A 1,286-Patient Trial Found Remote Monitoring Didn't Reduce Readmissions. Here Is What That Means
Years ago I watched a Stripe dashboard cross a million dollars in a single day, and I remember standing there with my hands on my head like an idiot, absolutely convinced I'd cracked something.
The product was bad.
Not fatally bad. Bad in the quiet way, the way where the returns show up ninety days later and the reviews get weird and you realise the number on the screen was measuring how well you bought attention rather than how well the thing worked. Scale hides problems. It's very good at it. And I thought about that dashboard again this summer when a randomised trial landed that most of the remote monitoring industry has been extremely quiet about.
Did remote patient monitoring reduce readmissions in the trial?
No. The ACCOMPLISH trial published in JAMA Network Open randomised 1,286 patients across 19 hospitals after hospitalisation for sepsis or lower respiratory tract infection, tested four different remote monitoring configurations against usual care, and found no improvement in days spent alive at home at ninety days. Readmission rates sat between 36% and 44% across every arm, including usual care. Among patients aged 65 and over, the monitoring arms did worse than doing nothing, with inferiority probabilities of 99.6% and 97.9%.
Worse. In the oldest, sickest group. That's not a null result, that's a finding.
Why would monitoring make outcomes worse for older patients?
Look at the enrolment number, because it's the whole story. Of 887 patients assigned to remote monitoring, 529 actually enrolled. That's 59.6%. Four in ten people randomised to the intervention never really received it, and the ones who did had to own a connected device, keep it charged, and fill in questionnaires while recovering from sepsis.
Then there's the alert side. More questionnaires means more flags, more flags means more precautionary escalation, and precautionary escalation for a frail 78 year old often ends in an emergency department at eleven at night. The monitoring didn't fail to detect. It detected plenty. It just detected into a workflow that had one lever, and the lever was send them back.
What separates programmes that do work?
Reach, and then judgement. UPMC is expanding an AI-native transitional care model to 18 hospitals this autumn after starting at a single campus, and the design detail that matters is this: automation reaches 100% of enrolled patients within 48 hours of discharge, including the ones who don't pick up first time and the ones no risk score flagged. Then it routes the subset who need a real conversation to a coordinator. Clinician time on non-clinical work fell 63% per encounter and triage removed 82% of non-actionable alerts.
Notice what's inverted. The machine does the reaching. The human does the deciding. Most failed programmes have it exactly backwards: humans grinding through call lists at the top of the funnel, then an algorithm quietly making the escalation call at the bottom.
So is the real variable engagement rather than monitoring?
I think so, and there's a cleaner natural experiment for it than anything in the sensor literature. In a post-discharge outreach programme for heart failure patients, the highly engaged group hit a 7.7% thirty-day readmission rate. The low engagement group, same programme, same messages, same clinicians, hit 21.6%. The intervention didn't change. The showing up did.
Which means the binding constraint on every transitional care programme in the country isn't detection sophistication. It's contact. And contact is a distribution problem dressed up as a clinical one, which is exactly why we spend most of our research effort on the boring end of the funnel.
What should a health system build first?
Contact rate first, and nothing else until it's above 80%. Not a dashboard, not a risk model, not a wearable pilot. Reach every discharged patient within 48 hours in their own language, on a channel that requires them to do nothing except answer, and get a structured response. Only then build triage, because triage on a 60% contact rate is just a nicer way of losing the same 40%. Then the escalation pathway, and it needs more than one option, because a system whose only response to a flag is come back to hospital will reproduce the ACCOMPLISH result at scale. The integration path is usually the easy part. The hard part is deciding what happens after somebody says they don't feel right, and our deployment write-ups spend more pages on that than on anything technical.
Key takeaways
A well-run randomised trial across 19 hospitals found remote monitoring didn't increase days at home after sepsis or respiratory infection, and actively reduced them for patients over 65. The most likely explanation isn't that monitoring is useless, it's that only six in ten assigned patients enrolled and that the alerts landed in a workflow whose only real response was readmission. Programmes that work invert the roles, letting automation carry the reach and humans carry the judgement, and the clearest evidence for that comes from engagement gradients rather than sensor accuracy: same programme, 7.7% readmission for engaged patients against 21.6% for disengaged ones. Build contact first, escalate into more than one option, and treat any monitoring investment that hasn't solved reach as a very expensive way to observe people getting worse. We run 85% weekly engagement against a 15 to 20% baseline across five countries and three languages, and the reason is that we treated reach as the product rather than as a precondition.
Frequently asked questions
Does this trial mean we should cancel our RPM programme?
No, it means you should audit your enrolment and escalation numbers before you expand it. If fewer than three quarters of eligible patients are actually enrolled and your only escalation path leads to the emergency department, you're running the trial's failure conditions.
Why does voice reach more patients than apps or portals?
Because it asks nothing of the patient beyond answering. No download, no password, no charged device, no literacy assumption. That gap widens exactly in the populations with the highest readmission risk.
How do you build a business case when the evidence is mixed?
On contact rate and staff hours returned, which are direct and measurable inside a quarter, rather than on readmission deltas that take a year to separate from noise. Model the clinical benefit as upside, not as the justification.
If you want to pressure test your own numbers, the cost side is published and the rest is a conversation. Grab twenty minutes and bring your discharge contact rate. That single number tells us more than any pilot deck.