Beyond the Transmission: Alert Burden & Triage

A CMO-led series on what cardiac data can teach us about better care. 

Welcome to Beyond the Transmission, a blog series led by Dr. Benjamin Steinberg, PaceMate's Chief Medical Officer, exploring what cardiac data can teach us about delivering better care.  

In this Q&A, he examines the challenge of alert burden and triage, looking at why the volume of remote monitoring data so often outpaces our ability to act on it and what well-designed triage really requires.

Q: Remote cardiac device monitoring generates an enormous volume of alerts. Before we talk about solutions, can you put the problem in concrete terms?

A: I've been living with this problem for a long time — longer than I've been at PaceMate. Running a busy EP practice and spending real time in the device clinic, one pattern keeps reasserting itself: the infrastructure we've built for remote monitoring is extraordinary at collecting information and genuinely underpowered at using it. Devices transmit daily, data accumulates in queues, and embedded in that data — somewhere — is clinical signal that could change what we do for a patient that week. The system just isn't built to find it.

That's the core of the alert burden problem. It isn't only volume, though volume is real. It's the mismatch between the rate at which the data is generated and the rate at which a clinical team can meaningfully engage with it. And when that mismatch becomes large enough, the queue stops functioning as a safety system. It becomes noise. Not because anyone failed, but because the system wasn't designed for the scale it's now operating at.

Q: How does that translate into patient risk? Who actually gets hurt?

A: I want to be precise about this, because I think the way alert burden usually gets framed misses the deeper clinical issue.

The conversation tends to stop at efficiency: response times slow, workflows back up, teams get overwhelmed. All true. But efficiency is a system metric. The question I care about — the question that should govern how we design these programs — is whether patients' outcomes are worse because of it. And that's a different question, and harder to answer.

Here's the mechanism that keeps me up at night. A meaningful proportion of patients with atrial fibrillation are minimally symptomatic or entirely asymptomatic — a finding that's consistent across multiple datasets and well-documented in the literature. Symptom-rhythm discordance is real: what the patient feels and what the device records often don't match. That means the device, in many cases, is the only witness to what's actually happening. There's no call from the patient, no emergency department visit. Just a transmission in a queue.

So consider a patient who is 70, has paroxysmal AF, a CHA₂DS₂-VASc of 4, and whose AF burden has been climbing steadily over three months of device transmissions. That trend is visible in the data — has been visible, accumulating, for weeks. But if the alerts aren't triaged in a way that surfaces it, that patient has no other path to the clinician's attention. And by the time something happens, the window for prevention has passed.

That isn't a failure of diligence on anyone's part. It's a failure of design.

Q: What does good alert triage actually require? What are the essential components?

A: The first thing I'd say is that triage has to be oriented toward patients, not toward alerts. That distinction matters more than it might seem, because most monitoring platforms present information the other way around — they surface the event, the alert, and then ask the clinician to reconstruct the patient's clinical context from there. That's backwards.

When a coordinator opens a transmission, they should already know: What is this patient's underlying substrate? What is their anticoagulation status, and are they actually taking it? Has their AF burden been trending upward? Have there been other device findings — fluid indices, activity levels — that together are telling a story larger than any single alert? Without that context, you're triaging the alert. With it, you're triaging the patient. Those lead to meaningfully different decisions.

The second requirement is explicit stratification. The same AF episode in a 45-year-old with no comorbidities is not the same clinical event as the same episode in a 72-year-old with prior stroke and compromised renal function. A system that treats them as identical in the queue is, at some level, a system that doesn't know the difference — and the clinical team is left to reconstruct that context from scratch on every review.

Third, and this is consistently underappreciated, are clear protocols. Who acts on which alert category, on what timeline, and with what documentation? Ambiguity in a triage workflow is how urgent findings slip through the seams between team members. Not through inattention. Through a system that didn't make the path obvious.

Q: How does AI fit into a well-designed triage system, and where does it fall short?

A: Trust is central to this. That's not a throwaway line. It's one of the primary constraints on what's achievable.

What AI can increasingly do in this context, when it's built and validated carefully, is what I'd call adjudication at scale. If you have longitudinal data across hundreds of thousands of patients with implanted devices, transmissions, alert histories, clinical trajectories, outcomes, you can train a model to recognize which patterns tend to precede adverse events, and use that to prioritize the queue. You can surface the patient whose trajectory is changing before they've had a reason to call. That work — aggregating findings across time, flagging combinations of signals that together warrant action — is not something any human team can do at the scale modern monitoring programs now operate at. That's real, and it matters.

But there's something the model cannot know: that the patient was hospitalized two weeks ago and her renal function dropped, or that she told you at her last visit she'd stopped anticoagulation because of a fall concern and you were still working through it together. Context that lives in the relationship, not in the device data. No algorithm currently captures that — and I'd be skeptical of one that claimed to.

So the question isn't whether AI can help. It's whether we build it honestly — with the understanding that it adds a layer of adjudication at a scale no team of humans can match, while the clinical decision remains where it belongs. The way you earn physician trust for AI tools is the same way you earn trust for any new therapy: published evidence, transparent methodology, honest acknowledgment of what the data don't show. Not efficiency metrics dressed up as outcomes. Actual outcomes.

The data we already collect from these devices is richer than our current ability to use it. That's genuinely the opportunity. But the technology only matters if the clinician trusts it enough to act on it, and that trust is earned slowly and lost quickly.

Q: If you could leave device clinic teams with one thing from this conversation, what would it be?

A: I'd ask them to sit with one question: can you close the loop?

Most programs can tell you their alert volume. Many can tell you response times — how quickly an alert was reviewed, how quickly a patient was contacted. That's workflow data, and it matters for operations. But it's not the same as knowing whether the surveillance is actually changing what happens to patients. Of the alerts your team reviewed last month, how many led to a clinical action? Of those, how many changed an outcome?

Most programs can't answer that. Not because the people running them aren't thoughtful, they are, but because the infrastructure wasn't built to close that loop. The connection between the alert and the downstream outcome isn't tracked in a way that supports learning.

That measurement gap is, I'd argue, the next frontier for remote monitoring. We've gotten very good at generating data. We've gotten better, incrementally, at reviewing it. What we haven't done rigorously is ask whether the reviewing is working, and for which patients, and under what conditions.

The monitoring infrastructure exists. The patients are there. The clinical questions are clear. What we owe them — the ones counting on this system to see what they can't feel — is the commitment to answer those questions honestly.