A novel method of adverse event detection can accurately identify venous thromboembolisms (VTEs) from narrative electronic health record data

Abstract

Background Venous thromboembolisms (VTEs), which include deep vein thrombosis (DVT) and pulmonary embolism (PE), are associated with significant mortality, morbidity, and cost in hospitalized patients. To evaluate the success of preventive measures, accurate and efficient methods for monitoring VTE rates are needed. Therefore, we sought to determine the accuracy of statistical natural language processing (NLP) for identifying DVT and PE from electronic health record data.

Methods We randomly sampled 2000 narrative radiology reports from patients with a suspected DVT/PE in Montreal (Canada) between 2008 and 2012. We manually identified DVT/PE within each report, which served as our reference standard. Using a bag-of-words approach, we trained 10 alternative support vector machine (SVM) models predicting DVT, and 10 predicting PE. SVM training and testing was performed with nested 10-fold cross-validation, and the average accuracy of each model was measured and compared.

Results On manual review, 324 (16.2%) reports were DVT-positive and 154 (7.7%) were PE-positive. The best DVT model achieved an average sensitivity of 0.80 (95% CI 0.76 to 0.85), specificity of 0.98 (98% CI 0.97 to 0.99), positive predictive value (PPV) of 0.89 (95% CI 0.85 to 0.93), and an area under the curve (AUC) of 0.98 (95% CI 0.97 to 0.99). The best PE model achieved sensitivity of 0.79 (95% CI 0.73 to 0.85), specificity of 0.99 (95% CI 0.98 to 0.99), PPV of 0.84 (95% CI 0.75 to 0.92), and AUC of 0.99 (95% CI 0.98 to 1.00).

Conclusions Statistical NLP can accurately identify VTE from narrative radiology reports.

INTRODUCTION

Venous thromboembolism (VTE), which includes deep vein thrombosis (DVT) and pulmonary embolism (PE), is one of the most common complications of hospitalization.^1,2 In the absence of thromboprophylaxis, the incidence of VTE ranges from 10–40% in medical and general surgical populations, to as high as 40–60% in patients who have undergone major orthopedic surgical procedures.¹ VTE is also a leading cause of mortality and morbidity. The 30-day VTE case fatality rate in hospitalized patients ranges from 5% to 15%.³ Moreover, it is estimated that 15–50% of VTE patients will experience post-thrombotic syndrome,^4,5 and that 4–5% will develop chronic thromboembolic pulmonary hypertension.^6,7 VTE is the second most common cause of excess length of hospital stay,⁸ and each case of hospital-acquired VTE results in an incremental inpatient cost of US $7000–$21 000.^8,9 Given the significant mortality, morbidity, and cost associated with VTE, its prevention has been ranked as the most important of 79 strategies aimed at improving patient safety in hospitals.¹⁰

Accordingly, regulatory agencies worldwide have launched VTE prevention initiatives that aim to encourage hospitals to assess VTE risk, and to institute appropriate thromboprophylaxis in high-risk patients.^11,12 An important requirement to evaluate the success of these initiatives is access to accurate, timely, and efficient methods for monitoring VTE rates. However, at present, there are no such methods.^13,14 Although manual chart review is the reference standard in many adverse events studies, it is a time-consuming, resource-intensive, and costly process.^13–15 As a consequence, it is an impractical means for routine VTE monitoring. While discharge diagnostic codes have the advantage of being readily available, relatively inexpensive, and easy to use,^15,16 previous studies have found that they generally have low to moderate sensitivity and positive predictive value (PPV) for identifying VTEs.^17,18 In addition, it can be difficult to determine from discharge diagnostic codes whether a VTE occurred before the patient was hospitalized or during the actual hospitalization.^19,20 With the increasing availability of electronic health records (EHRs), a far richer source of clinical information for identifying VTEs is becoming available. Moreover, with the advent of automated methods for encoding and classifying electronic narrative documents, such as natural language processing (NLP), an exciting opportunity has emerged to develop potentially more accurate, timely, and efficient methods for monitoring VTE rates.

NLP refers to automated methods for converting free-text data into computer-understandable format.²¹ NLP techniques have been divided into two broad categories: symbolic and statistical. Symbolic (or grammatical) techniques use the characteristics of the language (i.e., semantics, syntax, and the relationships among sentences) to interpret a narrative document to the extent necessary for encoding it into one of a set of discrete categories.²² Only a few studies have used symbolic NLP techniques to identify adverse events such as VTEs. While the results of these studies are promising, symbolic NLP techniques were found to have low to moderate sensitivity and PPV for identifying DVT and PE.^23–25 This could possibly be attributed to the characteristics of clinical narratives, which are often ungrammatical, composed of short telegraphic phrases, and replete with abbreviations, acronyms, and local dialectal shorthand phrases.^26,27

Alternatively, statistical NLP techniques use the frequency distribution of words and phrases to automatically classify a set of documents into one of a discrete set of predefined categories.²⁸ Among the various statistical NLP techniques, support vector machines (SVM), which is a type of supervised machine learning, have been widely used in pattern recognition and classification problems.^29,30 SVM models have demonstrated high performance in automatically classifying and detecting diseases,^31,32 and are one of the most effective models for automated text classification.³³ To our knowledge, no prior study has documented their accuracy in identifying VTEs from narrative clinical documents. We thus sought to determine the accuracy of SVM models for identifying DVTs and PEs.

METHODS

Setting

The study was conducted at the McGill University Health Centre (MUHC), a university health network located in the Canadian province of Quebec. The MUHC is composed of five adult-care hospitals and has more than 800 beds. It serves a population of 1.7 million people (22% of the provincial population), with an annual volume of approximately 865 000 ambulatory visits, 34 000 surgeries, and 38 000 hospitalizations. The research ethics committee of the MUHC approved this study.

Data sources

Data for this study were extracted from three electronic databases at the MUHC and were linked by unit, patient, and hospital admission date. The Discharge Abstract Database provided patient age and sex, and dates of hospital admission and discharge. The Admission, Discharge and Transfer Database was used to identify the unit where the patient was located at the time of the radiological examination. The Radiology Report Database provided data on all radiological examinations that were performed over the study period in patients suspected of having a VTE, including dates when these examinations were performed, a text description of the radiological findings, and the radiologist's interpretation. At the time of this study, no other clinical narratives were available in an electronic format at the MUHC.

Study design and data sample

To determine the accuracy of SVM models in identifying DVTs and PEs, we conducted a validation study. First, we randomly sampled 2000 narrative radiology reports among all radiological examinations that were performed at the MUHC between January 1, 2008 and December 31, 2012 for patients with a suspected DVT or PE. Then, two sets of alternative SVM models were trained and tested: one set predicting DVTs (including DVTs of the lower and upper extremities) and one set predicting PEs. To identify the best performing SVM model for predicting DVT and PE, the accuracy of the models within each set was compared. The decision to develop two sets of SVM models was based on the observation that several narrative reports described the results of a radiological examination (e.g., pulmonary embolus study with distal runoff) that was performed in a patient suspected of having both conditions. While it is possible to train a multi-class SVM model, most work on SVMs and most standard evaluation techniques are designed for binary classification problems.³⁴ As such, we opted for the two-model approach.

SVM model development and validation

To develop and validate the two sets of SVM models, four successive steps were followed: (a) reference standard development, (b) text pre-processing and feature generation, (c) feature selection, and (d) SVM training, testing, and validation.

Reference standard development

First, the 2000 radiology reports were manually coded by a clinical expert (CMR) to identify cases of DVT and PE. During the coding process, each report was assigned two codes: (a) positive or negative for DVT of the lower or upper extremities, and (b) positive or negative for PE.

Positive radiology reports for a DVT were those where a thrombus was identified in the proximal deep veins of the lower extremities (e.g., external iliac, common femoral, deep femoral, or popliteal), in the deep distal veins of the lower extremities (e.g., peroneal and posterior tibial), or in the deep veins of the upper extremities (e.g., brachial, radial, ulnar, axillary, or subclavian). Negative cases included those where no thrombus was identified or where a thrombus was identified in a superficial vein of the lower extremity (e.g., saphenous), in a superficial vein of the upper extremity (e.g., cephalic), or in a perforating vein of the lower extremity but not extending into a deep vein.¹⁷ Radiological examinations finding evidence of chronic thrombosis were coded as negative.

Similarly, positive radiology reports for a PE included those where a filling defect was identified in the central, segmental, or subsegmental pulmonary arteries. Radiological reports describing evidence of chronic PE were coded as negative, as were those finding no evidence of the disease. Lastly, a second clinical expert (TE) blindly recoded a 20% random sample of the radiology reports and inter-coder reliability was assessed using the κ statistic, yielding near perfect agreement (κ = 98).

Text preprocessing and feature generation

In preparation for SVM training and validation, the unstructured text data embedded in the narrative radiology reports were transformed into a corpus, which is a database of text documents.³⁵ Then, a series of transformations were applied to each radiology report within the corpus, including: (a) conversion of all words to lower case, and (b) the removal of punctuation marks and superfluous white spaces.³³ The transformed corpus was converted into a document-by-term matrix (bag-of-words),³⁶ a structured format holding radiology report IDs as rows, terms as columns, and term frequencies within a given radiology report as matrix elements. The original bag-of-words contained 7370 distinct features (i.e., words or unigrams). To introduce some elements of contextual knowledge, and preserve the local dependencies of each word, bigrams (which are combinations of two consecutive words) were introduced into the bag-of-words as additional features.³⁷ This resulted in a total of 62 416 distinct features (i.e., unigrams and bigrams). Text preprocessing was conducted in R with package tm.³⁸

Lastly, because feature generation is a key determinant of SVM model prediction performances,³⁹ we experimented with several potential enhancements to our proposed feature generation approach, including word stemming (i.e., reducing inflected words to their root form) and using higher order n-grams (e.g., trigrams). These potential enhancements were evaluated as part of the SVM model training and validation.

Feature selection and SVM training and validation

To assess the accuracy of the SVM models for identifying DVT and PE, we used a 10-fold cross-validation approach. To avoid biasing the accuracy of the SVM models by using information from the test sets, we first performed feature selection within each set of k−1 training folds.^34,40 We used the Pearson's correlation coefficient (ρ) to identify subsets of features significantly associated with DVT and PE.³⁹ Using a threshold value of |ρ|≥0.10, 118 unigrams and 218 bigrams were associated with DVT, and 301 unigrams and 1242 bigrams were associated with PE. The top 30 most informative features (highest values of ρ) are summarized in table 1. As can be seen from table 1, most of these features are either related to: (a) a relevant anatomical body part (e.g., vein, pulmonary artery, femoral vein, or segmental branches (of the pulmonary arteries)), or (b) pathological manifestations of DVT or PE (e.g., thrombus, emboli, or filling defect).

Table 1:

The top 30* most informative unigrams and bigrams for deep vein thrombosis and pulmonary embolism identification according to the Pearson's correlation statistic

Deep vein thrombosis		Pulmonary embolism
Unigrams	Bigrams	Unigrams	Bigrams
Vein	Length of	Filling	Pulmonary artery
Thrombus	The thrombus	Segmental	Filling defect
Occlusive	Thrombosis involving	Artery	Lower lobe
Peroneal	Popliteal vein	Lobe	Defect in
Length	Over a	Pulmonary	Pulmonary emboli
Patent	Non occlusive	Defect	A filling
Popliteal	Posterior tibial	Subsegmental	There are
Over	A length	Strain	With pulmonary
Femoral	Is deep	Emboli	Upper lobe
Tibial	Peroneal vein	Chest	Main pulmonary
Thrombosis	The mid	Lung	Segmental branch
Veins	Basilic vein	Main	Defects are
Involving	Femoral vein	Branches	Segmental branches
Entire	Entire length	Basal	Embolus in
Thrombosed	Reminder of	Small	Multiple filling

*Features were ranked using their Pearson correlation coefficient (ρ) and the top 30 unigrams and bigrams were selected.

Then, using these subsets of features, a nested 10-fold cross-validation was performed within each set of k−1 training folds to identify the optimal value of: (a) the cost parameter C, which controls the trade-off between false positives and false negatives (used with all kernel functions), and (b) the γ parameter, which controls the shape of the hyperplane (used with the radial basis function (RBF) kernel only).^34,41 The optimal value of C and γ (when applicable) were then automatically applied to the relevant set of k−1 training folds, and a parametrically optimized multivariate SVM model was trained using either a linear or an RBF kernel.⁴⁰ Lastly, this optimal SVM model was tested on the remaining kth test fold. During SVM training and validation, repeated radiology reports from the same patient over any given hospitalization (if any) were not allowed to cross the training and test sets to avoid inflating the performances of the SVM models.

Within each test fold, we estimated the sensitivity, specificity, PPV, negative predictive value (NPV), and the area (AUC) under the receiver operating characteristic (ROC) curve. These estimates were then averaged over the 10 folds, and their 95% CIs estimated.⁴² For each test fold, an ROC curve was generated by using the classification probabilities produced by the SVM models as the discrimination threshold.⁴³ Vertical averaging was used to generate the average ROC curve for each of the two models.⁴³ To account for the asymmetric class sizes (i.e., fewer positive cases of DVT and PE), and to avoid biasing the SVM predictions towards the majority class, each class was weighted by the inverse proportion.³⁴ Then, using this general approach, 10 alternative models were trained and tested. The first eight models were trained using the aforementioned selected subsets of features and were characterized by their kernel function (linear vs RBF), gram type (unigrams only vs unigram and bigrams), and the use or non-use of SVM parameter optimization (tuning vs no tuning). SVM models with no tuning used the default values for the cost (C = 1) and γ (γ = 1/number of features) parameters.³⁴ Two additional SVM models were trained and tested using all available features: (a) all features, linear kernel, uni-bigrams, and tuned parameters, or (b) all features, RBF kernel, uni-bigrams, and tuned parameters. To identify the best performing model, the AUCs of these 10 SVM models were compared using the Friedman test, and pairwise comparisons were performed using Tukey's test with the Bonferroni adjustment for multiple comparisons.³⁴ To determine if potential enhancements to our proposed feature generation approach (i.e., stemming and higher order n-grams) influenced SVM prediction performances, alternative SVM models were trained, tested, and compared using the aforementioned procedures. SVM models were implemented in R using package e1071.⁴⁴

RESULTS

Overall, 1649 patients contributed 1751 hospitalizations from which the 2000 randomly selected narrative radiology reports were generated. A total of 1544 (88.2%) hospitalizations contributed only one radiology report to the analyses, while 207 (11.8%) hospitalizations contributed between two and five radiology reports (table 2). The other characteristics of these hospitalizations are summarized in table 2.

Table 2:

Characteristics of the 1751 hospitalizations which contributed the 2000 narrative radiology reports

Hospitalization characteristics	Statistics (n = 1751)
Demographic characteristics
Sex
Male, n (%)	892 (50.9)
Female, n (%)	859 (49.1)
Age, mean ± SD	66.7 ± 16.1
Nursing unit at the time of the examination
Internal medicine, n (%)	643 (36.7)
Surgery, n (%)	537 (30.7)
Intensive care unit, n (%)	332 (19.0)
Other (e.g., geriatrics, neurology, short stay), n (%)	239 (13.6)
Length of hospital stay (days), median (IQR)	15 (28)
Number of radiology reports contributed to the analyses
One radiology report, n (%)	1544 (88.2)
Two radiology reports, n (%)	173 (9.9)
Three to five radiology reports, n (%)	34 (1.9)

The typical radiology report had a median length of 110.5 words, ranging from 8 to 727 words. On manual review, 454 narrative radiology reports (22.7%) described a VTE. Of these, 324 reports (16.2%) described a DVT, including 216 DVTs of the lower extremities and 108 DVTs of the upper extremities. In addition, 154 reports (7.7%) described a PE. Notably, 24 reports described 2 simultaneous events, the most prevalent combination being a lower extremity DVT and a PE (n = 19).

The average sensitivity, specificity, PPV, and NPV of the SVM models for identifying DVTs are shown in table 3. On average, the best performing model correctly identified 80% of the radiology reports describing a true DVT (sensitivity: 0.80 (95% CI 0.76 to 0.85)) and generated 11% of false positives (PPV: 0.89 (95% CI 0.85 to 0.93)). The average AUC of this model was 0.98 (95% CI 0.97 to 0.99), and the associated average ROC curve is presented in figure 1. These performances, which are highlighted in boldface in table 3, were achieved on an SVM model trained on the whole feature set using an RBF kernel for which both the cost and γ parameters were optimized. The performance of this SVM model did not statistically differ from that of similar SVM models trained using a linear kernel function on either the whole feature set or a selected subset of features (table 3). It was also not statistically significantly influenced by feature selection (table 3). Nonetheless, it performed statistically significantly better than any other alternative SVM models (table 3).

Table 3:

Average accuracy estimates of the SVM models for identifying deep vein thrombosis (DVT)

	Average estimates (95% CI)†
SVM models	Sensitivity	Specificity	PPV	NPV	AUC
Unigram only, linear kernel, no tuning	0.65 (0.58 to 0.71)	0.97 (0.96 to 0.98)	0.82 (0.77 to 0.87)	0.93 (0.90 to 0.96)	0.94 (0.92 to 0.96)‡
Unigram only, RBF kernel, no tuning	0.57 (0.45 to 0.68)	0.96 (0.95 to 0.97)	0.72 (0.66 to 0.79)	0.92 (0.89 to 0.95)	0.93 (0.91 to 0.96)‡
Unigram only, linear kernel, with tuning	0.69 (0.64 to 0.73)	0.97 (0.96 to 0.98)	0.82 (0.77 to 0.87)	0.94 (0.91 to 0.97)	0.95 (0.94 to 0.97)‡
Unigram only, RBF kernel, with tuning	0.70 (0.65 to 0.74)	0.97 (0.96 to 0.98)	0.81 (0.74 to 0.88)	0.94 (0.92 to 0.97)	0.95 (0.92 to 0.97)‡
Uni + bigrams, linear kernel, no tuning	0.67 (0.63 to 0.73)	0.98 (0.97 to 0.99)	0.87 (0.82 to 0.93)	0.94 (0.90 to 0.97)	0.95 (0.93 to 0.98)
Uni + bigrams, RBF kernel, no tuning	0.59 (0.49 to 0.69)	0.96 (0.95 to 0.97)	0.74 (0.66 to 0.82)	0.92 (0.89 to 0.96)	0.94 (0.92 to 0.96)‡
Uni + bigrams, linear kernel, tuning	0.70 (0.63 to 0.77)	0.98 (0.97 to 0.99)	0.85 (0.81 to 0.90)	0.94 (0.91 to 0.97)	0.96 (0.93 to 0.98)
Uni + bigrams, RBF kernel, tuning	0.77 (0.72 to 0.84)	0.97 (0.97 to 0.98)	0.84 (0.79 to 0.89)	0.96 (0.92 to 0.99)	0.97 (0.94 to 0.99)
Uni + bigrams, linear kernel, tuning, all features.	0.79 (0.74 to 0.84)	0.98 (0.97 to 0.99)	0.89 (0.84 to 0.94)	0.96 (0.93 to 0.98)	0.98 (0.97 to 0.99)
Uni + bigrams, RBF kernel, tuning, all features	0.80 (0.76 to 0.85)	0.98 (0.97 to 0.99)	0.89 (0.85 to 0.93)	0.96 (0.93 to 0.99)	0.98 (0.97 to 0.99)*

Bold typeface is used to highlight the characteristics of the best performing SVM model. *p<0.001; statistically significant difference in performance compared to alternative SVM models.

^†Averages correspond to the mean accuracy estimates obtained after 10 rounds of cross-validation.

^‡Statistically significantly different compared to the best performing SVM model (i.e., Uni + bigrams, RBF kernel, tuning, all features).

AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value; RBF, radial basis function kernel; SVM, support vector machine.

Figure 1:

Average receiver operator characteristic (ROC) curve associated with the deep vein thrombosis (DVT) model. The average ROC curve for the DVT model was estimated by vertically averaging the 10 ROC curves generated during 10-fold cross-validation. The best performances were achieved using an SVM model trained on the whole feature set using an RBF kernel for which both the cost and γ parameters were optimized.

Similarly, on average, the best SVM model predicting PE correctly identified 79% of the true PEs (sensitivity: 0.79 (95% CI 0.73 to 0.85)) and generated 16% of false positives (PPV: 0.84 (95% CI 0.75 to 0.92)) (table 4). This model had an AUC of 0.99 (95% CI 0.98 to 1.00), and the associated average ROC curve is shown in figure 2. These performances were achieved on an SVM model trained on the whole feature set using an RBF kernel for which both the cost and γ parameters were optimized (table 4). The performance of this SVM model did not significantly differ from that of similar SVM models trained using a linear kernel function on either the whole feature set or a selected subset of features (table 4). This model was also not statistically significantly influenced by feature selection (table 4). Nonetheless, this model performed statistically significantly better than any other alternative SVM models (table 4).

Table 4:

Average accuracy estimates of the SVM models for identifying pulmonary embolism (PE)

	Average estimates (95% CI)‡
SVM models	Sensitivity	Specificity	PPV	NPV	AUC
Unigram only, linear kernel, no tuning	0.51 (0.34 to 0.68)	0.99 (0.98 to 1.00)	0.75 (0.62 to 0.88)	0.96 (0.95 to 0.98)	0.92 (0.87 to 0.96)†
Unigram only, RBF kernel, no tuning	0.39 (0.32 to 0.46)	0.98 (0.97 to 0.99)	0.63 (0.45 to 0.81)	0.95 (0.94 to 0.96)	0.93 (0.92 to 0.95)†
Unigram only, linear kernel, with tuning	0.53 (0.36 to 0.70)	0.98 (0.97 to 0.99)	0.75 (0.65 to 0.85)	0.96 (0.94 to 0.98)	0.92 (0.88 to 0.96)†
Unigram only, RBF kernel, with tuning	0.55 (0.42 to 0.67)	0.98 (0.97 to 0.99)	0.70 (0.54 to 0.86)	0.96 (0.95 to 0.98)	0.95 (0.93 to 0.97)†
Uni + bigrams, linear kernel, no tuning	0.60 (0.45 to 0.75)	0.99 (0.98 to 1.00)	0.85 (0.77 to 0.93)	0.97 (0.95 to 0.98)	0.95 (0.90 to 1.00)
Uni + bigrams, RBF kernel, no tuning	0.40 (0.33 to 0.47)	0.98 (0.97 to 0.99)	0.67 (0.51 to 0.83)	0.95 (0.94 to 0.96)	0.95 (0.93 to 0.96)†
Uni + bigrams, linear kernel, tuning	0.61 (0.46 to 0.76)	0.99 (0.98 to 1.00)	0.84 (0.76 to 0.92)	0.97 (0.95 to 0.98)	0.95 (0.90 to 1.00)
Uni + bigrams, RBF kernel, tuning	0.66 (0.49 to 0.83)	0.99 (0.98 to 1.00)	0.80 (0.68 to 0.93)	0.97 (0.96 to 0.99)	0.96 (0.92 to 1.00)
Uni + bigrams, linear kernel, tuning, all features	0.78 (0.72 to 0.85)	0.99 (0.98 to 0.99)	0.84 (0.76 to 0.91)	0.98 (0.98 to 0.99)	0.99 (0.98 to 1.00)
Uni + bigrams, RBF kernel, tuning, all features	0.79 (0.73 to 0.85)	0.99 (0.98 to 0.99)	0.84 (0.75 to 0.92)	0.98 (0.98 to 0.99)	0.99 (0.98 to 1.00)*

Bold typeface is used to highlight the characteristics of the best performing SVM model. *p<0.001; statistically significant difference in performance compared to alternative SVM models.

^†Statistically significantly different compared to the best performing SVM model (i.e., Uni + bigrams, RBF kernel, tuning, all features).

^‡Averages correspond to the mean accuracy estimates obtained after 10 rounds of cross-validation.

AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value; RBF, radial basis function kernel; SVM, support vector machine.

Figure 2:

Average receiver operator characteristic (ROC) curve associated with the pulmonary embolism (PE) model. The average ROC curve for the PE model was estimated by vertically averaging the 10 ROC curves generated during 10-fold cross-validation. The best performances were achieved using an SVM model trained on the whole feature set using an RBF kernel for which both the cost and γ parameters were optimized.

As a final step, alternative SVM models predicting DVT and PE were trained and tested using potential enhancements to the proposed feature representation (i.e., word stemming and including trigrams). However, none of these approaches were superior to using a bag-of-words representation enhanced with bigram features. As such, they are not detailed any further.

DISCUSSION

In this study, we assessed the accuracy of a statistical NLP technique, based on SVM models, for the purpose of identifying VTEs from electronic narrative radiology reports. We found that DVTs and PEs can be accurately identified from a bag-of-words representation of narrative radiology reports that is augmented with bigram features.

To our knowledge, there have been few studies using NLP techniques for the purpose of identifying VTEs, and most relied on symbolic NLP. For instance, Melton and Hripcsak²³ assessed whether symbolic NLP techniques could accurately identify 45 adverse events, including DVTs, from narrative discharge summaries. For DVTs, a PPV of 0.51 was observed. In another study, Murff et al.²⁴ evaluated the accuracy of symbolic NLP techniques for identifying several postoperative complications, including VTEs, from an integrated EHR at six Veterans Health Administration (VHA) medical centers. The sensitivity of the NLP system for identifying VTEs was 0.59 (95% CI 0.44 to 0.72) and the specificity was 0.91 (95% CI 0.90 to 0.92). In a subsequent study based on a larger sample of VHA patients, FitzHenry et al.²⁵ used symbolic NLP techniques to detect nine post-operative complications, including DVT and PE. For DVT, sensitivity was 0.56 (95% CI 0.45 to 0.67), specificity was 0.94 (95% CI 0.93 to 0.95), and PPV was 0.15 (95% CI 0.11 to 0.20). For PE, a sensitivity of 0.80 (95% CI 0.66 to 0.89), a specificity of 0.97 (95% CI 0.96 to 0.98), and a PPV of 0.23 (95% CI 0.17 to 0.30) were observed. Overall, the results of our study add to this emerging body of literature and provide further evidence that automated methods based on NLP techniques can successfully be applied to EHR data for the purpose of identifying adverse events such as DVT and PE.

Interestingly, we were able to achieve relatively good prediction performances for both DVTs and PEs using a bag-of-words representation of the narrative radiology reports that was augmented with bigram features. While these performances compare to those noted in recent statistical text classification studies,^45–48 some characteristics of the models validated in this study should be emphasized. First, we found, in contrast to Bejan et al.⁴⁶ that feature selection did not significantly increase the performances of the SVM models beyond what can be achieved with bigrams and SVM parameter optimization. This observation provides evidence than when computational resources are limited, feature selection, as performed in this study, may provide acceptable results. Second, we observed, consistent with Hsu et al.⁴⁹ that when SVM parameters are optimized, both the linear and RBF kernel achieve similar performances. Lastly, we observed that several potential enhancements to the feature representation (e.g., using higher order n-grams) did not improve the SVM model predictive performances. Similar results were observed in a recent study where SVM models were used to categorize EHR progress notes pertaining to diabetes.⁴⁸ It is possible that more complex features, such as higher order n-grams, do not occur in a high enough frequency to significantly improve the performance of an SVM model. Moreover, it is possible that these alternative features do not contain the information needed to discriminate narrative reports from different classes. Recently, Garla and Brandt⁵⁰ used symbolic NLP techniques to encode semantic relationships between concepts in the narrative reports, and were able to improve the prediction performance of a machine learning clinical text classifier pertaining to obesity. Additional studies are needed to determine if this approach can be successfully replicated in other clinical contexts, including the identification of VTEs. Nonetheless, the combination of symbolic and statistical NLP techniques represents an interesting area of future research, and promises improved performance in adverse event detection.

Automated methods based on statistical or symbolic NLP techniques have a number of advantages over traditional adverse event monitoring methods, such as manual chart review. Because they are automated, they can rapidly scan large numbers of patient records and clinical data with minimal human effort and cost, potentially allowing surveillance of an entire healthcare organization's population rather than just subsamples. In addition to being scalable, another potential benefit of automation is that human resources traditionally assigned to manual adverse event monitoring could be more productively reassigned to the development, implementation, and follow-up of preventive interventions. Moreover, because NLP techniques use electronic data that are available in near-real time, they offer the potential for timely adverse event monitoring and for prompt intervention. This represents an important advantage over automated methods based on discharge diagnostic codes, which are only available several months after discharge.

An important strength of this study is the approach used to fit the SVM models. Indeed, feature selection performed on the whole dataset would have created a positive bias in prediction accuracy by indirectly using information from the test sets. To guard against this, we performed feature selection within each set of k−1 training folds, which ensured independence from the test sets. Similarly, we performed a nested 10-fold cross-validation within each set of k−1 training folds to identify the optimal values for the SVM parameters (C and γ). This approach penalized over-fit models during the parameter search process, therefore increasing generalizability. Lastly, we reported on the average accuracy estimates of the SVM models over the 10 test folds. This reduced the influence that any bias in the distribution of the data between the training and test sets could have on the predictive performance of the SVM models. It also provided more reliable insights on how the SVM models would perform on an independent data set drawn from the same population of radiology reports.

Another strength of this study was to train and validate the SVM models on narrative radiology reports from a network of five adult-care hospitals. Because these hospitals operate for the most part independently, and have their dedicated pool of radiologists, this provided us with a range of reporting practices and styles (as observed during the manual coding process). Moreover, by randomly sampling radiology reports over a 5-year period, we could also capture variations through time, if any, in reporting practices by the radiologists. For these reasons, we are confident that our SVM models are relatively robust to variations in reporting styles.

Despite these strengths, several limitations of this study should be acknowledged. First, the SVM models were trained and tested on a single source of data: electronic narrative radiology reports. It is thus possible that their accuracy would have been different if additional sources of clinical data had been included in the models. However, at the time of this study, there was no integrated EHR available at the MUHC, and discharge summaries were still in paper format. As a consequence, no other source of electronic clinical narratives was available to us. While the adoption of integrated EHRs is improving, only a small minority of institutions currently use them. As such, the approach taken in this study probably represents what could be realistically implemented in most hospitals across the USA and Canada. Second, we have focused on a single condition—VTEs. As such, we have not tested our proposed approach on other types of adverse events. Nonetheless, we have provided evidence that SVM models can accurately detect the two clinical manifestations of VTE (i.e., DVTs and PEs) with similar prediction performances. While we believe that the approach taken in this study can be generalized to other types of adverse events, this has yet to be demonstrated.

CONCLUSION

We found that SVM models based on narrative radiology reports can identify VTEs with high accuracy. The SVM models developed and validated in this study have many potential applications. Accurate identification of VTEs could assist hospital quality improvement staff monitor VTE rates, and evaluate the effectiveness of preventive interventions. Future studies should assess if the approach used in this study can be generalized to the detection of other types of adverse events, and if the combination of symbolic and statistical NLP techniques results in higher prediction performances than either method used alone.

CONTRIBUTORS

All listed authors: have made substantial contributions to conception and design, or acquisition of data, or analysis and interpretation of data; have been involved in drafting the manuscript or revising it critically for important intellectual content; and have given final approval of the version to be published.

FUNDING

This study was funded by the Canadian Institutes of Health Research (CIHR).

COMPETING INTERESTS

None.

ETHICS APPROVAL

McGill University Health Centre Research Ethics Committee approved this study.

PROVENANCE AND PEER REVIEW

Not commissioned; externally peer reviewed.