cs.HCMar 6, 2026

An Interactive LLM-Based Simulator for Dementia-Related Activities of Daily Living

Authors: Kruthika Gangaraju, Shu-Fen Wung, Kevin Berner, Jing Wang, Fengpei Yuan

Organizations: Worcester Polytechnic Institute Worcester, Massachusetts, USA · University of California Davis Sacramento, CA, USA · MGH Institute of Health Professionals Boston, MA, USA · University of New Hampshire Boston, MA, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Effective dementia caregiving requires training and adaptive communication, but assistive AI and robotics are constrained by a lack of context-rich, privacy-sensitive data on how people living with Alzheimer's disease and related dementias (ADRD) behave during activities of daily living (ADLs). We introduce a web-based simulator that uses a large language model (gpt-5-mini) to generate multi-turn, severity- and care-setting-conditioned patient behaviors during ADL assistance, pairing utterances with lightweight behavioral cues (in parentheses). Users set dementia severity, care setting (and time in setting), and ADL; after each patient turn they rate realism (1-5) with optional critique, then respond as the caregiver via free text or by selecting/editing one of four strategy-scaffolded suggestions (Recognition, Negotiation, Facilitation, Validation). We ran an online formative expert-in-the-loop study (14 dementia-care experts, 18 sessions, 112 rated turns). Simulated behavior was judged moderately to highly plausible, with a typical session length of six turns. Experts wrote custom replies for 54.5 percent of turns; Recognition and Facilitation were the most-used suggested strategies. Thematic analysis of critiques produced a six-category failure-mode taxonomy, revealing recurring breakdowns in ADL grounding and care-setting consistency and guiding prompt/workflow refinements. The simulator and logged interactions enable an evidence-driven refinement loop toward validated patient-caregiver co-simulation and support data collection, caregiver training, and assistive AI and robot policy development.

Figures & tables

Figure 1. Simulated person with Alzheimer’s dementia (PLWD) interacting with a human caregiver at a selected dementia stage.
Figure 2. End-to-end participant-facing workflow of the web-based dementia–ADL simulator. After consent and a brief background questionnaire, participants configure a scenario (dementia stage, care setting, ADL). Within the dashed region (turn-level interaction), the LLM-based patient simulator generates a PLWD response; participants provide a realism rating and optional critique, then respond as the caregiver either via free-text or by selecting a strategy-scaffolded suggestion (Recognition, Validation, Facilitation, Negotiation; edits optional). The loop continues until the participant terminates the scenario or reaches turn 10. Participants may then reset the scenario to start a new simulation or end the session. Interaction transcripts can be downloaded.
Figure 3. Prompt design and turn-level context flow for generating PLWD behaviors and caregiver responses. Scenario configuration (care setting and time in setting, living situation, dementia severity, ADL, and task progress context) forms a shared context. Each turn uses a windowed interaction history to condition (i) PLWD response generation and (ii) strategy-scaffolded caregiver response suggestions.
Figure 4. Simulation page as displayed to participants during interaction with the simulated PLWD. A) Conversation history updates as the interaction proceeds; B) Downloadable transcript of the entire interaction for participants records; C) Simulation settings chosen by the participant; D) Generated PLWD response (black text: verbal utterance; grey text in parentheses: nonverbal behavior); E) Realism rating (1-5) for the current generated PLWD response, with optional free-text critique; F) Caregiver response interface: participants may write a free-text response or select a strategy-scaffolded suggestion.
Figure 5. An example of PLWD-caregiver interaction during taking medicine in PLWD’s home. The PLWD is with middle-stage Alzheimer’s dementia, living at their home for more than 1 year. The text in parentheses represents PLWD’s nonverbal behavior. Other text represent PLWD’s and caregiver’s verbal response, correspondingly
CharacteristicCategoryn (%)
Age range18–241 (7.1%)
 25–341 (7.1%)
 35–446 (42.9%)
 45–545 (35.7%)
 55–641 (7.1%)
 65+0 (0.0%)
Table 1. Participant characteristics for the online evaluation with dementia care experts of the PLWD simulator ( N=14N=14 ). Values are reported as n (%).
Figure 6. Mean participant-rated realism (1–5 scale) of generated PLWD responses across ADL and dementia stage. Text in each cell shows the average turn-level realism rating provided by participants for the specified ADL × dementia stage combination. Cell shading indicates the number of rated occurrences contributing to the mean (yellow: 2 occurrences; purple: 1 occurrence). Blank cells indicate combinations not evaluated during the study.
Figure 7. Turn-by-turn realism ratings for the LLM-generated PLWD responses. Each colored dashed line corresponds to one simulation session (18 sessions total), showing the participant’s realism rating (1–5) at each turn. The thick black line plots the mean realism rating at each turn across sessions. The vertical dashed black line marks the typical stopping point (median session length).
Failure modeDefinitionExample excerptFreq.Design implication / mitigation
Stage mismatchComment suggests the response does not match the selected dementia stage (too coherent/able or too impaired/disoriented).Expected more resistance; response seems closer to early-stage dementia.5 (25%)Tighten stage constraints (memory/orientation/language) and cap planning/initiative; add guardrails for severe stages.
Task/ADL grounding errorComment flags implausible ADL steps, sequencing, assumptions about task setup, or multi-step planning beyond expected ability.Multi-step ADL planning (e.g., wiping + tucking napkin) may exceed moderate-stage capacity.9 (45%)Refine ADL step plans and enforce step-local responses; add object-location context and prevent skipping without caregiver guidance.
Care-setting mismatchComment indicates the PLWD behavior depends on or contradicts the selected environment (e.g., home vs facility; where objects are stored).Assisted-living workflows: medications are centralized (not stored with the resident).4 (20%)Add setting-specific constraints (storage policies, routines) and use duration to modulate familiarity; avoid inconsistent assumptions.
Overcompliance / low resistanceComment suggests the simulated PLWD is too agreeable/compliant and should show more refusal, anxiety, or resistance.PLWD is too agreeable given selected challenges (distraction/anxiety/refusal).5 (25%)Increase probability of hesitation/refusal in moderate/late stages and when challenges are present; require repeated cues before progress.
Language/utterance unnaturalnessComment critiques phrasing (overly well-formed, too long/complex, inappropriate word choice) relative to expected communication ability.Utterance is overly well-structured/verbose; clinicians suggested shorter, simpler phrasing.5 (25%)Constrain to simpler phrasing; discourage formal grammar; avoid infantilizing terms; permit more nonverbal actions instead of naming objects.
Needs more prompting / supportComment suggests the scenario should require repeated prompting, nudging, or more caregiver assistance across turns.Even when plausible, progress often requires nudging and repeated reminding across turns.4 (20%)Prompt the PLWD model to stall/ask for clarification; progress only with repeated caregiver prompts in moderate/late stages.
Table 2. Failure-mode taxonomy from optional free-text clinician comments. Comments were coded at the turn level and a turn may receive multiple codes; therefore, frequencies sum to more than N=20N=20 commented turns. Frequency is reported as # commented turns (percent of commented turns). Example excerpts are lightly edited for brevity.
Figure 8. Caregiver response strategy usage across all turns. “Custom” denotes participant-authored free-text replies or any strategy suggestion that was selected and subsequently edited before submission.
More in this topic

Dementia Care

See all 18 papers

Explore similar work

CardsList

  1. MEMOR-E: In-Context and Fine-Tuned LLM Personalization for Alzheimer's Assistive Robotics

    Apr 28, 2026Maissa Abir Smaili, Eren Sadikoglu, Ransalu SenanayakeAlzheimer's DiseaseLLM Personalization

    Istanbul Medipol University Istanbul, Türkiye · Arizona State University Tempe, USA

  2. TALK-Dem: Benchmarking Embodied Task Planning under Dementia-Associated Communication Patterns

    Sep 29, 2026Guangxin Zhao, Yiran Hu, Yuan Cao +7Large Language Model-Based Robot PlanningRobot Task Planning

    University of Notre Dame · Nanyang Technological University · Tohoku University

  3. When AI Tells You What You Want to Hear: Sycophantic Behavior of Large Language Models in Dementia Care Settings

    Apr 13, 2026Christian KolbDementia CareLLM Sycophancy