A Pilot Study Evaluating Traditional and Artificial Intelligence (AI)-Generated Bedside Art Interventions in Hospital Care

Abstract

The psychological burden of hospitalization adversely affects patient well-being and recovery trajectories, prompting an investigation into whether personalized artwork—created either traditionally or through artificial intelligence (AI)—offers a viable therapeutic approach. This pilot study evaluated traditional artist-created and AI-generated personalized artwork in an inpatient setting. Adult inpatients (n = 37) at Stanford Hospital underwent a personalized artwork intervention. Participants completed preintervention and postintervention surveys assessing changes in psychological measures, including pain, stress, anxiety, comfort, and attitudes toward art in healthcare. A 15-min structured interview exploring personal interests and meaningful life experiences informed the creation of personalized artwork, either by an artist or using a commercially available AI art generation platform. Altogether, the art interventions were associated with significant reductions in anxiety and stress, and increased comfort. Linear regression analysis showed no significant differences in effectiveness between traditional and AI-generated art modalities. Patient satisfaction was high, with 74% reporting the experience exceeded expectations, 72% indicating it meaningfully improved their hospital experience, and 85% of participants being “very likely” to recommend it to others. This pilot study suggests that personalized art interventions, whether created traditionally or via AI, can be effectively implemented in hospital settings and may improve patient well-being.

Introduction

Hospitalization is often a stressful, anxiety-provoking experience associated with uncertainty, discomfort, and feelings of diminished autonomy and personal identity. Recent systematic reviews have indicated that clinically significant anxiety symptoms affect approximately 28% of general hospital inpatients, while rates of depression can range from 28% to 30% depending on the underlying condition.^1,2 These rates can be even higher in specific patient populations, such as those hospitalized with COVID-19, in whom studies have reported anxiety rates up to 60% and depression rates reaching 81%. ³ In recent years, there has been growing interest in addressing this critical need by integrating the arts and humanities into healthcare settings. Contemporary systematic reviews and meta-analyses have demonstrated that art-based interventions can significantly improve anxiety, depression, distress, psychological well-being, and quality of life in hospitalized patients.^4,5 While traditional art interventions have shown promise, new technologies offer opportunities to expand and scale these approaches.

The advent of new and more accessible technologies, including generative artificial intelligence (AI), has broadened the possibilities for personalized art-related interventions in the hospital. Traditional art therapy typically involves an artist working with the patient to create custom pieces that reflect and explore participants’ interests, emotions, and life experiences. ⁶ Recent advances in AI have introduced tools that use machine learning algorithms trained on millions of images to create artwork based on text prompts and stylistic preferences, such as DALL-E (OpenAI) and Midjourney (Midjourney). While current studies have highlighted both the potential and limitations of AI-generated imagery in healthcare settings,^7,8 this technology offers a novel, potentially scalable approach to delivering art-based wellness interventions in hospital environments. ⁹ Digital approaches to traditional art therapy have already shown promise in promoting therapeutic engagement and accessibility, ¹⁰ suggesting that technology-enhanced art interventions may help address the growing need for scalable mental health support in healthcare settings. ¹¹ Importantly, these AI tools also broaden access to this type of activity, as they remove barriers for patients who lack formal artistic training or feel limited by perceived limitations in traditional artistic skill.

Previous studies have demonstrated the feasibility of bedside art interventions and in-house art programs in hospital settings.^12,13 However, despite the growing availability of digital and AI-based tools, to our knowledge, this is the first study assessing the impact of personalized AI-generated art as a bedside art intervention in the hospital setting. As such, in this pilot study, we assess the impact of traditional and AI-generated personalized art interventions by evaluating changes in patient-reported pain, stress, anxiety, mood, and overall satisfaction with their hospital experience. To foreground humanistic care and ground the intervention in principles of dignity, personhood, and meaning-making, patients were invited to reflect on personally meaningful experiences and values through a structured bedside conversation, with the artwork serving as a catalyst for this exchange. Our findings demonstrate significant reductions in stress and anxiety, with high patient satisfaction across both traditional and AI-generated art modalities. The results of this preliminary work will inform future larger-scale studies and guide best practices for integrating new technologies to support mental well-being in hospital settings, potentially offering a scalable solution to address the high prevalence of psychological distress among hospitalized patients.

Methods

Study Design and Institutional Review Board Approval

This pilot study was conducted at a single hospital over a 16-month period. Written informed consent was obtained from all participants prior to enrollment.

Inclusion and Exclusion Criteria

To be included, a participant must have been currently hospitalized (inpatient) within the study site, be English-speaking or able to utilize an interpreter throughout participation, and be ≥18 years old. Exclusion criteria included patients who were unable to complete baseline or postintervention surveys due to cognitive, language, or clinical constraints, and patients for whom participation would interfere with essential medical care.

Recruitment and Enrollment

Participants were identified by screening inpatient units and approaching individuals who were clinically stable, expressed interest in art-related activities, and met the inclusion criteria. Our study coordinator also received referrals of possible participants directly from hospitalists. Following consent and baseline assessment, participants were assigned to either traditional or AI-generated art groups based on artist availability rather than formal randomization, which was chosen to maximize operational efficiency and minimize patient wait times.

Patient and Public Involvement

Prior to data collection, several patients and patient advocates participated in a draft of the interview and art-generation process. They provided feedback on the structure and material involved in the interview process, suggested additional and improved questions, and informally evaluated the researcher-artists’ communication approaches.

Data Reporting

Participant demographics and clinical characteristics were collected at baseline. Psychological measures included the generalized anxiety disorder 7-item scale (GAD-7) for anxiety assessment, with scores ranging from 0 to 21 (higher scores indicating greater anxiety). We also collected self-reported ratings of pain, stress, anxiety, comfort, and boredom using numerical rating scales (0–10). The perception that art could help hospitalized patients was measured on a scale from 0 to 10.

Following the intervention, participants rated their experience on multiple dimensions (meaningfulness, enjoyment, boredom, frustration, and confusion) using 0–10 scales. They also provided qualitative feedback about their expectations, experience, and suggestions for improvement.

Intervention

Participants underwent a 15-min structured interview conducted by a researcher-artist. The interview prompts included questions about personal interests, meaningful parts of their life, color and style preferences, and whether they preferred abstract or realistic representations. For the traditional art group, an artist created a piece using a medium chosen by the participant or artist, such as watercolor or pencil sketch. For the AI group, artwork was made using a commercially available AI art generation platform. In both groups, the art was made iteratively and involved collecting real-time feedback from the patient.

Patients received the original artwork and a digital image was retained for study purposes. The interview sessions were audio recorded (with participant consent) for qualitative analysis. All personally identifiable information was removed from transcripts, and all paperwork was stored securely.

Composite Metrics

To comprehensively evaluate the intervention's impact, we developed two primary composite metrics:

Distress Improvement Score

We assessed overall psychological improvement by combining changes in multiple distress indicators from preintervention to postintervention. This included changes in pain, stress, anxiety, boredom, and comfort levels. The composite distress improvement score was calculated by summing the reductions in negative symptoms (pain, stress, anxiety, and boredom) and subtracting changes in positive symptoms (comfort), such that higher positive scores indicate greater overall improvement in psychological state. We also calculated well-being change, a variation of distress improvement that focused exclusively on mood and psychological measures, excluding the physical symptom of pain.

Experience Rating

The quality of the intervention experience was evaluated using a composite metric incorporating five dimensions: meaningfulness, enjoyability, boredom, frustration, and confusion. For negative dimensions (boredom, frustration, and confusion), scores were reversed (subtracted from 10) before averaging, ensuring that higher composite scores consistently indicate more positive experiences. This produced a single experience rating ranging from 0 to 10, with higher scores reflecting more positive overall experiences.

Statistical Analysis

All statistical analyses were performed using R (version 4.2.0). To evaluate intervention effects, we conducted paired t-tests comparing preintervention and postintervention measures of stress, anxiety, comfort, and pain levels. For the analysis of desired versus actual frequency of meaningful conversations, we used the Wilcoxon signed-rank test.

Linear regression was used to identify predictors of intervention response, with the primary outcome being the overall experience rating rather than the distress improvement score, as we were specifically interested in factors that influenced participants’ subjective experience of the intervention itself. Predictor variables included baseline pain levels, anxiety (GAD-7 scores), self-reported health status, and demographic factors.

A threshold of p < 0.05 was used to determine significance.

Results

Study Cohort

A total of 75 patients who met the inclusion criteria were recruited, and 37 (49.3%) were enrolled. Of the 38 patients who did not participate, 12 were not interested (31.6%), and the remaining 26 (68.4%) could not participate due to timing issues or fatigue.

The median age at the time of interview was 67 (range 24–97 years). Overall, 20 (54.1%) patients identified their gender as female, 16 identified as male (43.2%), and 1 identified as nonbinary (2.7%). The patients spent a median of 8 days in the hospital prior to their interview (range 0–79 days). Table 1 describes the baseline participant characteristics overall and by intervention group.

Table 1.

Baseline Participant Characteristics by Intervention Group.

Characteristic	All Participants (n = 37)	Only Traditional Art (n = 20)	Only Artificial Intelligence (AI) Art (n = 12)	Both Art Modalities (n = 5)
Demographics
Age, median (range), years	67 (24–97)	65 (24–88)	76 (31–94)	53 (30–97)
Gender, n (%)
Female	20 (54.1)	14 (70)	5 (41.7)	1 (20)
Male	16 (43.2)	6 (30)	6 (50)	4 (80)
Nonbinary	1 (2.7)	0 (0)	1 (8.3)	0 (0)
Days in hospital	8 (0–79)	9.5 (0–68)	5.5 (0–40)	11 (4–79)
Baseline measures, median (range)
Pain	2 (0–8)	2 (0–8)	0 (0–7)	2 (0–6)
Stress	3 (0–10)	2 (0–10)	3 (0–10)	2 (0–10)
Anxiety	2 (0–10)	2 (0–8)	3 (0–10)	1 (0–10)
Comfort	8 (0–10)	8 (0–10)	7 (3–10)	7 (0–10)
Boredom	5 (0–10)	5 (0–10)	6 (0–10)	3 (0–9)
Preintervention affect, n (%)
Positive	15 (40.5)	7 (35)	5 (41.7)	3 (60)
Neutral	8 (21.6)	4 (20)	2 (16.7)	2 (40)
Anxious	6 (16.2)	2 (10)	4 (33.3)	0 (0)
Other	8 (18.9)	7 (35)	1 (8.3)	0 (0)

Overall, participants reported a significant gap between actual and desired frequency of meaningful life conversations (Wilcoxon signed-rank test: V = 6, n = 26, p = 0.0004). The median frequency of actual meaningful conversations was weekly, while the median desired frequency was daily (mean difference = 1.08 points; interquartile range = 0.00–1.75). This discrepancy was particularly pronounced among the 8 patients reporting highly infrequent meaningful conversations (annually or never), with 62.5% (5 out of 8) of these participants desiring at least weekly meaningful interactions.

Intervention Overview

Three intervention modalities were provided: AI-generated art only (n = 12, 32.4%), traditional art only (n = 20, 54.1%), or both modalities (n = 5, 13.5%). The median duration of each patient intervention varied by modality: AI-only was fastest (26 min), followed by both modalities (61 min) and traditional-only (90 min).

To compare the effectiveness across intervention types, we examined changes in participant outcomes by modality (traditional art, AI-generated art, or both). All three modalities produced improvements in participants’ distress scores. To formally test whether any modality was superior, we conducted a linear regression analysis with intervention type as the predictor variable and composite distress improvement as the outcome. The model did not reach statistical significance (p = 0.18), suggesting that the type of art intervention had minimal impact on its effectiveness. Supporting the overall benefit of the intervention regardless of modality, both composite scores showed significant positive changes: the median distress improvement score was +8.5 points (p = 0.003), and the median well-being change score was +4.0 points (p = 0.002) by the Wilcoxon signed-rank test (Figure 1).

Figure 1.

Distributions of composite metrics testing overall intervention impact. Vertical line represents median value, and p-values are from Wilcoxon signed-rank tests comparing scores to zero, where positive values indicate improvement. The distress improvement score combines changes in pain, stress, anxiety, boredom, and comfort (higher scores indicate greater improvement), while the well-being change score reflects the net positive change across all psychological measures.

For the AI-generated art, themes and meaningful imagery from patient conversations were identified and incorporated into text prompts for the AI platform. For example, one patient enjoyed needlepoint embroidery, walking in nature, and drinking hot tea on their maroon lounge chair. These themes were reflected in the AI-generated piece given to the patient (Figure 2).

Figure 2.

Examples of traditional and artificial intelligence (AI)-generated artwork. (a) A watercolor painting made for a participant in the traditional group. (b) An AI-generated illustration made for a participant in the AI group.

Intervention Effects on Patient-Reported Outcomes

Pain, Stress, Anxiety, and Comfort

Participants reported significant improvements in multiple measures following the art intervention (Figure 3 and Supplemental Table S1). Stress levels decreased from a median of 3.0 to 1.0 (mean decrease = 2.0; p = 0.0003). Anxiety levels similarly improved from a median of 2.0 to 0.0 (mean decrease = 1.8; p = 0.002). Comfort scores increased from a median of 8.0 to 9.0 (mean increase = 1.3; p = 0.02). Pain scores showed minimal change from a median of 2.0 to 1.5 (mean decrease = 0.33; p = 0.154). These results suggest that while the intervention was particularly effective for psychological measures such as stress and anxiety, its impact on physical pain was more limited.

Figure 3.

Changes in patient-reported measures before and after art intervention. Violin plots show the distribution of scores for multiple measures, with overlaid box plots displaying median and interquartile range. P-values are from paired t-tests comparing preintervention and postintervention scores.

Predictors of Intervention Response

A linear regression model identified baseline pain levels as a significant positive predictor of the composite experience rating score (β = 0.111, p = 0.024), even after controlling for anxiety and self-reported health status. Anxiety showed no significant relationship with experience ratings (β = 0.020, p = 0.305). Patients with lower self-reported health scores showed a trend toward rating the experience more positively (β = −0.008, p = 0.086). These findings suggest that the intervention may be particularly valuable for patients experiencing higher levels of pain and those with worse health status—precisely the populations who might benefit most from additional support during hospitalization. The lack of relationship between anxiety and experience ratings indicates that the intervention was equally well-received regardless of patients’ baseline anxiety levels, suggesting broad applicability across different psychological states.

Mood States

Prior to the intervention, mood varied by participant, with 15 participants reporting a positive affect (eg, “optimistic” and “happy”), 8 reporting a neutral affect (“even” and “okay”), 6 reporting an anxious affect (eg, “anxious” and “confused”), 8 reporting an other affect (eg, “surviving” and “stimulated but bored”).

After the intervention, nearly all participants (35 out of 37) reported they enjoyed the experience, describing feelings of peace and contentment (eg, “calm” and “peaceful”) or engagement and enthusiasm (eg, “intrigued” and “creative”). They reported these feelings were maintained or enhanced postintervention, describing themselves as having an elevated mood (eg, “elated” and “very happy”) or a peaceful state (eg, “really mellow” and “comfortable, relaxed”).

Participants particularly valued the personal connection fostered during sessions, with one noting how the experience allowed them to “connect as a person and not just a broken body.” The combination of conversation and art created a therapeutic environment that provided both emotional relief and creative expression. As 1 participant expressed, it offered a rare opportunity to “reflect on what means a lot to me” beyond immediate health concerns. Suggestions for program improvement focused on customization, such as adjusting session duration based on pain levels and incorporating more varied art techniques.

Impact on Hospital Experience

When asked about the intervention's impact, 72.2% of participants (26 out of 36) reported it improved their hospital experience “a lot,” while 25% (9 out of 36) reported “a little” improvement. Only 2.8% (1 out of 36) were “unsure,” and no participants reported “not at all.” Similarly, 85.3% (29 out of 34) reported being “very likely” to recommend the intervention to another patient, with the remainder (14.7%, 5 out of 34) being “a little likely” to recommend it.

The intervention exceeded expectations for nearly three-quarters of participants (26 out of 35) and met expectations for those remaining (9 out of 35). No participants reported the experience as worse than expected.

Relative Contributions of Art and Conversation

To clarify perceived contributions of each component of the intervention, we asked participants to indicate preference for future interventions if limited to either conversation or personalized artwork. Participants were evenly split between those who would prefer just conversation (9 out of 18, 50%) and those who would prefer just receiving personalized artwork (9 out of 18, 50%). This even distribution suggests that both the interpersonal and artistic components of the intervention were equally valued by participants.

Discussion

In this pilot study, we found that a brief, personalized art intervention—whether provided by a traditional artist or generative AI—was well-received by hospitalized patients. Importantly, our findings suggest that the therapeutic impact of the intervention was driven not solely by the artwork itself, but also by the human-centered conversation that accompanied its creation. Our results suggest that discussing sources of hope and meaningfulness, combined with viewing a personalized piece of artwork, had statistically significant and clinically relevant beneficial effects on patient-reported mood and comfort, and reduced stress and anxiety levels.

The high completion rate (49.3% enrollment, with most nonparticipation occurring after expression of initial interest due to logistical constraints rather than lack of interest in the program) and strong satisfaction metrics indicate substantial patient receptivity to art-based interventions in the hospital setting. Furthermore, the intervention's design proved well-suited to the hospital environment—the brief duration and bedside delivery format were practical to implement, while the iterative creation process allowed for meaningful patient engagement without excessive burden.

The intervention showed significant positive effects on multiple psychological measures, with particularly strong improvements in stress and anxiety. Our composite distress improvement score and well-being change scores indicated substantial overall psychological benefit. Importantly, our regression analyses revealed that patients with higher baseline pain levels and lower self-reported health status tended to rate the experience more positively, suggesting the intervention may be particularly valuable for more severely ill patients. This is consistent with prior findings in which patients with higher baseline anxiety and distress experience greater benefits from mindfulness-based interventions. ¹⁴

There is a wealth of literature demonstrating the value of patients being engaged in meaningful, humanizing conversations with healthcare professionals.^15,16 Research by Sun et al ¹⁷ established a robust association between well-being and both the quantity and quality of social interactions, highlighting that deeper, more meaningful conversations are especially effective at improving social connectedness and reducing stress and anxiety. ¹⁷ This context underscores our compelling discovery of the profound disparity between patients’ current and desired frequency of meaningful conversations. The intervention directly addressed this unmet need, with participants specifically valuing the opportunity to “connect as a person and not just a broken body.” The equal division in preferences between conversation and artwork components suggests that both elements contribute meaningfully to the intervention's impact on patient well-being.

Our results suggest that AI-generated art can be an effective alternative or complement to traditional artist-created pieces in terms of patient experience and psychological benefit. The use of generative AI did not appear to detract from the patient experience and, in fact, may offer a more scalable approach to providing art-based interventions, particularly in settings with limited access to traditional art therapy. However, the extent to which AI-generated art compares to human-created art in terms of improving patient well-being remains to be fully explored. As use broadens, safeguards should include transparent disclosure of AI use and rigorous protection of patient privacy and data security. In this study, generative AI was intentionally positioned as a supportive tool rather than a replacement for human interaction. As AI-enabled approaches are considered for broader implementation, maintaining this emphasis on human presence and relational ethics will be essential to preserving the intervention's humanistic intent.

Several limitations should be noted. Our single-center study with 37 participants may not fully generalize to the diverse patient populations across different healthcare settings. The nonrandomized assignment to traditional versus AI art groups, while practical for this pilot, limits our ability to make definitive comparisons between modalities. Additionally, the lack of long-term follow-up prevents us from assessing the durability of the observed benefits. Notably, several interested patients were unable to participate due to specific logistical barriers—including sudden hospital discharges and scheduling conflicts with the researcher-artists—highlighting the potential for further study as well as a robust organization and referral system for participant identification and recruitment.

The demonstrated effectiveness of AI-generated art opens new possibilities for expanding access to personalized art interventions in healthcare settings. To realize these benefits at scale, implementation should preserve the central therapeutic role of conversation and human presence. Future work should explore how to optimize the balance between technological efficiency and personal engagement to maximize therapeutic outcomes.

Conclusion

This pilot study provides compelling evidence that personalized art interventions, whether traditional or AI-driven, can be effectively integrated into inpatient care with significant positive effects on patient well-being. The intervention demonstrated feasibility in a busy hospital setting, with high patient satisfaction and significant improvements in psychological measures. The comparable effectiveness of AI-generated and traditional artwork, combined with the strong positive response to both the conversational and artistic elements, suggests multiple viable pathways for scaling such interventions. While technology may provide efficiency and accessibility, the study underscores the fundamental importance of personal connection in the hospital experience. These findings contribute to a growing body of literature at the intersection of art, technology, and healthcare, while offering practical insights for implementing personalized wellness interventions in hospital settings. As healthcare systems increasingly recognize the importance of addressing patients’ psychological well-being, this work provides a promising model for integrating scalable, person-centered interventions into routine hospital care.