Deep Learning for Predicting Exacerbation of COVID-19 Patients Based on Time Series of Three Vital Signs

data set

All methods were performed in accordance with relevant guidelines and regulations. This retrospective study will use the “NYU Langone anonymized his COVID-19 database.”⁵³were collected from patients at the New York University Langone Facility between January 2020 and September 2022. The COVID-19 Anonymized Clinical Database is an anonymized dataset of clinical activity at NYU Langone Health acquired from Epic after January 1, 2020. These data have been stripped of unique identifiers and date shifted by an arbitrary number of days for each patient. In other words, IRB was exempt because these data are not subject to his HIPAA restrictions on research use.Learn more about⁵³.

Data preprocessing

Define inclusion and exclusion criteria. First, in case of multiple patient encounters, use the patient’s most recent encounter. Second, include patients who tested positive for COVID-19 at the facility or who had already been diagnosed with COVID-19 at the time of admission. Next, include inpatients with vital sign measurements. These patients’ vital signs such as her SpO2, temperature and heart rate are measured and recorded regularly approximately every 4-5 hours. Each patient’s age, gender, vaccination status and timing, and comorbidities such as obesity, diabetes, and hypertension are also recorded.

Same as previous work (reference)⁵² and references therein), we define exacerbation as the occurrence of a composite outcome of any of three events: death, ICU admission, or intubation. If a patient has had several adverse events, consider only the occurrence of the earliest exacerbating event. He notes that if a patient has multiple exacerbations of the same type recorded more than a week apart (e.g., ICU admission), only the most recent one is considered as the reference time of exacerbation for that patient. please. For those patients who experienced exacerbation, we extracted her 48-hour vital sign data before exacerbation.

Use this data to define a “positive” window for each forecast period. \(t=0\) represents the end of the window, \(t=-\,24\) represents the start of the window. For example, within a 24-hour forecast window, degradation occurs at: \(t=24\).

For patients discharged without exacerbation of symptoms, the 48-hour window prior to the last vital sign recording is used and similarly used to formulate a ‘negative’ window. Exclude all samples whose vital signs monitoring was less than 48 hours prior to either the time of deterioration or the time of discharge.

To preprocess the time series data, first normalize the data using Z-score normalization based on the mean and standard deviation of each vital sign. Since vital signs are measured at irregular intervals, we resample each time series to obtain regularly sampled inputs for the LSTM and TCN networks. This resampling is done by first interpolating the raw non-uniformly sampled data with cubic spline interpolation and then sampling the interpolated signal every 15 minutes. In the figure, 1shows a diagram summarizing the preprocessing of raw time series data (called SEQ data).

For non-SEQ data, encode patient gender, vaccination status, hypertension status, and obesity status in binary (0 or 1). For diabetic status, a one-hot encoding is used to represent whether the patient is non-diabetic ([1, 0, 0]), uncomplicated diabetic patients ([0, 1, 0]), or diabetes with complications ([0, 0, 1]). We grouped the ages into 18 different subgroups and replaced each age with the corresponding age subgroup (values ​​from 1 to 18). For vaccination timing, count the number of months that have passed between the date of the second COVID-19 vaccination and the predicted timing (\(t=0\)).

Data preprocessing pipeline. Encode non-SEQ data and preprocess SEQ data. (i) normalize by z-score normalization, (ii) model the time series using cubic spline interpolation, and (iii) resample every 15 min.

Architecture of LSTM-based deep neural networks. The SEQ vital signs network processes SEQ data through an LSTM-based network, while the non-SEQ module contains non-SEQ data.

Architecture of a TCN-based deep neural network. The SEQ vital signs network processes SEQ data through a series of residual TCN blocks followed by an FC layer, while the non-SEQ modules process non-SEQ data.

DL-based deterioration prediction model

Since these two model architectures are commonly used for modeling time series data, we chose to develop an LSTM-based model and a TCN-based model to compare their performance. Our proposed LSTM-based deep neural network architecture consists of TD-LSTM layers and fully connected (FC) layers. The diagram shows the overall architecture of this network. 2and call it like this: Repetitive Sequential Vital Signs Network (RSVS-Net), which consists of two modules. SEQ data is processed by a module consisting of an LSTM network and a single FC layer, and non-SEQ data is processed by a second module consisting of an independent FC layer. Final predictions are based on both modalities.

The TCN-based network proposed in this work contains residual TCN blocks and FC layers.We call this model as Convolutional sequential vital signs network (CSVS Net). The module processing the SEQ data in this network consists of a series of five remaining TCN blocks with exponentially growing kernel extensions and one FC layer.non-SEQ data processing modules in CSVS Net is the same as that of RSVS-Net.the architecture of CSVS Net shown in the illustration. 3. For brevity, the remainder of this document refers to the remaining TCN blocks as R blocks.

Details of the RSVS-Net architecture

LSTM networks are well-known for their ability to learn from SEQ data and are widely used in research where time series data are essential for training systems and predicting future events.^55,56. The temporal module consists of a temporally extended LSTM that takes three vital signs as input at each time step and has three layers each containing 32 hidden units. The final hidden state of the LSTM network consists of 32 dimensions and is processed by an FC layer of output dimension 16. Non-SEQ data with dimensionality 9 is processed by a single FC layer. Compute the 16-dimensional output.

Finally, the latent representations of the two modalities are concatenated as a 32-dimensional vector and processed by an FC fusion layer of output dimension 8. Then follows a single FC layer with sigmoidal activations and output dimension 8. One represents the prediction that the sample precedes degradation or is not within the specified time period. Hyperbolic tangent activation is used in all FC layers except the last layer which uses sigmoidal activation.of RSVS-Net The model has 22,209 trainable parameters.

Details of the CSVS-Net architecture

TCN models have recently received more and more attention in time series data analysis^{49,50,57,58,59}. In TCN, convolution is causal.That is, the convolutional output in time is t rely only on the signal of the hour t before⁴⁹.in the CSVS Net, the SEQ data module consists of five serialized R blocks. Each block has two extended causal 1D convolutional (DC-Conv) layers. Additionally, each R block has a residual connection, which makes the network more efficient at learning the extracted features and the modifications applied to the input signal by the DC-Conv layer. At the end of each R-block structure, the output of the last DC-Conv layer and the input of the R-block are summed. Therefore, the dimensions of these signals must match. To verify that this criterion is met for the entire model, in the first residual R block, \(1 \times 1\) A convolutional layer with a linear activation function is implemented in the residual connection to match the dimensions of the input data to the output of the DC-Conv layer.

of CSVS Net, every DC-Conv layer in the R block has 28 filters and uses a hyperbolic tangent activation function. Therefore, the output dimension of all R blocks is (\(N \times 28\)), and N. is the number of samples for each feature in the vital signs time series signal. For a 24-hour observation period for each feature, there are 4 samples per hour, so the signal length is: \(N = 1 + (4 \times 24) = 97\). To analyze the full length of this signal, CSVS Net It should be able to have a receptive field (\(R_{field}\)) is large enough to analyze the length of the entire signal.of \(R_{field}\) Parameters can be calculated using formulas. (1).

In the expression, (1), K. represents the size of the filter, B. is the number of R blocks in the architecture and \(d_i\) is the kernel extension used by each R block.by choosing \(K=3\), \(B=5\)and \(d_i = 2^{(i-1)}\)Note that choosing other numbers of R blocks may result in an insufficient or unnecessarily large receptive field for the model. Finally, a “slicer” layer is implemented at the end of the R block chain, passing only the final array of outputs of the final R block. Therefore, the dimension of the output of the slicer layer is (\(1\×28\)).then this \(1\×28\) Vectors are processed by a 16-dimensional FC layer. CSVS Net is the same as that of RSVS-Net.of CSVS Net The model has 22,709 trainable parameters.

3-step training strategy

To optimize the performance of the proposed network, our training strategy consists of three phases described below.

• Phase 1: SEQ module training In the first phase, we only train the SEQ module and freeze all other layers. The output of the FC layer of the SEQ module is connected to another FC layer that computes predictions. After this training phase, the weights are used to pre-initialize the SEQ module in the next phase, and his second his FC layer, which was used to compute predictions during pre-training in the first phase, is removed. increase.

• Phase 2: Fusion layer training The second phase initializes the relevant modules using the weights obtained in the first phase, computes a representation of the SEQ data, and then freezes the SEQ module. Next, train the FC, FC fusion, and FC output prediction layers using the non-SEQ data.

• Phase 3: End-to-end fine-tuning of RSVS-Net and CSVS-NetThe final phase initializes the network-wide parameters using the weights obtained in the first two phases. The network is trained end-to-end to improve overall network performance.

Model training and evaluation

To train and evaluate a model, we split the entire dataset into training, validation, and test sets. \(66.6\%\), \(16.6\%\)and \(16.6\%\), Each. Each dataset has the same distribution of positive and negative samples. Final performance results are obtained by evaluating the trained model on the test dataset.

We train the model using a training set of 200 epochs using a three-stage training strategy. For the first and second phases, we choose a learning rate of 0.0001 based on our initial experiments, and a learning rate of 0.00001 for the final finetuning stage. All phases use the ADAM optimizer.⁶⁰ and \(\beta_1 = 0.9\), \(\beta_2 = 0.999\), \(\epsilon = 10^{-8}\). After 200 epochs of training, the best model is selected based on the epoch with the lowest validation loss across all epochs. Evaluate the model on the test set. Use binary focal cross-entropy loss.⁵¹ To manage class imbalance in the dataset. Measure model performance using three commonly used metrics for binary classification: accuracy (0.5 threshold for converting predicted probabilities to binary), AUROC, and area under the accuracy recall curve (AUPRC). Evaluate.