Physiopathological Insights into Atrial Fibrillation Onset through Heart Rate Variability Correlations

Our study identified five key HRV parameters primarily responsible for the accuracy of short-term prediction of paroxysmal AF episodes. Each parameter reflects a distinct class of pathophysiological mechanisms that trigger such episodes. An important point to note is that PACs are known to trigger AF episodes, their number increases before these episodes, and this has already been used in models to predict AF episodes [6]. In this study, HRV calculations were performed on the first hour of recordings, when the number of PACs is low and, moreover, changes in HRV occur before the increase in PACs. Additionally, HRV calculations were performed using cubic spline interpolation of RR intervals, thereby excluding PACs from the predictions. As shown in Figure 2, our approach highlights groups of HRV features with closely related variability patterns, enabling clear cluster identification and interpretation across physiological domains (e.g., time-domain, frequency-domain, and non-linear metrics). Using Spearman’s method ensures robustness to non-linear relationships and non-normal data distributions, both of which are common in HRV datasets.. As illustrated in Figure 3, the HRV feature set exhibits substantial variability. Notably, the five most important predictive parameters are distributed across different regions of the PCA space, indicating minimal redundancy among them. According to the SHAP scores, the parameters used by the decision tree, in order of importance, are: RMSSD, PAS, SODP Q1, SODP CTM100, and SD1/SD2. This indicates that the model identified these parameters as the most predictive. The dimensionality reduction analyses (t-SNE, UMAP, PCA) performed on the absolute Spearman correlation matrix revealed consistent and physiologically meaningful clustering patterns among the HRV parameters. Notably, the projections highlighted distinct groupings centered around five key parameters: RMSSD, PAS, SODP_Q1, SODP_CTM_100, and SD1/SD2. These clusters likely reflect different underlying physiological domains contributing to heart rate dynamics. RMSSD corresponds to the time-domain statistical HRV metrics computed over 5-min segments. It is a well-established marker of short-term vagal activity [13]. PAS is an index belonging to the cluster of fragmentation parameters that capture disruptions in the neuroautonomic-electrophysiological regulation of cardiac rhythm, a phenomenon frequently associated with aging and the progression of cardiovascular disease[14]. Altered PAS values have been linked to ANS dysfunction and may reflect underlying sick sinus syndrome, a condition often implicated in the pathogenesis of atrial fibrillation [15]. SODP_Q1 corresponds to the first quadrant of the SODP derived from RR interval variations. This quadrant captures sequences in which both the current and previous RR interval differences are positive, reflecting a progressive slowing of the heart rate over three consecutive beats [16]. SODP_Q1 is thus considered a marker of very short-term vagal modulation [17]. Its relevance in the model underlines the importance of rapid, transient parasympathetic influences detectable at the beat-to-beat level. Interestingly, both PAS and SODP_Q1 belong to the same cluster in our analysis, suggesting a potential physiological link between neuroautonomic fragmentation and short-term vagal modulation. This association supports the hypothesis that episodes of abrupt changes in cardiac rhythm variability and transient vagal surges may share common pathophysiological substrates in the development of AF. SODP_CTM_100 belongs to the cluster of central tendency measures (CTMs) derived from the SODP, which are designed to quantify the overall degree of variability within the plot. This parameter captures global variability, encompassing short-term, long-term, and non-linear components of the RR interval dynamics [18]. SD1/SD2, the fifth factor corresponds to the balance between short- and long-term variability in RR intervals. Specifically, the width of the Poincaré plot (SD1) reflects the magnitude of short-term heart rate variability (HRV), while its length (SD2) represents long-term variability. The SD1/SD2 ratio, derived from these two measures, captures the relationship between these distinct temporal components of HRV. This parameter appears isolated from the other clusters in our analysis, suggesting that it reflects a distinct, predominantly non-linear aspect of heart rate dynamics [19]. To the best of our knowledge, no prospective or randomized study has yet addressed the short-term prediction of AF onset using HRV or other ECG-derived metrics. Existing work in this field is limited to a few studies that identified an AF “signature” from 12-lead ECGs recorded months before the arrhythmia, which fundamentally differs from our objective of predicting AF onset within a 24-h time window [5,20,21,22]. Based on two-week Holter monitoring, only one retrospective analysis reported AF prediction several days in advance [23]. Our approach is therefore distinct in targeting an actionable, short-term prediction horizon (<24 h) using Holter-derived HRV features, with the potential to support timely, preventive clinical interventions. It should be noted that variations in HRV are not specific to the condition under investigation and can be observed in a broad range of clinical contexts. Any disorder affecting the autonomic nervous system has the potential to modify HRV. Cardiac conditions such as acute or chronic heart failure, particularly during decompensation, and coronary artery disease are commonly associated with a global reduction in HRV, reflecting an imbalance with increased sympathetic and reduced parasympathetic activity. Beyond cardiac disorders, several non-cardiac conditions, including diabetes mellitus and other forms of dysautonomia, can also impair autonomic regulation and induce significant alterations in HRV patterns [17]. Apart from providing insight into the underlying pathophysiology, the identification of the five most significant uncorrelated HRV parameters associated with AF onset offers the opportunity to predict paroxysmal AF episodes using models relying solely on these parameters, thereby reducing model complexity while maintaining predictive performance. This enhances their suitability for integration into wearable devices, such as smartwatches. Ultimately, this could enable a ‘pill-in-the-pocket-like’ preventive strategy, distinct from the conventional approach, which involves taking medication after AF onset, whereby the patient would take a class I antiarrhythmic or a beta-blocker before the arrhythmia occurs, when the model predicts an autonomic state indicating a high likelihood of imminent AF.