Neurologic Outcomes of Carotid and Other Emergent Interventions for Ischemic Stroke over Six Years with Analysis Enhanced by Machine Learning
Philip A Rivera, Bethany Jennings, Jeffrey Burton, Aaron Hayson, Faith Mason, Jaron Pettis, Adam Berenson, Sam Money, Waldemar C Sternbergh, III, Daniel Fort, Hernan A Bazan
Ochsner Health, New Orleans, LA
Background: Despite continued evolution in treatment, stroke represents one of the most common and debilitating diseases suffered. Approximately 795,000 strokes occur annually in the United States, with 85% ischemic in origin. Tissue plasminogen activator (tPA, thrombolysis) along with other urgent stroke interventions, such as carotid endarterectomy (uCEA), carotid artery stenting (uCAS), and mechanical endovascular thrombectomy/reperfusion (MER), are the current mainstays of treatment for ischemic stroke. However, scarce data describe the presenting stroke severity and neurologic outcomes for these acute stroke interventions. We created a novel machine-learning natural language processing (NLP) algorithm to build upon data collected from a Comprehensive Stroke Center database. This method enhanced our ability to determine neurological outcomes for all urgent stroke interventions accurately. Here, we demonstrate stroke severity and functional neurologic outcomes for all ischemic stroke patients undergoing: 1) uCEA/uCAS, 2) tPA alone, 3) MER alone, 4) tPA + MER, and 5) no intervention over a six-and-a-half-year period.
Methods: A master database was created by first importing data from a prospectively gathered Comprehensive Stroke Center dataset encompassing telemedicine transfers from over 50 referring hospitals and urgent care centers in three states. This was synthesized with data collected from an electronic medical record data warehouse. Records gathered from 10,975 patient encounters from January 1, 2015, through July 31, 2021, included stroke etiologies, presenting stroke severity (National Institutes of Health Stroke Scale, NIHSS), interventions for stroke performed, patient demographics and comorbidities, medication use, and length of stay (LOS). The presenting stroke severity was determined by vascular/stroke neurologists using the NIHSS. An NIHSS score of <4 represents a ‘minor' stroke, 5 - 15 a ‘moderate' stroke, 15 - 20 a ‘moderate-severe' stroke, and 21 - 42 a ‘severe' stroke.
Functional neurological outcomes were ascertained using the discharge modified Rankin scale (mRS) score, a clinical outcome measure that quantifies the degree of neurologic disability. This scale ranges from functional independence (mRS 0 - 2) to severe neurologic disability with complete functional dependence (mRS= 5) and death (mRS= 6). An mRS < 3 is regarded as a good outcome with functional independence, whereas an mRS ≥ 3 will require increasing levels of assistance with activities of daily living.However, mRS values were only available for 3,627 encounters in the original dataset. Hence, to increase the number of patients with a neurologic outcome measurement, an NLP algorithm was created to review physician discharge notes and physical/occupational therapy documentation to measure and assign an mRS value. All data was then combined into the master database for outcomes analysis.
We began by using a simple text-scraping method to extract a gain of 645 mRS values from discharge documentation by the vascular neurology service. To train our NLP models, we exported and reconstructed physicians' and therapists' documentation from an electronic medical record (EMR) into plain text for all patients in the master database. Utilizing a custom-built, secure, online portal, several physicians performed manual classification of features applicable to identifying outcomes data. These NLP models were used to identify documentation useful in predicting a discharge mRS. An additional NLP model was then created and validated by analyzing 116,614 documents for those patients with existing mRS assignments available, resulting in a gain of 1,082 mRS values. A text-scraping algorithm was also used to improve the accuracy of presenting pre-procedure NIHSS. The incoming NIHSS was used for patients in the no intervention cohort. Next, 5,805 encounters were excluded due to either lack of endpoint data, documentation, or hemorrhagic etiology of stroke. A total of 5,170 ischemic stroke encounters were included in the final statistical analysis. Encounters were then divided into 5 cohorts: uCEA/uCAS, (n=189), tPA only (n=1,053), MER (n=418), tPA followed by MER (n=199), and no intervention (n=3,311).
P values were calculated using appropriate statistical tests and are reported along with summary statistics in the Table. Logistic regression was used to investigate the association between presenting stroke severity and functional independence (mRS <3), and the analysis was stratified by interventional cohort. The resulting odds ratios (OR) represent odds of functional independence within each cohort for patients with different levels of stroke severity.
Results: Patient demographics, presenting stroke severity (NIHSS), and discharge functional neurologic outcomes (mRS) are shown in the Table. Thrombolysis prior to an uCEA/uCAS occurred in 21% of patients in the uCEA/uCAS cohort (39/189); 3.2% (6/189) underwent MER followed by uCEA/uCAS; and 1.6% (3/189) were treated with combined tPA, MER, and uCEA/uCAS. The 30-day stroke, death, and myocardial infarction (MI) for the uCEA/uCAS cohort were 2.6% (5/189), 2.6 % (5/189), and 4.2 % (8/189), respectively. Four strokes occurred post-uCEAs (2.9% 4/139), and 1 stroke occurred in a patient who underwent tPA followed by CEA (3.9% 1/26). There was only one stroke in the uCAS group (2%; this patient had not received tPA). Patients undergoing an uCEA/uCAS following an ischemic stroke had a higher proportion of peripheral arterial disease (P = .005) and active tobacco use (P < .001) compared to patients in the other cohorts. The mean time to intervention for uCEA/uCAS following an ischemic stroke was 3.2 days (0-14).
The LOS for patients undergoing an uCEA/uCAS was 7.2 days (SD 5.9). LOS was significantly longer for those in the MER (10.2d) and tPA + MER (8.8d) cohorts compared to the other cohorts (P <.001). The 30-day stroke, death, and MI rates were, respectively: tPA only cohort, 2.2%, 3.7% and 5.2%; MER cohort, 1.2%, 12.2% and 7.4%; tPA + MER 3.5%, 4.5% and 8.0%; and in the no intervention cohort, 2.3%, 5.3% and 4.6%.
Presenting NIHSS was two- to three-fold higher in patients undergoing MER and tPA + MER than all other interventions (P <.001; Figure, panel A). The MER and tPA + MER cohorts presented with a mean NIHSS of 15.9 (SD 7.5) and 16.0 (SD 6.8), respectively. In contrast, patients undergoing uCEA/uCAS presented with a mean NIHSS of 4.9 (SD 5.3). Neurologic outcomes were best in uCEA/CAS cohort (mRS=1.7), followed by the tPA only (mRS=1.8), and no intervention (mRS=1.8) cohorts (Figure, panel B). Patients undergoing MER (mRS=2.6) and tPA + MER (mRS=2.3) following an ischemic stroke had a slightly worse functional independence (P <.001).
We ascertained the odds ratio (OR) and 95% confidence intervals (CI) for discharge with functional independence (mRS < 3) at various presenting stroke severities of NIHSS ≤4 vs. >4; ≤10 vs. > 10, and ≤15 vs. >15 using logistic regression (Table). Patients in the uCEA/uCAS cohort are over three times more likely to maintain functional independence post-procedure when they present with an NIHSS of ≤ 10 (OR 3.11 [95% CI 1.3 - 7.3]). Patients in the tPA only cohort also maintained significant functional independence up to an NIHSS of 15 (OR 3.13 [95% CI 2.2 - 4.4]). There are overall lower rates of functional independence seen in the MER and tPA + MER cohorts across all strata.
Conclusion: We describe the presenting stroke severity (NIHSS) and neurologic outcomes (mRS) of uCEA/uCAS and other ischemic stroke treatments in a large stroke referral center over six and a half years. This study is the largest report of presenting stroke severity and neurologic outcomes that combines all stroke treatment cohorts. Patients who undergo uCEA/uCAS following an ischemic stroke have largely favorable outcomes. In contrast, MER and MER + tPA have poorer functional outcomes, likely related to their high presenting stroke severity.
Despite tissue plasminogen activator approval for treatment of ischemic stroke in 1996, current utilization of tPA for ischemic stroke in the U.S. is 2 - 10%, though up to 25% are eligible. As a result of the coordinated stroke treatment, a thrombolysis rate > 20% was achieved for patients undergoing uCEA/uCAS, which is two to four-fold higher than nationally reported thrombolysis rates. Notably, tPA followed by uCEA/uCAS did not have an increased risk of post-procedure complications. Presenting stroke severity was the lowest for patients undergoing uCEA/uCAS (mean NIHSS of 4.9), and these patients have a high degree of functional independence at discharge when presenting with an NIHSS of 0-10.
The tPA only cohort presented mainly with minor to moderate stroke severity and displayed strong odds (OR, 2.99) of functional independence. Likewise, a high degree of functional independence is achieved in the no intervention cohort (OR, 6.91) for those presenting with NIHSS 0-4. MER is increasingly being offered in those presenting with higher stroke severity with the hopes of improved functional outcomes. There is a small but significant benefit to tPA administration before MER over MER alone for moderate stroke severity. We could not demonstrate significance in odds of functional outcome for any strata of presenting NIHSS for the tPA + MER cohort.
Using natural language processing, we enhanced and validated a highly granular prospectively collected stroke database and by combining a procedural data, and natural language processing, we exponentially increased the yield of usable data, making the data processing faster to generate neurologic outcomes. Though this technique has been used for other types of medical documentation review, this is the first report of its use in predicting mRS to aid in assessing neurologic outcomes.
Back to 2022 Abstracts