Revolutionizing Dementia Care: NUS' Rememo AI Emerges from arXiv Field Study in Singapore Nursing Homes
Singapore's rapid aging has thrust dementia care into the spotlight, with one in 10 seniors aged 60 and above affected.
Unlike conversational agents that sideline human facilitators, Rememo empowers care staff, addressing Singapore's unique challenges like chronic understaffing and cultural-linguistic barriers in nursing homes. Led by design researchers including Celeste Seah, Assoc Prof Jung-Joo Lee, Assoc Prof Ching-Chiuan Yen, Asst Prof Clement Zheng, and collaborator Yoke Chuan Lee from ECON Healthcare Group, the project culminated in a two-week field study across two private nursing homes, involving 21 residents and generating profound insights into human-AI collaboration in eldercare.
Singapore's Demographic Imperative: Why AI-Driven Memory Aids Matter Now
By 2030, Singapore's elderly population will surge, straining nursing home capacities—from 14,918 beds in 2017 to even greater demands today.
The government's push for AI in eldercare aligns perfectly. Budget 2026 allocates resources for AI missions in healthcare, including free premium AI tools for workers via SkillsFuture.
Unpacking Reminiscence Therapy: A Proven Yet Under-Resourced Practice
Reminiscence Therapy (RT), first formalized in the 1950s, involves structured recall to validate life stories, reduce agitation, and foster social bonds. In Singapore nursing homes like those of ECON Healthcare Group—the nation's largest private operator—RT is standard but hampered by logistics. Therapists spend hours curating images from archives or stock photos, often mismatched to residents' Singaporean-Chinese, Malay, Indian, or Eurasian backgrounds.
- Visual cues excel as memories are predominantly visual.
- Group vs. one-to-one sessions demand adaptable prompts.
- Cultural specificity boosts resonance—e.g., 1980s Chinatown bicycles or Padang soccer fields.
Prior tech like AR/VR apps or chatbots (e.g., Remihaven) automate but erode the 'relational facilitation'—empathic improvisation—that makes RT effective. Rememo flips this, augmenting therapists.
Rememo AI: Anatomy of a Therapist-Tool Memory Prosthesis
Rememo is a mobile webapp bridging physical and digital. Therapists select from illustrated prompt cards categorized by Who/What/Where/When/How/Why, across eight Singapore-centric themes: Festivals, Food, Places, etc. Optical Character Recognition (OCR) scans cards, crafting prompts like "smiling durian sellers in the market, 1980s Singapore hawker center."
Generative AI engines—Stable Diffusion XL (SDXL), Flux.1 (LoRA-fine-tuned for local accuracy), Google Imagen—produce photorealistic images in ~20-27 seconds. Therapists preview, refine, and print via instant photo printers for tangible mementos. Large Language Models (LLMs) generate five tailored questions, e.g., "What did the durian smell like?" Multilingual support (English, Mandarin, Malay, Tamil) aids inclusivity.
Read the full arXiv field study for technical architecture.
Photo by Markus Winkler on Unsplash
From Concept to Deployment: NUS' Two-Year Research-Through-Design Journey
Initiated as Celeste Seah's award-winning thesis (NUS CDE Innovation Research Award 2024), Rememo evolved via sociotechnically-aware Research-through-Design (RtD).
NUS Division of Industrial Design faculty—Jung-Joo Lee (Deputy Head Research), Ching-Chiuan Yen (CUTE Center), Clement Zheng (Interactive Materials Lab)—mentored, emphasizing human-centered AI. This aligns with NUS' role in Singapore's AI ecosystem, fostering research jobs in design-AI hybrids.
Field Study Insights: Real-World Testing in ECON Nursing Homes
August 2025: 5 staff (OT principal, assistants) at two ECON sites ran 26 sessions—19 one-to-one, 7 groups (2-4 residents)—with 21 participants (ages 51-90, mild-moderate impairment per Pool Activity Level scores). 151 images generated, 50 printed.
| AI Engine | Images Generated | Print Rate |
|---|---|---|
| Imagen | 83 (55%) | 34.9% |
| Flux.1 | 53 (35.1%) | 34.0% |
| SDXL | 15 (9.9%) | 20.0% |
Sessions averaged 32 minutes, yielding vivid recalls—quiet residents engaged more, groups fostered listening like "coffee shop chats."
Staff Perspectives: Easing Burdens, Sparking Joy
"Therapy staff don’t really need to think what to talk about," noted one. Rememo cut prep time, bridged languages: "Some don’t know how to read, so we read to them." Residents: "Really enjoyed... want to share more." Images deemed 70-80% accurate, evoking nostalgia without perfectionism.
- Benefits: Structured prompts, richer narratives, mood uplift.
- Risks: Latency disrupting flow (mitigated post-session).
One-to-one shone; groups needed curation. Staff rejected full AI replacement: humans provide empathy.
Explore careers in AI-health design at NUS-like institutions.Challenges Overcome: Latency, Accuracy, and Cultural Fit
Key hurdles: AI latency prompted hybrid use; variable quality required previews/iterations. Dialects challenged OCR, solved via illustrations. Cost (printers) led to projections. Findings underscore AI as augmentor, not replacer, respecting care's relational core.
Singapore's context—migrant caregivers, diverse residents—amplified needs, but solutions generalize.
Photo by Markus Winkler on Unsplash
Broader Impacts: Paving AI Eldercare's Future in Singapore
Rememo exemplifies 'AI-in-the-loop' care, aligning with Budget 2026's AI missions.
For higher ed, it highlights interdisciplinary roles—design, AI, therapy—fueling Singapore university jobs. As super-aging looms, such innovations promise dignity in memory loss.
Interested in pioneering research? Check higher-ed jobs, rate your professors, or career advice.