Heart rate variability helps classify phenotype in systemic sclerosis

Abstract

Abstract We aimed to develop a systemic sclerosis (SSc) subtypes classifier tool to be used at the patient’s bedside. We compared the heart rate variability (HRV) at rest (5-min) and in response to orthostatism (5-min) of patients (n = 58) having diffuse (n = 16, dcSSc) and limited (n = 38, lcSSc) cutaneous forms. The HRV was evaluated from the beat-to-beat RR intervals in time-, frequency-, and nonlinear-domains. The dcSSc group differed from the lcSSc group mainly by a higher heart rate (HR) and a lower HRV, in decubitus and orthostatism conditions. Stand-up maneuver lowered HR standard deviation (sd_HR), the major axis length of the fitted ellipse of Poincaré plot of RR intervals (SD2), and the correlation dimension (CorDim) in the dcSSc group while increased these HRV indexes in the lcSSc group (p = 0.004, p = 0.002, and p = 0.004, respectively). We identified the 5 most informative and discriminant HRV variables. We then compared 341 classifying models (1 to 5 variables combinations × 11 classifier algorithms) according to mean squared error, logloss, sensitivity, specificity, precision, accuracy, area under curve of the ROC-curves and F1-score. F1-score ranged from 0.823 for the best 1-variable model to a maximum of 0.947 for the 4-variables best model. Most specific and precise models included sd_HR, SD2, and CorDim. In conclusion, we provided high performance classifying models able to distinguish diffuse from limited cutaneous SSc subtypes easy to perform at the bedside from ECG recording. Models were based on 1 to 5 HRV indexes used as nonlinear markers of autonomic integrated influences on cardiac activity.

Keywords

• Systemic sclerosis
• Heart rate variability
• Nonlinear dynamics
• Cardiovascular function
• Machine learning