Predicting Intraoperative Hypoxemia with Hybrid Inference Sequence Autoencoder Networks

During anesthesia procedures (i.e., surgeries), a set of patient’s physiological signals such as vital signs (e.g., heart rate, SpO₂, $\text{CO}_{2}$) and ventilator parameters (e.g., respiration rate, $\text{FiO}_{2}$) are recorded at one-minute intervals. These intraoperative time series track the patient’s physical status during surgery and may contain information associated with potential complications and adverse surgical events (meyer2018machine, ).

The data used in this work were collected from 79,142 pediatric surgical encounters spanning from 2014 to 2018 with approximately 118,000 hours of surgeries in total (89 min per case in average) at the St. Louis Children’s Hospital, a free-standing tertiary care pediatric hospital, and St. Louis Children’s Specialty Care Center, an outpatient pediatric surgical center. The institutional review board of Washington University approved this study with a waiver of consent (IRB #201906030).

We use a stringent clinical definition of hypoxemia to assess a patient encounter during surgery. We follow the guideline recommended by World Health Organization (WHO2011pulse, ) and use the emergency level of SpO₂ ($\leq 90\%$) as the threshold for hypoxemia. We define two types of hypoxemia events based on two different severeness levels and assign labels for the two prediction problems following the criteria as shown in Figure 3.

• •

  General hypoxemia: Any low SpO2 ($\leq 90\%$) instance (similar to prior work (lundberg2018explainable, ), light purple region).

• •

  Persistent hypoxemia: Low SpO₂ ($\leq 90\%$) instance that consecutively lasts for 5 minutes and more (dark purple region).

In clinical practice, temporary drops in SpO₂ (i.e., general hypoxemia) are common (24.0% patient encounters in our dataset) and less concerning which is usually rapidly correctable with simple maneuvers with no short-or long-term sequelae. Persistent hypoxemia (1.9% patient encounters) is more clinically relevant and can rapidly become life-threatening and also much more difficult to predict since we need to anticipate deeper into the future.

Figure 3 shows how labels are assigned for model development. For each prediction problems, given a 5-minute prediction horizon, we assign positive labels to the time steps within this 5 min predictive window right before the start of a hypoxemic event (persistent or general hypoxemia, respectively). Other time steps prior to this predictive window are assigned with negative labels. We leave the samples during the time window when the patient is already in a hypoxemic event unlabeled, as it has little clinical benefit to predict hypoxemia when hypoxemia already occurs.

Our dataset consists of $N$ independent surgeries, denoted as $\mathcal{D}=\{\mathcal{V}_{i}\}_{i=1}^{N}$, where $i$ is the index of surgeries, and $\mathcal{V}_{i}$ is a set of time series inputs. We assume the time span is divided into equal-length time intervals. The multi-channel time series $\mathcal{V}_{i}=[\mathbf{v}_{i}^{1},\mathbf{v}_{i}^{2},...,\mathbf{v}_{i}^{T_{i}}]\in\mathbb{R}^{V\times T_{i}}$, where $V$ is the number of time series channels, $T_{i}$ is the length of surgery $i$, and $\mathbf{v}_{i}^{t}\in\mathbb{R}^{V}$ is the vector of recorded data values at timestep $t$. For timestep $t$, we have a binary label $y_{i,t}\in\{0,1\}$, where $1$ indicates that a hypoxemic event will occur anytime within the next fixed-length time window $[t+1,t+W_{h}]$, otherwise $y_{i,t}=0$. $W_{h}$ denotes the prediction horizon. We aim to solve the following:

Recently, several attempts have been made targeting hypoxemia using data-driven approaches. For instance, (elmoaget2016multi, ) used a linear auto-regressive (AR) model on SpO₂ forecast. (erion2017anesthesiologist, ) sought to use deep learning models such as LSTM to directly classify past SpO₂ data. (nguyen2018reducing, ) used AdaBoost to identify false SpO₂ alarms, without directly targeting prediction. Recently, a more comprehensive approach, Prescience (lundberg2018explainable, ), employed GBM to predict general hypoxemia based on both preoperative static variables and intraoperative time series. All these approaches aimed at either forecasting SpO₂ (regression) or predicting only general hypoxemia.

Autoencoder (AE) (davis2006relationship, ), as shown in Figure 2(a), is widely used for representation learning. AE is trained to simultaneously optimize an encoder that maps input into latent representation, and a decoder that recovers the input by minimizing the reconstruction error. However, AE cannot directly handle sequential data. Recently, (dai2015semi, ) proposed a seq2seq AE by instantiating the encoder and decoder with LSTM, referred to as LSTM-AE, shown in Figure 2(b). (srivastava2015unsupervised, ) further extended it to the composite LSTM autoencoder (LSTM-CAE) that additionally trains another decoder fitting future data as regularization. The LSTM-AE based methods have shown promising performance in learning representation for sequential data. For instance, (laptev2017time, ; zhu2017deep, ) proposed to use pretrained LSTM-AE to extract deep features from time series for Uber trips and rare event forecasting. Recent clinical applications can be found in (suresh2017use, ; ballinger2018deepheart, ; baytas2017patient, ) that used LSTM-AE to extract patient-level representation for phenotyping and cardiovascular risk prediction. These works are related to our approach, as they all use representation learning of pretrained LSTM-AE to facilitate a classification task. However, with the goal to continuously provide real-time prediction, instead of extracting patient-level representation, we use a sequence AE to aggregate local data sequence and learning representation of a data window sliding on each surgical trajectory. Unsupervised pretraining tends to learn general task-agnostic underlying structure, and thus the greedy layer-wise optimization in separate steps can lead to a suboptimal classifier (zhou2014joint, ). Instead, our approach relies on building an end-to-end model that jointly optimizes the classification and latent representation learning while balancing them more delicately.

When applying a sequence AE on streaming time series using a sliding data window, the sequence encoder tends to learn mainly the local temporal patterns within the window. Recently, memory network (sukhbaatar2015end, ; gong2019memorizing, ) has shown promising results in data representation. Generally, a memory network updates an external memory consisting of a set of basis features for look-up to preserve general patterns optimized on the whole dataset (santoro2016meta, ). To help capture the global dynamics of physiological time series, we design novel a memory augmented sequence AE with the dual-level embedding of the input sequences.

Reconstructor helps learn latent representation that improves model robustness against class imbalance. However, Reconstructor may not provide enough future-indicative patterns that the prediction task relies on. Motivated to learn more future-related contextual representation, we propose the hybrid inference Network (hiNet) that incorporates both generative and discriminative components, and simultaneously makes inference to both sequence of future low SpO₂ instances and hypoxemia event outcome. The overall architecture of hiNet is shown in Figure 3.

Examples with recorded SpO₂ of less than 60% were considered aberrant and excluded. We further excluded 5,606 cardiac surgery cases (with ICD-9 codes 745-747 and ICD-10 codes Q20-Q26) and 1,449 surgery cases with initial persistent low SpO2 (most likely cardiac related surgeries), in which SpO₂ levels were likely affected by the surgical procedure. After extensive data cleaning, there are 72,018 surgeries and 18 channels of time series variables minutely sampled during surgical procedures. We aim to build a hypoxemia prediction system with 1-minute resolution. However, not all variables were originally observed and recorded at every minute during surgery. We use carry-forward imputation for the missing values in a gap between two observations less than 20 min, and fill those that are never observed or haven’t been observed for the past 20 min with zeros. In addition, we concatenate a binary mask matrix with the time series input to indicate variable missingness. All variables are standardized to zero mean and unit variance.

To further explore the effect of incorporating preoperative features (e.g., age, sex, weight) as part of the empirical analysis (see section 5.3 and 6.1), we also collected a set of preoperative variables for modeling. Table 1 lists the 18 intraoperative variables and 9 preoperative variables used in this study.

We randomly select 70% of all the surgery cases for the model training, setting aside 10% as a validation set for hyperparameter tuning and the other 20% for model testing. We make sure that all data points from the same surgery case stay in the same subset of data. We follow (lundberg2018explainable, ) and set the prediction horizon as $W_{h}=5$ min, given that it is short enough for a model to capture relevant and predictive information of potential future hypoxemia but long enough for clinicians to take actions.

We compare hiNet to the following classical models, deep sequential models, and unsupervised pretraining based methods:

• •

  LR: Logistic Regression. Since LR cannot directly process time series, for fair comparison, we follow (li2020deepalerts, ; fritz2019deep, ) to extract a series of summary statistics (e.g., min, max, trend, energy, kurtosis) that capture temporal patterns of history time series within the window of the same length as hiNet.

• •

  GBM: Gradient Boosting Machines, employed by the state-of-the-art hypoxemia prediction system (lundberg2018explainable, ), which is implemented using XGBoost (chen2016xgboost, ). We use the same statistical features as in LR.

• •

  LSTM: Using stacked bi-LSTM for feature gathering and a FC block as the classifier, with layers configured the same way as the Event Predictor in hiNet.

• •

  TCN (bai2018empirical, ): Temporal Convolutional Network with causal convolutions and exponentially increased dilation. It is configured the same way as the TCN module in hiNet.

• •

  FCN (wang2017time, ): Full Convolutional Networks, a deep CNN architecture with Batch Normalization, shown to have outperformed multiple strong baselines on 44 benchmarks for time series classification.

• •

  LSTM-AE (zhu2017deep, ): A deep LSTM classifier with the weights pretrained on an LSTM-AE.

• •

  LSTM-CAE (srivastava2015unsupervised, ): A deep LSTM classifier with the weights pretrained on an LSTM-CAE that jointly reconstructs input and forecasts future input.

• •

  TCN-AE and FCN-AE: Replacing the RNN encoder in LSTM-AE (zhu2017deep, ) with TCN and FCN for comparison.

• •

  hiNet-l, hiNet-t and hiNet-f: The hiNet variants with the sequence encoder implemented by LSTM, TCN, and FCN.

• •

  GBM w/ PreOp and hiNet w/ PreOp: For GBM w/ PreOp, the preoperative variables are added to the input of GBM as in (lundberg2018explainable, ). For hiNet w/ PreOp, the preoperative features are directly concatenated with the data representation $\mathbf{p}_{i,t}$ in Eq. (7).

For TCN and hiNet-t, we use 3 TCN blocks with number of filters set as 64 and dilation of each block set as 2, 4, and 8. For FCN and hiNet-f, we use 3 blocks for the FCN and use the filter sizes $\{64,128,64\}$ for each block. Each FC block in hiNet has only one hidden layer with rectified linear unit (ReLU) as the activation function. The number of neurons for each hidden layer in both LSTM and FC block, and the number of basis $M$ in the memory are all set as 128. We select the regularizer coefficient $\lambda$ from $\{10^{-4},10^{-3},10^{-2},10^{-1},10^{0},10^{1}\}$. The best $\lambda$ is 0.1 for persistent hypoxemia and 0.01 for general hypoxemia.

We use Adam as the optimizer with default learning rate and train the model with mini batches. For each batch, we feed 32 independent surgeries containing on average about 2,880 extracted examples into the model. We use the same settings for all deep baselines and all variants of hiNet. All of them are trained for 50 epochs with early stopping and drop-out applied to prevent overfitting. The model with the lowest epoch-end classification loss at each run is saved and evaluated with test data. The proposed model is implemented using TensorFlow 2.4.0 and Keras 2.4.3 with Python 3.8, and trained using NVIDIA GeForce RTX 3080 Ti GPUs and Intel Core i9-10850K 3.60GHz CPUs.

As shown in Table 3, we analyze the contribution of each component by removing it from hiNet. Note that for the variant w/o Memory, the memory layers are replaced with two stacked independent linear layers to maintain similar model complexity for fair comparison. We can see that for both the two prediction tasks (persistent & general hypoxemia), each component in hiNet plays essential role in improving the prediction performance.

As shown in Figure 7, we extracted the latent representations learned by LSTM-AE and various layers of hiNet for general hypoxemia prediction, and visualize these vectors in a 2D space using t-SNE (van2014accelerating, ). Considering the extreme class imbalance, we randomly select 50 surgeries where persistent hypoxemia occurred and 50 hypoxemia-free surgeries from the test set for visualization purpose. As shown in Figure 7, the representation of the same class in the latent space tends to group together in hiNet, which enlarges the partitioning margin and makes them easier to classify. More explicit grouping patterns can be observed at layers closer to the Event Predictor output with stronger supervisory signal. In contrast, the representation learned by unsupervised LSTM-AE is structured but shows much less salient grouping patterns. We observe that, hiNet is able to learn powerful task-specific representation via joint training, where supervisory signal is propagated to fine-tune latent representation towards task-specific effectiveness.

When producing labels for clinical outcomes such as hypoxemia, anaesthesiologist interventions can indirectly affect the prediction outcome (lundberg2018explainable, ). As these interventions may affect certain vital parameters including SpO₂, models that use these parameters can learn when a doctor is likely to intervene and hence lower the risk of a potential high-risk patient. The ideal solution to this issue is to remove all samples where clinicians have intervened for model training. But this is difficult in practice, since fully identifying when clinicians are taking hypoxemia-preventing interventions is not possible. Hence, our model as all other previous learning based approaches (lundberg2018explainable, ; erion2017anesthesiologist, ) to this problem, must be based on the natural assumption that the model predicts hypoxaemia when clinicians are following standard procedures, including (possibly) taking interventions to prevent potential hypoxemia anticipated based on clinician’s professional knowledge.