Electronic health records contain valuable temporal information for sepsis prediction. However, irregular time intervals between neighboring events are typically neglected. Besides, transparency and interpretability of deeplearning models with increasing complexity and superior performance has become a barrier to the models' clinical adoption. To this end, we propose an interpretable deep-learning model that better captures time information and achieves promising performance on sepsis prediction in the emergency department.
Sepsis, a life-threatening illness caused by the body's response to an infection, is the leading cause of death worldwide and has become a global epidemiological burden. Sepsis occurs at all ages and increases mortality rates. In the United States, for example, over 1.7 million adults develop sepsis and nearly 270,000 patients die as a result of sepsis each year. 1 Besides, sepsis is the costliest among all disease states and accounted for $24 billion of United States hospital costs in 2013. 2 Without timely and adequate treatment, sepsis can progress to severe sepsis and septic shock, which lead to higher mortality rates. 3 Several studies suggest that early prediction of sepsis enables early treatment and is able to significantly improve patient outcomes. 4,5 However, common signs and symptoms of sepsis, such as fever, chills, rapid respiration, and high heart rate, are the same as in other conditions, making sepsis difficult to diagnose in its early stages. Besides, it is clinically meaningless to predict sepsis minutes before onset even with high prediction accuracy. A good predictive model should be able to trigger THE BIGGER PICTURE Sepsis is the leading cause of death worldwide and has become a global epidemiological burden. Early prediction of sepsis enables early treatment and increases the likelihood of survival for septic patients. The broad adoption of electronic health records (EHRs) provides an opportunity for sepsis prediction. However, most existing prediction approaches do not consider irregular time intervals between neighboring clinical events in EHRs. Besides, many deep-learning models suffer from black-box problems and are not trusted in clinical settings. We propose a deep-learning model with time encodings, offering both high accuracy and high transparency as well as clinical interpretability. We have already made our code and its detailed documentations publicly available, enabling colleagues to apply it to their applications and eventually make clinical impacts.
Proof-of-Concept: Data science output has been formulated, implemented, and tested for one domain/problem alerts as early as possible and present increasingly stronger signals as it approaches the actual event.
Electronic health records (EHRs) are longitudinal electronic records of patients' health information. The rapid growth in volume and diversity of EHRs during the last decades makes it possible to apply machine-learning and data-mining methods to the early prediction of sepsis. Screening tools have been used clinically to recognize sepsis, including quick Sequential (Sepsis-Related) Organ Failure Assessment (qSOFA), 4 Modified Early Warning Score (MEWS), 6 National Early Warning Score (NEWS), 7 and Systemic Inflammatory Response Syndrome (SIRS). 8 However, those tools were designed to screen existing symptoms as opposed to explicitly predicting sepsis prior to its onset, and their efficacy in sepsis diagnosis is limited. For example, prior studies show that qSOFA had low sensitivities in identifying sepsis in both prehospital and emergency department (ED) settings. 9,10 With recent advances and success, machine-learning methods have shown great potential in unlocking insights from EHRs. Various methods have been developed for accurate sepsis prediction. 11,12 Faisal et al. 13 developed a logistic regression model (CARS) to predict the risk of sepsis using a patient's firstly recorded vital signs and blood test results, which are usually available within a few hours of emergency admission. Horng et al. 14 constructed a machine-learning model using a linear support vector machine and demonstrated the incremental benefit of using free text data in addition to vital signs and demographic data for sepsis clinical decision support at the ED. Mollura et al. 15 trained a bagged tree classifier using the recorded electrocardiogram and arterial blood pressure waveforms, showing that the waveform monitoring information may help in detecting sepsis within the first hour of stay in the intensive care unit (ICU). Kamaleswaran et al. 16 showed that artificial intelligence can be used to predict the onset of severe sepsis as early as 8 h ahead using physiomarkers in critically ill children. Lyra et al. 17 used an optimized random forest to predict sepsis for imbalanced clinical data from ICUs in the PhysioNet Computing in Cardiology Challenge 2019. 12 Mao et al. 18 validated a machine-learning algorithm with gradient-boosting trees, InSight, which used only six vital signs for the prediction of sepsis, severe sepsis, and septic shock and showed that InSight outperformed existing sepsis-scoring systems. Using 65 features from a combination of EHRs and high-frequency physiological data, Nemati et al. 19 developed and validated an interpretable machinelearning model based on a modified Weibull-Cox proportional hazards algorithm for making an accurate and interpretable prediction of sepsis. Recently, deep-learning methods have achieved improving performances over traditional models and have shown unprecedented potential in the healthcare domain. 20 Deep-learning models automatically learn the data representation with improved performance and do not require conventional feature-extraction steps. Recurrent neural networks (RNNs) are commonly used network architectures in modeling multivariate series prediction. [21][22][23] Kam and Kim 21 proposed a sepsis-detection model with long short-term memory (LSTM), which showed better performance than InSight and superior capability for sequential patterns. However, deep-learning models usually suffer from black-box problems and are not trusted in clinical settings. RETAIN 24 and Dipole 25 proposed to introduce attention mechanisms and interpret the models' output risks based on the learned attention weights, which is helpful for models' application to real-world clinical settings.
Most existing approaches 11,12,17,21 focus on the sepsis prediction for ICU settings and may suffer from performance decrease for predicting sepsis onset for patients in EDs with low resolution of medical observations, while many patients have been diagnosed with sepsis at ICU admission. 26 Moreover, most of the aforementioned existing methods do not or only consider the relative order of events and ignore the irregular time intervals between neighboring events while modeling time-series EHR data. Besides, the increasing complexity of deep-learning models has brought superior model performances at the price of lack of transparency and interpretability, which has become a barrier to the models' clinical adoption. To this end, we address these problems with our proposed interpretable LSTM-based deep-learning model that can achieve state-of-the-art sepsis-onset prediction in the ED.
Our proposed deep-learning model handles irregular time intervals with time encodings, and leverages attention mechanism and global max pooling techniques to help interpret the model's behavior. Our team, BuckeyeAI, participated in the 2019 DII challenge with the proposed deep-learning method and ranked second out of 30 teams on the early prediction of sepsis onset in the ED, with an average area under the receiver-operating characteristic curve (AUC) score of 0.892. The goal of the 2019 DII challenge is the early prediction of sepsis using a patient's demographic and physiological data in the ED. Different from the PhysioNet Computing in Cardiology Challenge 2019 on sepsis prediction in the ICU, 12 the 2019 DII challenge focused on sepsis prediction in the ED where the environment is more chaotic. 27 In this paper, we present our methods, results, and analyses. To summarize, the contributions are as follows.
d We present benchmark results of sepsis-onset prediction in the ED. We show that our model outperforms four early-warning scores and three baseline machine-learning models. d We propose an LSTM-based model for sepsis-onset prediction, which handles irregular time intervals with time encodings. d We leverage the attention mechanism and global max pooling techniques to help interpret our model.
Definition of Sepsis-2, the presence of proven or suspected infection together with two or more SIRS criteria, 28 is used to define ground truth in the ED. The inclusion and exclusion diagram of the 2019 DII challenge data preparation pipeline is shown in Figure 1. A summary of patient characteristics is provided in Table We implemented and evaluated four early-warning scores, three traditional machine-learning methods, and four deeplearning models as baselines. The four early-warning scores comprised MEWS, 6 NEWS, 7 SIRS, 8 and qSOFA. 4 For traditional machine-learning methods, we considered logistic regression, random forest, and gradient-boosting trees. Because these standard machine-learning methods cannot work directly with multivariate time-series sequences, the element-wise aggregation (i.e., count, mean value, minimum value, maximum value, and standard deviation of events) of temporal features are used as model inputs. For the deep-learning baselines, two classical RNN models (i.e., GRU 29 and LSTM 30 ) and two state-ofthe-art interpretable RNN models (i.e., RETAIN 24 and Dipole 25 ) are selected. The RNN models cannot handle the missing values of EHR data. We mapped the feature variables into vectors via an embedding layer. The concatenation of the embedding vectors and the observed feature values were then input to the RNN models.
Table 2 summarizes the performance of various models for sepsis-onset prediction. From Table 2, our model outperforms baseline models. The main reasons why our model works better are 2-fold: (1) our model can automatically learn better patient representations as the network grows deeper and yield more accurate predictions with sufficient data; (2) our LSTM-based model can better capture temporal information, while logistic regression, random forest, and gradient-boosting trees simply aggregate time-series features and hence suffer from information loss.
We found that machine-learning-based algorithms outperformed early-warning scores on both cases. All three machinelearning methods achieved similar performance on both Case 1 and Case 2. MEWS and NEWS were shown to perform better than SIRS and qSOFA on Case 2. However, the result suggested little discrimination of four scores on Case 1 with low AUC scores. The deep-learning models outperformed the early-warning scores and performed comparably with the machine-learning algorithms. We speculate the reason for this is that the feature engineering (e.g., minimum and maximum feature values) is effective, and both machine-learning and deep-learning methods can capture the abnormal values from EHRs. With the help of attention mechanisms, RETAIN and Dipole can focus on the abnormal values better, and thus outperform GRU and LSTM.
On the private test dataset, our proposed model achieved AUC scores of 0.940 and 0.845 for two use cases, respectively. The official score is ð0:940 + 0:845Þ=2 = 0:892. Compared with attention-based models (i.e., RETAIN and Dipole), the proposed model still achieves better prediction accuracy. Our model considers the whole history of a patient's EHRs with a global pooling operation rather than attention, which is useful for relieving the long-term dependency problem of RNN. Moreover, the time embedding can capture the temporal information more efficiently, which further improves the proposed model's performance.
To measure the effectiveness of different components (i.e., event embeddings, time encodings, and global max pooling), we adopt an ablation study to gain a better understanding of the proposed model by removing one component each time. The results of ablation study on Case 1 sepsis-onset prediction are reported in Table 3. Based on the results from Table 3, the most influential component is event embeddings. By removing event embeddings, the AUC score decreases by 0.11. By handling irregular time intervals using time encoding, the model performance increases from 0.89 to 0.94. Moreover, incorporating global max pooling causes an AUC score increase of 0.03.
Generally, linear models and tree-based models can be easily interpreted because of their intuitive way of predicting output from inputs, but these models are quite simple. Although deeplearning models can usually yield more accurate predictions, they usually operate as black boxes and make it unclear why the models make specific predictions. However, due to the attention mechanism and global max pooling operation, our deep-learning model is interpretable as shown in Figure 6. At patient level, we are able to calculate the contribution rate of each medical event for sepsis risk according to Equation 5. Medical events with higher contribution rates contribute most to the clinical outcome (i.e., sepsis onset in the next 4 h).
While patient-level interpretation reveals medical events that are most influential to sepsis onset for an individual patient, population-level analysis is needed to determine the most influential medical events as well as clinical features over the entire EHR dataset. Therefore, to better understand the model's behavior, we
As we can calculate the contribution rate of each medical event of each patient, we can compute each medical event's importance at the population level. For each medical event, event importance is calculated by averaging its contribution rates for all patients whose EHR data contain this event.
Figure 4 shows the medical event importance (average contribution rate) over time for all patients. This plot shows an overall upward trend, which meets our expectation that the medical events closer to sepsis onset are more important for our model to make predictions.
Apart from medical event importance, we also want to know which clinical features are most important for sepsis-onset prediction. Similar to medical event importance, for each clinical feature we compute its importance over all medical events across the entire population according to Equation 6. The top influential features found by the deep-learning model are shown in Figure 5. The full contribution rate list of clinical features can be found in Table S1. The clinical features with the highest contribution to sepsis prediction are easily attainable clinical values. Thus, our model suggests that the development of sepsis can be predicted easily based on items within the EHR. Interestingly, lab values traditionally associated with sepsis prediction (e.g., white blood cell count and renal function) were not predictive.
From this perspective, we compare the model performance across various subpopulations and report the results on The concatenation of the medical event embedding vectors (v 1 ; v 2 ; /; v n ) and the corresponding time encoding vectors (e 1 ; e 2 ; /; e n ) are inputs to the BiLSTM model, which generates output vectors (h 1 ; h 2 ; /; h n ). All the output vectors are concatenated, then a global max pooling operation is performed to produce the patient representation vector. Finally, a fully connected layer and the sigmoid function are used to predict the probability of sepsis onset in the next 4 h. Case 1 sepsis prediction as an example in Table 4. The results show that our model achieves high prediction performance (AUC R0.929) across all subpopulations. Confidence intervals are calculated at the 95% level. We also test paired p values for model performance between subgroups, the results of which are reported in Table S2. Concerning gender, the model seems to perform better on female patients compared with male patients, with higher AUC scores (p = 0.025). For race subgroups, performance on the African American patients is the most discriminatory, with relatively lower p values compared with other combinations. The model's AUC on Asian patients is lower with large variance, perhaps because the proportion of Asian patients is small. With respect to age subgroups, the model achieves higher performance for patients whose age is lower than 20 years while the result shows large variance due to the low proportion of such patients. Model performances on patient pairs aged 20-30 and 30-40, 50-60, and 60-70 years are quite similar. The reason for this could be that the distributions of features of these pairs are closer.
Our team, BuckeyeAI, participated in the 2019 DII Challenge and ranked #2 out of 30 teams on the early prediction of sepsis onset task. In this paper, we present our solution to sepsis-onset prediction 4 h before it occurs. For sepsis-onset prediction, our proposed deep-learning model achieved an AUC score of 0.892 and outperformed four early-warning scores and three baseline machine-learning models. By incorporating event embeddings, time encodings, and global max pooling, our model yields more accurate predictions. Time encodings help to handle irregular time intervals. The global pooling operation enables the model to associate the contribution of each medical event with the final clinical outcome, paving the way for interpretable clinical risk predictions.
Although we mainly focus on sepsis-onset prediction in this challenge, our model is general and can be applied to other multivariate time-series prediction problems. In addition to the superior performance, our proposed model is interpretable from an individual patient to the whole population.
Lead contact Further information and requests for resources should be directed to and will be fulfilled by the Lead Contact, Ping Zhang, PhD (zhang.10631@osu.edu).
This study did not generate any new materials.
Protected Health Information restrictions apply to the availability of the 2019 DII Challenge dataset. As a result, the dataset is not publicly available. The source code is provided and is available at https://github.com/ yinchangchang/DII-Challenge.
The challenge data are extracted from the Cerner Health Facts database. All challenge entrants signed an enforceable data use agreement as part of the competition registration process. Regarding the use of Cerner Health Facts, all challenge publications authors are covered under IRB protocol HSC-SBMI-13-0549, approved by the UT Health Committee for the Protection of Human Subjects.
The challenge data are extracted from the Cerner Health Facts database. Cerner Health Facts is a database that comprises de-identified EHR data from over 600 participating Cerner client hospitals and clinics in the United 1. Descriptions and statistics of clinical features are available in Table S1.
In this challenge, we aim to predict sepsis 4 h before onset for hospitalized adult patients. There are two use cases, as demonstrated in Figure 2. Case 1 In this case, patients are sampled from septic patients, and the goal is to find out whether a model can tell if a patient is likely to have high sepsis risk a few hours before the onset. For each patient, the patient records is split into two segments at the middle point, segment close to sepsis onset (= 4 h) is labeled as 1, another segment (>4 h before sepsis onset) is labeled 0. We randomly pick either the former or latter segment to build the Case 1 cohort. The introduction of case 1 is to measure the model in terms of time-sensitive prediction to ensure models are indeed clinically useful and relieve warning fatigue as alarm burden. Given patient records either from Tadmission to T middle or from T middle to T onset À 4, our model is required to distinguish these two kinds of records.
In this case, case and control segments are from different patients who have sepsis onset in the next 4 h, as well as those who do not have sepsis. Given patient records from T admission to T onset À 4, we are going to predict whether sepsis occurs in the following 4 h.
The proposed neural network architecture is shown in Figure 3. This model is inspired by DG-RNN. 31 Although we focus on the early prediction of sepsis onset in this challenge, our proposed model is general and can be applied to other multivariate time-series prediction tasks, such as mortality prediction for septic patients. embeddings For each temporal feature, we sort the values from low to high and use the order to replace the original values. We then divide the orders into ten groups (i.e., 0.0-0.1, 0.1-0.2, ., 0.9-1.0) and each event is then embedded into a 512-day vector.
When modeling time-series EHR data, most existing LSTM-based models do not or only consider the relative order of events. However, these methods typically ignore the irregular time intervals between neighboring events. Similar to position encodings in Transformer, 32 we infuse time information using time encodings. Time encodings are sent to LSTM together with event embeddings. We compute each event's relative time to the criterion operation date and the time interval relative to the last event. We then use sine and cosine functions of the different time intervals to represent the time encoding for the t th event:
(Equation 1)
where date o denotes the criterion operation date, date t denotes the t th event's date, p t ˛R2d denotes the time encoding vector, and j is the dimension of EHRs event embeddings. Both the event embeddings and time encodings are then input to LSTM.
To better align patient records at their last recorded medical event, the time of each event is mapped from ½0; T lastevent to ½ À T lastevent ; 0.
RNNs are popular and suitable for sequential EHR data modeling. Given medical event embedding and time encoding vectors, we build our model based on LSTM 30 for its ability to recall long-term information. The LSTM model can be described as follows: Global max pooling RNN-based models are sometimes inefficient due to their long-term dependency. When the input sequence is too long, it is easy for the models to forget the earlier data. Therefore, we adopt a global pooling operation to shorten the distance between the earlier inputs and the final outputs. As is shown in Figure 3, all the outputs of the LSTM are concatenated, then a global pooling operation is followed. The output og is fed through the fully connected layer to produce the clinical risk of patient i, which is defined as
where W s ˛Rk and b s ˛R are the learnable parameters and y i denotes the predicted probability for sepsis onset. Because of the shortened distance between the inputs and the outputs, the pooling operation makes it more efficient to propagate the gradients. Besides, the global pooling operation is useful to compute the contribution rates of the outputs and their corresponding input medical events.
For binary classification, the objective function is defined as the binary crossentropy loss between ground truth y à and predicted probability y: Here we display four medical events (e 1 ; e 2 ; e 3 ; e 4 ) and their corresponding output vectors (h 1 ; h 2 ; h 3 ; h 4 ). After a global max pooling layer and a fully connected layer, the model predicts the risk of sepsis onset in the next 4 h for an individual patient. Each output vector's contribution is then calculated by summing the corresponding dimensions' contribution risks. Finally, the contribution of each medical event is calculated according to Equation 5.
Patterns 2, 100196, February 12, 2021
Patterns 2, 100196, February 12, 2021 3
Patterns 2, 100196, February 12, 2021
Patterns 2, 100196, February 12, 2021 5
Patterns 2, 100196, February 12, 2021
Patterns 2, 100196, February 12, 2021
Patterns 2, 100196, February 12, 2021 9