AI in Cardiac Remote Monitoring Earns Its Keep in the Workflow

This is the first in a regular series from Sean Shoffstall, Chief Product Officer at PaceMate, sharing his perspective on where cardiac remote monitoring and healthcare AI are headed. Look for new posts from Sean on the product thinking and industry shifts shaping the field. 

 A response to "Why healthcare AI fails without workflow redesign," Becker's Health IT, by Ramesh Yapalparvi, PhD  

Watch a cardiac device clinic on the Tuesday after a long holiday weekend. Three days of transmissions have stacked up in the queue. Two nurses are working it. Every manufacturer portal is open in its own tab, each with its own login, its own layout, its own idea of what matters. Somewhere in that stack is the patient whose ICD recorded a run of fast VT on Saturday night, and the clinical skill to recognize it is sitting right there in the room. What stands between the patient and the intervention is not expertise. It's the workflow.

Ramesh Yapalparvi published a piece in Becker's that gets this right. Stop asking how accurate the model is. Start asking how it changes the way your people work. Accuracy demos well. Workflow is where AI either earns its place in a clinic as a trusted tool or quietly gets abandoned.

I'd add one thing from where we sit. The clinics we serve arrived at this problem years before the rest of healthcare did. CIED remote monitoring has been drowning in continuous data since the first pacemaker started phoning home. Volume grew, staffing didn't, and no amount of predictive accuracy made a difference to a nurse buried in thousands of transmissions needing clearing.

The queue is the workflow

Alert fatigue in cardiac device clinics is not a perception problem. It is the direct product of an architecture that treats every transmission as equally deserving a human's attention. Most transmissions are routine. Some are urgent. The clinic learns which is which by opening all of them, one at a time.

An AI model that scores those transmissions more accurately and then presents the results in a new dashboard has made the day slightly worse. It added a click. The value shows up only when alert prioritization changes what actually lands in front of a clinician, which means the routine transmissions close themselves, and the exceptions rise to the top of the queue with the context already attached.

That distinction is measurable. Rules-based triage in our platform reduced alert burden by 61 percent in peer-reviewed work covering 132,000 patients, and physician alert burden specifically by 30 percent. The number I find more persuasive is the one from Ascension Jacksonville: 543 hours returned annually. That is not a productivity statistic. That is roughly a third of a clinical FTE, handed back to a device clinic that was already short-staffed, and spent instead on the patients who needed a person.

Automating a task and automating a decision are two different things

I say this often enough inside PaceMate® that people finish the sentence for me. Automating a task is documentation, data aggregation, transmission reconciliation, billing capture, the mechanical work that consumes a device clinic's day and requires no clinical judgment. Automating a decision is deciding that a patient's arrhythmia burden doesn't warrant a call. Those are different acts, and they deserve different levels of scrutiny.

There's a version of clinical decision support that gets this right, and it starts earlier than most people expect. Before AI narrows anything, it should assemble the patient. A single transmission is a snapshot. The clinical question is almost always longitudinal: how this episode compares to the last eleven, what the arrhythmia burden has done since spring, whether lead impedance has been drifting, what changed in the medication list, what the last in-clinic interrogation showed. Today an EP reconstructs that picture by hand, opening transmissions one at a time in one portal after another, holding the history in working memory long enough to make a call.

That reconstruction is the work AI should absorb. Pull the relevant history forward, show which prior events and trends it's drawing on, and the physician starts from a complete patient rather than hunting for one across a dozen sessions. Every data point, at your fingertips, and the decision still sits with the clinician who carries the license and the relationship. We keep a person in the loop on the decisions that matter, because judgment is what we're there for. It's also how AI earns durable trust from an EP team, and trust is what determines whether your carefully validated model is still in use in month seven.

Trust has an infrastructure dimension too, and it's worth being plain about it. HIPAA and SOC 2 are floors. Every platform meets them. ISO/IEC 27001:2022 and HITRUST r2 are what we hold, and we hold them because security posture is a precondition for putting AI anywhere near a clinical workflow, not a compliance checkbox to satisfy afterward.

Scaling a monitoring program without adding staff

The operational version of Yapalparvi's argument is the one device clinic leaders live with. Remote monitoring volume compounds. Implant rates rise, indications broaden, patients live longer with their devices, and every one of those patients transmits for the rest of their life. Hiring does not compound at the same rate. Neither does reimbursement.

Program growth without proportional staff growth is only possible if the unit of work changes. Not the number of clicks per transmission, the number of transmissions that need a human at all. That's workflow redesign in the specific, unglamorous sense: standardizing how episodes are triaged, letting the platform close what's routine, routing exceptions to the right person the first time, and writing the encounter into the EHR without a nurse retyping it. Sentara Health did this at system scale across facilities, which matters because a single clinic optimizing locally is a pilot, and pilots are where healthcare AI goes to stall.

The metrics a device clinic should actually track

Yapalparvi's four categories translate cleanly to remote monitoring, and I'd encourage device clinic leaders to write their own version on a whiteboard. Under technical performance, whether transmissions arrive and reach the record reliably. Under workflow adoption, documentation time per episode. Under clinical and operational outcomes, time from transmission to clinical action. And under strategic business value, revenue capture against what the program is entitled to bill.

If your AI investment can't move at least one of those four, it isn't integrated yet. It's adjacent to the workflow, which is a polite way of saying it's optional.

What I'm still working out

I don't yet know how far exception-based review can safely go. There's a ceiling somewhere, a point past which removing human eyes from routine transmissions stops being efficiency and starts being a risk, and I don't think anyone in this category has established where it sits with the rigor the question deserves. We've published 33 peer-reviewed papers and built a research-grade dataset of more than 2.3 million patients partly to work on questions like that one, and I'd rather answer it with evidence than with a product claim.

What I'm confident about is narrower. The device clinics that get the most out of AI over the next five years will be the ones that redesign the day around their queue, keep clinicians deciding, and measure whether the work actually got easier. The models will keep improving on their own. The workflow won't.

Your patients don't get to choose their monitoring platform. You do.