Natural Language Processing Enhances Prediction of Functional Outcome After Acute Ischemic Stroke.

Background Conventional prognostic scores usually require predefined clinical variables to predict outcome. The advancement of natural language processing has made it feasible to derive meaning from unstructured data. We aimed to test whether using unstructured text in electronic health records can improve the prediction of functional outcome after acute ischemic stroke. Methods and Results Patients hospitalized for acute ischemic stroke were identified from 2 hospital stroke registries (3847 and 2668 patients, respectively). Prediction models developed using the first cohort were externally validated using the second cohort, and vice versa. Free text in the history of present illness and computed tomography reports was used to build machine learning models using natural language processing to predict poor functional outcome at 90 days poststroke. Four conventional prognostic models were used as baseline models. The area under the receiver operating characteristic curves of the model using history of present illness in the internal and external validation sets were 0.820 and 0.792, respectively, which were comparable to the National Institutes of Health Stroke Scale score (0.811 and 0.807). The model using computed tomography reports achieved area under the receiver operating characteristic curves of 0.758 and 0.658. Adding information from clinical text significantly improved the predictive performance of each baseline model in terms of area under the receiver operating characteristic curves, net reclassification improvement, and integrated discrimination improvement indices (all P<0.001). Swapping the study cohorts led to similar results. Conclusions By using natural language processing, unstructured text in electronic health records can provide an alternative tool for stroke prognostication, and even enhance the performance of existing prognostic scores.

acute ischemic stroke

Acute Stroke Registry and Analysis of Lausanne

area under the receiver operating characteristic curve

bidirectional encoder representations from transformers

bag‐of‐words

Chia‐Yi Christian Hospital

electronic health record

history of present illness

integrated discrimination improvement

machine learning

National Institutes of Health Stroke Scale

net reclassification improvement

National Taiwan University Hospital

preadmission comorbidities, level of consciousness, age, and neurological deficit

Clinical Perspective

What Is New?

Using natural language processing, it is feasible to develop machine learning models for predicting functional outcome after acute ischemic stroke based on unstructured clinical text stored in electronic health records.

Machine learning models using deep learning techniques outperformed those based on the traditional “bag‐of‐words” text representation model.

The machine learning model based on the “history of present illness” in the admission note performed nearly as well as the National Institutes of Health Stroke Scale score and achieved an adequate discriminatory ability in both within‐site and across‐site validations.

What Are the Clinical Implications?

The developed machine learning models could not only provide an alternative method of stroke prognostication but also could enhance the predictive performance of conventional risk models.

The process of model development can be reproduced in individual hospitals to build customized versions of similar prognostic models.

Stroke is a common disabling neurologic condition. About one quarter of adults aged ≥25 years will develop a stroke in their lifetime. Even though the acute treatment of strokes has advanced substantially, more than half of patients who have had strokes still have poor outcomes such as permanent functional dependence or even death. Thus far, several prognostic risk models have been developed to predict functional outcomes following an acute stroke. Most of them use similar input variables to make predictions, such as age, initial stroke severity, and comorbidities. While most of the risk models were validated to have reasonable prognostic accuracy, they still are not widely adopted into clinical practice, probably because of implementation issues. Hence, having a readily available digital tool that provides automated prognostication is beneficial for clinical decision‐making and resource allocation.

Artificial intelligence–aided prediction has been introduced to improve diagnostic precision and streamline clinical decision‐making. With the advances in machine learning (ML) and deep learning, it has become feasible to integrate various types of structured data for the data‐driven prediction of clinically meaningful outcomes in patients with stroke. , Furthermore, by using natural language processing (NLP) to extract hidden but valuable information stored in textual data, it is possible to automate the detection of acute ischemic stroke (AIS) or the classification of stroke subtypes from neuroimaging reports, , , and even improve the prognostication of patients with critical illness using clinical notes. ,

Supposedly, all hospitalized patients with stroke would have their corresponding admission note, in which the history of present illness (HPI) comprises the most essential textual data regarding the clinical features of the index stroke event. Furthermore, patients with stroke would also undergo baseline neuroimaging, especially a head computed tomography (CT) scan. The textual component of the CT report may disclose relevant information about the extent of cerebrovascular diseases. Considering the points above, we aimed to develop and validate ML models to investigate whether unstructured clinical text in the HPI and CT report can improve the prediction of functional outcome at an early stage after AIS.

METHODS

Data Source

Data that support the study findings are available from the corresponding author on reasonable request. The Ditmanson Medical Foundation Chia‐Yi Christian Hospital (CYCH) is a 1000‐bed teaching hospital in southern Taiwan. The National Taiwan University Hospital (NTUH) is a university‐affiliated medical center with a capacity of >2000 beds in northern Taiwan. The study protocol was independently approved by the CYCH Institutional Review Board (CYCH‐IRB No. 2020090) and Research Ethics Committee B of NTUH (202104028RINB). Study data were maintained with confidentiality to ensure the privacy of all participants.

The stroke centers of both hospitals have maintained their stroke registries since 2007 and 1995, respectively. The stroke registries prospectively registered all cases of stroke by daily screening of all patients receiving head CT or those with a diagnosis of stroke at the emergency department or during hospitalization, as well as screening for a diagnosis at discharge using the International Classification of Diseases, Ninth and Tenth Revisions (ICD ‐9 and ICD‐ 10 revisions with clinical modification) codes. Data regarding the demographics, cause, risk factor profiles, intervention, and outcomes of patients with stroke were collected. Stroke severity was assessed using the National Institutes of Health Stroke Scale (NIHSS) and functional outcome was assessed by the modified Rankin Scale.

Study Design

We followed the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis guidelines for the methods and reporting of prediction models. All consecutive adult patients with first‐ever AIS admitted to CYCH from October 2007 to December 2019 and those admitted to NTUH from January 2016 to December 2020 were identified using the institutional stroke registries. Patients who had an in‐hospital stroke or whose clinical data included a missing admission NIHSS score were excluded. Those who did not provide consent for follow‐up or who were lost to follow‐up at 90 days were also eliminated. For each patient, we retrieved the HPI upon admission and the initial CT report from the electronic health record (EHR) database. Both types of documents were written in English in the study hospitals. Patients whose EHRs were unavailable were excluded.

The CYCH cohort (Cohort A in Figure 1) were randomly split into a derivation set that consisted of 75% of the patients, and an internal validation set comprising the remaining 25% of the patients, who were withheld from all ML models during the training process. The NTUH cohort (Cohort B in Figure 1) comprised the external validation set. In addition, to test the generalizability of the ML approach, we did another experiment where the NTUH cohort was used for derivation and internal validation and the CYCH cohort was used for external validation.

Figure 1

Process of model development and validation.

AIS indicates acute ischemic stroke; and ML, machine learning.

Outcome Variable

The outcome variable was a poor functional outcome at 90 days poststroke, which was defined as a modified Rankin Scale score of 3 to 6.

Baseline Risk Models

We used 4 prognostic models that used clinical variables available upon admission as the baseline risk models for comparison. The first risk model was the NIHSS score because stroke severity is the most important determinant for poststroke functional outcome. The second model, consisting of age and NIHSS score within the first 6 hours of the onset of AIS, was useful in predicting 3‐month mortality and functional outcome. The third model was preadmission comorbidities, level of consciousness, age, and neurologic deficit (PLAN) score, which was developed to predict 30‐day and 1‐year mortality and a modified Rankin Scale score of 5 to 6 at discharge. In the PLAN score, preadmission comorbidities refer to preadmission dependence, cancer, congestive heart failure, and atrial fibrillation, whereas neurologic focal deficits indicate weakness of the leg or arm, aphasia, and neglect. The fourth model was derived using a cohort of patients from the Acute Stroke Registry and Analysis of Lausanne (ASTRAL). The ASTRAL score, comprising age, NIHSS score, time from stroke onset to admission, range of visual fields, acute glucose level, and level of consciousness, was designed to predict 3‐month unfavorable outcome (modified Rankin Scale >2) poststroke upon hospital admission.

NLP and ML Models

Figure 1 illustrates the process of model development and validation. ML models were trained separately using HPI (model HPI) and CT reports (model CT). Experiments were done with 2 approaches of text representation: a simple “bag‐of‐words” (BOW) approach and a more sophisticated deep learning approach using the bidirectional encoder representations from transformers (BERT).

Free text was preprocessed as follows: (1) words were spell‐checked, and misspelled words were automatically corrected using the Jazzy spell checker (https://github.com/kinow/jazzy); (2) acronyms and abbreviations were expanded to their full forms by looking up a list of common clinical acronyms and abbreviations used locally; (3) non‐ASCII characters and nonword special characters were deleted; (4) words were converted to lowercase; (5) words were lemmatized to their root forms; and (6) stop words were removed. Only step 1 through step 3 were needed for the BERT approach.

In the BOW approach, we built a document‐term matrix, where each column stood for a unique word from the text corpus, the rows represented each document (HPI or CT report for each patient), and the cells indicated whether each word appeared within each document. The random forest algorithm was used to build the ML models. After fitting the ML models, the permutation importance method was used to assess the impact of each word on the prediction results, thus aiding in interpreting the ML models.

Clinical notes are generally lengthy, and their words are usually dependent on each other. However, the BOW approach does not consider the sequence of words and cannot capture the meaning of words in their context. Even conventional word‐embedding methods such as Word2Vec disregard the long‐range dependency and may not fully capture clinical meaning from clinical notes. Therefore, the BERT approach was used for representation of clinical notes. BERT is a deep neural network–based contextualized word‐embedding model that is pre‐trained using bidirectional transformers based on masked language modeling and next sentence prediction. The original BERT is pre‐trained using the general BooksCorpus and English Wikipedia corpus.

For the current study, a clinical domain‐specific BERT model (ie, ClinicalBERT) was used. ClinicalBERT is pre‐trained on the Medical Information Mart for Intensive Care III clinical notes, which is a large collection of clinical notes from >40 000 patients who stayed in the intensive care unit. The embeddings learned from the ClinicalBERT model were directly fed to a feed‐forward neural network with a softmax function for classification. The weights of the pre‐trained CliniclBERT model along with the classification layer were updated simultaneously during the fine‐tuning stage of the model. The preprocessed text was split into BERT tokens. Because the BERT model can only deal with 512 tokens and most of the documents contained <512 tokens (Table 1), input documents were truncated to 512 tokens to simplify the analysis.

Table 1

Characteristics of the Study Cohorts

	CYCH (n=3847)	NTUH (n=2668)	P value
Age, mean (SD)	69.5 (12.3)	69.8 (13.9)	0.528
Female	1583 (41.1)	1118 (41.9)	0.543
Hypertension	3098 (80.5)	2090 (78.3)	0.031
Diabetes	1602 (41.6)	1024 (38.4)	0.008
Hyperlipidemia	2195 (57.1)	1369 (51.3)	<0.001
Atrial fibrillation	684 (17.8)	790 (29.6)	<0.001
Congestive heart failure	196 (5.1)	223 (8.4)	<0.001
Cancer	249 (6.5)	424 (15.9)	<0.001
Preadmission dependence (mRS >2)	419 (10.9)	407 (15.3)	<0.001
Onset‐to‐admission delay >3 h	2763 (71.8)	1913 (71.7)	0.915
NIHSS, median (IQR)	5 (3–10)	5 (2–13)	0.267
Glucose, mean (SD), mg/dL	163.2 (82.6)	146.8 (67.7)	<0.001
PLAN score, median (IQR)	8 (6–12)	9 (7–12)	0.001
ASTRAL score, median (IQR)	21 (18–27)	22 (18–30)	0.178
Word count in HPI, median (IQR)	132 (109–161)	268 (209–342)	<0.001
BERT tokens in HPI, median (IQR)	192 (156–240)	420 (329–535)	<0.001
Word count in CT reports, median (IQR)	127 (93–189)	42 (34–52)	<0.001
BERT tokens in CT reports, median (IQR)	225 (164–351)	86 (68–106)	<0.001
Poor outcome (mRS >2)	1674 (43.5)	1118 (41.9)	0.196

Data are expressed in number (percentage) unless specified otherwise. ASTRAL indicates Acute Stroke Registry and Analysis of Lausanne; BERT, bidirectional encoder representations from transformers; CT, computed tomography; CYCH, Chia‐Yi Christian Hospital; HPI, history of present illness; IQR, interquartile range; mRS, modified Rankin Scale; NIHSS, National Institutes of Health Stroke Scale; NTUH, National Taiwan University Hospital; and PLAN, preadmission comorbidities, level of consciousness, age, and neurological deficit.

Only data from the derivation set were used in the training (fine‐tuning) process, which was implemented using Python 3.8.5 on a Windows 10 system with graphics processing unit. Hyperparameter optimization was performed using 10‐fold cross‐validation. The number of trees for the random forest classifier was varied from 10 to 200 with an increment of 10. During the fine‐tuning of the ClinicalBERT model for the prediction task, the batch size of the neural network was set at 16. The learning rate of the Adam optimizer was varied from 1×10−5 to 5×10−5 with an increment of 1×10−5, and the number of epochs from 2 to 4 with an increment of 1. Model error was minimized in terms of the area under the receiver operating characteristic curve (AUC). Once the optimal hyperparameters were determined, the ML models were fitted with the full derivation set.

Statistical Analysis

Categorical variables were expressed as counts (percentages) while continuous variables were expressed as means (SDs) or medians (interquartile ranges). Differences between 2 groups were tested by χ2 tests for categorical variables and t tests or Mann–Whitney U tests for continuous variables, as appropriate.

Model performance was evaluated on the internal and external validation sets separately. For each patient in the validation sets, the probability of a poor functional outcome was estimated using the ML models. Model discrimination was assessed with AUCs. The AUCs between the BOW and BERT approaches were compared using the DeLong method. The approach that resulted in higher AUCs was used in the following analysis.

For each baseline model, a logistic regression model was fitted by entering the risk score as a continuous variable. To construct a “text‐only” risk model, named model HPI+CT, a logistic regression model was fitted by entering the probabilities of a poor functional outcome predicted separately by model HPI and model CT as continuous variables. To assess the incremental value of adding information from clinical text to the baseline models, the probabilities of a poor functional outcome predicted by model HPI and model CT were introduced to the logistic regression model as continuous variables to construct “text‐enhanced” risk models. Model discrimination was assessed with AUCs. Model calibration was evaluated by visual inspection of the calibration plot, which depicts the observed risk versus the predicted risk. In addition, the added predictive ability of clinical text was evaluated by calculating the continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) indices. , Unlike categorical NRI, the continuous NRI does not require established risk categories. It quantifies upward and downward changes in the predicted probabilities of an event. The IDI is equivalent to the difference in discrimination slopes, which measure the difference between mean predicted probabilities of an event for those with events and the corresponding mean for those without events. , Higher values of NRI and IDI indicate superior discrimination.

All statistical analyses were performed using Stata 15.1 (StataCorp, College Station, TX) and R version 4.0.5 (R Foundation for Statistical Computing, Vienna, Austria). The calibration plot and the analysis of NRI and IDI were performed using R package “PredictABEL”. Two‐tailed P values of 0.05 were considered significant.

RESULTS

A total of 3847 eligible patients with AIS were recruited from CYCH. The derivation and internal validation sets consisted of 2885 and 962 patients, respectively. The external validation set comprised 2668 patients with AIS from NTUH. The characteristics of the study cohorts are listed in Table 1. The CYCH and NTUH cohorts were similar in age, sex, onset‐to‐admission delay, NIHSS, ASTRAL score, and the proportion of a poor functional outcome. However, they significantly differed in the prevalence of comorbidities and preadmission dependence, glucose level, PLAN score, and word counts in the HPI and CT reports.

In the internal validation set, the AUC of model HPI of the BERT approach was not significantly different from that of the BOW approach (0.820 versus 0.802, P=0.111), whereas model CT of the BERT approach achieved a higher AUC than that of the BOW approach (0.758 versus 685, P<0.001). Model HPI+CT of the BERT approach yielded a higher AUC than that of the BOW approach (0.840 versus 0.819, P=0.042). In the external validation set, model HPI of the BERT approach had a significantly higher AUC than that of the BOW approach (0.792 versus 0.761, P<0.001), whereas the AUCs of model CT of both approaches were similar (0.658 versus 662, P=0.690). Model HPI+CT of the BERT approach achieved a higher AUC than that of the BOW approach (0.798 versus 0.778, P=0.009). Therefore, ML models built by the BERT approach were used in the following analysis. Figure S1 shows the top 20 most influential words for model HPI and model CT.

Figure 2 shows the AUCs of the baseline and ML models. In the internal validation set, model HPI+CT yielded an AUC of 0.840, which was comparable to those of NIHSS (0.811, P=0.062), age and NIHSS (0.841, P=0.935), PLAN score (0.837, P=0.830), and ASTRAL score (0.840, P=0.995). Model HPI achieved an AUC of 0.820, which was comparable to those of NIHSS (P=0.581), age and NIHSS (P=0.111), PLAN score (P=0.194), and ASTRAL score (P=0.134). Model CT yielded an AUC of 0.758, which was significantly lower than those of NIHSS (P=0.005) and the other 3 baseline models (all P<0.001). In the external validation set, the AUCs of model HPI+CT, model HPI, and model CT were 0.798, 0.792, and 0.658, respectively. Model HPI+CT and model HPI performed equally well with NIHSS (0.807, P=0.359 and 0.134, respectively), but were inferior to the other 3 baseline models (all P<0.001). Model CT had a significantly lower AUC than the 4 baseline models (all P<0.001).

Figure 2

Receiver operating characteristic curves for predicting a poor functional outcome in the internal (A) and external (B) validation sets.

ASTRAL indicates Acute Stroke Registry and Analysis of Lausanne; AUC, area under the receiver operating characteristic curve; CT, computed tomography; HPI, history of present illness; NIHSS, National Institutes of Health Stroke Scale; and PLAN, preadmission comorbidities, level of consciousness, age, and neurological deficit.

Table 2 lists the comparison of performance between the baseline and text‐enhanced risk models. The AUCs of the text‐enhanced models were significantly higher than those of the baseline models in both the internal and external validation sets (all P<0.001). The NRI and IDI indices also indicated a statistically significant improvement (all P<0.001) in predictive performance when the baseline models were enhanced by the information from the clinical text. The calibration plots of the baseline and text‐enhanced models are displayed in Figure 3. It shows that the text‐enhanced models were generally well calibrated over the entire risk range because all points lie close to the 45‐degree line.

Table 2

Comparison of the Predictive Ability of Baseline Models With or Without Adding Information From Clinical Text

	Baseline AUC (95% CI)	Text‐enhanced AUC (95% CI)	P value	NRI (95% CI)	P value	IDI (95% CI)	P value
Internal validation
NIHSS	0.811 (0.783–0.839)	0.869 (0.846–0.891)	<0.001	0.766 (0.648–0.884)	<0.001	0.109 (0.089–0.129)	<0.001
Age and NIHSS	0.841 (0.815–0.866)	0.872 (0.850–0.895)	<0.001	0.514 (0.391–0.637)	<0.001	0.065 (0.049–0.080)	<0.001
PLAN score	0.837 (0.811–0.863)	0.870 (0.847–0.893)	<0.001	0.593 (0.471–0.715)	<0.001	0.061 (0.046–0.077)	<0.001
ASTRAL score	0.840 (0.814–0.866)	0.871 (0.849–0.894)	<0.001	0.527 (0.405–0.650)	<0.001	0.070 (0.054–0.086)	<0.001
External validation
NIHSS	0.807 (0.790–0.823)	0.843 (0.828–0.858)	<0.001	0.719 (0.648–0.791)	<0.001	0.089 (0.078–0.100)	<0.001
Age and NIHSS	0.838 (0.823–0.853)	0.854 (0.840–0.868)	<0.001	0.556 (0.482–0.630)	<0.001	0.043 (0.035–0.052)	<0.001
PLAN score	0.834 (0.818–0.849)	0.852 (0.838–0.867)	<0.001	0.561 (0.488–0.635)	<0.001	0.045 (0.037–0.054)	<0.001
ASTRAL score	0.839 (0.824–0.854)	0.854 (0.840–0.868)	<0.001	0.572 (0.499–0.646)	<0.001	0.052 (0.043–0.061)	<0.001

ASTRAL indicates Acute Stroke Registry and Analysis of Lausanne; AUC, area under the receiver operating characteristic curve; IDI, integrated discrimination improvement; NIHSS, National Institutes of Health Stroke Scale; NRI, net reclassification improvement; and PLAN, preadmission comorbidities, level of consciousness, age, and neurological deficit.

Figure 3

Calibration plots of the baseline and text‐enhanced models.

ASTRAL indicates Acute Stroke Registry and Analysis of Lausanne; NIHSS, National Institutes of Health Stroke Scale; and PLAN, preadmission comorbidities, level of consciousness, age, and neurological deficit.

In the additional experiment where the 2 study cohorts were exchanged, similar results were obtained. Figure S2 shows the AUCs of the baseline and ML models. In the internal validation cohort, model HPI+CT achieved an AUC of 0.818, which was comparable to those of NIHSS (0.815, P=0.867), age and NIHSS (0.842, P=0.115), PLAN score (0.837, P=0.214), and ASTRAL score (0.847, P=0.056). Model HPI yielded an AUC of 0.818, which was comparable to those of NIHSS (0.815, P=0.889), age and NIHSS (0.842, P=0.134), PLAN score (0.837, P=0.234), and ASTRAL score (0.847, P=0.065). Model CT achieved an AUC of 0.674, which was significantly lower than the 4 baseline models (all P<0.001). In the external validation set, the AUCs of model HPI+CT, model HPI, and model CT were 0.778, 0.772, and 0.662, respectively. However, all of the 3 models had significantly lower AUCs than the 4 baseline models (all P<0.001). The calibration plots of the baseline and text‐enhanced models are displayed in Figure S3. Table S1 gives the comparison of performance between the baseline and text‐enhanced risk models. Significant improvements in AUCs, NRI, and IDI indices (all P<0.001) were observed in the text‐enhanced models versus the baseline models.

DISCUSSION

This study demonstrates that ML models trained from clinical text could not only provide an alternative method of stroke prognostication but also enhance the predictive performance of conventional risk models in terms of the prediction of 90‐day functional outcome. In general, the deep learning NLP approach (BERT) outperformed the simpler NLP approach (BOW paired with random forest classifiers). Based on the BERT approach, the text‐only model based on HPI and CT reports and the model based on HPI alone both achieved an adequate discriminatory ability in within‐site and across‐site validations and they performed nearly as well as the NIHSS score. Moreover, the text‐enhanced risk models demonstrated a considerably higher discriminatory ability than the baseline risk models as well as acceptable model calibration.

The functional outcome of AIS is largely determined by stroke severity, which is closely related to the location and size of ischemic brain lesions. Therefore, the free‐form text in the HPI and CT reports, which may implicitly contain information regarding stroke severity, can be used to predict stroke outcomes. However, the predictive performance of model CT was worse than model HPI. One possible explanation might be that the initial unenhanced CT scan generally does not reflect the final extent of brain ischemia. In this regard, magnetic resonance imaging studies are more sensitive than CT studies for detection of acute ischemia, and magnetic resonance imaging reports seemed to be promising for predicting outcome after AIS. Nevertheless, magnetic resonance imaging is not as widespread and readily available for emergency situations as CT.

On the other hand, clinical text may contain much richer information than that captured by conventional stroke prognostic models. Most of the existing models tended to base their predictions on the same concepts, such as demographics, initial stroke severity, pre‐stroke functional status, and comorbidities, and thus shared a similar degree of prognostic accuracy. It might reasonably be expected that incorporating other less traditional factors, such as the degree of frailty, emotional wellness, strength of social support, or even the clinician’s clinical gestalt could improve the utility of prognostic models. The text in the HPI could complement such information and thereby enhance the predictive performance of the baseline models. Even though model HPI based on the BERT approach performed and generalized well across hospitals, such deep learning NLP models are often considered a “black box” model lacking interpretability. In situations where model interpretability is given a high priority, simpler NLP approaches such as the BOW approach may be reasonable alternatives despite their lower predictive ability. , Furthermore, influential features identified from free text might be collected and used to develop new prognostic models.

ML methods have been applied to develop models for prognostication of AIS. , , , , , , The ML models in the existing studies generally had a comparable or even higher discriminatory ability than conventional logistic regression models. , , One of the reasons may be that ML algorithms can handle potential nonlinear relationships and model complex interactions between variables. , , However, these ML models were seldom externally validated, undermining their utility in other populations or health care settings. Like any other diagnostic or prognostic tool, a ML model should be validated in an independent data set by assessing its discrimination and calibration. , In particular, in order to improve model performance, ML models usually undergo hyperparameter optimization where the hyperparameters are tuned on a tuning set independent of the validation set. , This study followed these recommendations by tuning the hyperparameters by cross‐validation within the derivation set and validating the ML models on both a holdout test set (within‐site validation) and a completely independent data set (across‐site validation).

Although ML methods are gaining popularity, textual data have rarely been analyzed or used in previous ML prognostic models of stroke. Despite this, NLP has been applied in the field of stroke medicine, such as building ML models to identify AIS , , or automating AIS subtype classification. , The merit of using textual data is that clinical notes are generated within EHRs in the process of medical care, thus saving the extra effort required for data collection and coding. Furthermore, the nuances of symptoms across patients are more likely to be preserved in unstructured textual data. The study findings supported the incremental value of unstructured clinical text over the conventional prognostic models.

With the routine use of EHRs in clinical practices, a large amount of health care data, either structured or unstructured, has not only accumulated rapidly but also has become more available for downstream use. Big data analytics is now increasingly used in diverse health care applications such as disease surveillance, health management, and clinical decision support. In addition, by directly drawing data from EHRs, conventional prognostic models can be integrated into the EHR system to provide automated outcome prediction. , Nevertheless, this approach would miss the opportunity to capture meaningful information embedded in clinical notes. By contrast, the methods used in this study harnessed both unstructured and structured data to generate prognostic models, which can be easily implemented as an electronic decision support tool to help health care professionals to establish a prognosis.

This study has some limitations that need to be addressed. First, although data‐driven ML methods have their own advantages, the relationships discovered from the data do not mean any causal inference, and prediction accuracy should not be interpreted in any way as causal validity. Second, the vocabulary and style used in clinical documentation may vary across hospitals and regions, thereby affecting the performance of the ClinicalBERT model, which was pre‐trained on clinical notes from a US hospital. A BERT model pre‐trained on clinical text from the local health system is likely to further improve the predictive performance of NLP models but requires a larger computational cost. Despite this, the study results showed that the developed models generalized well in the external validation group from a geographically distant hospital. Therefore, we believe that the influence by the variation in clinical documentation is not substantial. Furthermore, the process of model development can be reproduced in individual hospitals to build customized versions of similar prognostic models. Third, ML models based on features directly derived from neuroimaging data are promising for predicting various stroke outcomes. , Although this issue is out of the scope of the current study, future studies may explore the value of alternative sources of unstructured data such as imaging data in the prediction of poststroke functional outcome.

CONCLUSIONS

By using NLP and ML methods, information derived from clinical text has the potential to prognosticate patients with AIS. This study developed and validated text‐enhanced prognostic models to aid in the early prediction of functional outcome after AIS. However, further studies are needed to confirm the generalizability of this approach and the clinical usefulness in routine practice.

Sources of Funding

This research was funded by the Ditmanson Medical Foundation Chia‐Yi Christian Hospital (grant number R109‐37‐1). The funder of the research had no role in the design and conduct of the study, interpretation of the data, or decision to submit for publication.

Disclosures

None.

Table S1

Figures S1–S3

Click here for additional data file.