Feature space and hybrid neural network model for estimating human blood pressure based on video data

The subject of this study is neural network algorithms for intelligent video data analysis in the task of contactless assessment of the level of human physiological parameters. The task of assessing blood pressure levels based on video plethysmographic recordings of a face is being investigated. Since blood pressure is characterized by systolic and diastolic values, the task is formulated as regression reconstruction for two target variables. The work uses a representative dataset containing video recordings of 2,251 subjects. For the analysis, a method for selecting the area of interest, extracting the green channel, and constructing time-frequency representations using a continuous wavelet transform is proposed and implemented. A hybrid architecture has been developed that provides efficient processing of non-stationary blood flow signals, reducing the dimension of features and increasing the accuracy of forecasting. Special attention is paid to the model's resistance to external interference. For the task of restoring blood pressure regression, a hybrid neural network model based on one-dimensional convolutional and recurrent LSTM networks was used. The proposed method significantly reduces the data dimension by extracting amplitude-energy, frequency, dynamic and phase parameters of blood flow fluctuations in four frequency ranges corresponding to various regulatory mechanisms of the body. The values of the average absolute error of 15.2 mmHg for systolic and 9.3 mmHg for diastolic pressure were achieved. The obtained indicators are comparable to their foreign counterparts, although somewhat worse, despite using a much more representative sample. However, this model does not meet the clinical accuracy requirements (5 mmHg) and requires further improvement. Particular attention should be paid to expanding the dataset with cases with extreme pressure values, as well as checking the algorithm's stability to variable lighting and random camera movements.