Distinguishing cardiac catheter ablation energy modalities by applying natural language processing to electronic health records

This was a retrospective analysis of existing de-identified EHR data and an IRB committee waived the requirement for informed consent. For this analysis two datasets were provided by Optum. The first dataset is an Optum EHR dataset. This EHR dataset is payor agnostic but relies on the facility to be an Optum-EHR platform contributing clinic. Patient's data availability is limited to only visits that occurred at Optum contributing sites though historical documentation is sometimes available. The regional breakdown of the contributing hospitals are as follows, 46% are Midwest region 12% Northeast region, 30% Southern region, 8% Western region, and 3% with region unknown. The EHR dataset is 43% female, 57% male, and 0.05% unknown sex. The second dataset is an Optum generated NLP table. This table was generated with the aim of categorizing cardiac ablation patients and therefore the NLP table inclusion criteria was set to only patients with cardiac ablations. Optum generated the NLP table at the request of Medtronic. The NLP table is also payor agnostic and relies on the patient being in the Optum EHR dataset. Medtronic has an on-going contract with Optum to utilize this database. Standard SQL coding using Cloudera-ODBC Driver for Impala was used for this analysis. Data were also transferred to a SAS 9.4 server for further analytical support.

NLP's purpose is to extract key information from EHR records. Optum's NLP system was developed using vocabulary from the Unified Medical Language System (UMLS) that includes multiple medical dictionaries. These include the Logical Observation Identifiers Names and Codes (LOINC), the Systemized Nomenclature of Medicine-Clinical Terms (SNOMED-CT) and RxNorm terms within these dictionaries are distinguished as terms of interest within the NLP System (tokens) so that the presence of these terms in the free text will result in their inclusion in the semi-structured data. Development and validation of NLP programming involves manual review of a sample of narrative texts to identify appropriate targets for the NLP process, and to confirm the correct output. The algorithms require complex computer programming to search text fields and recognize syntax and formatting in EHRs to convert facts into a structured format. The NLP System uses an architecture based on pipelines and standoff annotations as described in the Unstructured Information Management Architecture (UIMA) The specific stages of the pipeline are proprietary to Optum but are like other UIMA based systems described by Savova et al. [11]. The primary objective of this analysis was to specifically evaluate pulmonary vein isolation (PVI) for patients with atrial fibrillation (AF). Additionally, patients who presented with atrial flutter (AFl) unrelated to pulmonary vein triggers during their ablation procedure and required a supplementary SVT ablation were categorized as a distinct group [12,13]. These individuals are referred to as ‘SVT add-on ablation’ patients. The hypothesis was that, if initially treated with cryoablation, these patients would subsequently need a second radiofrequency catheter ablation to address their AFl, potentially constituting a substantial portion of the broader ‘ablation’ cohort.

The NLP process does not utilize medical codes for its patient inclusion criteria therefore it was not possible to create these PVI/SVT distinctions or sub-categories of procedure types with the NLP table alone. To both validate the NLP table and create PVI/SVT patient cohorts, an internal validation table was created using CPT codes, ICD-10-CM diagnostic codes, and ICD-10-PCS codes (Supplementary Tables 1–3) from the overarching EHR data. Coding definitions of PVI for AF ablation and SVT add on ablation can be found in Supplementary Tables 1–3. This internal validation table was then linked to the Optum NLP table on available data variables (Supplementary Table 4). Patients from the NLP table were linked to the internal validations table by an encrypted patient identifier and their NLP note date must have occurred within three days post their ablation procedure date. Specific healthcare facilities are not identifiable in this dataset therefore it was not possible to trace the procedures back to a single patient (Table 1).

To analyze patients' index (their first) cardiac ablation procedures, the internal validation table was limited to index ablation procedures only and any accompanying SVT add-on procedures. Due to the structure of Optum EHR data, it was possible a patient's index ablation procedure was not available in the data because it was performed at a non-Optum contributing hospital prior to their inclusion in the Optum dataset. Unlike claims data, EHR data do not have continuous enrollment information to capture a patient's enrollment status. To better assess if the index procedure identified in this study was truly Index, we analyzed patients with a history of an NLP note term greater than 30 days prior to their indicated index procedure.

If a patient had an NLP note term greater than 30 days prior to their indicated index procedure, it is likely the patient had their first ablation procedure at a non-Optum clinic, entered the Optum database on a later date and their NLP terms were referencing historical documentation. For this reason, patients with note terms >30 days prior to their index procedure were excluded for not meeting index criteria. A sensitivity analysis was performed assessing patients with note terms >60 days prior to their index procedure and no meaningful differences were found (Supplementary Data). It is possible for a single procedure to have more than one note term associated. All note terms per procedure were included in this analysis.

Of the remaining patients, a frequency analysis was conducted to determine the NLP note terms associated with their index ablation procedure. The three-note term possibilities were (1) ‘ablation’, (2) ‘cryoablation’, (3) ‘radiofrequency ablation’. Since terms were distinct, repetitions of the same terms were not possible – creating a maximum amount of three terms per procedure. Note, term combinations were then further categorized via the method below (Table 2). Term combinations from all index procedures were summed to categorize the frequency of note terms.

To analyze all ablation procedures, the internal validation table was not limited to index procedures. Therefore, patients could contribute multiple ablation procedures. Note, SVT ablation for atrial flutter is defined as an add-on procedure and must be accompanied by a PVI for AF procedure indicated on the same day. Sequential procedures required at minimum a PVI for AF code but could include the SVT atrial flutter add on codes if indicated.

It was possible for more than one note term to be associated with a procedure. All note terms associated with their respective procedure were included. It was also possible a patient could have sequential ablation procedures within three days of each other. Frequency analysis was performed on procedures that occurred within three days of each other to determine if the note terms associated with both procedures were the same. Out of the procedures that occurred within three days of each other, 100% of terms were categorized as the same terms, and therefore only the initial procedure and note terms were retained. Any procedures that occurred outside of a three-day window were included as sequential procedures. The maximum number of ablation procedures a single patient had was four procedures. The same process used to categorize the index procedures was used to categorize frequency of note terms per procedure in the multiple analysis (Table 2).

Data availability was limited to the number of index ablation procedures that could be identified and linked to a note term(s). b. Of the 40,810 patients identified with medical codes only 13% (n = 5350) had linked note terms within three days after their procedure date (Table 1). The patient baseline characteristics are displayed in Table 3. Furthermore, another 1573 patients were excluded for likely having an index ablation procedure at a non-Optum contributing hospital (Figure 1). The remaining analytical sample size for the index cohort was then 9% (n = 3777) (Figure 1). This 9% represents the proportion of cardiac ablation procedures identified via medical codes who had available note terms in the NLP table. This indicates many procedures do not utilize note documentation.

Of the 3,777 index ablation procedures with available NLP note terms, 22% (n = 844) had an ‘ablation’ term associated, 27% (n = 1016) had a cryoablation’ term associated, 49% (n = 1855) had a ‘radiofrequency’ term associated, and 2% (n = 62) had both ‘cryoablation’ and ‘radiofrequency’ terms associated (Figure 2).

Of the 844 ablation terms, 45.3% (n = 382) were categorized as SVT add-on procedures. Of the 1,016 cryoablation terms, 25.5% (n = 255) were categorized as SVT add-on procedures, and of the 1855 radiofrequency terms, 41.5% (n = 769) were categorized as SVT add-on procedures (Figure 2).

For the multiple procedures analysis, patients were allowed to contribute more than one ablation procedure. Of the 40,810 cardiac ablations identified with medical codes,14% (n = 5691) had note terms available within three days after a procedure date (Figures 3 & 4) in the NLP table. The slight increase in patient count with available note terms is likely due to a patient's index ablation not meeting inclusion criteria, but a later procedure date did meet inclusion criteria and had associated terms found in the NLP table.

The final multiple procedures with available notes in the NLP table had 5435 patients and 5691 ablation procedures. 4.5% (n = 242) of all patients had two ablation procedures, 0.22% (n = 12) had three ablation procedures and 0.04% (n = 2) had four ablation procedures (Figure 3). Of the 5691 ablation procedures with available NLP not terms, 24% (n = 1362) had an ‘ablation’ term associated, 27% (n = 1516) had a ‘cryoablation’ term associated, 47% (n = 2691) had a ‘radiofrequency’ term associated, and 2% (n = 122) had both ‘cryoablation’ and ‘radiofrequency’ terms associated (Figure 4).

Of the 1362 ablation terms, 46% (n = 626) were categorized as SVT add-on procedures. Of the 1516 cryoablation terms, 23.2%% (n = 352) were categorized as SVT add-on procedures, of the 2691 radiofrequency terms, 43% (n = 1147) were categorized as SVT add-on procedures, and of the 122 both terms, 4.8% (n = 59) were categorized as SVT add-on procedures (Figure 4).

Our study has several limitations. First, it is limited to only hospitals with an Optum contributing platform. A large portion of these hospitals (46%) are in the Midwest and a very small portion 93%) in the Western region. Potential differences in regional locations should be considered as a limitation of this analysis. Moreover, only 9% (n = 3777) of cardiac ablation procedures identified via medical codes had available note terms in the NLP table. This limited our small sample size and indicated many procedures do not utilize documentation. It is possible there are meaningful differences in patients with and without documentation. No additional analysis was undertaken to consider these potential differences and future work should seek to understand these differences. The small sample size may also limit the results from being extrapolated beyond this analysis.

Second, NLP methods do not utilize medical codes and therefore as Optum created the NLP table a second-stage internal validation was required. As previously stated, the internal validation table relied on ICD-10-CM diagnostic and procedural codes. This validation was time consuming and limited the use of the NLP table to only those patients that had a matched link thus, limiting generalizability. Third, there were limitations surrounding medical codes used. No inpatient PVI specific ablation code existed, limiting the specificity. Moreover, EHR and NLP data are inherently limited by their completeness and quality [5,8,9]. With varying provider practices, if specific words, terms, or notes are not regularly input into the patient's record, these data will be absent from the NLP table as seen in these analyses [5,8,9].