The Challenge of Long-Term Holter Analysis
Automatic analysis of long-term Holter recordings has always been challenging. One of the main reasons is variable signal quality—noise and artifacts that remain even after filtering can significantly affect how well algorithms perform. Simply put, the longer the recording, the more opportunities there are for such distortions to appear: electrode disconnections, patient motion, and interactions with the recorder all leave traces in the signal.
Fortunately, Holter recorders often capture multiple ECG channels. Using all high-quality channels helps improve arrhythmia detection, but relying on all available channels also comes with a trade-off. Noisy channels can introduce false alarms, especially in segments affected by artifacts. At the same time, noise in some leads can hide clinically important arrhythmias that are visible only in the single cleanest channel.
A Modular AI Approach to Better ECG Analysis
Our recently published patent application presents a comprehensive solution to this challenge—a modern, modular ECG analysis system powered by artificial intelligence.
At its core, the system is built from a set of cooperating modules, each responsible for a specific stage of the analysis. Together, they significantly improve R-peak detection and QRS classification in multichannel Holter recordings by intelligently managing channels based on their signal quality.
It all starts with automatic signal quality assessment. This module, trained on a large dataset of manually annotated ECG recordings, can accurately identify both short artifacts and longer segments of poor signal quality—independently for each channel.
Next comes intelligent channel management. The system continuously decides which channels and time segments are reliable enough to use and which should be ignored. By focusing only on high-quality data, it improves the robustness of the analysis—even in situations where only a single channel is usable.
The R-peak detection stage builds on this foundation. Based on a modified version of the Pan–Tompkins algorithm adapted for multichannel data, it uses only the segments identified as high-quality. The detection is intentionally tuned for very high sensitivity, ensuring that all potential heartbeats are captured—even at the cost of additional false positives, which are handled later in the pipeline.
Finally, the QRS classification module brings everything together. It combines information from both channel management and R-peak detection. In addition to standard heartbeat classes (normal, ventricular, and supraventricular), it introduces a dynamically generated category called “non-QRS.” This class represents falsely detected R-peaks, allowing the system to learn how to eliminate detection errors in an adaptive, data-driven way.
Real-World Impact
The result is a clear and measurable improvement: fewer false detections, fewer missed beats, and more reliable analysis overall. In practice, this leads to more accurate calculation of key diagnostic parameters such as heart rate variability (HRV), arrhythmia detection, and ectopic beat identification—even in the presence of noisy channels.