Your Patients Aren't Ignoring You. They're Ignoring Your Questionnaire.

I spent eighteen months building a mental health app for people with bipolar disorder.

It was good. Genuinely. Mood tracking, sleep logs, med reminders, a clean interface a designer friend made me pay for twice. We launched it into a small clinic population and waited for the data to tell us we were geniuses.

Fifteen percent weekly engagement. That's what we got. Fifteen. Which in app terms is not catastrophic, and in clinical terms is a rounding error, because the eighty five percent who stopped opening it were exactly the people whose sleep was collapsing at three in the morning.

I threw it out.

Then we called them instead. AI on the phone, asking how they slept, whether they took the lithium, whether anything felt off. Eighty five percent engagement, week after week. Same patients, same clinic, same questions. Different door.

That flip is basically why HANA exists, and I think about it every time somebody publishes another trial proving remote monitoring doesn't work.

Does remote monitoring actually reduce hospital readmissions?

Mostly no, and there's now a big trial saying so out loud. The ACCOMPLISH randomized trial published in JAMA Network Open followed 1,286 adults discharged after sepsis or a lower respiratory tract infection across 19 hospitals, tested four different remote monitoring strategies against usual care, and found none of them increased days spent alive at home. Readmissions came in at 37.8% with usual care and between 36.3% and 44.2% across the monitoring arms.

In patients over 65, monitoring did worse than doing nothing special.

Read that again. The population Medicare reimburses you to monitor got fewer days at home when you monitored them.

Why didn't it work?

Because the intervention was a questionnaire, and questionnaires need a well person on the other end. Of the 887 patients assigned to monitoring, only 529 actually enrolled. Of 10,561 questionnaires sent, 56% came back. The nurses were fine. They responded to over 94% of alerts. The University of Pittsburgh summary is careful about this: the bottleneck wasn't alert neglect.

The bottleneck was the fifty six percent.

You cannot triage a symptom nobody reported. Every remote monitoring architecture I've seen quietly assumes the patient does the work of initiating, and then measures the clinicians. It's like grading a restaurant on kitchen speed when half the orders never got taken.

What makes a phone call different?

The call inverts who has to act. Nobody downloads anything, remembers a password, charges a device, or decides at 9pm that today is the day they'll finally fill in the form. The phone rings. They answer. Somebody asks about their breathing.

We've run over a million of those conversations across five countries and three languages, with zero critical adverse events, and the thing that surprises clinicians every single time is who picks up. Eighty two year olds pick up. People who have never installed an app in their life pick up. The use cases where this lands hardest are exactly the ones the trials keep failing on: post-discharge, post-op, chronic med adherence, older patients living alone.

A questionnaire selects for the healthy and the digitally comfortable. A phone call doesn't select at all.

Isn't an AI call worse than a nurse call?

Worse than a nurse call, yes, obviously. Better than the nurse call that never happens, which is the actual comparison.

Look at what usual care meant in that trial. A post-discharge phone call from a nurse and continued primary care management. That's a genuinely strong control arm, and honestly the more interesting finding in ACCOMPLISH is how hard usual care was to beat. The phone call worked. It reached 64% of patients, and 78% of those answered.

So the question for your clinic isn't whether AI beats a nurse. It's whether you can call every single patient, every week, in their language, at a time they answer, forever. If you can, do that. If you're like every clinic I've ever sat in, you're reaching maybe a fifth of them, and the ones you skip are the ones you'd have most wanted. We publish the outcomes and cost side of that math because it's usually the part people don't believe.

So what should a clinic do differently?

Stop buying monitoring and start buying contact. Concretely: pick the smallest cohort where a missed deterioration is expensive and obvious, post-op or post-discharge, call them on a fixed rhythm instead of waiting for them to report, route only the exceptions to a human, and measure percentage of patients actually reached rather than percentage of alerts closed.

That last metric change is the whole thing. Reach is upstream of everything else, and it's the number nobody puts on the dashboard because it's embarrassing.

Then put it where the work already lives, which is your EHR and not another tab. Our docs cover that part.

Key Takeaways

The largest trial of its kind found remote symptom monitoring after sepsis or pneumonia didn't increase days at home, and made things worse for patients over 65. That isn't a story about bad technology. It's a story about a design that makes sick people do the reaching, which they don't, which is why only 56% of questionnaires came back. Voice flips the burden: the system reaches out, the patient just answers, and engagement goes from the teens to eighty five percent without asking anybody to learn anything new. If you're evaluating post-discharge programs this year, ask the vendor what share of patients they actually reach. Not alerts. Reach.

FAQ

Does remote patient monitoring reduce readmissions?

The evidence is much weaker than the marketing. The 2026 ACCOMPLISH trial found no reduction in readmissions across four monitoring strategies for sepsis and respiratory infection patients, and worse outcomes in adults over 65. Earlier positive results tend to come from small cohorts in heart failure and COPD with heavy human coordination attached.

Why do older patients engage with phone calls but not apps?

Because a call requires nothing but answering a ringing phone. Apps and questionnaires require enrollment, device literacy, and patient-initiated action, all of which drop off sharply with age, illness severity, and social isolation. Our own research and deployment data show the reach gap widening exactly where clinical risk is highest.

Can AI phone calls be safe for post-discharge patients?

They can be, with the right guardrails: fixed protocols, no diagnosis, clear escalation to a human on any red flag, and full transcripts in the record. Across more than a million interactions we've logged zero critical adverse events. If you want to pull apart how that's built rather than take my word for it, book a call and I'll walk you through it.