Prediction of Brain Metastases Development in Patients With Lung Cancer by Explainable Artificial Intelligence From Electronic Health Records

PURPOSE

Early detection of brain metastases (BMs) is critical for prompt treatment and optimal control of the disease. In this study, we seek to predict the risk of developing BM among patients diagnosed with lung cancer on the basis of electronic health record (EHR) data and to understand what factors are important for the model to predict BM development through explainable artificial intelligence approaches accurately.

MATERIALS AND METHODS

We trained a recurrent neural network model, REverse Time AttentIoN (RETAIN), to predict the risk of developing BM using structured EHR data. To interpret the model's decision process, we analyzed the attention weights in the RETAIN model and the SHAP values from a feature attribution method, Kernel SHAP, to identify the factors contributing to BM prediction.

RESULTS

We developed a high-quality cohort with 4,466 patients with BM from the Cerner Health Fact database, which contains over 70 million patients from more than 600 hospitals. RETAIN uses this data set to achieve the best area under the receiver operating characteristic curve at 0.825, a significant improvement over the baseline model. We also extended a feature attribution method, Kernel SHAP, to structured EHR data for model interpretation. Both RETAIN and Kernel SHAP can identify important features related to BM prediction.

CONCLUSION

To the best of our knowledge, this is the first study to predict BM using structured EHR data. We achieved decent prediction performance for BM prediction and identified factors highly relevant to BM development. The sensitivity analysis demonstrated that both RETAIN and Kernel SHAP could discriminate unrelated features and put more weight on the features important to BM. Our study explored the potential of applying explainable artificial intelligence for future clinical applications.

INTRODUCTION

Brain metastases (BMs) develop when cancer cells migrate from their original sites to the brain. Lung cancer, especially non–small-cell lung cancer, accounts for 77%-88% of all BMs.^1-4 Often associated with cerebral edema, BMs cause neurologic morbidity and affect the quality of life. Therefore, early detection of BM would allow the implementation of minimally invasive treatments.^4,5 However, magnetic resonance imaging, although highly effective in detecting small BMs, is not currently incorporated within the standard of care for the initial staging of patients with newly diagnosed lung cancer because of the concern of cost-effectiveness. Therefore, a predictive model with high accuracy to identify patients with lung cancer with an increased risk of developing BM would maximize the value of early magnetic resonance imaging screening.

CONTEXT

• Key Objective

• Early detection of brain metastases (BMs) from lung cancer is critical for implementing prompt and minimally invasive treatments. Therefore, we seek to predict the risk of developing BM among patients diagnosed with lung cancer on the basis of electronic health record (EHR) data through explainable artificial intelligence approaches.

• Knowledge Generated

• A list of BM-related diagnoses and medications was automatically identified and then manually reviewed by clinical experts to accurately define cases and controls. With a series of well-designed case-control matching criteria, we achieved decent performance using all features, including demographic characteristics, diagnosis, medication, surgical, laboratory test, and clinical event. In addition, a list of important features contributing to the BM prediction was also identified.

• Relevance

• Considering the current low-quality EHR data used in this study, we can expect that accurate prediction and novel biomarker discovery for early BM detection are achievable with more high-quality EHR data that include cancer staging, clinical histology, and molecular testing.

The widely deployed electronic health record (EHR) systems present an unprecedented opportunity to leverage an enormous amount of historical patient data for predictive model development.^6-10 Recent advances in deep learning technologies demonstrated successes in EHR-based prediction tasks, for example, predicting heart failure risk,^8,11 hospital readmission,¹² and early sepsis detection.¹³ In addition, the recurrent neural network and attention mechanism were used to develop the REverse Time AttentIoN (RETAIN) model to gain high accuracy without losing interpretability.¹⁴ Recurrent neural network–based models also show good generalizability on large data sets, including various hospitals for heart failure prediction.¹¹

However, deep learning is often challenged by its capability to answer why a model makes a certain prediction.^15,16 Therefore, various methods have been developed to explain deep learning models. For example, the attention mechanism is usually used to identify the contributing factors that are most concerning for the outcome by model prediction^14,17 but cannot be applied to explain models with no attention mechanism. Other model explanation methods include (1) intrinsic explanations embedding biomedical knowledge into the structure design of the neural network,¹⁸ (2) post hoc explanation methods computing the attributions for all input features by tracking the changes in each layer via backpropagation,¹⁹ and (3) additive feature attribution methods explaining the deep learning model without relying on specific architectures and therefore being more generalizable.^15,16,20-22 The additive feature attribution method was used to explain the predictions for preventing hypoxemia during surgery and other mortality data sets, demonstrating its applicability to clinical problems.¹⁶ It is worth noting that both the attention mechanism and the additive feature attribution methods explain the behavior of the predictive model itself, which may or may not necessarily elucidate the data.²⁰

In this study, we used an explainable deep learning–based disease prediction model, RETAIN, to predict the risk of developing BM among patients diagnosed with lung cancer on the basis of EHR data and to understand what factors are important for the model to predict BM development. In addition, we extended the feature attribution method, Kernel SHAP, to structured EHR data to make our study generalizable. Furthermore, Kernel SHAP can be applied to any model for an explanation by changing the inputs and calculating the difference between including and excluding a specific feature. Therefore, our algorithm can be used for any model built for structured EHR data.

To our knowledge, our work is the first deep learning application to predict BM development using structured EHR data. Predicting BM from patients diagnosed with lung cancer needs a data-driven approach as a large cohort of patients with lung cancer is needed to identify a sufficient number of patients with BM to train the deep learning model. An EHR database from a single or few centers is not likely to contain sufficient data samples for this purpose. Therefore, we queried a large nationwide EHR database, Cerner Health Facts, to identify sufficient samples to develop a deep learning model. We carefully design the cohort construction criteria by automatically detecting high-risk factors and manual review by an experienced neurosurgical oncologist. The results show that the area under the receiver operating characteristic curve (AUC) on the separated testing data set reaches approximately 82.5%. Compared with logistic regression, the deep learning–based disease risk prediction model achieved a 7.4% improvement. Meanwhile, the attention mechanism and feature attribution approach identified important features contributing to the BM prediction.

MATERIALS AND METHODS

Data Set

Cerner Health Facts²³ is a deidentified EHR database that contains over 70 million unique patients from over 600 clinical organizations in the United States. More than 10 years of patient records were extracted with the latest encounters up to 2017. Most encounters are outpatient, clinic, inpatient, recurring, and emergency, accounting for more than 70% of all encounters. Each encounter may include admission, billing, diagnoses, medications, procedures, surgical, and laboratory microbiology with admitted and discharged time stamps. International Classification of Diseases (ICD) codes and Logical Observation Identifiers Names and Codes are used to normalize diagnoses and laboratory microbiology.

In this study, we identified 203,504 patients with lung cancer with ICD-9 code "162.*" and ICD-10 code "c34.*." Among these patients with lung cancer, we identified 26,923 patients with BM with ICD-9 code “198.3” and ICD-10 code “C79.3.” However, further analysis showed that more than half of these patients with BM were inappropriate for BM development prediction because the BM and lung cancer diagnoses were made at the same encounter. This also indicates that many patients present with BM at the time of the original lung cancer diagnosis. After we removed these samples, the total number of patients with BM for the following analysis was 12,086.

Sample Definition

A good selection criterion for building a case-control cohort must ensure the credibility of each case and control sample. In this study, relying on a single ICD code is sufficient for labeling a patient with BM but not for a patient who should have lung cancer but not BM. However, many patients with lung cancer without BM ICD codes were observed with evident diagnosis and medication clues of BM, such as seizures and the medications for treating seizures. Moreover, some inconsistency may occur if a patient was first diagnosed and treated in one hospital but then transferred to another. To tackle these issues, we performed a series of screening experiments to identify these factors automatically by using all the features available in the raw data. After further review by the clinical expert, the evident codes, which we named BM facts, were selected to purify the cohort. We repeated this process until no more BM facts occurred.

After we obtained the list of BM facts, the following operations were conducted to purify the samples of BM after lung cancer (case) and lung cancer without BM (control) groups. A valid BM case should meet two conditions: (1) the diagnosis date of BM must be after at least 3 months of lung cancer and (2) BM facts must not occur before BM diagnosis. In addition, the cases with BM facts but no BM diagnosis were excluded to ensure validity. Then, for decent controls, BM diagnosis and BM facts should not be observed in the whole medical history of patients with lung cancer. In addition, considering the potential loss of follow-up because of a new diagnosis, a patient with lung cancer with a discharge disposition as hospice was excluded. Finally, a valid control sample must have at least a 1-year follow-up to exclude BM. Figure 1 shows the overall process from raw data extraction to data cleaning for model development.

FIG 1.

The flowchart of sample processing from raw data extracted from the database to model development. BM, brain metastasis.

A case-control matching analysis was performed with the purified data to identify 10 matched controls for each valid case. We then randomly split these samples into training, validation, and test sets with a ratio of 7:2:1. The model was first trained using all the samples in the training set. Next, the validation set was used to select the best-performed model. Finally, the performance on the test set was reported throughout the paper. All the samples underwent the same processing pipeline and split randomly to ensure the same distribution between the training and test set. Data Supplement shows the balanced distributions of demographic variables on training, evaluation, and test set at the time of lung cancer diagnosis.

Feature Definition

The following features were used to predict BM development: (1) demographic information like age, sex, race, and marital status and (2) medical codes in each encounter. For the latter, we considered diagnoses, medication, surgical procedures, laboratory tests, and clinical events in one encounter. All the ICD-9 codes in diagnosis and clinical events were converted to the ICD-10 version for reduced parameters. Although the numerical results may exist for laboratory tests and clinical events, we still used Logical Observation Identifiers Names and Codes concatenated with result indicators (eg, low, high, and within range) because of the inconsistency of units and device biases in different hospitals. Moreover, the generic names for medications are used to eliminate the issue of multiple IDs for the same drug. Finally, 7,726 unique medical codes were included to build the models. All these features were learned by an embedding layer to get the numerical representations.

Baseline Models and RETAIN for BM Risk Prediction

We used two baseline models, logistic regression and single-layer gated recurrent unit (GRU), as benchmarks for performance evaluation. In addition, RETAIN was used to predict BM in patients diagnosed with lung cancer. Details of the baseline models and the RETAIN model are provided in the Data Supplement.

Kernel SHAP for Structured EHR Data

Shapley value is a method from coalitional game theory that satisfies certain properties for feature attribution.^20,22 It can assign an importance score (attribution) to each input feature by measuring the difference in the prediction with/without this specific feature. However, the importance score is calculated using all possible combinations of the input features, raising a computation bottleneck. Lundberg and Lee¹⁵ developed Kernel SHAP to approximate the Shapley value via a sampling strategy. Compared with other additive feature attribution methods, Kernel SHAP is model-agnostic and makes no assumptions about the model type. Meanwhile, the approximation of Shapley value via sampling also benefits the structured EHR data, which usually contains thousands of unique medical codes. Specifically, Kernel SHAP aimed to fit a locally linear model $g$ with simplified input $Z$,

$$\begin{array}{c}g\left(Z\right)={\varphi }_0+{\displaystyle \sum }_{i=1}^m{\varphi }_i{Z}_i,\left(1\right)\end{array}$$

where $Z$ is a binary vector and ${\varphi }_i$ is the effect of each feature. Summing all the effects can approximate the prediction from the original model $f$. To calculate the effects $\varphi$, it will minimize a locally weighted least squares loss function plus a LASSO penalization,

$$\begin{array}{c}\xi =\arg \min \sum {\left[f\left(h_x\left(Z\right)\right)-g\left(Z\right)\right]}^2\pi \left(Z\right)+\mathrm{\Omega }\left(g\right),\left(2\right)\end{array}$$

$$\begin{array}{c}\pi \left(Z\right)=\frac{\left(M-1\right)}{\left(M\hspace{0.17em}\mathrm{c}\mathrm{h}\mathrm{o}\mathrm{o}\mathrm{s}\mathrm{e}\hspace{0.17em}\left|Z\right|\right)\left|Z\right|\left(M-\left|Z\right|\right)},\left(3\right)\end{array}$$

where $M$ is the total features in a prediction and $h_x\left(Z\right)$ is a function that can transform the simplified features into the original feature space.

However, because of the hierarchical structure of EHR data, the original Kernel SHAP model cannot be applied directly. To construct the simplified input while maintaining its ability to transform back to the original feature space, we first flattened all the encounters of a patient. We then calculated all the unique medical codes to obtain the features that needed to be interpreted. Next, the simplified input was constructed by randomly masking some features. Finally, the simplified input was transformed into a patient's EHR data by replacing all masked features as the background code in the original patient's EHR data. The remaining operation adopted the process of Kernel SHAP as described in Appendix Table A1.

RESULTS

Descriptive analysis for the cohort of BM after lung cancer and lung cancer without BM is shown in Table 1. The number of patients with BM after lung cancer is reduced from 12,086 to 4,782 after the patient selection workflow in Figure 1. Overall, case samples have more features than control samples.

TABLE 1.

Data Statistics of BM Data

BM Facts Identified by RETAIN Using All Features

To identify the codes directly related to BM, we trained a RETAIN model with all samples labeled by ICD codes of BM. Each code's contributing score was calculated by averaging its weight across encounters and patients. Next, a clinical expert reviewed the top 50 codes to identify a list of codes considered BM facts (Data Supplement). ICD codes 191.9/C71.9, 239.6/D49.6, and V10.85/Z85.841 indicate neoplasm of the brain, and we conjecture that these codes might have occurred because of the inaccurate operations by the billers. ICD codes 348.5/G93.6 and 780.39/R56.9 and medications levetiracetam and phenytoin are very obvious clues that show that a patient with lung cancer had developed brain-related complications and was treated by these medications. A patient with lung cancer with any of these codes or medications will be filtered out to make a clean cohort, reducing the case numbers from 4,782 to 4,466. Then, a case-control matching experiment was conducted to identify controls for each valid case. Finally, 4,466 cases and 36,078 controls were obtained to train the RETAIN model for prediction.

RETAIN Outperforms Baseline Models

We benchmarked RETAIN and baseline models using the test set in Table 2. Overall, RETAIN achieved higher performance than logistic regression and single-layer GRU. Comparing experiment 1 with diagnoses and medication as input (AUC 0.797) and experiment 2 with laboratory tests and clinical events as input (AUC 0.688), we found that patients' historical diagnoses and medications contribute more to the prediction of the disease development. Using all the features, RETAIN achieved the best AUC of 0.825, which is 7.4% better than logistic regression alone (AUC 0.751). The logistic regression model can be enhanced by the embeddings from RETAIN to achieve an AUC of 0.788 (Data Supplement). The result indicates that concept representation learned by deep learning is very informative. Furthermore, RETAIN outperformed logistic regression with embedding using other evaluation metrics, that is, precision, recall, and accuracy, as shown in the Data Supplement. Meanwhile, RETAIN achieved AUC 0.6% higher than single-layer GRU with all data as input. Interestingly, single-layer GRU performed better than RETAIN if not including laboratory tests and clinical events.

TABLE 2.

AUC Scores on the Separated Test Set

Identifying High-Contribution Factors for BM Prediction

We identified high-contribution factors using both RETAIN and Kernel SHAP. The Kernel SHAP was applied to single-layer GRU, and both approaches were applied to the data set without laboratory tests and clinical events. For RETAIN, we averaged the contribution weights of each code across encounters and patients. Meanwhile, we consider effect size by only keeping the codes with more than 20 patients. The top 10 risk and protective contribution factors from each approach were identified, as shown in Tables 3 and 4. Although the BM facts were filtered out, the high-contribution factors left were still related to the development of BM. RETAIN and Kernel SHAP have five overlapped features in their top 10 list, showing the consistency between these two methods. Some codes in Tables 3 and 4 show the secondary malignant neoplasm from the primary lung cancer, indicating that metastasis has already been observed in other organs like bone and liver, and these patients are in their late stage of lung cancer. RETAIN and Kernel SHAP can still infer important information like cancer stage from the symptoms and medications although they are absent in the Cerner database. Meanwhile, several codes for brain disorders (eg, R51-ICD code for headache) indicate some brain-related diseases observed in patients with lung cancer with a high probability of BM onset.

TABLE 3.

High-Contribution Factors for BM Prediction Identified by RETAIN

TABLE 4.

High-Contribution Factors for BM Prediction Identified by Kernel SHAP

Interpretation Sensitivity for RETAIN and Kernel SHAP

To test the validity of these two interpretation approaches, we performed a sensitivity analysis by randomly inserting controlled dummy variables into the data set. We introduced 10 variables into the data set with the probability of 0.15 and 0.20 (at 0.99 quantiles for all patients) for patient and encounter levels, respectively. The overall contribution score for each feature was calculated to indicate the feature's importance in the test set. Figure 2A shows that no dummy features were ranked in the top 300 contribution code list for RETAIN and Kernel SHAP. RETAIN put less attention on these dummy variables than Kernel SHAP. Figure 2B shows the interpretation of Kernel SHAP for a specific prediction of a patient with BM. The interpretation method can identify features closely related to brain diseases like headaches and other cerebrovascular diseases as the top contributing factors for BM prediction. The important features identified are also consistent with the clinical knowledge of BM development evaluated by clinical experts.

FIG 2.

The interpretation result of BM prediction on structural EHR data. (A) The rank of importance for dummy variables in the BM data set. The rank of each dummy variable is calculated by using the absolute value of its contribution score in the whole test set. (B) The bar plot of important features of a specific prediction. BM, brain metastasis; EHR, electronic health record.

DISCUSSION

The interpretable RETAIN model achieved reasonably good performance, with the AUC reaching 82.5%. We also extended the model-agnostic feature attribution method, Kernel SHAP, to structured EHR data. The sensitivity analysis demonstrates that the attention mechanism and feature attribution approach can recognize unrelated features and identify high-contribution features for BM prediction. Furthermore, the factors identified by the attention mechanism and feature attribution method explain how the model makes a decision (Data Supplement).

Figure 2B shows that malignant neoplasm of the lung is the top contributing factor for predicting BM development, which may be puzzling because both the cases and the controls included in this study are patients with lung cancer with this code. However, models like RETAIN consider all the data, including temporal information. A careful examination of data shows that although the code malignant neoplasm of the lung is used to identify both cases and controls, the distribution/temporal information of this code is different between patients with and without BM (Data Supplement). In the data we used, the temporal feature of each code includes the frequency, the order of appearance with regard to other codes, the time this feature appears, and so on. Therefore, our model identified the malignant neoplasm of the lung as the most important feature for this specific patient on the basis of the temporal characteristics of this code.

One main challenge for this study is to design good criteria for constructing a high-quality cohort. Because of the inconsistency in the Cerner EHR database that collected data from many hospitals, even a clinical expert has difficulty in listing an exclusive cohort construction criterion. For example, many patients with lung cancer do not have the ICD code of secondary neoplasm of the brain. However, we observe many BM-related medications, like LEVETIRACETAM and PHENYTOIN, and symptoms, like cerebral edema and convulsions (Data Supplement). Therefore, if we rely on a single BM ICD code to construct the cohort, those patients would be mislabeled as the control sample. So, one of the contributions in this study is the hybrid approach we adopted to identify a list of BM facts, which can help to filter out these ambiguous samples and make a clean case and control group.

A limitation of this study is the absence of some important information in the Cerner EHR database. For example, it is well-studied that BM from non–small-cell lung cancer accounts for the majority of BM cases and late-stage lung cancer has a much higher probability of developing BM. However, the cancer type and stage information are not collected in the Cerner EHR database. Despite this limitation, the data still contain signals directly or indirectly related to cancer staging. For example, the model shows that patients with advanced disease (eg, metastases in other organs such as liver, bone marrow, bone, and skull) or certain neurologic symptoms (headaches or patients labeled as unclear/other disorders of the brain) are more likely to have BMs. One strength of the deep learning model is to learn from these subtle signals and incorporate them into BM risk prediction, which is reflected in the very good performance.

We used explainable deep learning models in this study to predict BM development in patients with lung cancer using a large EHR data set. A list of BM-related diagnoses and medications was automatically identified and manually reviewed by a clinical expert to accurately define cases and controls. With a series of well-designed case-control matching criteria, the overall performance using RETAIN with all features, including demographic, diagnosis, medication, surgical, laboratory test, and clinical event, reaches 82.5% for AUC. We also extended a model-agnostic feature attribution method, Kernel SHAP, to structured EHR data for model interpretation. The sensitivity analysis shows that both RETAIN and Kernel SHAP can identify important features related to BM prediction. However, some essential information missed in the Cerner EHR database, such as the cancer stage and the cancer subtype, limits the performance of the prediction models.

APPENDIX

TABLE A1.

Kernel SHAP for Structured EHR Data

DATA SHARING STATEMENT

The code is available via https://github.com/WJZheng-group/lung_brain_meta_prediction. Since the Cerner database is commercial, anyone interested can subscribe to Cerner Health Facts to access the data.

AUTHOR CONTRIBUTIONS

Conception and design: Zhao Li, Laila Rasmy, Xiaoqian Jiang, Hua Xu, Yoshua Esquenazi, W. Jim Zheng

Financial support: Xiaoqian Jiang, W. Jim Zheng

Collection and assembly of data: Zhao Li, Yujia Zhou, Ping Zhu

Data analysis and interpretation: Zhao Li, Rongbin Li, Degui Zhi, Antonio Dono, W. Jim Zheng

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).