Does Brain Training Last for People with Mild Memory Loss?
A deep dive into one paper: what was known before, what the researchers did, what the numbers say and what they don't.
Chapters
Transcript
The script the two AI voices read from. The times are worked out from the script, not heard, so a line can start a few seconds early or late.
Read the whole episode
Cold open
Host 1If someone with mild memory problems does brain-training games on a computer for a few months, is their memory still any better a year later?
Host 2Across the randomised trials that measured it, no: three months or more after the training stopped, the average benefit was close to zero. Right after training, the averages leaned slightly positive, but they were small, unstable and mostly driven by one unusual trial.
Host 1Welcome to the Hastrup Review of Science, where artificial intelligence reads real science, so you don't have to pretend you understood the paper.
Host 2Today, one paper: a systematic review and meta-analysis from Hao Luo, Yong Liu and colleagues in China, published on the fifteenth of October twenty twenty six in Frontiers in Aging Neuroscience. It asks whether computerised cognitive training for older adults with mild cognitive impairment reaches beyond the game itself, and whether it lasts.
Why anyone trains a brain
Host 1Before we get to the games, what exactly is mild cognitive impairment, and how is it different from dementia?
Host 2It is the zone between normal ageing and dementia: tests show a real drop in one or more mental abilities, such as memory or attention, but the person still manages daily life largely on their own. The paper stresses that the course varies; some people go on to dementia, some stay stable, and some drift back towards normal.
Host 1So it's a kind of warning window, where something might still be done before people lose their independence?
Host 2That is exactly how the authors frame it, a potentially modifiable window. And they note that no drug currently available can reliably reverse its course, which is why safe, cheap, non-drug options get so much attention.
Host 1And computerised cognitive training is basically brain-training apps, the puzzles on a tablet?
Host 2Broadly, yes: structured exercises repeated on a computer, tablet, phone or in virtual reality, aimed at memory, attention, planning or processing speed. Many programmes adjust the difficulty to how well you are doing and give instant feedback, and they can be done at home.
Host 1Where did the idea come from that practising puzzles would protect the brain in the first place?
Host 2In the wider field, it grew from the observation that people who stay mentally active tend to keep their thinking longer, plus large trials in healthy older adults in the early two thousands that trained speed and reasoning. Then came a commercial boom in brain-training products in the two thousand tens, followed by a public argument among scientists about whether the claims were justified.
What was known
Host 1What was the argument actually about, if everyone agreed people get better at the games?
Host 2That is precisely the point of contention. Nobody doubts you get better at the trained task; the question is whether that improvement spreads to anything else, which researchers call transfer.
Host 1Could you spell out the kinds of transfer the paper is talking about?
Host 2Three kinds. Near transfer means improving on an independent test of the same ability you trained, say a different memory test after memory training; far transfer means improving on overall cognition or on abilities you did not train; and functional far transfer means everyday life, like managing money, medication or shopping.
Host 1So if I practise a memory game and then do better on a new memory test, that's near, and if I suddenly balance my chequebook better, that's the far, functional kind?
Host 2Yes, with one refinement: the same test can count as near in one trial and far in another, depending on what that trial actually trained. The authors insist on classifying by the link between training and test, not by the test's name.
Host 1What did earlier reviews conclude about all this for people with mild impairment?
Host 2Several earlier meta-analyses suggested benefits for global cognition, memory or working memory. But this paper argues they had three weak spots that could make the benefit look bigger than it is.
The open question
Host 1What were those three weak spots?
Host 2First, they often lumped together practice on the trained task with genuine near and far transfer. Second, they mixed tests taken right after training with tests months later, so a short-lived bump and a lasting effect got blended.
Host 1And what was the third problem with those earlier reviews?
Host 2Double counting. One trial can produce several papers, many outcome measures and several time points, and unless you tie them back to one trial, the same volunteers can be counted more than once and given too much weight.
Host 1So the question this paper sets itself is, once you clean all that up, does any benefit spread beyond the game, and does it survive after people stop?
Host 2Exactly that, with real-world functioning as a separate question, and adherence, dropouts and side effects as secondary questions.
How they did it
Host 1How did they go about finding the trials?
Host 2They searched seven databases, including PubMed, Embase and the Cochrane Library, from their start up to the first of July twenty twenty six, and also chased reference lists and later citations. They registered the review in advance in the international register for systematic reviews, called PROSPERO.
Host 1How many papers turned up, and how many survived?
Host 2Five thousand and fifty records, which became two thousand nine hundred and twenty seven after removing duplicates. One hundred and seventeen full texts were assessed, and in the end eleven reports describing eight independent trials were included.
Host 1Eight out of five thousand sounds tiny, so were they unusually strict?
Host 2Deliberately so. A trial had to be truly randomised, enrol people with a clinical diagnosis of mild impairment rather than just a low screening score, isolate the effect of the computer training, and use an independent outcome measured in a defined time window.
Host 1Who were the people in those eight trials, and what did the training look like?
Host 2Seven hundred and twenty two people were randomised in total, in trials of twenty two to one hundred and seventy five people each, published between twenty fourteen and twenty twenty five, in the United States, Australia, Israel, Singapore, Italy, Canada and China. Training lasted from about three weeks to six months, one to five sessions a week, thirty to one hundred minutes a session, at clinics, at home or both.
Host 1And what were the training groups compared against, was it people doing nothing at all?
Host 2It varied: usual care, health education, general mental stimulation, non-adaptive computer games, or sham training. In two trials everyone also exercised, and the comparison was real training plus exercise against sham training plus exercise, so those measure what the computer training adds on top.
Host 1That seems important, because a control group playing other games is getting some stimulation too, right?
Host 2Right, and the authors make that point: active comparisons estimate the extra effect of the training, not its effect compared with nothing. That tends to make effects smaller, and it is one reason trials disagree.
Host 1How did they draw the line between a quick effect and a lasting one?
Host 2Tests within four weeks of the last session counted as immediate; tests at least twelve weeks after training stopped counted as sustained. Anything in between, more than four but under twelve weeks, was set aside and not used in either main analysis.
Host 1Why throw away the in-between measurements rather than use them?
Host 2To keep the two questions clean, so a lingering short-term bump is not mistaken for durability. One trial's thirty-day follow-up, for example, fell in that gap and was left out of the sustained analysis.
Host 1Who decided whether a given test counted as near or far, and couldn't that be bent to suit the result?
Host 2Two reviewers coded each outcome independently using a written manual, without looking at the effect sizes, and a third settled disagreements. Tasks from inside the training platform itself were excluded entirely as mere practice.
Host 1How do you combine results from trials that used completely different tests with different scoring scales?
Host 2They converted everything to a standardised effect size called Hedges g, which expresses the difference between groups in units of spread of the scores. By convention, zero point two is a small effect, roughly the training group shifting a fifth of a standard deviation, and the authors use that only as a rough guide, not as a threshold of clinical importance.
Host 1So a Hedges g of zero point two means a fairly small nudge, not something you'd notice in your day?
Host 2A small nudge on average, yes, though whether a given person notices depends on the test and the person. Positive numbers here always mean the training group did better.
Host 1And how did they stop one trial with many tests from counting several times?
Host 2Every trial got a single identifier, and all its measures for one type of transfer and one time window were merged into one number per trial. When they did not know how strongly those measures were related, they assumed a correlation of zero point five, and rechecked with zero point three and zero point seven.
Host 1Then the actual pooling, what kind of statistics did they use to combine eight trials?
Host 2A random-effects model, which assumes the true effect genuinely differs from trial to trial, with a correction designed for when you only have a handful of studies. They built six separate models: near and far cognitive transfer, plus everyday function, each at the immediate and the sustained time point, and only pooled when at least three trials were available.
Host 1You mentioned they also report something called a prediction interval, what's that, as opposed to the usual confidence interval?
Host 2The confidence interval is the plausible range for the average effect across trials. The prediction interval is the plausible range for the effect in the next similar trial, so it shows how much results bounce around between settings.
Host 1Did they also judge how trustworthy each trial was?
Host 2Yes, with the Cochrane risk-of-bias tool, which checks randomisation, deviations from the plan, missing data, measurement and selective reporting. And then GRADE, a standard system that rates how much confidence the whole body of evidence deserves, from high down to very low.
Host 1And what did the risk-of-bias check show?
Host 2Not one of the eight trials was rated at low risk overall: five raised some concerns, and three were at high risk. Missing outcome data was the biggest problem; in one trial only thirty one of sixty one people completed the six-month follow-up.
A word from our sponsor
Host 2This episode is brought to you by the Rubik's Cube. Twist it for twelve weeks and the benefits transfer far beyond the cube.
Host 1After my course I could read Finnish, a domain I never trained.
Host 2Effects are durable. Eighteen months after you stop, the cube keeps practising alone in its drawer, adaptively raising its own difficulty.
Host 1The makers of the Rubik's Cube have not paid for this, were never asked, and are unaware it is happening.
Host 2Transfer is not limited to the owner. In the sustained window, every hedge on our street came out solved, one colour per side.
Host 1The Rubik's Cube. Benefits that have left the platform.
What they found
Host 1Back to the trials: right after the training ended, did the brain-training groups do better on tests of the abilities they practised?
Host 2Six trials with three hundred and seventy six people fed that analysis, and the pooled Hedges g was zero point one four, with a confidence interval from minus zero point two three to plus zero point five zero. So a small lean towards training, but the range includes no effect at all, and the trials disagreed a lot with each other.
Host 1When the range runs from slightly negative to moderately positive, does that mean they just can't tell?
Host 2It means this evidence cannot tell a small benefit from none. The prediction interval was even wider, from minus zero point seven four to plus one point zero two, so a new trial could land almost anywhere.
Host 1What about the far kind straight after training, overall thinking and untrained abilities?
Host 2That was the largest model, eight trials and five hundred and eighty six people, with a Hedges g of zero point two three, and a confidence interval from minus zero point three five to plus zero point eight one. Heterogeneity was very high, an I squared of about eighty four percent, meaning most of the variation came from real differences between trials rather than chance.
Host 1Zero point two three is the biggest number so far, so is that the encouraging bit?
Host 2It looks like it, until you look at one trial. A Chinese hospital study reported a Hedges g of one point nine seven, roughly ten times its neighbours, while most of the others sat near zero or pointed in different directions.
Host 1What happens to the average if you take that one study out?
Host 2It drops from zero point two three to zero point zero zero. Their statistical influence checks also flagged it as the single most influential trial, and only fifty of its sixty randomised people were analysed.
Host 1So is the encouraging average for far transfer really resting on a single trial?
Host 2For the immediate far-transfer estimate, largely yes, though the authors kept it in, because you should not drop a study only for having a big effect. They tested its influence openly instead.
Host 1And if you remove all three trials that were at high risk of bias?
Host 2Immediate near transfer falls to zero point zero zero, and immediate far transfer to zero point zero four, with the heterogeneity dropping from about eighty four to forty seven percent. So the immediate benefits shrink to nothing when you keep only the more careful trials.
Host 1Now the question from the very start, what about three months or more after people stopped training?
Host 2Sustained near transfer, from four trials and two hundred and eighty seven people, was minus zero point zero seven, ranging from minus zero point three one to plus zero point one eight. Sustained far transfer, from five trials and three hundred and ninety seven people, was zero point zero two, ranging from minus zero point two six to plus zero point two nine.
Host 1Those numbers sit almost on zero, but they also look tighter than the earlier ones, is that right?
Host 2Yes, the trials agreed much more at the later time point; for near transfer the heterogeneity was zero. And removing the high-risk trials barely moved them, which is why GRADE rated these two results at moderate certainty, the best rating in the paper.
Host 1What about everyday life, the money, the medicines, the shopping?
Host 2Only three trials and three hundred and five people measured that right after training, and the estimate was minus zero point one six, slightly favouring the control groups, with a wide range from minus zero point six six to plus zero point three three. For the lasting version, only two trials reported anything, minus zero point two seven and plus zero point zero five, too few to pool.
Host 1Did any of the trials check whether the training kept people from developing dementia?
Host 2Essentially no. One American trial noted that over one year six participants developed functional difficulties suggestive of dementia, with no clear difference between groups, and its own authors said a year was too short to judge.
Host 1Were there any signals about side effects, or people dropping out because the training was too hard?
Host 2The trials reported this so inconsistently, with different definitions and denominators, that the authors could not pool it. They are careful to say that not reporting harms is not the same as having none.
Host 1Did any type of training look better than another, say supervised in a clinic versus alone at home?
Host 2In a check they added after seeing the data, estimates were somewhat larger for supervised, centre-based training and near zero for home or remote training. But they stress it is hypothesis-generating only, with wide intervals and no proper test, so it cannot show one format is better.
What it means
Host 1So if my mum has just been diagnosed with mild cognitive impairment, should she bother with the brain-training app?
Host 2This review does not show it gives her a lasting benefit on independent tests or in daily life, and the immediate effects are uncertain. The authors suggest it could be one part of an individual support plan, but not a replacement for physical activity, managing blood pressure and other vascular risks, social contact, sleep and emotional health.
Host 1Does this mean brain training is proven useless for people like her?
Host 2No, and the authors are firm about that. Moderate certainty here means moderate confidence that the average lasting effect is near zero, but the ranges still allow a modest benefit or a modest harm, and a benefit for some subgroup cannot be ruled out.
Host 1What's the general lesson here, beyond this one kind of training?
Host 2That getting better at what you practise is not the same as getting better at life. In the wider field this is a very old finding in learning research: skills tend to stay close to where they were trained, and here the trials did not show them spreading or lasting.
Host 1Why does this review land lower than the earlier, more hopeful ones?
Host 2The authors attribute it mostly to the question asked and the counting rules: classifying transfer by what was actually trained, separating immediate from lasting results, and counting each trial once. They say their estimates may be more conservative, but closer to what patients actually care about.
The caveats
Host 1Where are the weak points in the review itself?
Host 2The biggest is size: every analysis rests on two to eight trials, which is too few for subgroup analyses or for the usual tests of publication bias. And none of those trials were at low risk of bias.
Host 1Were there problems in how the reviewers themselves worked?
Host 2Some, and they report them openly. The original independent coding sheets were not kept, so they could not calculate how often the two reviewers agreed, and the near-versus-far system itself relies on judgement and has not been independently validated.
Host 1Are there any other limitations the authors admit to?
Host 2They only included published studies in English, which could miss unfavourable or foreign trials. Their registration record had inconsistencies with the final methods, and a full log of every exclusion decision from early screening was not available, so not every exclusion can be checked independently.
Host 1You said earlier they had to assume how closely related the tests were, could that assumption sway things?
Host 2It could affect the weights; that is why they reran the merging with correlations of zero point three and zero point seven. And beyond that, the trials themselves differed in diagnosis, training content, dose, delivery and comparison groups, too much to untangle with eight studies.
Host 1And did people actually do the training they were assigned, or do we not know?
Host 2We largely do not know. The trials described the planned schedule, but few reported the actual minutes completed, so the dose people really received remains unclear.
What to watch
Host 1What would a study need to look like to really settle this?
Host 2The authors ask for larger, multicentre trials with clear diagnostic criteria, outcomes registered in advance, and follow-up long enough to see whether booster sessions or continued access help effects last. They also want everyday function measured properly, not just test scores.
Host 1Could the training apps themselves help measure how much people really trained?
Host 2That is one of their suggestions: use the platform's own records of time spent, tasks completed and difficulty reached, and report technical problems and dropouts consistently. With enough data, pooling individual participants' results could show whether particular subtypes or doses respond.
Sign-off
Host 1So in one sentence, where does this leave us?
Host 2For older adults with mild cognitive impairment, computer brain training has not been shown to produce benefits that spread to other abilities or daily life and last after training stops, though eight imperfect trials cannot rule out a modest effect. The paper is open access in Frontiers in Aging Neuroscience, and the link is in the show notes.
Host 1Every weekend this same feed carries the roundup, the week's new papers in one short episode, should you prefer your uncertainty in smaller doses.
More from this week
- Parachutes, Eye Drops and a Comet That Fell Into the SunEpisode 13 · 10 min
- Does Sprinting With a Parachute Make You Faster?Deep dive · 18 min
- Does Sleeping Next to Your Dog Change Your Heart?Deep dive · 18 min
- Stronger Drops for Very Shortsighted ChildrenDeep dive · 14 min
- Can Qi Gong Online Lift Depression as Well as MoodGym?Deep dive · 17 min
- The Comet That Puffed Cyanide, Then Fell ApartDeep dive · 21 min