Machine learning-enabled behavioural and psychological symptoms of dementia management intervention for dementia caregivers: protocol for a hybrid factorial SMART–MRT trial

Background
Behavioural and psychological symptoms of dementia (BPSD) affect over 90% of people living with dementia and are a major contributor to stress and adverse health outcomes among family caregivers. Caregiver‑led behavioural management interventions based on the Antecedents–Behaviour–Consequences (ABC) framework are effective, yet their overall impact is modest. This may be due to the highly individualised and fluctuating nature of BPSD and insufficient adaptive support to sustain engagement. Mobile health interventions offer scalable support; however, engagement and adherence remain challenging. Recent advances in adaptive trial designs and machine‑learning methods offer opportunities to optimise interventions and deliver personalised just‑in‑time support.

Methods
This hybrid experimental trial integrates a five‑arm randomised controlled trial (RCT), a two‑stage sequential multiple assignment randomised Trial (SMART), and three embedded micro‑randomised trials (MRTs). A total of 550 family caregivers of community-dwelling people with dementia will be recruited and randomised to one of five conditions: (1) BPSD management intervention with combined human‑ and app‑based coaching; (2) human coaching only; (3) app‑based coaching only; (4) no coaching; or (5) psychoeducation control. The intervention includes telephone‑based psychoeducation and BPSD management training followed by app‑based support.

Participants with inadequate engagement after six weeks will be re‑randomised to intensified human coaching or a booster session in the next four weeks. MRTs will be conducted to evaluate the proximal effects of push‑notification timing, personalised motivational messages, and symptom‑specific guidance on engagement.

The primary outcome measure is caregiver stress. Secondary outcome measures include caregiver burden, psychological well‑being, anxiety, depression, sleep quality, health‑related quality of life, caregiving self‑efficacy, and severity of BPSD in care recipients. Linear mixed‑effects models will be used to evaluate intervention effectiveness, coaching strategies, and adaptive sequences. MRT data will be analysed using weighted regression methods to estimate time‑varying causal effects of just‑in‑time intervention components on next‑day engagement. Supervised machine‑learning models (e.g., random forests, support vector machines), reinforcement learning, including contextual multi‑armed bandit approaches, will be trained to predict caregiver receptivity and optimise notification timing, frequency, and content.

Discussion
This study design will inform the development of scalable, personalised interventions for dementia caregivers that are feasible for real‑world implementation.