TriageHF® in Real-World Practice: The Limits of “No News Is Good News” Authors: Masataka Kawana Published: 2026 Heart failure (HF) remains one of the most resource-intensive conditions in cardiovascular medicine, and ambulatory monitoring has moved rapidly from single-parameter surveillance (weight, symptom diaries, thoracic impedance) toward multiparametric, sensor-fused algorithms. Multisensor algorithms embedded in cardiac implantable electronic devices (CIEDs) have shown promise across retrospective and prospective cohorts for anticipating decompensation before it reaches the emergency department,1,2) reflecting a simple logic: HF hospitalizations are common, costly, and often preventable if risk is identified early. A large proportion of HF patients already carry CIEDs—implantable defibrillators, permanent pacemakers, and resynchronization devices—that continuously capture diagnostic data repurposable for HF risk stratification at no additional cost or procedure. This is the appeal of algorithms such as TriageHF® (Medtronic, Minneapolis, MN, USA), which combines intrathoracic impedance, activity level, nocturnal heart rate, heart rate variability, atrial fibrillation burden, percentage of resynchronization pacing, and detection of ventricular arrhythmias into a daily estimate of 30-day HF event risk.1) Because the hardware is already implanted, CIED-based risk scores are among the most immediately scalable remote monitoring tools available today. In this issue of the journal, Pereira Santos et al.3) reported a real-world, modestly sized, retrospective cohort study evaluating the performance of TriageHF®. Their cohort had an event rate of 7.9% per month during high-risk alert days, compared with 0.8% per month during standard-risk days; TriageHF® yielded a sensitivity of 35%, a specificity of 71%, a positive predictive value (PPV) of 25%, and a negative predictive value (NPV) of 80% for predicting a composite of HF hospitalization and unplanned emergency room visits. Notably, rather than using the maximal monthly risk score employed in other studies, they defined high-risk exposure by the cumulative duration of continuous days under a high-risk alert. This is a coarser unit of analysis: with fewer independent observations per patient than a day- or month-level analysis provides, the resulting estimates carry more statistical noise, and they answer a different question—the probability that a patient who ever crosses the high-risk threshold ever has an event during follow-up, rather than the probability that a specific high-risk day or month is followed by an event within a defined window. The two are not directly interchangeable, and this difference in granularity should be kept in mind when comparing this cohort’s figures with the day- or evaluation-level estimates below. Even with that caveat, the low sensitivity and PPV reflect a pattern that sits within the wide range of algorithm performance established by prior work. The original report by Cowie et al.1) that developed TriageHF® showed a high-risk PPV of 6.8% (sensitivity 46%) in its 1,310-patient validation cohort; Zile et al.,4) in 22,901 patients, found an even lower PPV of 2.23% (sensitivity 39%), a threshold the authors deliberately set conservative to limit false alarms; García Iglesias et al.5) reported a higher PPV of 27.27% (sensitivity 30%). High-risk PPV has thus ranged from roughly 2% to 27% across cohorts, and much of that spread likely reflects differences in event prevalence rather than in the algorithm itself, since PPV moves with the base rate just as NPV does. HF event prevalence in turn depends on the endpoint definition used and the patient population studied: García Iglesias et al.’s5) composite endpoint, for instance, additionally captured New York Heart Association class deterioration requiring any treatment modification, a broader trigger than the hospitalization- or emergency-care-based endpoints used elsewhere, which would tend to raise both the event rate and PPV independent of the algorithm’s underlying discrimination. That same base-rate dependence cuts the other way for NPV. Cowie et al.’s original data1) showed that even the standard high-risk threshold missed roughly half of true events (46% sensitivity), yet the separately defined low-risk group still posted a 99.4% NPV, simply because the overall 30-day HFH rate was low (1.5%). The same held in Zile et al.,4) 39% sensitivity at the high-risk threshold still coincided with a 99.75% NPV in the low-risk group, whose event rate was just 0.25%, and in García Iglesias et al.,5) where 30% sensitivity for high-risk alerts still produced a 97.67% NPV. Paired with the low sensitivity at 35% reported by Pereira Santos et al.,3) this means a meaningful share of true events likely occurred without a preceding alert. Cowie et al.1) themselves cautioned against acting unilaterally on a high-risk state given its modest absolute risk—a caution now echoed by real-world PPVs in a similarly low range. Neither “no news is good news” nor “bad news demands action” holds without discriminative performance behind it; in this cohort, neither alert carried it. In the US, avoiding 30-day HF readmissions carries added weight under the Hospital Readmissions Reduction Program, which penalizes hospitals for excess risk-adjusted readmission rates.6) A monitoring algorithm supports that goal only if its risk signal is reliable enough to act on, and acting on it changes outcomes. Randomized trials of CIED-based remote management have already raised doubts about the second condition—despite sound risk detection, translating alerts into outcome benefit has not been straightforward, and results have not uniformly shown reductions in hospitalization or mortality.7) The present study suggests the first condition cannot be assumed either: real-world practice is exactly the setting where sensitivity and PPV from earlier validation cohorts may not reproduce, and here they largely did not. None of this diminishes the study’s contribution—if anything, it is the more useful message. Rather than confirming TriageHF® as a ready-made triage instrument, the data show that real-world diagnostic performance needs to be established directly in the population where an alert pathway will be used, not assumed from earlier validation cohorts. A “no alert” status should not deprioritize a patient from surveillance unless the underlying sensitivity has been demonstrated locally; here it had not. Nor can a high-risk flag serve as an efficient trigger for intervention when roughly 92% of such flags did not predict an event (given 7.9% event rate during high-risk alert days as above). On the other hand, the large real-world cohort by Zile et al.4) did find a clean 4-year mortality gradient (14% low-risk to 38% high-risk), suggesting the score may hold up better for longer-term risk stratification than for 30-day event triage—a distinction worth preserving rather than treating every TriageHF® application as equally uncertain. Larger, prospective, multicenter cohorts that is powered to estimate sensitivity and PPV with reasonable precision are needed before TriageHF® alerts are built into staffing or formal triage protocols. In conclusion, the present findings add a valuable, real-world data point to the TriageHF® literature. A reassuring NPV in this cohort was accompanied by modest sensitivity and a PPV for high-risk alerts under 10%—squarely within the roughly 2%–27% range reported across the algorithm’s other validation cohorts. Together, these results suggest TriageHF®’s real-world performance for high-risk PPV is inherently variable and often modest, and that its application in everyday practice warrants caution, particularly before alerts are used to guide surveillance intensity or staffing decisions. Confirming its discriminative performance in the population where it will be deployed remains a worthwhile next step for the field. Footnotes Conflict of Interest: The author has no financial conflicts of interest. References