified in the equations since it is automatically implied by the structure of our model. First, the model allows both constructs to be simultaneously affected by the task settings. This in itself induces marginal correlations between the observed values of these contrasts. Furthermore, the measurements of these constructs within a subject result in a within-subject association. The latter is accounted for by the random effects introduced in the model.

The results of the estimation of the reduced models in the form of (11), but with only one random-effects term across individuals for $c=1,2$, are provided in

Table A2. Effects of tasks and demographic variables on perceived usefulness (PU).

Perceived Usefulness
Predictors	Estimates	CI	p-Value
Task 2	0.56 ***	0.31–0.80	<0.001
Task 3	0.67 ***	0.42–0.92	<0.001
Gender (Male)	0.57 *	0.03–1.11	0.042
Race (Black)	−0.00	−0.37–0.37	0.996
Race (Asian)	−0.02	−0.40–0.43	0.934
Age	0.0047	−0.04–0.05	0.826
Random Effects
${\sigma }^2$	0.46
${\tau }_{00,\mathrm{ID}}$	0.17
ICC	0.28
$N_{\mathrm{ID}}$	58
Marginal $R^2$/Conditional $R^2$	0.151/0.385

* $p<0.05$, *** $p<0.001$.

and

Table A3. Effects of tasks and demographic variables on perceived assistance (PA).

Perceived Assistance
Predictors	Estimates	CI	p-Value
Task 2	0.39 ***	0.18–0.60	<0.001
Task 3	0.50 ***	0.30–0.71	<0.001
Gender (Male)	−0.17	−0.63–0.28	0.456
Race (Black)	0.01	−0.31–0.33	0.936
Race (Asian)	0.05	−0.30–0.41	0.765
Age	0.00	−0.03–0.04	0.841
Random Effects
${\sigma }^2$	0.32
${\tau }_{00,\mathrm{ID}}$	0.13
ICC	0.29
$N_{\mathrm{ID}}$	58
Marginal $R^2$/Conditional $R^2$	0.101/0.361

*** $p<0.001$.

in Appendix A. They suggest that participants rated the robot’s performance in terms of perceived usefulness and perceived assistance in ambulation tasks as higher than with the human and gait belt alone, since all coefficients for robot tasks were positive and highly significant. It was found that variables such as race or age did not have any effect on PU or PA. Gender (male) had a positive, significant effect on PU (${\widehat{\beta }}_g^{PU}=0.57$, p-value = 0.042).

Table 3. Hierarchical mixed-effects model results.

Perceived Satisfaction
Predictors	Estimates	CI	p-Value
PU	0.28 ***	0.13–0.43	<0.001
PA	0.39 ***	0.21–0.57	<0.001
Task 2 × PU	0.56 ***	0.36–0.76	<0.001
Task 2 × PA	0.39 ***	0.20–0.59	<0.001
Task 3 × PU	0.67 ***	0.47–0.87	<0.001
Task 3 × PA	0.51 ***	0.31–0.71	<0.001
Gender (Male)	0.20	–0.17–0.58	0.277
Random Effects
${\sigma }^2$	0.30
${\tau }_{00,\mathrm{ID}}$	0.12
${\tau }_{00,\mathrm{i}3}$	2.32
ICC	0.89
$N_{\mathrm{ID}}$	58
$N_{\mathrm{i}3}$	3

PU = perceived usefulness; PA = perceived assistance; CI = confidence interval. *** $p<0.001$.

summarizes the results of the estimation of model (12) with only one demographic variable: gender (age and race were excluded from the analysis due to insignificance). It can be seen that the effects of PU and PA on PS (Hypotheses H2 and H3) were positive and highly significant (${\widehat{\gamma }}_{3,PU}=0.28$, p-value < 0.001; and ${\widehat{\gamma }}_{3,PA}=0.39$, p-value < 0.001). The estimates of the interaction terms of PU and PA with all tasks for the ARNA (Hypothesis H1) were positive and highly significant, with all p-values < 0.001. The gender variable became insignificant (${\widehat{\beta }}_g^{PS}=0.2$, p-value = 0.277). The total model’s explanatory power was substantial, with $R^2$ = 0.93. The high ICC = 0.89 indicates the appropriateness of including random effects. Thus, both intervention scenarios employing the ARNA were perceived as more useful and as providing higher assistance levels in ambulation tasks compared to the non-robot scenario.

4. Discussion

The present study examined the use of an ARNA in patient ambulation tasks. As previously shown, the integration of robots into direct nursing care not only provides a possible solution to mitigate the physical demands of healthcare workers but also reduces the risk of work-related injuries [4,13,47,48]. The present study further highlighted the ARNA’s capabilities and model-free control architecture (NAC), demonstrating its potential as a robotic assistant. Observations by nursing student participants revealed seamless and natural interactions between humans and the robot, underscoring the ARNA’s effectiveness in supporting ambulation tasks.

The first objective of the study was to demonstrate that nursing students perceived the ARNA as more useful and as providing a higher level of assistance in a simulated patient walking task (with either a gait belt or a harness attached to the robot) than when using a human and gait belt alone (Hypothesis H1). The second objective was to test a set of hypotheses (H2 and H3) that examined the relationship between perceived usefulness (PU), perceived assistance (PA), and perceived satisfaction (PS) as a proxy variable for behavioral intention (BI) to use the technology. Finally, we proposed the use of a hierarchical linear mixed-effects regression model with two random effects to capture the correlations both within subjects and within constructs for a particular subject to produce parameter estimates. This allowed for the use of a smaller sample size than in standard approaches [31,33,36] for the estimation of the interrelationships between different variables and constructs.

Hypothesis H1 was fully supported by our data, indicating that perceived usefulness (PU) and perceived assistance (PA) were significantly higher when using the ARNA in ambulation tasks compared to manual controls. These findings are consistent with previous ARNA studies [4,5,6,13] that highlight the potential positive impact of introducing the robot into nursing practices to support ambulation tasks.

Hypotheses H2 and H3 were also fully validated by our data. Our findings demonstrated the presence of direct and positive relationships between perceived usefulness (PU) and perceived satisfaction (PS), as well as between perceived assistance (PA) and perceived satisfaction (PS). Although the construct of perceived assistance (PA) is not widely used, we incorporated it into our TAM in place of perceived ease of use (PEOU), as our experimental design emphasized the extent to which the robot provides meaningful assistance. These results are consistent with findings from previous studies on the potential importance of perceived assistance (PA) [36,49] and user satisfaction [50,51] in the extended TAM framework that may affect behavioral intention to use a new technology.

With regard to demographic variables, we demonstrated that in ambulation tasks, men perceived the ARNA as more useful than women in a reduced model with one random effect. This finding is consistent with previous research based on gender role and stereotype theories [52,53] and can be explained by the tendency to view robots as tools for tasks traditionally linked to masculine roles, which can enhance perceptions of their usefulness among males [54,55]. However, the limited gender variability in our sample (53 females vs. 5 males) constrained the statistical power of this result in the integrated model, rendering the effect non-significant. None of the other demographic variables—such as race or age—were found to significantly affect the variables’ mean scores.

The hierarchical mixed-effects linear model used in this research demonstrated some advantages over traditional SEM techniques, which require large sample sizes [56], or over multivariate regression models, which are unsuitable for repeated-measures experimental designs. Our approach can be readily applied to the TAM framework to evaluate interrelationships among constructs and external variables.

The current study highlights some common limitations and important directions for future research. The first limitation is the size of the current ARNA, which is too large for use in tight spaces. Although the harness and frame offer protection during potential falls, implementing an immediate automatic stop mechanism is necessary for an effective fall response. To address this, we recently designed a new handlebar equipped with pressure sensors to detect hand presence and grip force. This enables the robot to halt immediately during a fall, enhancing patient safety during ambulation tasks.

Next, the use of a single simulated patient and a sample of nursing students may limit the study’s generalization to broader clinical settings, where various healthcare professionals work with diverse patient populations. Future research should include patients with different profiles to enhance the generalization of the findings and evaluate potential effects related to user experience. Nursing students may not accurately reflect the perspectives, clinical decision-making, and judgment of experienced nursing professionals, primarily due to their limited exposure to real-world healthcare settings, especially in matters related to patient safety. As a result, in our experiments, perceived satisfaction was used only as a proxy for actual intent to use the robot. Further research is needed in this area to evaluate perceptions of the technology by both patients and practicing nurses across various clinical roles to validate the acceptance of the ARNA in more complex and dynamic care environments.

At this stage of the research, this study was exploratory in nature, aiming to assess system feasibility and gather preliminary data on user interactions and physiological responses. As such, a formal sample size calculation was not performed, and an ideal number of participants could not be predetermined due to the novelty of the setup. However, the chosen sample size was consistent with similar pilot studies in the field [4,13] and was sufficient to identify practical challenges and inform future findings. Generally, as demonstrated previously in similar studies, a larger sample size with fifty-one participants per group in repeated-measures experimental designs to detect effect sizes of 0.4 or greater may enhance statistical power and improve the reliability and robustness of the findings [57].

Although this study did not find significant effects for race, age, or gender-based differences in an integrated model, we acknowledge that these and other contextual factors (for example, technology anxiety, work experience, or other subjective norms [31,58]) may still warrant further exploration through other instruments, such as focus groups, interviews, or open-ended survey questions, to enrich the understanding of participants’ attitudes and perceptions, particularly regarding robotic assistants.

5. Conclusions

The present study explored nursing students’ attitudes (perceived usefulness, perceived assistance, and perceived satisfaction) toward accepting the ARNA when performing walking tasks with fall-prevention devices (a gait belt and a full-body protection harness). The overall findings showed that participants perceived walking tasks with the robot as more useful and as offering a higher level of assistance compared to tasks performed with a human alone. This, in turn, resulted in a greater level of satisfaction from using the technology. These results support previous findings on attitudes and factors that affect users’ acceptance of new technology. They can provide a basis for ongoing discussions regarding the acceptance of service robots by medical personnel engaged in daily caregiving tasks [4,11,13].

The effects of demographic variables (sex, race, and age) on perceived satisfaction and assistance in a full model were found to be non-significant, likely due to the limited variability in these characteristics within the sample and the relatively small overall sample size in our experiment, which aimed to evaluate the impact of such variables on psychological constructs.

To analyze the survey-based quantitative data, we used an extended TAM with external variables, replacing the standard perceived ease-of-use construct with perceived assistance within the proposed relationship framework. The use of the perceived assistance construct relates to how much help the robot provides in a specific context, for example, in any ambulation, lifting, or guidance tasks. That perception of usefulness and assistance, as shown in previous studies [11,12,36], may, in turn, influence perceived satisfaction, which can serve as a proxy for the behavioral intention to adopt the new technology. This framework can be applied to similar contexts involving other assistive robots designed for specific functions. A hierarchical mixed-effects model with two random terms was used to obtain the parameter estimates. This modeling approach enabled the analysis of interrelationships among TAM variables in an integrated way, while also accommodating a relatively small sample size, controlling for individual differences, and facilitating efficient within-subject comparisons across different conditions.

We believe our work can contribute to a better understanding of the factors that affect the acceptance of new technology, ultimately supporting the creation of a healthier, safer, and less burdensome work environment.