Results: Automated extraction required significantly less time (<1 day) than manual extraction (∼225 person-hours). Collection of demographic data (age, sex, diagnosis) was highly accurate and concordant with both methods (96-100%). Accuracy (for either extraction approach) and concordance were lower for unstructured data elements in EHR, such as performance status, date of diagnosis, and smoking status (NLP accuracy: 88-94%; Manual accuracy: 78-94%; Concordance: 71-82%). Concurrent medications (86-100%) and comorbid conditions (96-100%), were reported with high accuracy and concordance. Treatment details were also accurately captured with both methods (84-100%) and highly concordant (83-99%). Detection of whether biomarker testing was performed was highly accurate and concordant (96-98%), although detection of biomarker test results was more variable (accuracy 84-100%, concordance 84-99%). Features with syntactic or semantic variation requiring clinical interpretation were extracted with slightly lower accuracy by both NLP and manual review. For example, metastatic sites were more accurately identified through NLP extraction (NLP: 88-99%; Manual: 71-100%; Concordance: 70-99%) with the exception of lung and lymph node metastases (NLP: 66-71%; Manual: 87-92%; Concordance: 58%) due to analogous terms used in radiology reports not being included in the gold standard definition.
Conclusion: Automated data abstraction from EHR is highly accurate and faster than manual abstraction. Key challenges include poorly structured EHR and use of analogous terms beyond the gold standard definition. Application of NLP can facilitate real-world evidence studies at a greater scale than could be achieved with manual data extraction.
Real-world data describe patient health and experiences outside of a structured clinical trial setting. As patients receive medical care, large quantities of healthcare data are generated through the maintenance of health records, which has been accelerated by the widespread adoption of electronic health record (EHR) systems over the past decade. The current gold standard of generating real-world data from structured and unstructured EHR fields is manual data abstraction. While this approach has proven effective, there are drawbacks, such as being time-consuming, labour-intensive, and expensive, making it an arduous process that is highly susceptible to human error. These drawbacks often limit the scale and scope of real-world evidence studies.
To overcome these barriers, natural language processing (NLP) has been explored as an alternate method of data extraction from health records [1, 2]. NLP-based data extraction can provide results more rapidly and on a larger scale than could be achieved through manual extraction. However, uncertainty remains surrounding the validity of NLP-based extraction results especially in the context of free-text or dictated clinical notes [3-6]. Recently, the commercially available artificial intelligence (AI) engine, DARWEN™, was evaluated against manual extraction of EHR data from a tuberculosis clinic, successfully extracting data from free-format clinical notes. The AI NLP method generated rapid results that were also accurate . Extracted features were grouped to evaluate their accuracy based on linguistic and clinical complexity into groups of “simple”, “moderate”, and “complex” variables. To answer clinical questions, however, it is important to be able to investigate each of these features individually or grouped based on the research question at hand. To this end, we compared NLP-based extraction with manual data extraction of clinical features from EHRs of patients with advanced lung cancer at a feature level.
A cohort of 1,209 patients diagnosed with advanced lung cancer (documented as stage IIIB or stage IV at diagnosis) was identified through an institutional cancer registry. DARWEN™ identified a subset of patients who were diagnosed and treated at Princess Margaret Cancer Centre (PM) between January 2015 and December 2017, allowing for a minimum of 2 years’ follow up. The resulting study cohort consisted of 333 adult patients with advanced lung cancer who had received any systemic treatment at PM during this time (Figure 1). DARWEN™ extracted data from EHRs of these patients between their dates of diagnosis until March 30, 2019. This study was conducted in alignment with the University Health Network (UHN) Research Ethics Board (REB)-approved protocol. As a retrospective review of patient records, individual patient consent was waived.
Natural Language Processing Approach
Pentavere’s commercially available AI engine, DARWEN™, was used for NLP-based data extraction [7-10]. This AI engine combines linguistic (lexical, syntactic, and semantic) rules-based algorithms, machine learning models, and neural networks to extract relevant data from structured and unstructured EHR fields. DARWEN™’s capabilities have been previously described in detail . Significant innovations since then include the use of transformer-based models for classification and named entity recognition, as well as new techniques to facilitate and accelerate model training with low volumes of training data.
Establishing Ground Truths
All feature definitions and the ground truth were developed and modified through an iterative process whereby initial definitions were established in partnership with an expert clinical team from PM. These definitions were then manually tested using a subset of data to identify any discrepancies between the definitions and the actual text, which were resolved with further input from the clinical team. This process allowed for multiple points of clinical input and resulted in a comprehensive final set of definitions that captured clinically relevant language for each feature (Table S1).
Training and Fine-Tuning Algorithms
DARWEN™’s algorithms were pre-trained on other datasets and were then fine-tuned using a subset of patients from the present cohort of 333 patients. Algorithms were tuned based on the feature definitions until accuracy, precision (positive predictive value), recall (sensitivity), and F1-score targets were achieved. Accuracy measured the overall effectiveness of the NLP algorithm by calculating the ratio of correctly predicted outputs as a proportion of the total. F1 score is the harmonic mean of precision and recall, and was used to evaluate the performance of the algorithm (𝐹1𝑠𝑐𝑜𝑟𝑒 = 2 ∙ 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 ∙ 𝑟𝑒𝑐𝑎𝑙𝑙 / 𝑝𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛+ 𝑟𝑒𝑐𝑎𝑙𝑙 ). Algorithm precision and recall were balanced on a case-by-case basis, favoring precision over recall, when necessary, as dictated by research needs. Once the stability of these targets was confirmed on other subsets of unseen data from the study cohort, the algorithms were run on the entire study cohort (n=333) and the results were independently validated as described below.
Independent Manual Validation
The independent manual validation of these algorithms included 100 randomly selected patient records from the study cohort, which were not used for algorithm training or fine-tuning. Two trained manual abstractors from PM used the final set of feature definitions to extract clinical features from these same 100 patients. A third expert adjudicator reviewed the EHRs in the case of disagreements between abstractors or discrepancies between NLP and manual methods.
Data extracted (described further below and in Table S1) included patient demographics, smoking status, date of diagnosis, Eastern Cooperative Oncology Group (ECOG) performance status closest to date of diagnosis, tumour histology, biomarker testing and results, comorbid conditions, number and location of metastases, types of systemic therapy, and line of therapy (grouped by first line or any line), and concomitant immunosuppressive medication. DARWEN™ extracted patient demographics, including date of birth and sex, from unstructured clinical notes. Date of birth and date of diagnosis, extracted from pathology reports, were used to calculate age at diagnosis. All mentions of smoking status were extracted from unstructured clinical notes and contextualized to date of diagnosis. ECOG was extracted longitudinally, with corresponding time stamps, from the date of diagnosis to the end of the study period. For the analyses included in this study, the ECOG status documented the closest to each patient’s date of diagnosis was used. All tumour histologies identified as lung-related within pathology reports or unstructured clinical notes were extracted, grouped into adenocarcinoma, large cell, non-small cell, small cell, or squamous carcinomas (based on AJCC 8th Edition). All documented biomarker tests and results were extracted from pathology reports or clinical notes, specifically for ALK, BRAF, EGFR, KRAS, PD-L1, and ROS1. PD-L1 results were extracted based on explicit mention of positive or negative findings (e.g., “patient is PD-L1 positive”) as well as the tumour proportion score (TPS) (PD-L1 < 1%, 1-49%, and ≥ 50%); both features were not always present in the EHRs simultaneously. Any diagnosis or positive history of the comorbid conditions of interest were extracted from clinical notes (see Table S2). Any mention of metastases from the predefined anatomical locations were extracted from radiology reports at any time point after date of diagnosis (Table 3). Systemic therapies were extracted from clinical notes at any time point after date of diagnosis and were grouped into chemotherapy, immunotherapy, or targeted therapy. First line therapy was identified as the first treatment(s) a patient received after diagnosis. For analyses, line of therapy was categorized as first line or any line. Immunosuppressive medications specified as being of interest by the study investigators were extracted from the “current medications” section of the clinical notes (Table S3).
The results of NLP and manual data extraction were compared for accuracy against the expert adjudicator’s final response. The concordance rate was calculated as the percentage of agreement between the two extraction methods. Where applicable, sensitivity and specificity were calculated for both methods.
Prior to extracting data through either method, clinical definitions were developed and used to train both manual abstractors and NLP algorithms. Once trained, NLP-based data extraction for this study cohort (n=333) took less than 1 day, while manual data extraction (n=100) took approximately 225 hours.
Of the 100 patients included in the validation, 54% were male, 66% had adenocarcinoma, 70% had an ECOG of 0-1 at diagnosis, and 35% were non-smokers. All 100 patients received systemic therapy, with 59 on chemotherapy, 36 on targeted therapy, and 6 on immunotherapy in the first line setting (Table 1).
Demographics and Disease Characteristics
In general, patient demographics were reported with high accuracy and concordance across extraction methods as expected, given these elements were captured with low linguistic complexity (Table 1). Age at diagnosis and sex were extracted from unstructured sources with high accuracy by NLP (100% for both features) and manual extraction methods (99% and 100%, respectively). NLP and manually extracted data were 99% concordant for age and 100% concordant for sex (Figure 2).
Many disease characteristics were reported with high accuracy by NLP extraction as they were described using a limited number of terms. Histologic subtype was 98% accurate across methods and was highly concordant between NLP and manual extraction (96%). NLP was more accurate than manual extraction for date of diagnosis (94% vs. 83%) and ECOG performance status (93% vs. 78%) with concordance between methods of 77% and 71%, respectively.
The dynamic nature of some features adds further complexity. For example, smoking status can change over time with strict definitions of ex-smokers and non-smokers. However, detail in clinical notes may not accurately categorize patient smoking status. Given these challenges, manual extraction was slightly more accurate than NLP for smoking status (94% vs. 88%) with concordance of 82% between methods (Table 1).
A total of 16 comorbidities were investigated, 11 of which were found, extracted, and validated from the study cohorts’ EHRs (Table S2). Comorbidities are reported in EHRs in a less content-rich and more straight-forward manner. Synonymous terms for each comorbidity are incorporated into feature definitions and are therefore captured highly accurately by NLP. Comorbidities were reported with 96-100% accuracy for both extraction methods with concordance ranging from 93-100%. NLP extraction of
more frequent comorbidities was more sensitive than manual extraction (50-100% vs. 20-100%, respectively; Table S2). Specificity was more similar between methods (97-100% vs. 99-100%, respectively). In the case of less frequent comorbidities (i.e., an occurrence of ≤1), specificity was 100% and sensitivity ranged from 0-100% for both NLP and manual extraction.
Detailed cancer treatment information in clinical notes was extracted. These treatments were then expressed as Boolean variables capturing whether chemotherapy, immunotherapy, or targeted therapy were received as first line therapy or were ever received throughout the course of the patient’s treatment. By expressing therapeutic information in such a manner, the variability and complexity surrounding documentation of cancer treatments in clinical notes was mitigated. Type of first line treatment (chemotherapy, immunotherapy, and targeted therapy) was extracted by both methods with high accuracy (NLP: 95-99%; manual: 96-100%) and concordance (92-99%; Table 1). Sensitivity and specificity were high for both methods of extraction, where sensitivity ranged from 95% to 100% (NLP: 97-100%; manual: 95-100%) and specificity ranged from 93% to 100%
(NLP: 93-100%; manual: 100%). Treatments received at any line were also evaluated (Table 1). Both NLP and manual methods performed well when extracting chemotherapy (n=69) or immunotherapy (n=12) received at any point (94-98% accuracy across methods; 88-96% concordance; 92-94% sensitivity; 94-99% specificity). However, for patients receiving targeted therapy ever (n=40), manual extraction either missed or incorrectly reported 16 cases, resulting in lower accuracy (84% vs. 99%, respectively) and specificity (73% and 100%, respectively) than NLP-based extraction, with 83% concordance.
The study cohort was screened for 11 different immunosuppressive medications, 7 of which were received by patients (Table S3). While all patients received systemic therapy (n=100), only 66 patients received concurrent immunosuppressive treatment. Dexamethasone (n=54) and prednisone (n=7) were the most frequently administered immunosuppressive treatments, with the remaining 5 medications each only prescribed to one patient. When screening for the use of dexamethasone, outputs between extraction methods were 76% concordant. Manual extraction of dexamethasone was more accurate (96% vs. 80%) than NLP extraction, but similar specificity was observed across methods (100% and 97.8%, respectively). Manual extraction was also more sensitive (92.6% vs. 64.8%) than NLP extraction of dexamethasone. This is likely due to the inferred use of dexamethasone as part of chemotherapy treatment protocols, despite a lack of explicit mention of inclusion within the medical records. As NLP did not have this clinical insight as part of the feature definition, this contributed to missed cases. Prednisone data was reported with 90% concordance and was more accurately detected by NLP than by manual extraction (100% vs. 90%). NLP extraction of prednisone data was also more sensitive (100% vs. 57.1%), and more specific (100% vs. 92.3%) than manual extraction. While
cyclosporine, eculizumab, hydrocortisone, hydroxychloroquine, and methotrexate were taken by one patient each, it is worth noting that manual extraction of hydrocortisone data resulted in 14 false positive results (NLP specificity 100% vs. manual 85.9%).
EHRs were screened for gene and protein alterations often observed in lung cancer patients. These biomarker reports are content-rich, and the report structure can vary both between test types as well as over time. However, whether a biomarker was tested is documented relatively clearly and consistently in the clinical records. NLP detected whether biomarker testing was performed with 98-99% accuracy for ALK (n=71), BRAF (n=19), EGFR (n=72), KRAS (n=19), PD-L1 (n=29), and ROS-1 (n=4; Table 2). Concordance between the methods ranged from 96-98% across biomarkers. NLP extraction across all biomarkers for whether testing was performed resulted in high sensitivity (94.7-98.6%, except for ROS-1 with 50%) and high specificity (96.6-100%). Manual extraction reported biomarker testing with 97-100% accuracy and was highly sensitive (89.5-100%) and specific (96.5-100%).
Compared to whether a test has been performed, the biomarker results may be recorded in multiple locations within a pathology report, adding a source of variability to the extraction of these data. NLP extraction of biomarker test results was highly accurate for ALK, BRAF, EGFR, KRAS, and ROS-1 (95-100%) and was slightly less accurate for PD-L1 status (86%); accordingly, concordance between methods varied across biomarkers (86-100%). Biomarker status for ALK (n=8), EGFR (n=29), and ROS-1
(n=1) was reported with high sensitivity (NLP and manual: 100%) and specificity (NLP: 98-100%; manual: 98-100%) across both extraction methods (Figure 3A). Similarly, PD-L1 results were reported with high sensitivity (n=20; NLP: 94%; manual: 100%) but with varying specificity (NLP: 73%; manual: 100%). Both BRAF (n=1) and KRAS (n=3) status were reported with low sensitivity (BRAF: 0%; KRAS: 67%) and high specificity (BRAF and KRAS:100%) by NLP extraction. Manual extraction of these same features was highly sensitive (BRAF and KRAS:100%) and specific (BRAF: 100%; KRAS: 95%). With few patients testing
positive for BRAF, KRAS, or ROS-1, there is expected variability in the sensitivity of these extracted data.
Metastatic sites were detected with varying concordance (70-99%) between NLP and manual data extraction methods (Table 3). NLP extraction was more accurate than manual extraction for the detection of adrenal (96% vs. 77%), brain (99% vs. 71%), and bone (95% vs. 81%) metastases. NLP-based extraction less accurately detected metastases in the lymph node (66% vs. 92%) and lung (71% vs. 87%) compared to manual extraction. For all other metastatic sites, NLP and manual data extraction were comparably accurate: abdominal (88% vs. 86%), liver (96% vs. 95%), pericardium (99% vs. 100%), renal (99% vs. 99%), and spleen (99% vs. 97%). While NLP-extracted data were reported with high specificity (97-100%), sensitivity varied widely (33-100%; Figure 3B). Similarly, manually extracted data was more specific (69-100%) than sensitive (10-100%). Metastases are sometimes reported vaguely in radiology reports, with findings frequently being reported as being suspicious (and all the various ways of saying this) but not confirmed. As such, it can be difficult to identify from a passage of text alone whether a mass is explicitly considered to be metastatic. Clinical interpretation by manual abstractors can increase the accuracy of some extracted features but can also present an opportunity for incorrect interpretation of the text. In this study, clinical judgement exercised by manual abstractors when reviewing metastases resulted in low sensitivity. While sensitivity of metastases extracted by NLP also varied widely, NLP was able to more consistently capture reported metastases based solely on the established definitions.
This study demonstrates the validity of a commercially available NLP tool to extract feature-level data from the EHRs of patients with advanced lung cancer. Many previous studies either grouped features based on clinical and linguistic complexity, or extracted a single feature from clinical documentation[6, 11-13]. This study implemented DARWEN™ to extract clinical features through an automated NLP-based method. These features were validated against a manually extracted dataset compiled by two extractors and reviewed by an expert adjudicator with extensive clinical knowledge. The results of NLP-based data extraction were largely comparable to those of the expert manual extraction team, with a few exceptions where NLP outperformed manual review, or, conversely, was challenged by features requiring clinical interpretation. Sensitivity, specificity, accuracy, and concordance of both extraction methods were evaluated for all extracted features, however from a clinical perspective, accuracy and concordance are more important. Regardless of methodology, extracting this data from EHRs is critical for real-world evidence studies and is also necessary for identifying patient subgroups for respective analyses; NLP-based extraction achieves this more rapidly and at a larger scale than could be accomplished with manual review alone.
Despite a single set of feature definitions used across both methods of data extraction, there is opportunity for interpretation from the set definitions by manual reviewers, leading to variability in extracted results. In some cases, this benefits manual review, as clinical judgement outside of the established feature definitions can be used to identify cases not explicitly documented in the EHR. NLP-based extraction, however, will identify features based on how they are described in the established feature definitions and explicitly captured in clinical notes. As certain metastatic sites are reported with richer syntactic and semantic variation in clinical notes, these features have slightly lower accuracy by both NLP and manual extraction. Specifically, NLP extracted lymph node metastases less accurately than manual review due to analogous terms used in radiology reports not included in the feature definitions. Similarly, it is often difficult to determine whether a lung mass is metastatic, resulting in unclear documentation within imaging reports. Here, clinical judgement allowed the manual reviewer to identify lymph node or lung metastases that were not explicitly documented as metastases. Our iterative process used to define features attempts to account for this complexity found across clinical documents, but clinical documentation is often not explicit and varies significantly in content and quality.
Beyond linguistic complexities and unclear documentation, some clinical characteristics rely on knowledge-based inference more than others. For example, dexamethasone was extracted more accurately by manual review than NLP due to clinical knowledge that many chemotherapy regimens include dexamethasone without explicit mention of this in the EHR. This unique characteristic of dexamethasone administration as part of chemotherapy was not incorporated into the feature
definitions for either manual or NLP review. However, manual reviewers with clinical knowledge naturally deviated from the definition to identify cases where dexamethasone was administered based on concomitant therapies. Another feature requiring clinical interpretation was PD-L1 immunohistochemistry results. During the study timeframe, PD-L1 testing was a relatively newer addition to routine biomarker testing in advanced lung cancer patients, with rapidly evolving guidelines defining criteria for positive or negative PD-L1 status. In 2015, at the beginning of this study period, optimal immunohistochemistry cut-offs were uncertain, and it was unclear which patients would benefit from anti-PD-L1 agents. Subsequent studies introduced various cut-offs for PD-L1 expression that would determine whether a patient was labelled as “positive” for PD-L1, ranging from a TPS of >1% to ≥50%[15, 16]. More recently it has been suggested that both PD-L1 positive and negative patients may benefit from therapies targeting PD-L1. Given the evolution of PD-L1 threshold requirements, the way these results have been reported in the EHR has shifted over time. To reflect this, two features were developed for PD-L1 in this study: explicit mention of “positive” or “negative” for PD-L1, and TPS (<1%, 1-49%, and ≥50%) of PD-L1. These two features were not always simultaneously recorded, and where PD-L1 status was not explicitly documented, DARWEN™ did not infer positivity or negativity based on TPS alone. This resulted in slightly lower accuracy and specificity of NLP-extracted PD-L1 results when compared to manual extraction, which was supplemented by clinical interpretation.
Dynamic variables are also a challenge to capture accurately over time. For example, accurate capture of smoking status goes beyond identifying the terms “smoker” or “non-smoker” in a patient’s record. The specific definition of smoker status used in this study requires that a “former smoker” has quit for at least one year prior to their date of diagnosis. This, in turn, requires not only identifying the smoking status as above, but also determining whether the patient stopped smoking and when. These ideas are often fragmented across multiple notes throughout the patient record, and may be repeated inconsistently or stated imprecisely, with only approximate relative time (e.g., “patient is a smoker who quit about 3 to 4 years ago”). Compound error is the consequence of this fragmentation, imprecision, and inconsistency; error accumulates at multiple levels, creating a messy “picture” of the patient’s true smoking status.
This study has several limitations, including some inherent to the structure of EHRs as well as the content captured in these documents. EHRs as a source of real-world clinical data include both structured and unstructured fields. Unstructured notes provide clinicians with the opportunity to record clinical information using their preferred language, which can vary widely over time. These unstructured notes can contain semi-structured fields, which are formatted to capture clinical data with relatively low
syntactic variability. On the other hand, unstructured fields that are unformatted can result in linguistic variability, presenting a challenge to manual and NLP-based extraction alike. As this study only includes EHRs of advanced lung cancer patients from a single cancer centre in Canada, it may not be representative of national or global EHR documentation, necessitating varying degrees of tuning for different cohorts of patients. The algorithms evaluated in this study were applied to another hospital
site in Alberta, Canada and achieved comparable results after fine-tuning. Sensitivity and specificity were variable across rare biomarkers in this cohort, emphasizing the value of larger sample sizes for training and implementing NLP and potential benefit of purposefully selected validation cohorts. Finally, despite our iterative process of developing, testing, and modifying feature definitions with input from clinical experts at each stage, unanticipated language was encountered in some patient records. In rare circumstances (e.g., lung and lymph node metastases), this led to relatively lower accuracy and sensitivity for extraction by NLP compared to manual extractors, who could exercise clinical judgement to interpret as they reviewed patient records. Where possible, subsequent work should translate this clinical judgement into additional feature definition requirements to improve NLP accuracy. However, clinical judgement can be subjective, and clinicians may disagree. Regardless, NLP extracted data in a
consistent, objective, and accurate manner and at a much faster and larger scale than can be achieved manually.
NLP-based data extraction from structured and unstructured fields of EHRs is highly accurate and
produces results faster than manual methods. Key challenges remain, including inconsistently structured
EHRs and the use of complex, variable, and vague terms to describe clinical information. Despite these
challenges, the use of NLP offers a practical alternative to traditional manual extraction, enabling real-
world evidence studies at a larger scale than ever before.