Method for assessing the severity index of insomnia based on electrocardiogram analysis

Abstract

This invention discloses a method for estimating the severity index of insomnia based on electrocardiogram (ECG) analysis. It includes an input signal step and an insomnia severity index estimation step. In the insomnia severity index estimation step, an optimized deep learning technique, trained, validated, and tested, is employed. First, noise removal, extraction of multiple 30-minute ECG signals, missing value compensation, and signal standardization are performed on the entire night's ECG input from the input signal step. Then, a multi-layered computational approach using a deep learning model architecture is used to analyze the preprocessed multiple 30-minute ECGs, further automating the estimation of the insomnia severity index and generating an output result. This effectively and objectively improves the efficiency of insomnia severity index estimation, providing a mechanism for relevant personnel to consider and intervene in the process.

Claims

1. A method for estimating insomnia severity index based on electrocardiogram (ECG) analysis, executed by computer software, to objectively estimate and analyze the ECG of a subject throughout the night. The method includes: an input signal step, inputting an ECG signal, which is the ECG measured throughout the night by the subject; and an insomnia severity index estimation step, which uses an optimized deep learning technology that has been trained, validated, and tested, along with hyperparameter settings and performance evaluations calculated in conjunction with the deep learning technology, to estimate the insomnia severity index of the input ECG; wherein... The hyperparameter setting algorithm consists of a loss function, optimizer, batch size, and number of training epochs. The performance evaluation algorithm consists of root mean squared error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The deep learning model technology includes: establishing a dataset using various sleep physiological signal data from different subjects provided by the Stanford Technology Analytics and Genomics in Sleep (STAGE) database. 10,000 30-minute ECGs (including the corresponding Insomnia Severity Index (ISI)) were extracted from the entire night's ECGs of all subjects in STAGE as data, and divided into a training set, a validation set, and a test set. An analysis and estimation model was established, comprising a signal preprocessing program and an insomnia severity index estimation program. The signal preprocessing program removes baseline drift and high-frequency noise from the input overnight electrocardiogram (ECG), extracts signals from multiple 30-minute ECGs, performs missing value compensation, and normalizes the data to generate multiple preprocessed 30-minute ECGs. The insomnia severity index estimation program consists of an input layer, a feature extraction module, a temporal modeling layer module, and an output module. It utilizes extracted feature information and temporal relationship modeling and learning to perform calculations on the aforementioned preprocessed multiple 30-minute ECGs and generate an insomnia severity index. The analysis and estimation model is trained, validated, and tested using the aforementioned dataset to establish the training, validation, and test sets. Hyperparameter settings and performance evaluations are used in conjunction with the model to learn electrocardiogram (ECG) patterns of varying degrees of insomnia. This allows for the adjustment of hyperparameters and monitoring of model performance on unseen data to prevent overfitting and improve generalization on new data. The goal is to achieve an optimized analysis and estimation model. The model automatically estimates the subject's ISI score by inputting the ECG processed by the aforementioned signal preprocessing procedure and finally generates the estimation result output.

2. The insomnia severity index estimation method based on electrocardiogram analysis as described in request item 1, wherein, The signal preprocessing procedure includes: a removal step, which uses a bandpass filter and a bandstop filter. The bandpass filter uses a passband of 0.5 to 20 Hz to remove low-frequency baseline drift and high-frequency muscle noise from the overnight ECG, and the bandstop filter uses a passband of 60 Hz to eliminate 60 Hz power supply interference; a truncation step, which truncates the filtered overnight ECG into multiple 30-minute ECGs; a missing value processing step, which uses a long short-term memory network missing value prediction model to predict subsequent data based on the input ECG signal after the aforementioned truncation, in order to fill in missing data; and a z-score standardization step, which uses z-score standardization to standardize the input ECG signal so that the amplitude of the standardized ECG signal has a consistent scale.

3. The insomnia severity index estimation method based on electrocardiogram analysis as described in request item 1, wherein, The feature extraction module consists of three one-dimensional convolutional layers, a concatenated layer, a batch normalization layer, a residual network (ResNet) module, and a global average pooling (GAP) layer. The three one-dimensional convolutional layers are arranged in parallel, and the filter sizes used in the three one-dimensional convolutional layers are 3, 5, and 7, respectively. The number of feature maps in each of them is 64, and they are all used with the ReLU (Rectified Linear Unit) activation function for calculation. The residual network module consists of a convolutional layer, a skip connection, a ReLU activation function, and a batch normalization layer.

4. The insomnia severity index estimation method based on electrocardiogram analysis as described in request item 1, wherein, The temporal modeling layer module consists of two connected layers of Long Short-Term Memory (LSTM) networks. The first layer of the LSTM network contains 128 neurons and uses the tanh activation function to learn temporal features. The second layer of the LSTM network contains 64 neurons and uses the tanh activation function to enhance the capture of dynamic signal features.

5. The insomnia severity index estimation method based on electrocardiogram analysis as described in request item 1, wherein, The output module is designed as a dense layer, which has a single neuron and uses a linear activation function for computation.

Images