Machine Learning Techniques for Arterial Pressure Waveform Analysis

Abstract

The Arterial Pressure Waveform (APW) can provide essential information about arterial wall integrity and arterial stiffness. Most of APW analysis frameworks individually process each hemodynamic parameter and do not evaluate inter-dependencies in the overall pulse morphology. The key contribution of this work is the use of machine learning algorithms to deal with vectorized features extracted from APW. With this purpose, we follow a five-step evaluation methodology: (1) a custom-designed, non-invasive, electromechanical device was used in the data collection from 50 subjects; (2) the acquired position and amplitude of onset, Systolic Peak (SP), Point of Inflection (Pi) and Dicrotic Wave (DW) were used for the computation of some morphological attributes; (3) pre-processing work on the datasets was performed in order to reduce the number of input features and increase the model accuracy by selecting the most relevant ones; (4) classification of the dataset was carried out using four different machine learning algorithms: Random Forest, BayesNet (probabilistic), J48 (decision tree) and RIPPER (rule-based induction); and (5) we evaluate the trained models, using the majority-voting system, comparatively to the respective calculated Augmentation Index (AIx). Classification algorithms have been proved to be efficient, in particular Random Forest has shown good accuracy (96.95%) and high area under the curve (AUC) of a Receiver Operating Characteristic (ROC) curve (0.961). Finally, during validation tests, a correlation between high risk labels, retrieved from the multi-parametric approach, and positive AIx values was verified. This approach gives allowance for designing new hemodynamic morphology vectors and techniques for multiple APW analysis, thus improving the arterial pulse understanding, especially when compared to traditional single-parameter analysis, where the failure in one parameter measurement component, such as Pi, can jeopardize the whole evaluation.

Cardiovascular risk analysis by means of pulse morphology and clustering methodologies

“…This is a simple, yet reliable tool, that can be used in scoring trials. As this approach needs some developments, a computational tool to integrate these results with other machine learning techniques such as classification algorithms, is currently being developed [15]. A larger sample will also be studied in future trials for the stratification by medication use, age, diabetes duration, and/or gender.…”

“…Inferences about CVD progressive development can be assessed by the analysis of the mechanical properties of arteries through a variety of indices based on the Pulse Wave Analysis (PWA) [13,14]. The analysis is based on the identification of the key features in the arterial pressure wave profile, such as systolic wave transit time (SWTT), reflected wave transit time (RWTT) and dicrotic notch (evaluated by left ventricular ejection time (LVET)), and can include time or amplitude considerations, as well as variability based parameters [15]. The wave reflections are often addressed, in terms of the augmentation index (AIx), which expresses the ratio of the "augmented pressure" assigned to the reflected wave towards each overall pulse.…”

“…Their feasible implementation in Computer-Aided Diagnosis (CAD) methodologies has given new insights in the development of innovative and effective decision support systems for CVD premature risk assessment [15,[17][18][19]. An interesting approach is the exploration of different classifiers, as it was proposed by Jovic and Bogunovic [20].…”

Forecasting intracranial hypertension using multi-scale waveform metrics

“…After the full feature matrix is built, we standardize all feature columns to have zero mean and unit standard deviation, using statistics from the training data-set. Missing values are replaced by zero, which corresponds to global mean Energy in frequency bands [0,1], [1,2], [2,3], [3,6], [6,9], [9,12], [12,15] Hz Autoregulation indices on time series (1 Hz sample rate) AmpIndex(ICP,ABPm), AmpIndex(ICP,CPP), AmpIndex(CPP,ABPm) [31] PaxIndex(ICP,CPP,ABPm) [31] PrxIndex(ICP,CPP,ABPm) [63,64] RapIndex(ICP,CPP) [65,66] SlowWaveIndex(ICP) [67] TFIndex(ICP,ABPm), TFIndex(ICP,CPP), TFIndex(CPP,ABPm) [68] Autoregulation indices on waveforms (125 Hz sample rate) AmpIndex(wICP,wABP) [31] SlowWaveIndex(wICP) [67] TFIndex(wICP,wABP) [68] IaacIndex(wICP,wABP) [37] Morphological pulse metrics on waveforms wABP pulse descriptor (17 metrics) [61]: A, UpstrokeTime, TimeAtΠ, TimeAtDw, DownstrokeTime, SysDiasTimeDifference, HeightSysPeak, HeightInflPoint, HeightDicroticWave, R1, R2, R3, R4, R5, R6, Aix wICP pulse descriptor (20 metrics) [20]: Mean, Dias, DP1, DP2, DP3, DP12, DP13, DP23, L1, L2, L3, L12, L13, L23, Curv1, Curv2, Curv3, Slope, DecayTimeConst, AverageLatency imputation. For machine learning models that can deal with missing data natively, like decision trees or tree ensembles, missing data imputation/normalization was not performed.…”

Morphology variability of radial pulse wave during exercise

“…Thus accurate interpretation of pulse wave is of great significance. In recent years, work on how to quantify the state of the cardiovascular system from the parameters of pulse wave has been reported [2][3][4][5][6], one of which is to change the working condition of cardiovascular system by taking exercise, and study the variability of the pulse wave. A. Figueroa studied the influence of acute exercise with whole-body vibration on wave reflection and leg arterial stiffness [4].…”