Use of artificial intelligence algorithm to increase productivity in ILR monitoring: a multi-centre observational study

Background Implantable loop recorders (ILRs) enable precise diagnostic evaluation of syncope and atrial fibrillation (AF) but generate large volumes of false-positive alerts, creating substantial additional workload. Remote monitoring can amplify this burden. Deep learning algorithms aim to suppress false alerts and improve efficiency. To date, no robust, multicentre, real-world evidence across a fixed consistent cohort has been published. Purpose To quantify the effect of Medtronic AccuRhythm algorithm on alert burden and estimated productivity in a consistent ILR cohort across multiple NHS Scotland centres Methods Retrospective, multicentre cohort analysis of consecutive ILR patients monitored for at least 12 months before and at least 2 months after AI activation ("switch-on"). All transmitted episodes were counted in each period. Between-period differences in alert counts (overall and by site) were tested using paired T-test to derive p-values. Time savings were estimated using published time-and-motion data for ILR transmission review. Results Across four sites, 445 patients (Reveal LINQ n = 438; LINQ II n = 7) were included. 440 (99%) were inserted for syncope (mean age 67 [SD ± 15] years, male 218 (49.6%). The overall intervention effect was highly significant; total alert volume decreased from 4,261 to 2,509, representing a 29% reduction (P < 0.05; paired t-test). Mean paired difference per patient was - 3.9 alerts (95 % CI −7.5 to −0.4), indicating a statistically significant within patient reduction (p = 0.026, paired t-test). The magnitude of the false episode filtering effect differed between sites, indicating heterogeneity in implementation effect. Safety bypass rules means the full potential implementation effect is likely underestimated (and may improve with further algorithm iteration). Alert generation was highly concentrated: 7% of patients accounted for >90% of alerts. 104 patients were intermittently disconnected from remote monitoring, indicating an opportunity to optimise connectivity solutions. Applying established workflow timings, the alert reduction translates into a material saving in physiologist review time, approximating four hours each week of "virtual physiologist" capacity gain (table 1 [1]) Conclusions In a consistent, real-world cohort, deep learning algorithms can significantly reduce ILR alert burden, with multicentre consistency and meaningful productivity gains. Concentration of alerts in a small subgroup suggests targeted management may yield additional efficiency. In resource-constrained services, AI-enabled filtering can function as a "virtual physiologist," improving throughput while preserving clinical oversight.