Artful Path to Healing: Using Machine Learning for Visual Art Recommendation to Prevent and Reduce Post-Intensive Care Syndrome (PICS)

Literature Title

Artful Path to Healing: Using Machine Learning for Visual Art Recommendation to Prevent and Reduce Post-Intensive Care Syndrome (PICS)

Literature Information

• Subject Area: Interdisciplinary research between artificial intelligence (recommendation systems) and healthcare (PICS treatment)
• Keywords: recommendation system, personalization, art therapy, user experience, machine learning, intensive care unit, rehabilitation, health

Research Background and Problem

• Problems and Challenges:

  1. Experiences in the intensive care unit (ICU) often cause psychological trauma to patients, potentially leading to Post-Intensive Care Syndrome (PICS), which includes long-term physical, cognitive, and mental health impairments.
  2. Existing PICS interventions are limited and primarily focus on medical treatments, such as ICU follow-ups and patient diaries, lacking widely available and diverse psychological intervention methods.
  3. While studies suggest that visual art may benefit the mental health of PICS patients, there is a need to explore how to achieve personalized art therapy.

• Significance of the Research:

  1. Visual art has been proven to alleviate stress, anxiety, and pain, showing unique potential for intervention.
  2. Personalization is a key direction for art therapy, but existing methods are limited in their ability to support diverse and tailored art selection.
  3. By integrating machine learning-based recommendation technology, personalized art therapy may open new pathways for PICS patient rehabilitation.

• Motivation and Related Work:

  • Previous studies have shown that recommendation systems have strong potential in uncovering complex semantic relationships and personalizing content. While recommendation systems have seen successful applications in entertainment, their use in therapeutic contexts remains unexplored.
  • Personalized interventions in health and psychological therapy still require further research and technical support.

Solution

• Methods and Innovations:

  1. This paper proposes the development of a machine learning-based visual art recommendation system (VA RecSys) to provide PICS patients with personalized art therapy experiences.
  2. Four recommendation models (ResNet, LDA, BERT, BLIP) were compared for effectiveness:
     • ResNet: Based on image content.
     • LDA (Latent Dirichlet Allocation): Based on textual semantics in art descriptions.
     • BERT: Utilizes deep semantic features, combining text for art interpretation.
     • BLIP (Multimodal Model): Integrates features from both images and text.
  3. This is the first exploration of recommendation systems in the context of PICS treatment, with evaluations conducted through expert assessments and user testing.

• Implementation Steps and Key Technologies:

  1. Extract features from artwork images and text using pre-trained models (ResNet, BERT, etc.) to construct a unified representation space.
  2. Use cosine similarity to calculate recommended artwork sets based on initial user selections.
  3. Validate recommendation quality and therapeutic efficacy through expert and user experience evaluations.
  4. Provide text-based guidance (e.g., imagining entering the painting's scene) to further enhance participants' immersive art experiences.

Research Outcomes

• Specific Results:

  • Developed and tested four VA RecSys models, demonstrating that image-based (ResNet) and multimodal (BLIP) models outperform expert-curated recommendations in providing therapeutic suggestions.
  • Proposed and validated a personalized guided art therapy process, effectively improving patients' emotional states.

• Advantages Compared to Existing Solutions:

  • Machine learning-supported recommendation systems excel in recommendation quality (diversity, novelty, and serendipity) and emotional impact (mood enhancement, well-being).
  • The BLIP multimodal approach captures diverse and complex semantic features better than other unimodal techniques.

• Experimental and Evaluation Results:

  1. User study: Participants reported significantly increased immersion with recommended artworks (average rating above 4/5).
  2. In experiments involving 120 participants, most showed psychological improvement in emotional surveys, particularly in reducing negative emotions such as fear and depression.
  3. Expert evaluations indicated that text-driven models (e.g., BERT and LDA) may generate negative recommendation content, requiring cautious handling.

• Limitations and Future Directions:

  • The emotional dimensions described in current textual data may be insufficient, requiring higher-quality semantic descriptions.
  • The method of collecting user preferences (e.g., single-image selection) may be overly simplistic; multi-dimensional rating mechanisms could be explored.
  • The transparency of VA RecSys recommendations needs further improvement to meet the interpretability requirements in medical contexts.

Conclusion

This paper innovatively applies recommendation systems to the medical field, opening up new possibilities for personalized art therapy for PICS patients. The study demonstrates that image-based and multimodal recommendation systems excel in therapeutic efficacy and recommendation quality. This research not only extends art and entertainment technologies into the health intervention domain but also provides new technical support for personalized medicine. However, further optimization and exploration are needed in areas such as safety, scalability, and human-machine collaboration.