Abstract 13997: Identification of Novel Spectral Densities in Arterial Pressure Time Series for Cardiovascular Risk Stratification in Early Prediabetes: Machine Learning Beyond Conventional Frequency Domain Analysis

Abstract

Introduction: Despite the extensive use of average BP for diagnosis of cardiovascular disease (CVD), research has shown that changes in BP variability can better reflect CV deterioration. As patients with type 2 diabetes present to the clinic with pre-established CVD, there is a pressing need for early identification of aberrant BP fluctuations useful for risk appraisal. Hypothesis: We hypothesized that deep learning models with superior capacity in feature extraction would enable comprehensive characterization of arterial pressure (AP) time-series allowing CV risk stratification in early prediabetes. Methods: We trained a convolutional neural network (CNN, Figure 1A) to classify spectrograms generated using short-term Fourier transforms of beat-to-beat AP signals collected from control vs. prediabetic rats of both sexes fed a high-calorie diet for 12 or 24 weeks, representing young and old rats. Model performance was assessed on different sampling rates for binary and multi-level classifications of sex, age, and diet. Saliency maps were generated by back-propagation of model scores to extract frequency values crucial for classification. Conventional frequency domain analysis was performed on the original AP time-series and the power spectral densities (PSD) were determined for the frequencies identified for each classification task. PSD were compared among sex, age, and diet groups using 3-way ANOVA and principal component analysis. Results: The model successfully classified AP spectrograms according to age, sex, and diet revealing sex- and age-specific CV deteriorative processes in early prediabetes (Figure 1B-E). Saliency maps identified frequencies of 2, 13, 19, and 43 Hz as crucial for DietxAgexSex classification. Comparison of PSDs for these frequencies was not sufficient for CV risk separation (Figure 1F, G). Conclusions: Employing a CNN overcomes the shortcomings associated with CV risk stratification using conventional methods.