Age- and Sex- Specific Aberrations of Arterial Pressure Control in Prediabetes: What does Machine Learning reveal about Frequency Domain Analysis of Blood Pressure?

Abstract ID 95800 Poster Board 068 The prediabetic state is associated with insidious cardiovascular (CV) disease which often goes uncharacterized due to its subclinical nature. Early CV risk appraisal in prediabetes remains an unmet goal because of the absence of reliable risk assessment scores. As such, ensuing CV deteriorative processes remain unidentified and unmanaged until clinical signs of decompensated CV control, like elevated blood pressure (BP), manifest. Changes in BP dynamics, as quantified by BP variability (BPV) analysis, provide insight about the status of systems of CV control. Particularly, frequency domain analysis of BPV, carried out by fast Fourier transformation of BP signals, computes the power of cardio-autonomic firing. However, the utility of frequency domain analysis remains limited by its ability to provide comprehensive risk stratification in early disease stages. In an attempt to unveil novel characteristics of arterial pressure (AP) fluctuations sufficient for cardiovascular risk appraisal, we employed a convolutional neural network (CNN) to classify beat-to-beat AP signals according to age, sex, and prediabetic state and extract characteristic features crucial for risk separation. To do so, we used a pre-established prediabetic rat model based on 12- or 24 weeks (young vs. old) of mild hypercaloric feeding in 4-week-old male and female Sprague Dawley rats. Invasive 5-minute AP recordings were acquired using carotid catheterization, at a frequency of 1000 Hz. AP time series were then downsampled to 50 Hz and sliced to 1000 data points per bin. Short-term Fourier transformation was performed to obtain spectrograms to serve as fingerprints to identify and distinguish between the different rats. The data was divided into 80% training, 10% validation, and 10% testing sets. A CNN consisting of 5 convolutional layers separated by batch normalization, ReLU activation, and max-pooling (Figure 1) was trained for 10 epochs. The output of the last layer was dropped by 40% and delivered to a fully connected linear layer. The model successfully classified AP signals according to age, sex, and prediabetic state with over 90% test accuracies. In order to extract salient features used for spectrogram classification, backward propagation of model scores was performed. The model picked up novel frequencies in ranges much higher than those typically considered in traditional frequency domain analysis. Interestingly, the powers at these frequencies are much lower than those corresponding to parasympathetic and/or sympathetic firing. As such, we performed fast Fourier transformation to obtain the power spectral densities at the identified frequencies. Intriguingly, none of the salient frequencies could establish comprehensive risk appraisal across age, sex, and metabolic groups upon comparison or clustering with multivariate linear regression or principal component analysis, respectively, possibly indicating the presence of non-linear interactions. For further feature extraction, Morlet wavelet transform was performed to capture time-dependent interactions. The same CNN was trained with the resulting scalograms. Similar test accuracies and areas under the receiver operating and the precision recall curves were obtained. Validation using human data is underway. Utility of machine learning models carries the promise of revealing hidden patterns of BP control necessary for early diagnosis during a window of therapeutic responsiveness.