Does Technology Reduce Dementia Caregiver Burden?

There is a great deal of enthusiasm for technology aimed at family caregivers of people living with dementia, and a much smaller quantity of evidence that any of it works. We build in this field. We think the honest thing is to set out what the trial literature actually shows — including, and especially, the results that are unflattering to the premise.

The scale of the problem is not in dispute

The Alzheimer’s Association’s 2026 Facts and Figures puts 7.4 million Americans aged 65 and over as living with Alzheimer’s dementia, cared for by nearly 13 million unpaid family caregivers who provided some 19 billion hours of care in 2025. The report values that care at $446.3 billion.

A note on that dollar figure. It is hours multiplied by an assumed wage rate, not a measured quantity. It is a useful way to convey scale and a poor way to convey precision, and we cite it as the former.

Sixty-six percent of dementia caregivers live with the person they care for. Longitudinal data from the National Health and Aging Trends Study shows dementia caregivers providing 31.0 hours a week in 2022, up from 21.4 hours in 2011 — a substantial increase in a decade.

Caregiver burden: real, and more complicated than usually stated

A meta-analysis of 43 studies covering 16,911 participants puts the prevalence of depression among informal dementia caregivers at 31.2 percent (95% CI 27.7–35.0). That is a large number and a traceable one, with a confidence interval attached — which is more than can be said for several figures that circulate in this space.

The comparison usually drawn next is between dementia and non-dementia caregivers, and it is here that the evidence is weaker than the rhetoric. The cleanest head-to-head study we could find compared 117 dementia caregivers with 134 non-dementia caregivers:

Measure	Dementia	Non-dementia	Significance
Hours per week	49.7	37.7	p = .001
Said caregiving worsened their health	33.9%	15.4%	p = .003
Depression score (CES-D)	7.59	6.17	not significant (p = .072)
Mental health quality of life (SF-12)	48.47	50.54	not significant (p = .125)

REGARDS Caregiving Transitions Study; data collected 2013–2016. Adjusting for the care recipient’s total problem load eliminated most of the significant differences — suggesting burden tracks overall problem load rather than a dementia-specific effect.

So: dementia caregivers demonstrably provide more hours and are far more likely to say caregiving has damaged their health. The claim that they are measurably more depressed than other caregivers is, on this evidence, not established.

What the intervention trials show

This is the part that matters most, and it is the part least often quoted.

−0.06 Standardised mean difference in caregiver burden for remotely delivered information, training and support versus usual care (95% CI −0.35 to 0.23). The interval crosses zero. Cochrane systematic review, 26 randomized trials, 2,367 participants, search to April 2020. Conclusion: probably little or no effect.

The same review found a standardized mean difference of −0.05 for depressive symptoms (95% CI −0.22 to 0.12) and 0.10 for quality of life (95% CI −0.13 to 0.32) against usual care. Compared against information alone rather than usual care, one result did reach significance: depressive symptoms, at −0.25 (95% CI −0.43 to −0.06). Burden did not.

The finding we think is most important. Caregivers assigned to these interventions were more likely to drop out than controls. The reviewers’ suggested explanation is that taking part in the intervention was itself a burden to some caregivers.

Any product intended to reduce caregiver load has to reckon with that directly. A daily update that requires the family to log in, learn something, or keep up with it is not obviously an improvement on the situation it replaces.

An umbrella review published in 2022, covering 21 systematic reviews and 119 primary studies, reached a similar place from a different direction: depression improved in five of seven meta-analyses and anxiety in three of four, but caregiver burden was inconsistent — 20 percent of analyses showed improvement, 60 percent found no effect or an improvement in the control group, and 20 percent were mixed. The authors rated only 9.5 percent of the reviews as high quality; 47 percent were low or critically low.

An earlier systematic review covering 31 interventions recorded explicit failures worth naming: decision-making tools were poorly used and not appreciated by caregivers; peer support did not significantly reduce social isolation when measured with validated scales; forum usage was variable and sometimes minimal.

What has worked

The benchmark trial in this field, REACH II, enrolled 642 caregiver–recipient dyads and delivered twelve in-home and telephone sessions over six months. It found clinical depression at 12.6 percent in the intervention group against 22.7 percent in controls, and quality-of-life improvements for Hispanic, White, and Black spousal caregivers — but not for Black non-spousal caregivers. It found no statistically significant difference in institutionalization at six months.

Two things follow. REACH II was intensive, human-delivered, and in person; it is not evidence that a software product will do the same thing. And its benefit was unevenly distributed across the population it served, which is a warning to anyone assuming a single design serves everyone.

The measurement problem

The Zarit Burden Interview is the standard instrument for caregiver burden, and it is the one we plan to use. It deserves a clear-eyed description.

The revised version has 22 items scored 0–4, giving a range of 0–88. Shorter forms exist with 13, 12, 7, 6, and 4 items. A validation study in 394 dementia caregivers found the 6-, 7-, and 12-item versions performed comparably to the full instrument, with sensitivity around 77–85 percent and specificity around 60–80 percent, and recommended the 6-item version as the practical choice, with a cutoff of 9 or above on its 0–24 range.

Three caveats belong alongside that:

There is no consensus on the instrument’s dimensionality. Published analyses support one, two, three, four, and five factor solutions.
In a study of 194 dementia caregivers, the 12-item version resolved into “personal strain” and “role strain” — and only personal strain predicted caregiver psychological distress. A total score can therefore move on items that do not track the outcome anyone cares about.
The published cutoffs were validated against caregiver depression as the reference standard, not against an independent measure of burden. They are best understood as depression-screening thresholds.

One thing we will not do. Severity bands for the 22-item Zarit — 0–21 as little or no burden, 21–40 mild to moderate, and so on — appear on a great many clinical websites and in a great many papers. We could not trace them to any primary validation study. This is a separate matter from the screening cutoffs derived in individual validation studies, which are real; curated instrument databases publish those and do not publish the four-band scheme. We do not use the bands and we would treat any result reported against them with caution. The distinction is set out in full in what Zarit Burden Interview scores mean.

Where this leaves us

An even-handed reading of this literature says: caregiver burden is real and large; the population providing this care is enormous and providing more of it every year; and remote and digital interventions have not been shown to reduce burden, with the best-quality evidence pointing to little or no effect and some indication that participation can add to the load.

That is not an argument against building. It is an argument for building differently — toward something that requires nothing of the family beyond reading it, and for measuring the result honestly against instruments whose limitations you have already admitted to.

Our 90-day pilot with a Medicare-certified home health partner enrolls 40 to 60 clients living with dementia under full informed consent and measures caregiver burden, engagement quality, and real-world usability. We expect to publish what it finds, including if what it finds is nothing.

Common questions