Are Arthroplasty Patients Ready for Artificial Intelligence–Integrated Care?

Abstract

Background

Total hip arthroplasty and total knee arthroplasty patients generate high volumes of perioperative communication, yet the capacity for timely, personalized responses is constrained by growing surgical volumes. Generative artificial intelligence (AI)–powered digital medical assistants (DMAs) may help address this gap, but patient perceptions and willingness to use them remain poorly understood. This study evaluated perceptions and preferences regarding a generative AI-powered DMA among total hip arthroplasty and total knee arthroplasty patients.

Methods

A prospective cross-sectional survey was administered to 101 arthroplasty patients at a single institution. The 13-item questionnaire assessed prior generative AI experience, healthcare priorities, attitudes toward DMA usefulness and trustworthiness (Likert scale), and scenario-based preferences for addressing common postoperative concerns. Composite DMA attitude scores and subgroup differences were analyzed using nonparametric methods and logistic regression.

Results

Patients prioritized accuracy of health information and provider relationships over rapid assistance. Willingness to use the DMA (44.5% agree/strongly agree) and trust in its accuracy (44.5%) were mixed, while preference for DMA over human staff was low (9.1% agree/strongly agree). Receptivity was highest for supportive functions such as reminders and encouragement (54.6%). Across postoperative scenarios, calling the office was the dominant choice (56.6%-72.7%). Prior generative AI experience was the strongest predictor of more favorable DMA attitudes. Older age was associated with a lower likelihood of choosing the DMA for pain and medication concerns.

Conclusions

Arthroplasty patients showed mixed but generally neutral-to-slightly positive attitudes toward a generative AI-powered DMA, with a consistent preference for human-mediated communication.

Introduction

Total hip arthroplasty and total knee arthroplasty (THA and TKA) are among the most common and rapidly growing surgical procedures worldwide. Patients undergoing these joint replacements typically have numerous questions and concerns spanning the entire perioperative period, from preoperative preparation through postoperative recovery, including pain management, wound care, activity limitations, and expectations for rehabilitation. Effective postoperative communication has been shown to reduce patient anxiety and improve clinical outcomes [1]. However, delivering comprehensive guidance and reassurance to patients across this continuum of care can be challenging in practice, especially as surgical volumes increase and healthcare providers face time constraints. Throughout the postoperative period, patients often experience new symptoms and heightened anxiety while trying to follow discharge instructions, and surgeons must promptly address concerns to prevent complications, avoidable readmissions, and dissatisfaction [2,3]. Together, these factors may lead to a high volume of patient-initiated contacts before and after surgery (eg, phone calls, portal messages), highlighting the importance of prompt, accurate, and helpful responses at every stage of care.

Recent advances in generative artificial intelligence (AI) offer a potential approach to improving perioperative communication. Generative AI chatbots powered by large language models like OpenAI's ChatGPT have emerged as widely accessible tools that can provide instant, conversational health information throughout the perioperative continuum [4]. These platforms have the potential to function as digital medical assistants (DMAs), providing continuous 24/7 availability, delivering consistent responses, and, as the models continue to improve, requiring less direct day-to-day oversight from clinicians [5]. Early studies suggest that generative AI chatbots can answer common patient questions with high accuracy and clarity [6,7]. For example, in the context of orthopaedic arthroplasty, ChatGPT has demonstrated the ability to generate detailed answers to patients' perioperative questions about surgery and recovery, with performance often comparable to expert counseling [8].

Despite the technological potential of DMAs, patient attitudes, trust, and preferences regarding these tools remain unclear. Healthcare is inherently personal, and many patients may be hesitant to rely on a DMA for medical advice in place of direct physician communication, particularly during the vulnerable perioperative period when concerns can feel urgent and consequential. Given these gaps, it is crucial to understand patient perspectives before integrating generative AI-powered DMAs into routine perioperative care. This study evaluated patient perceptions and preferences regarding the use of a generative AI-powered DMA for perioperative communication in TKA and THA. We aimed to assess patients' prior experience with generative AI, their attitudes toward the usefulness and trustworthiness of a DMA in the perioperative orthopaedic setting, and their preferred approach (traditional communication pathways vs DMA) when facing common postoperative issues, including wound problems, pain management, medication concerns, activity guidance, or possible infection.

Material and methods

Study design and setting

Institutional review board approval was obtained at our institution prior to the commencement of the study. This was a prospective cross-sectional survey study conducted at a single institution and involving patients treated by a single arthroplasty surgeon. Eligible participants were adults aged 18 years or older seen in our clinic either during the preoperative evaluation period for primary THA or TKA or during routine postoperative follow-up within 8 weeks of primary THA or TKA. By surveying patients during both the preoperative and postoperative periods, we aimed to capture a broader range of patient perceptions and assess how perspectives may differ by the stage of care. Once eligible participants were identified, inclusion criteria solely required completion of the full survey instrument. Patients selected for participation who submitted incomplete surveys were excluded from the analysis. All surveys were completed in person during the clinic visit.

Survey instrument

Participants completed a 13-item, study-specific questionnaire that evaluated prior experience with generative AI, patient priorities in healthcare delivery, attitudes toward the usefulness and trustworthiness of a DMA, and preferences for addressing common postoperative concerns through traditional communication pathways vs a DMA (Appendix 1). Prior experience with generative AI was assessed as a binary variable, yes/no item, which did not capture the type of platform participants had used or to what extent. In the context of our study, generative AI referred to widely recognized platforms such as ChatGPT, Google Bard, and Copilot. This is also highlighted in Appendix 1. Collected demographic variables included age, gender, highest level of education, surgical procedure type (THA vs TKA), and stage of care (preoperative vs postoperative). Attitudinal items (Questions 6-9) used a 5-point Likert scale (strongly disagree to strongly agree) to assess willingness to use a DMA, trust in its accuracy, preference for DMA communication relative to human staff, and perceived value of postoperative reminders or encouragement. Before answering the attitudinal items, all study participants were instructed to read a brief description of the DMA's premise, which was provided within the survey instrument (Appendix 1). In this description, participants were asked to assume that their surgeon had a DMA that was trained on the surgeon's knowledge and preferences as well as being validated to answer commonly asked questions before and after hip or knee replacement surgery. They were also informed that the DMA operates 24 hours a day and can respond within seconds.

Questions 10-13 consisted of scenario-based preference items assessing patient responses to common postoperative concerns, including wound care, infection, pain and medication management, and activity guidance. Response options—including contacting the office, using a DMA, conducting an internet search, visiting a clinic, presenting to the emergency department, or taking no action—were selected to reflect common real-world care-seeking behaviors following arthroplasty and to compare traditional pathways with emerging digital alternatives. To our knowledge, this survey is among the first instruments developed to assess patients' perceptions of generative AI-powered DMAs in arthroplasty care. However, it has not been formally validated and was only intended for exploratory assessment of patient perceptions and preferences.

Statistical analysis

Demographic characteristics and baseline survey responses were summarized using descriptive statistics. Continuous variables are presented as means with standard deviations, while categorical variables are reported as frequencies and percentages. Question 5 was analyzed using the Friedman test. Likert-scale DMA attitude items (Q6-Q9) were analyzed as ordinal variables using nonparametric methods. Internal consistency was assessed with Cronbach's α, and a composite DMA attitude score was compared across groups using the Mann-Whitney U test or Kruskal-Wallis test, as appropriate. Individual responses to Q6-Q9 were also compared by prior generative AI use, using Mann-Whitney U tests. Spearman's rank correlation coefficient was calculated to assess the relationship between age and DMA attitude responses (Q6-Q9) and to examine correlations among the individual DMA attitude items.

For questions 10-13, responses were dichotomized into DMA vs non-DMA. Chi-square P values are reported for Q10, Q12, and Q13, while Fisher's exact test was used for Q11 due to an expected cell count of <5. Age differences between DMA and non-DMA response groups for questions 10-13 were analyzed using the Mann-Whitney U test. Age, gender, education, procedure type, stage of care, and prior generative AI use were included as covariates in adjusted ordinal logistic regression models for Q6-Q9 and adjusted binary logistic regression models for Q10-Q13. A P value <.05 was considered statistically significant. Statistical analyses were conducted with IBM SPSS Statistics (IBM Corp., Armonk, NY, USA).

Results

A total of 101 surveys were collected, and 99 were included in the final analysis after exclusion of 2 incomplete responses. Baseline demographic and clinical characteristics are presented in Table 1. The cohort had a mean age of 66.6 years and included 37 preoperative and 62 postoperative patients, with most undergoing TKA. A minority of respondents reported prior use of generative AI.

Table 1.

Participant characteristics.

Variable	n (%) or mean ± SD
Age (y)	66.56 ± 9.17
Gender
Female	40 (40.4%)
Male	59 (59.6%)
Education
High school/general educational development	30 (30.3%)
College/university	44 (44.4%)
Professional/vocational school	4 (4.0%)
Master's	14 (14.1%)
Doctorate	7 (7.1%)
Prior generative AI use
Yes	37 (37.4)
No	62 (62.6)
Stage of care
Preop	37 (37.4)
Postop	62 (62.6)
Procedure type
THA	31 (31.3)
TKA	68 (68.7)

SD, standard deviation.

Values are presented as mean ± standard deviation for continuous variables and n (%) for categorical variables.

Patient ranking of healthcare-related priorities (Q5) demonstrated a significant difference in perceived importance across the 3 ranked domains (Friedman χ² = 43.414, P < .001). Mean ranks showed that accuracy of health information was the highest patient priority, followed by a personal relationship with the healthcare provider, whereas quick assistance ranked lowest. This pattern was consistent with the response distributions shown in Table 2. For Q5, chi-square analyses showed no significant associations between respondent characteristics and which item was ranked most important (all P > .05).

Table 2.

Ranked healthcare-related priorities (Q5) and overall Friedman test.

Q5. Priority item	Rank 1 n (%)	Rank 2 n (%)	Rank 3 n (%)	Overall priority order
Accuracy of health information provided	48 (48.5)	40 (40.4)	11 (11.1)	1 (highest)
Personal relationship with healthcare provider	41 (41.4)	32 (32.3)	26 (26.3)	2
Quick assistance in my medical care	10 (10.1)	27 (27.3)	62 (62.6)	3 (lowest)
Friedman test (χ2 = 43.414)	P < .001

Q5 was a forced-ranking item (1 = most important; 3 = least important). Mean ranks are from the Friedman test; lower mean rank indicates higher priority. Friedman χ² and P values test whether rank distributions differ across the three domains.

Across respondents, willingness to use the DMA for accurate care information (Q6) and trust in the DMA to provide accurate information (Q7) were mixed, with 44.5% agreeing/strongly agreeing for both items and roughly one-third remaining neutral (32.3% and 31.3%, respectively) (Table 3). Preference to use the DMA rather than contact human office staff for medical education (Q8) was notably low, with 60.6% disagreeing/strongly disagreeing and only 9.1% agreeing/strongly agreeing. In contrast, respondents were more receptive to supportive functions, with 54.6% agreeing or strongly agreeing that DMA encouragement and reminders would increase engagement in recovery (Q9) (Table 3). A composite DMA attitude score was computed from Q6-Q9. The 4-item scale demonstrated good internal consistency (Cronbach's α = 0.803) (Table 4). Prior use of generative AI was associated with more favorable DMA attitudes, as indicated by the composite score (P = .003). Prior generative AI use was associated with higher agreement for individual DMA attitude items: Q6 (P = .002), Q8 (P = .031), and Q9 (P = .016), whereas Q7 was not statistically significant (P = .075) (Table 5). No significant differences in the composite attitude score were observed by gender, procedure type, or education level. Age was not significantly correlated with any of the Q6-Q9 item responses (all P > .05). DMA attitudes did not differ significantly between preoperative (n = 37) and postoperative (n = 62) patients for any individual item (Q6-Q9, all P > .05) or the composite attitude score (P = .450).

Table 3.

Attitudes toward the digital medical assistant (DMA) (Q6-Q9).

Item	Strongly agree, n (%)	Agree, n (%)	Neutral, n (%)	Disagree, n (%)	Strongly disagree, n (%)
Q6. Would use DMA for accurate care information	15 (15.2)	29 (29.3)	32 (32.3)	16 (16.2)	7 (7.1)
Q7. Trust DMA for accurate care information	9 (9.1)	35 (35.4)	31 (31.3)	17 (17.2)	7 (7.1)
Q8. Prefer DMA over human staff for education	2 (2.0)	7 (7.1)	30 (30.3)	39 (39.4)	21 (21.2)
Q9. DMA reminders increase engagement	19 (19.2)	35 (35.4)	22 (22.2)	16 (16.2)	7 (7.1)

Values are n (%) for each Likert response category. Full item wording: Q6 “I would use the digital medical assistant to provide me with accurate information about my care.” Q7 “I would trust the digital medical assistant to provide me with accurate information about my care.” Q8 “I would prefer to use the digital medical assistant instead of contacting a human medical assistant or office staff for medical education.” Q9 “Would encouragement and reminders from a DMA encourage you to be more engaged in your recovery?” Percentages may not sum to 100 due to rounding.

Table 4.

DMA attitude composite (Q6-Q9): reliability, descriptives, and subgroup associations.

Scale	Items	Cronbach’s α	Minimum-maximum
Q6, Q7, Q8, Q9 as single “DMA acceptance”	4	0.803	1-5

Subgroup comparisons for the DMA acceptance scale
Predictor	Test	Group (n)	P value
AI-use	Mann-Whitney	Yes (37)	.003
		No (62)
Gender	Mann-Whitney	Female (40)	.137
		Male (59)
Stage of care	Mann-Whitney	Preop (37)	.450
		Postop (62)
Procedure type	Mann-Whitney	THA (31)	.970
		TKA (68)
Education level	Kruskal-Wallis	(5 groups)	.423

DMA attitudes were calculated as a composite score from Q6 to Q9. Internal consistency is reported as Cronbach's alpha (α). Group comparisons used nonparametric tests (Mann-Whitney U for 2-group comparisons; Kruskal-Wallis for education). Values are mean ± standard deviation unless otherwise indicated.

Table 5.

Association of prior generative AI use with individual DMA attitude items (Q6-Q9).

Outcome	AI-use group (n)	P value
Q6. Would use DMA for accurate care info	Yes (37)	.002
	No (62)
Q7. Trust DMA for accurate care info	Yes (37)	.075
	No (62)
Q8. Prefer DMA over human staff for education	Yes (37)	.031
	No (62)
Q9. DMA reminders increase engagement	Yes (37)	.016
	No (62)

Prior generative AI use was compared with responses to Q6-Q9 using Mann-Whitney U tests. P values reflect between-group differences in Likert item responses.

Furthermore, Spearman correlations showed no association between age and DMA attitudes (all P > .05). In contrast, DMA attitude items were positively intercorrelated (ρ = 0.343-0.755, all P < .001), with the strongest correlation between willingness to use the DMA and trust in the DMA (ρ = 0.755) (Table 6). Kruskal-Wallis testing showed no differences in Q6-Q9 responses or the composite DMA attitude score across education levels (all P > .05). Ordinal logistic regression showed that prior generative AI use was significantly associated with higher agreement for Q6 (estimate = 1.140, P = .008), Q7 (estimate = 0.872, P = .041), and Q9 (estimate = 0.921, P = .030). Gender was also significantly associated with Q7 (estimate = −0.829, P = .040), indicating that males had lower odds of reporting higher trust than females. Gender was the only significant predictor for Q8 (estimate = −0.843, P = .040), indicating lower odds of preferring the DMA over human staff among males compared with females. Across all models, age, education, stage of care, and procedure type were not significant covariates.

Table 6.

Spearman correlations among DMA attitude items (Q6-Q9) (N = 99).

Item	Q6	Q7	Q8	Q9
Q6. Use DMA for accurate care info	1.000	0.755	0.517	0.412
Q7. Trust DMA for accurate care info	0.755	1.000	0.539	0.378
Q8. Prefer DMA over human staff for education	0.517	0.539	1.000	0.343
Q9. DMA reminders increase engagement	0.412	0.378	0.343	1.000

Values are Spearman's rho (ρ) (two-tailed). All pairwise correlations among Q6-Q9 were statistically significant (P < .001).

Across postoperative concern scenarios (Q10-Q13), calling the office was the most common first choice (56.6%-72.7%), while chatting with a DMA was the second most common option in most scenarios (8.1%-28.3%), with the highest rate for Q13 (Table 7). In-person clinic visits and emergency department utilization were selected less frequently overall (Table 7). Prior generative AI use was associated with a greater likelihood of choosing the DMA option for wound care (Q10; P = .025), pain/medication concerns (Q12; P = .040), and activity concerns (Q13; P = .036), but was not significantly associated with choosing the DMA for infection concerns (Q11; P = .467) (Table 8). Stage of care was significantly associated with selecting the DMA option only for infection concerns (Q11; P = .024), with DMA selection observed only among postoperative respondents, while no significant associations were observed for wound care, pain/medication, or activity concerns (Q10, Q12, and Q13; all P > .05) (Table 8).

Table 7.

Preferred first action for common postoperative concerns (Q10 - Q13).

Scenario	Call the office, n (%)	Chat with DMA, n (%)	Internet search, n (%)	Drive to ortho clinic, n (%)	Drive to ED, n (%)
Q10. Wound care concern	72 (72.7)	14 (14.1)	-	10 (10.1)	3 (3.0)
Q11. Infection concern	56 (56.6)	8 (8.1)	-	18 (18.2)	17 (17.2)
Q12. Pain/medication concern	72 (72.7)	19 (19.2)	1 (1.0)	6 (6.1)	1 (1.0)
Q13. Activity concern	67 (67.7)	28 (28.3)	2 (2.0)	2 (2.0)	-

ED, emergency department.

Values are n (%) for each response option. Full item wording: Q10 “If you had concerns about postop wound care…” Q11 “If you had concerns about postop infection…” Q12 “If you had concerns about pain and medication…” Q13 “If you had concerns about postop activity…” Percentages may not sum to 100 due to rounding.

Table 8.

Patient-factor associations with choosing DMA vs non-DMA for postoperative concern scenarios (Q10-Q13).

Predictor	Q10. Wound care concern	Q11. infection concern	Q12. pain/medication concern	Q.13 activity concern
Used generative AI (yes/no)	0.025 / 0.036	0.441 / 0.467	0.040 / 0.063	0.036 / 0.042
Gender (male/female)	0.840 / 1.000	0.632 / 0.711	0.725 / 0.779	0.755 / 0.793
Stage of care (postop/preop)	0.054 / 0.087	0.023 / 0.024	0.561 / 0.584	0.256 / 0.293
Procedure type (THA/TKA)	0.474 / 0.539	0.351 / 0.429	0.978 / 1.000	0.911 / 1.000
Education (5 levels)	0.227	0.343	0.099	0.053

Q10-Q13 were dichotomized as DMA vs non-DMA (DMA defined as selecting “Chat with a digital medical assistant”; non-DMA defined as all other choices). For 2 × 2 comparisons, values are reported as Pearson χ²P value / Fisher's exact (2-sided) P value. Education comparisons are 5 × 2; Pearson χ²P values are shown (Fisher's exact not reported in this output).

Across these scenario questions, gender, procedure type, and education level were not significantly associated with DMA selection (all P > .05). Age differed by response group only in the pain/medication scenario, with older patients more likely to choose a non-DMA option (Q12; P = .026), whereas no significant age differences were observed in the wound care, infection, or activity scenarios (Table 9). Binary logistic regression models for DMA vs non-DMA did not identify significant predictors for Q10, Q11, or Q13 (all P > .05). For Q12, age was significantly associated with DMA selection (B = −0.069, P = .042), indicating that increasing age was associated with a lower likelihood of choosing the DMA option for pain/medication concerns; all other covariates were not significant.

Table 9.

Age differences by choosing DMA vs non-DMA for postoperative concern scenarios (Q10-Q13).

Scenario	Group sizes (DMA vs Non-DMA)	P value
Q10. Wound care concern	14 vs 85	.872
Q11. Infection concern	8 vs 91	.472
Q12. Pain/medication concern	19 vs 80	.026
Q13. Activity concern	28 vs 71	.524

Age was compared between DMA and non-DMA groups for each scenario using the Mann-Whitney U test. DMA indicates selecting “Chat with a digital medical assistant”; non-DMA indicates all other response options.

Discussion

This cross-sectional survey of THA and TKA patients revealed that, although attitudes toward a generative AI-powered DMA were mixed and generally neutral to slightly positive, patients in this predominantly older cohort consistently preferred traditional, human-mediated communication throughout the perioperative period. When asked to rank healthcare priorities, patients placed the greatest importance on accuracy of health information, followed by maintaining a personal relationship with their provider, whereas rapid assistance ranked lowest. This pattern suggests that, within this cohort, accuracy and interpersonal aspects of care were prioritized over speed alone. Consistent with this, most patients preferred to contact office staff rather than engage a DMA for medical education, and across all postoperative concern scenarios, calling the office remained the dominant first choice. However, this finding is best interpreted as a first step rather than a definitive answer, as patients may take several actions at once, such as calling the office while also using a DMA or searching the internet or going through the options in a stepwise manner. Although willingness to use the DMA and perceived value of its supportive functions, such as encouragement and reminders, were somewhat favorable, patients still showed a clear preference for human involvement. This preference was especially evident when concerns were perceived as higher risk, such as possible infection. These patterns suggest that the current arthroplasty population, which skews older and may have limited prior exposure to AI tools, has not yet reached a comfort threshold for routine DMA adoption.

The most consistent factor associated with more favorable DMA attitudes in our cohort was prior experience with generative AI, which was associated with an increased willingness to select a DMA across wound care, pain management, and activity scenarios. In adjusted ordinal logistic regression, prior generative AI use was also associated with higher agreement for trust (Q7), suggesting that familiarity with generative AI may be linked to greater confidence in DMA-provided medical information after accounting for other covariates. Overall, DMA attitudes were broadly similar across patient subgroups, with no significant associations between the composite attitude score and age, education, gender, stage of care, or procedure type. Likewise, ordinal logistic regression showed no significant effects of age, education, stage of care, or procedure type on Q6-Q9 responses, whereas prior generative AI use was associated with more favorable responses for Q6, Q7, and Q9, and male sex was associated with lower trust and lower preference for the DMA over human staff. The main exception was that older patients were less likely to select the DMA for pain and medication concerns. These findings align with the broader pattern in our cohort, in which patients across nearly all demographic strata continued to prefer direct human contact over AI-mediated communication, particularly in scenarios perceived as higher risk. For the typical arthroplasty patient, a DMA may currently be best positioned as a supplementary tool for lower-acuity informational and supportive needs, such as postoperative activity guidance and rehabilitation reminders, rather than as a replacement for established communication pathways in higher-stakes situations, such as suspected infection, where patients prefer direct human contact or in-person escalation.

Prior work suggests that patient trust in AI-generated responses varies, with many patients reporting changes in comfort once AI involvement is known and others expressing baseline uncertainty about relying on AI for postoperative questions [9,10]. Similar patterns have also been observed in studies examining patients' perceptions of AI-generated responses to frequently asked questions about THA and TKA. Patients were often satisfied with the initial response, but satisfaction declined when they became aware that the responses were AI-generated, suggesting that disclosure of AI involvement can affect feelings of reassurance and trust even when the message itself remains unchanged [11]. Another study similarly found that labeling content as AI-generated influenced individuals' evaluation of the information, with perceptions of quality and credibility shifting despite the content being identical [12]. How a DMA is introduced and disclosed to arthroplasty patients may be just as important as the quality of its responses in shaping willingness to engage with it. Despite this hesitation, AI has demonstrated its ability to provide accurate surgeon-approved responses in THA and TKA care settings. Prior studies evaluating AI-generated answers to commonly asked questions about TKA have found that these responses can be accurate and useful as a patient education resource [13,14]. It is possible that, as time goes on and patients become more comfortable and confident in generative AI's capabilities, adoption will increase.

In the clinic, a DMA could benefit workflows by providing patients with timely information and reducing communication delays for common preoperative and postoperative concerns. Arthroplasty surgeons manage large patient volumes and may not always be able to respond promptly via patient portals or phone lines; a DMA could serve as an additional layer of support for addressing routine questions between visits. From a perioperative perspective, chatbots have generally been well received by patients and have been associated with high satisfaction and perceived knowledge gains, supporting the idea that these tools can potentially play a role in perioperative communication when used appropriately [15]. In the postoperative setting, an AI-assisted phone follow-up system achieved call connection and follow-up completion rates comparable to manual follow-up while saving substantial staff time by automating calls and report generation [16]. Patient attitudes toward a DMA may also vary depending on what the patient values most in their healthcare experience. Given that patients prioritized accuracy of information and maintaining a personal relationship with their provider over rapid assistance, implementation strategies should reflect those preferences. With this in mind, a DMA can be introduced in a way that supports each priority. For patients who prioritize accuracy, a DMA can be framed around surgeon-approved education and consistent messaging, reducing confusion that often arises from online searches or mixed advice. For patients who value maintaining a connection with their care team, a DMA can be introduced as a surgeon-trained extension of the practice rather than a replacement for human contact. For those who also value faster access, a DMA can provide timely responses to routine questions between visits. Moving forward, if DMAs are implemented, they should be introduced in a patient-friendly manner that supports education and efficient communication while preserving access to human contact when needed. Our findings suggest that greater familiarity with generative AI may increase willingness to engage with these tools, as prior generative AI experience was consistently associated with more favorable attitudes toward the DMA and greater selection of the DMA in several postoperative scenarios. As patient exposure to generative AI continues to grow, perceptions of this technology and comfort may also increase, allowing DMAs to play a more meaningful supportive role in patient care.

Limitations

This study has several limitations to consider when evaluating the results. Our study was conducted at a single orthopaedic institution and included patients treated by a single arthroplasty surgeon. Although this ensured consistency within our patient sample, it may limit the extent to which these findings can be generalized to other populations. Patient perceptions of generative AI may differ across practice settings, geographic regions, or patient populations with varying levels of baseline technology exposure and therefore may not fully reflect perspectives in community practices or among patients with lower digital literacy. Another limitation is that prior generative AI experience was only measured as a yes/no response, and this did not allow us to capture important variables such as the generative AI model used, frequency of use, and context of use. This limited our ability to distinguish between those who would have reported occasional use from those who reported more regular use. Furthermore, participants may have misclassified prior AI exposure, either by mistaking non-generative automated tools for generative AI or by failing to recognize that a tool they had used was generative AI. This may have led to inaccurate reporting of prior generative AI experience.

Additionally, although the sample size was adequate for analysis, it may not have been sufficient to capture the full range of viewpoints within the broader THA and TKA population. A larger sample size would have improved power and our data analyses. Moreover, patient responses may have been influenced by unmeasured confounders, such as health literacy, prior experience with technology, baseline anxiety, or overall health status, which could independently shape perceptions and decision-making, regardless of the factors analyzed. It is also important to recognize that surveys themselves simplify complex decision-making. Since the survey required patients to select a single option, such as calling the office or using a DMA, the results may not fully reflect how decisions are made in real clinical scenarios. In practice, patients may weigh several options or escalate concerns in stages rather than selecting a single action.

Another limitation of our study is that the survey instrument was study-specific and has not undergone formal validation. While the internal consistency of questions was assessed, the questionnaire was created for exploratory purposes. In addition, responses were based on hypothetical postoperative scenarios rather than real-world interaction with an active DMA. Participants evaluated a hypothetical surgeon-trained and surgeon-validated DMA rather than currently available generative AI systems. These findings may not be fully generalizable to perceptions of other AI platforms with different capabilities.

Also, responses reflect a single time point and are based on stated preferences regarding the described DMA, which may not align with real-world behavior once patients use the tool. Patients may respond differently when faced with an actual concern, particularly one involving pain, wound complications, or fear of infection. Furthermore, prior experience with generative AI, overall comfort with technology, and individual expectations of physician communication may have influenced responses in ways that are difficult to fully account for. Patients who regularly use digital tools or other forms of AI may feel more comfortable interacting with a DMA, whereas those who place greater value on direct physician involvement may be less inclined to adopt AI in their healthcare. Despite these limitations, this study provides an early look at how arthroplasty patients perceive generative AI-powered communication tools and offers a foundation for future prospective and implementation-based research.

Conclusions

Arthroplasty patients in this cohort showed mixed but generally neutral-to-slightly positive attitudes toward a generative AI-powered DMA, while still consistently preferring traditional human-mediated communication. Patients were most receptive to DMA use for lower-acuity educational and supportive functions, such as activity guidance, encouragement, and reminders, and least receptive in higher-risk situations, such as suspected infection. Prior generative AI experience was the factor most consistently associated with more favorable DMA attitudes and greater willingness to use the tool across several scenarios. Overall attitudes were otherwise similar across most demographic and clinical subgroups, although older patients were less likely to choose the DMA for pain and medication concerns. Given the low preference for the DMA relative to human assistance, these findings should be viewed as early-stage and perception-based. DMAs may be best introduced as supplementary tools that enhance perioperative education and communication while preserving clear access to human care, though real-world evaluation of actual DMA use should confirm this before broader adoption is considered.

CRediT authorship contribution statement

Jonathan Brutti: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Andres Wong: Writing – review & editing, Writing – original draft, Formal analysis, Data curation. Ivan Garces: Writing – review & editing, Writing – original draft, Formal analysis, Data curation. Zachary Grand: Writing – review & editing, Writing – original draft, Formal analysis. Charles M. Lawrie: Writing – review & editing, Visualization, Validation, Supervision, Resources, Methodology, Investigation, Conceptualization.

Conflicts of interest

C.M. Lawrie holds stock or stock options in Fios Health; receives royalties from Zimmer Biomet; is a paid consultant for Zimmer Biomet, Smith & Nephew, Mizuho OSI, Medtronic, Ignite Orthomotion, Ganymed Robotics, and Fios Health; and serves on the AAHKS Fellowship Oversight Committee and is President of the Anterior Hip Foundation. All other authors declare no potential conflicts of interest.

For full disclosure statements refer to https://doi.org/10.1016/j.artd.2026.102152.

Appendix 1.

Patient perceptions of AI research survey

Please answer the following questions. Shade in your choice for questions with circles and write your answer for questions with underlines. We appreciate your participation!

• 1Age: ____
• 2
  Gender identity:

  • ○
    Male
  • ○
    Female
  • ○
    Other: _________________________________
• 3
  What is your highest level of education?

  • ○
    High school or GED
  • ○
    College/University
  • ○
    Master's
  • ○
    Doctorate
  • ○
    Professional/vocational school
• 4
  Have you ever used generative artificial intelligence (ChatGPT, Google Bard, Copilot, etc.,) in the past?

  • ○
    Yes
  • ○
    No
• 5
  Please rank the following in order of importance to you when receiving healthcare (with 1 being the most important and 3 being the least important):

  • ○
    Personal relationship with healthcare provider: _____
  • ○
    Accuracy of health information provided to me by medical staff: _____
  • ○
    Quick assistance in my medical care (short wait times, quick phone call answers etc.,): _____

For questions 6-8, assume your surgeon has an artificial intelligence powered digital medical assistant. It has been trained on your surgeon's knowledge and preferences. It has been validated by your surgeon to give correct answers to commonly asked questions and issues before and after hip or knee replacement surgery. It is available 24 hours a day, 7 days a week, and can respond within seconds.

• 6
  I would use the digital medical assistant to answer my questions.

  • ○
    Strongly agree
  • ○
    Agree
  • ○
    Neutral
  • ○
    Disagree
  • ○
    Strongly disagree
• 7
  I would trust the digital medical assistant to provide me with accurate information about my care:

  • ○
    Strongly agree
  • ○
    Agree
  • ○
    Neutral
  • ○
    Disagree
  • ○
    Strongly disagree
• 8
  I would prefer to use the digital medical assistant instead of contacting a human medical assistant or office staff for medical information.

  • ○
    Strongly agree
  • ○
    Agree
  • ○
    Neutral
  • ○
    Disagree
  • ○
    Strongly disagree
• 9
  Would encouragement and reminders from a digital medical assistant regarding your postoperative care and physical rehab encourage you to be more engaged in your recovery?

  • ○
    Strongly agree
  • ○
    Agree
  • ○
    Neither agree nor disagree
  • ○
    Disagree
  • ○
    Strongly disagree
• 10
  If you had concerns about postoperative wound care would you rather…?

  • ○
    Call the office
  • ○
    Chat with a digital medical assistant
  • ○
    Internet search
  • ○
    Drive to the orthopaedic clinic
  • ○
    Drive to the emergency room
  • ○
    Do nothing
• 11
  If you had concerns about postoperative infection would you rather…?

  • ○
    Call the office
  • ○
    Chat with a digital medical assistant
  • ○
    Internet search
  • ○
    Drive to the orthopaedic clinic
  • ○
    Drive to the emergency room
  • ○
    Do nothing
• 12
  If you had concerns about pain and medication would you rather…?

  • ○
    Call the office
  • ○
    Chat with a digital medical assistant
  • ○
    Internet search
  • ○
    Drive to the orthopaedic clinic
  • ○
    Drive to the emergency room
  • ○
    Do nothing
• 13
  If you had concerns about postoperative activity would you rather…?

  • ○
    Call the office
  • ○
    Chat with a digital medical assistant
  • ○
    Internet search
  • ○
    Drive to the orthopaedic clinic
  • ○
    Drive to the emergency room
  • ○
    Do nothing