In:
Circulation, Ovid Technologies (Wolters Kluwer Health), Vol. 146, No. Suppl_1 ( 2022-11-08)
Abstract:
Introduction: Atrial fibrillation (AF) alerts from clinical monitoring systems are often inaccurate. Recently reported AF detectors utilize curated, publicly-available databases for model training. Real-world data rather than curated databases would be better suited to improve upon these inaccuracies, but labeling high-frequency waveform data over ICU stays would require significant clinician hours. Hypothesis: We hypothesize that AF can be accurately and efficiently labeled in stored, real-world telemetry data through clinician-made heuristics describing AF that inform a weakly supervised AF detection model. Methods: We examined high-frequency telemetry data from 72 beds in a tertiary care center using a commercial monitoring system. Previously described time- and frequency-domain heart rate variability features centered over 10-second segments were calculated. K-means clustering was applied over RR interval Poincaré plots (RR n vs RR n+1 ) to describe rhythmic patterns. Over 20 features and associated clinician-driven heuristics delineating AF from other rhythms informed a weakly supervised AF labeling algorithm. Performance was evaluated on ~100 independent, hand-labeled segments. Results: Over 10,000 10-second telemetry segments from 657 ICU patients were used to train our weakly supervised model. The model had an AUC of 0.91 and accuracy of 88%. Positive predictive value and sensitivity were 0.67 and 0.82 for AF segments and 0.95 and 0.89 for non-AF segments. Compared to our commercial monitoring system, our model demonstrated significant improvement in AF F1 score (0.73 vs 0.56) with the same F1 score for non-AF rhythms (0.92). The most influential features on AUC included a standard measure of heart rate variability (pNN20) and rhythm clustering pattern (sum of squared error). Conclusions: We demonstrated a time-saving means of labeling AF in a large, real-world telemetry database through weak supervision. Features describing heart rate variability and RR interval irregularity through clustering techniques led to intuitive, clinically sound rules for identifying AF. Our contribution permits the creation of much larger, more clinically relevant datasets with AF clearly labeled for improved study and care of this population.
Type of Medium:
Online Resource
ISSN:
0009-7322
,
1524-4539
DOI:
10.1161/circ.146.suppl_1.10198
Language:
English
Publisher:
Ovid Technologies (Wolters Kluwer Health)
Publication Date:
2022
detail.hit.zdb_id:
1466401-X
