Application of Machine Learning Techniques to Development of Emergency Medical Rapid Triage Prediction Models in Acute Care.

Given the critical and complex features of medical emergencies, it is essential to develop models that enable prompt and suitable clinical decision-making based on considerable information. Emergency nurses are responsible for categorizing and prioritizing injuries and illnesses on the frontlines of the emergency room. This study aims to create an Emergency Medical Rapid Triage and Prediction Assistance model using electronic medical records and machine learning techniques. Patient information was retrieved from the emergency department of a large regional teaching hospital in Taiwan, and five supervised learning techniques were used to construct classification models for predicting critical outcomes. Of these models, the model using logistic regression had superior prediction performance, with an F1 score of 0.861 and an area under the receiver operating characteristic curve of 0.855. The Emergency Medical Rapid Triage and Prediction Assistance model demonstrated superior performance in predicting intensive care and hospitalization outcomes compared with the Taiwan Triage and Acuity Scale and three clinical early warning tools. The proposed model has the potential to assist emergency nurses in executing challenging triage assessments and emergency teams in treating critically ill patients promptly, leading to improved clinical care and efficient utilization of medical resources.

A Machine Learning-Based Fall Risk Assessment Model for Inpatients.

Falls are one of the most common accidents among inpatients and may result in extended hospitalization and increased medical costs. Constructing a highly accurate fall prediction model could effectively reduce the rate of patient falls, further reducing unnecessary medical costs and patient injury. This study applied data mining techniques on a hospital's electronic medical records database comprising a nursing information system to construct inpatient-fall-prediction models for use during various stages of inpatient care. The inpatient data were collected from 15 inpatient wards. To develop timely and effective fall prediction models for inpatients, we retrieved the data of multiple-time assessment variables at four points during hospitalization. This study used various supervised machine learning algorithms to build classification models. Four supervised learning and two classifier ensemble techniques were selected for model development. The results indicated that Bagging+RF classifiers yielded optimal prediction performance at all four points during hospitalization. This study suggests that nursing personnel should be aware of patients' risk factors based on comprehensive fall risk assessment and provide patients with individualized fall prevention interventions to reduce inpatient fall rates.

Two Decades of Research Using Taiwan's National Health Insurance Claims Data: Bibliometric and Text Mining Analysis on PubMed.

BACKGROUND: Studies using Taiwan's National Health Insurance (NHI) claims data have expanded rapidly both in quantity and quality during the first decade following the first study published in 2000. However, some of these studies were criticized for being merely data-dredging studies rather than hypothesis-driven. In addition, the use of claims data without the explicit authorization from individual patients has incurred litigation. OBJECTIVE: This study aimed to investigate whether the research output during the second decade after the release of the NHI claims database continues growing, to explore how the emergence of open access mega journals (OAMJs) and lawsuit against the use of this database affect the research topics and publication volume and to discuss the underlying reasons. METHODS: PubMed was used to locate publications based on NHI claims data between 1996 and 2017. Concept extraction using MetaMap was employed to mine research topics from article titles. Research trends were analyzed from various aspects, including publication amount, journals, research topics and types, and cooperation between authors. RESULTS: A total of 4473 articles were identified. A rapid growth in publications was witnessed from 2000 to 2015, followed by a plateau. Diabetes, stroke, and dementia were the top 3 most popular research topics whereas statin therapy, metformin, and Chinese herbal medicine were the most investigated interventions. Approximately one-third of the articles were published in open access journals. Studies with two or more medical conditions, but without any intervention, were the most common study type. Studies of this type tended to be contributed by prolific authors and published in OAMJs. CONCLUSIONS: The growth in publication volume during the second decade after the release of the NHI claims database was different from that during the first decade. OAMJs appeared to provide fertile soil for the rapid growth of research based on NHI claims data, in particular for those studies with two or medical conditions in the article title. A halt in the growth of publication volume was observed after the use of NHI claims data for research purposes had been restricted in response to legal controversy. More efforts are needed to improve the impact of knowledge gained from NHI claims data on medical decisions and policy making.

Constructing Inpatient Pressure Injury Prediction Models Using Machine Learning Techniques.

The incidence rate of pressure injury is a critical nursing quality indicator in clinic care; consequently, factors causing pressure injury are diverse and complex. The early prevention of pressure injury and monitoring of these complex high-risk factors are critical to reduce the patients' pain, prevent further surgical treatment, avoid prolonged hospital stay, decrease the risk of wound infection, and lower associated medical costs and expenses. Although a number of risk assessment scales of pressure injury have been adopted in various countries, their criteria are set for specific populations, which may not be suitable for the medical care systems of other countries. This study constructs three prediction models of inpatient pressure injury using machine learning techniques, including decision tree, logistic regression, and random forest. A total of 11 838 inpatient records were collected, and 30 sets of training samples were adopted for data analysis in the experiment. The experimental results and evaluations of the models suggest that the prediction model built using random forest had most favorable classification performance of 0.845. The critical risk factors for pressure injury identified in this study were skin integrity, systolic blood pressure, expression ability, capillary refill time, and level of consciousness.

Validation of a novel claims-based stroke severity index in patients with intracerebral hemorrhage.

BACKGROUND: Stroke severity is an important outcome predictor for intracerebral hemorrhage (ICH) but is typically unavailable in administrative claims data. We validated a claims-based stroke severity index (SSI) in patients with ICH in Taiwan. METHODS: Consecutive ICH patients from hospital-based stroke registries were linked with a nationwide claims database. Stroke severity, assessed using the National Institutes of Health Stroke Scale (NIHSS), and functional outcomes, assessed using the modified Rankin Scale (mRS), were obtained from the registries. The SSI was calculated based on billing codes in each patient's claims. We assessed two types of criterion-related validity (concurrent validity and predictive validity) by correlating the SSI with the NIHSS and the mRS. Logistic regression models with or without stroke severity as a continuous covariate were fitted to predict mortality at 3, 6, and 12 months. RESULTS: The concurrent validity of the SSI was established by its significant correlation with the admission NIHSS (r = 0.731; 95% confidence interval [CI], 0.705-0.755), and the predictive validity was verified by its significant correlations with the 3-month (r = 0.696; 95% CI, 0.665-0.724), 6-month (r = 0.685; 95% CI, 0.653-0.715) and 1-year (r = 0.664; 95% CI, 0.622-0.702) mRS. Mortality models with NIHSS had the highest area under the receiver operating characteristic curve, followed by models with SSI and models without any marker of stroke severity. CONCLUSIONS: The SSI appears to be a valid proxy for the NIHSS and an effective adjustment for stroke severity in studies of ICH outcome with administrative claims data.

Comorbidity study of borderline personality disorder: applying association rule mining to the Taiwan national health insurance research database.

BACKGROUND: Borderline personality disorder (BPD) is a complex clinical state with highly polymorphic symptoms and signs. Studies have demonstrated that people with a BPD diagnosis are likely to have numerous co-occurring psychiatric disorders and physical comorbidities. The aim of our study was to obtain further insight about the associations among comorbidities of BPD and to demonstrate the practicality of using association rule mining (ARM) technique in clinical databases. METHODS: A retrospective case-control study was conducted on information of 1460 patients (292 BPD patients and 1168 control patients) selected from the Taiwan National Health Insurance Research Database. Information on physical and psychiatric comorbidities, which were diagnosed within 3 years before and after enrollment, was collected. A logistic regression model was used to calculate the odds ratios of comorbidities between patients with and without BPD. ARM technique was used to study the associations of BPD and two or more psychiatric comorbidities. RESULTS: We classified physical comorbidities into 13 categories according to the International Classification of Diseases, Ninth Revision, Clinical Modification system, and the results indicated that the 12 categories were more common in the BPD patients than in the control patients (except congenital anomalies). However, psychiatric comorbidities, including depressive disorder, bipolar disorder, anxiety disorder, sleep disorder, substance use disorder, and mental retardation were more common in the BPD patients than in the control patients. Furthermore, the associations of BPD and two or more comorbidities were evaluated. CONCLUSION: Most physical and psychiatric disorders were more common in the BPD patients than in the control patients. Because the failure to remit from BPD is associated with suffering from chronic physical conditions and because psychiatric comorbidities may lead to delays in diagnosis of BPD, clinicians caring for people with BPD should be aware of possible comorbidities.

Applying the Temporal Abstraction Technique to the Prediction of Chronic Kidney Disease Progression.

Chronic kidney disease (CKD) has attracted considerable attention in the public health domain in recent years. Researchers have exerted considerable effort in attempting to identify critical factors that may affect the deterioration of CKD. In clinical practice, the physical conditions of CKD patients are regularly recorded. The data of CKD patients are recorded as a high-dimensional time-series. Therefore, how to analyze these time-series data for identifying the factors affecting CKD deterioration becomes an interesting topic. This study aims at developing prediction models for stage 4 CKD patients to determine whether their eGFR level decreased to less than 15 ml/min/1.73m2 (end-stage renal disease, ESRD) 6 months after collecting their final laboratory test information by evaluating time-related features. A total of 463 CKD patients collected from January 2004 to December 2013 at one of the biggest dialysis centers in southern Taiwan were included in the experimental evaluation. We integrated the temporal abstraction (TA) technique with data mining methods to develop CKD progression prediction models. Specifically, the TA technique was used to extract vital features (TA-related features) from high-dimensional time-series data, after which several data mining techniques, including C4.5, classification and regression tree (CART), support vector machine, and adaptive boosting (AdaBoost), were applied to develop CKD progression prediction models. The results revealed that incorporating temporal information into the prediction models increased the efficiency of the models. The AdaBoost+CART model exhibited the most accurate prediction among the constructed models (Accuracy: 0.662, Sensitivity: 0.620, Specificity: 0.704, and AUC: 0.715). A number of TA-related features were found to be associated with the deterioration of renal function. These features can provide further clinical information to explain the progression of CKD. TA-related features extracted by long-term tracking of changes in laboratory test values can enable early diagnosis of ESRD. The developed models using these features can facilitate medical personnel in making clinical decisions to provide appropriate diagnoses and improved care quality to patients with CKD.

A comparison of stroke severity proxy measures for claims data research: a population-based cohort study.

PURPOSE: Confounding by disease severity has been viewed as an intractable problem in claims-based studies. A novel 7-variable stroke severity index (SSI) was designed for estimating stroke severity by using claims data. This study compared the performance of mortality models with various proxy measures of stroke severity, including the SSI, in patients hospitalized for acute ischemic stroke (AIS). METHODS: Data from the Taiwan National Health Insurance Research Database (NHIRD) were analyzed. Three proxy measures of stroke severity were evaluated: Measure 1, the SSI; Measure 2, intensive care unit admission and length of stay; and Measure 3, surgical operation, mechanical ventilation, hemiplegia or hemiparesis, and residual neurological deficits. We performed logistic regression by including age, sex, vascular risk factors, Charlson comorbidity index, and one of the proxy measures as covariates to predict 30-day and 1-year mortality after AIS. Model discrimination was evaluated using the area under the receiver-operating characteristic curve (AUC). RESULTS: We identified 7551 adult patients with AIS. Models using the SSI (Measure 1) outperformed models using the other proxy measures in predicting 30-day mortality (AUC 0.892 vs 0.851, p < 0.001 for Measure 2; 0.892 vs 0.853, p < 0.001 for Measure 3) and 1-year mortality (AUC 0.816 vs 0.784, p < 0.001 for Measure 2; 0.816 vs 0.782, p < 0.001 for Measure 3). CONCLUSIONS: Using the SSI facilitated risk adjustment for stroke severity in mortality models for patients with AIS. The SSI is a viable methodological tool for stroke outcome studies using the NHIRD.

Validity of a stroke severity index for administrative claims data research: a retrospective cohort study.

BACKGROUND: Ascertaining stroke severity in claims data-based studies is difficult because clinical information is unavailable. We assessed the predictive validity of a claims-based stroke severity index (SSI) and determined whether it improves case-mix adjustment. METHODS: We analyzed patients with acute ischemic stroke (AIS) from hospital-based stroke registries linked with a nationwide claims database. We estimated the SSI according to patient claims data. Actual stroke severity measured with the National Institutes of Health Stroke Scale (NIHSS) and functional outcomes measured with the modified Rankin Scale (mRS) were retrieved from stroke registries. Predictive validity was tested by correlating SSI with mRS. Logistic regression models were used to predict mortality. RESULTS: The SSI correlated with mRS at 3 months (Spearman rho = 0.578; 95 % confidence interval [CI], 0.556-0.600), 6 months (rho = 0.551; 95 % CI, 0.528-0.574), and 1 year (rho = 0.532; 95 % CI 0.504-0.560). Mortality models with the SSI demonstrated superior discrimination to those without. The AUCs of models including the SSI and models with the NIHSS did not differ significantly. CONCLUSIONS: The SSI correlated with functional outcomes after AIS and improved the case-mix adjustment of mortality models. It can act as a valid proxy for stroke severity in claims data-based studies.

Exploring the impact of intravenous thrombolysis on length of stay for acute ischemic stroke: a retrospective cohort study.

BACKGROUND: Understanding the factors that influence the hospital length of stay (LOS) for patients with stroke will help in discharge planning and stroke unit management. We explored how intravenous thrombolysis (IVT) affects LOS in an acute-care hospital setting. METHODS: We analyzed adult patients with ischemic stroke who presented within 48 h of onset from a hospital-based stroke registry. The relationship between IVT and prolonged LOS (LOS ≥ 7 days) was studied by both multivariate logistic regression and the classification and regression tree (CART) analyses. RESULTS: Among the study population of 3054 patients, 1110 presented within 4.5 h. The median LOS (interquartile range) was 7 (4 to 11) days, and 1619 patients had prolonged LOS. Multivariate logistic regression revealed that IVT (odds ratio, 0.53; 95 % confidence interval 0.38-0.74) was an independent factor that reduced the risk of prolonged LOS, whereas age, National Institutes of Health Stroke Scale (NIHSS) score, diabetes mellitus, and leukocytosis at admission predicted prolonged LOS. CART analysis identified 4 variables (NIHSS score, IVT, leukocytosis at admission, and age) as important factors to partition the patients into six subgroups. The patient subgroup that had an NIHSS score of 5 to 7 and received IVT had the lowest probability (19 %) of prolonged LOS. CONCLUSIONS: IVT reduced the risk of prolonged LOS in patients with acute ischemic stroke. Measures to increase the rate of IVT are encouraged.

Predicting Factors and Risk Stratification for Return Visits to the Emergency Department Within 72 Hours in Pediatric Patients.

OBJECTIVES: A return visit (RV) to the emergency department (ED) is usually used as a quality indicator for EDs. A thorough comprehension of factors affecting RVs is beneficial to enhancing the quality of emergency care. We performed this study to identify pediatric patients at high risk of RVs using readily available characteristics during an ED visit. METHODS: We retrospectively collected data of pediatric patients visiting 6 branches of an urban hospital during 2007. Potential variables were analyzed using a multivariable logistic regression analysis to determine factors associated with RVs and a classification and regression tree technique to identify high-risk groups. RESULTS: Of the 35,435 visits from which patients were discharged home, 2291 (6.47%) visits incurred an RV within 72 hours. On multivariable analysis, younger age, weekday visits, diagnoses belonging to the category of symptoms, signs, and ill-defined conditions, and being seen by a female physician were associated with a higher probability of RVs. Children younger than 6.5 years who visited on weekdays or between midnight and 8:00 AM on weekends or holidays had the highest probability of returning to the ED within 72 hours. CONCLUSIONS: Our study reexamined several important factors that could affect RVs of pediatric patients to the ED and identified high-risk groups of RVs. Further intervention studies or qualitative research could be targeted on these at-risk groups.

Integrating Fuzzy Multiobjective Programming and System Dynamics to Develop an Approach for Talent Retention Policy Selection: Case on Health-Care Industry.

The demand for medical services has been increasing yearly in aging countries. Medical institutions must hire a large number of staff members to provide efficient and effective health-care services. Because of high workload and pressure, high turnover rates exist among health-care staff members, especially those in nonurban areas, which are characterized by limited resources and a predominance of elderly people. Turnover in health-care institutions is influenced by complex factors, and high turnover rates result in considerable direct and indirect costs for such institutions (Lo and Tseng 2019). Therefore, health-care institutions must adopt appropriate strategies for talent retention. Because institutions cannot determine the most effective talent retention strategy, many of them simply passively adopt a single human resource (HR) policy and make minor adjustments to the selected policy. In the present study, system dynamics modeling was combined with fuzzy multiobjective programming to develop a method for simulating HR planning systems and evaluating the suitability of different HR policies in an institution. We also considered the external insurance policy to be the parameter for the developed multiobjective decision-making model. The simulation results indicated that reducing the turnover rate of new employees in their trial period is the most effective policy for talent retention. The developed procedure is more efficient, effective, and cheaper than the traditional trial-and-error approaches for HR policy selection.

An analysis of Chinese nursing electronic medical records to predict violence in psychiatric inpatients using text mining and machine learning techniques.

BACKGROUND: The prevalence of violence in acute psychiatric wards is a critical concern. According to a meta-analysis investigating violence in psychiatric inpatient units, researchers estimated that approximately 17% of inpatients commit one or more acts of violence during their stay. Inpatient violence negatively affects health-care providers and patients and may contribute to high staff turnover. Therefore, predicting which psychiatric inpatients will commit violence is of considerable clinical significance. OBJECTIVE: The present study aimed to estimate the violence rate for psychiatric inpatients and establish a predictive model for violence in psychiatric inpatients. METHODS: We collected the structured and unstructured data from Chinese nursing electronic medical records (EMRs) for the violence prediction. The data was obtained from the psychiatry department of a regional hospital in southern Taiwan, covering the period between January 2008 and December 2018. Several text mining and machine learning techniques were employed to analyze the data. RESULTS: The results demonstrated that the rate of violence in psychiatric inpatients is 19.7%. The patients with violence in psychiatric wards were generally younger, had a more violent history, and were more likely to be unmarried. Furthermore, our study supported the feasibility of predicting aggressive incidents in psychiatric wards by using nursing EMRs and the proposed method can be incorporated into routine clinical practice to enable early prediction of inpatient violence. CONCLUSIONS: Our findings may provide clinicians with a new basis for judgment of the risk of violence in psychiatric wards.

Early Prediction of Functional Outcomes After Acute Ischemic Stroke Using Unstructured Clinical Text: Retrospective Cohort Study.

BACKGROUND: Several prognostic scores have been proposed to predict functional outcomes after an acute ischemic stroke (AIS). Most of these scores are based on structured information and have been used to develop prediction models via the logistic regression method. With the increased use of electronic health records and the progress in computational power, data-driven predictive modeling by using machine learning techniques is gaining popularity in clinical decision-making. OBJECTIVE: We aimed to investigate whether machine learning models created by using unstructured text could improve the prediction of functional outcomes at an early stage after AIS. METHODS: We identified all consecutive patients who were hospitalized for the first time for AIS from October 2007 to December 2019 by using a hospital stroke registry. The study population was randomly split into a training (n=2885) and test set (n=962). Free text in histories of present illness and computed tomography reports was transformed into input variables via natural language processing. Models were trained by using the extreme gradient boosting technique to predict a poor functional outcome at 90 days poststroke. Model performance on the test set was evaluated by using the area under the receiver operating characteristic curve (AUC). RESULTS: The AUCs of text-only models ranged from 0.768 to 0.807 and were comparable to that of the model using National Institutes of Health Stroke Scale (NIHSS) scores (0.811). Models using both patient age and text achieved AUCs of 0.823 and 0.825, which were similar to those of the model containing age and NIHSS scores (0.841); the model containing preadmission comorbidities, level of consciousness, age, and neurological deficit (PLAN) scores (0.837); and the model containing Acute Stroke Registry and Analysis of Lausanne (ASTRAL) scores (0.840). Adding variables from clinical text improved the predictive performance of the model containing age and NIHSS scores, the model containing PLAN scores, and the model containing ASTRAL scores (the AUC increased from 0.841 to 0.861, from 0.837 to 0.856, and from 0.840 to 0.860, respectively). CONCLUSIONS: Unstructured clinical text can be used to improve the performance of existing models for predicting poststroke functional outcomes. However, considering the different terminologies that are used across health systems, each individual health system may consider using the proposed methods to develop and validate its own models.

Predicting the Risk of Future Multiple Suicide Attempt among First-Time Suicide Attempters: Implications for Suicide Prevention Policy.

Suicide is listed in the top ten causes of death in Taiwan. Previous studies have pointed out that psychiatric patients having suicide attempts in their history are more likely to attempt suicide again than non-psychiatric patients. Therefore, how to predict the future multiple suicide attempts of psychiatric patients is an important issue of public health. Different from previous studies, we collect the psychiatric patients who have a suicide diagnosis in the National Health Insurance Research Database (NHIRD) as the study cohort. Study variables include psychiatric patients' characteristics, medical behavior characteristics, physician characteristics, and hospital characteristics. Three machine learning techniques, including decision tree (DT), support vector machine (SVM), and artificial neural network (ANN), are used to develop models for predicting the risk of future multiple suicide attempts. The Adaboost technique is further used to improve prediction performance in model development. The experimental results show that Adaboost+DT performs the best in predicting the behavior of multiple suicide attempts among psychiatric patients. The findings of this study can help clinical staffs to early identify high-risk patients and improve the effectiveness of suicide prevention.

Automated risk assessment of newly detected atrial fibrillation poststroke from electronic health record data using machine learning and natural language processing.

Background: Timely detection of atrial fibrillation (AF) after stroke is highly clinically relevant, aiding decisions on the optimal strategies for secondary prevention of stroke. In the context of limited medical resources, it is crucial to set the right priorities of extended heart rhythm monitoring by stratifying patients into different risk groups likely to have newly detected AF (NDAF). This study aimed to develop an electronic health record (EHR)-based machine learning model to assess the risk of NDAF in an early stage after stroke. Methods: Linked data between a hospital stroke registry and a deidentified research-based database including EHRs and administrative claims data was used. Demographic features, physiological measurements, routine laboratory results, and clinical free text were extracted from EHRs. The extreme gradient boosting algorithm was used to build the prediction model. The prediction performance was evaluated by the C-index and was compared to that of the AS5F and CHASE-LESS scores. Results: The study population consisted of a training set of 4,064 and a temporal test set of 1,492 patients. During a median follow-up of 10.2 months, the incidence rate of NDAF was 87.0 per 1,000 person-year in the test set. On the test set, the model based on both structured and unstructured data achieved a C-index of 0.840, which was significantly higher than those of the AS5F (0.779, p = 0.023) and CHASE-LESS (0.768, p = 0.005) scores. Conclusions: It is feasible to build a machine learning model to assess the risk of NDAF based on EHR data available at the time of hospital admission. Inclusion of information derived from clinical free text can significantly improve the model performance and may outperform risk scores developed using traditional statistical methods. Further studies are needed to assess the clinical usefulness of the prediction model.

Developing a stroke alert trigger for clinical decision support at emergency triage using machine learning.

BACKGROUND: Acute stroke is an urgent medical condition that requires immediate assessment and treatment. Prompt identification of patients with suspected stroke at emergency department (ED) triage followed by timely activation of code stroke systems is the key to successful management of stroke. While false negative detection of stroke may prevent patients from receiving optimal treatment, excessive false positive alarms will substantially burden stroke neurologists. This study aimed to develop a stroke-alert trigger to identify patients with suspected stroke at ED triage. METHODS: Patients who arrived at the ED within 12 h of symptom onset and were suspected of a stroke or transient ischemic attack or triaged with a stroke-related symptom were included. Clinical features at ED triage were collected, including the presenting complaint, triage level, self-reported medical history (hypertension, diabetes, hyperlipidemia, heart disease, and prior stroke), vital signs, and presence of atrial fibrillation. Three rule-based algorithms, ie, Face Arm Speech Test (FAST) and two flavors of Balance, Eyes, FAST (BE-FAST), and six machine learning (ML) techniques with various resampling methods were used to build classifiers for identification of patients with suspected stroke. Logistic regression (LR) was used to find important features. RESULTS: The study population consisted of 1361 patients. The values of area under the precision-recall curve (AUPRC) were 0.737, 0.710, and 0.562 for the FAST, BE-FAST-1, and BE-FAST-2 models, respectively. The values of AUPRC for the top three ML models were 0.787 for classification and regression tree with undersampling, 0.783 for LR with synthetic minority oversampling technique (SMOTE), and 0.782 for LR with class weighting. Among the ML models, logistic regression and random forest models in general achieved higher values of AUPRC, in particular in those with class weighting or SMOTE to handle class imbalance problem. In addition to the presenting complaint and triage level, age, diastolic blood pressure, body temperature, and pulse rate, were also important features for developing a stroke-alert trigger. CONCLUSIONS: ML techniques significantly improved the performance of prediction models for identification of patients with suspected stroke. Such ML models can be embedded in the electronic triage system for clinical decision support at ED triage.

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.

EMR-Based Phenotyping of Ischemic Stroke Using Supervised Machine Learning and Text Mining Techniques.

Ischemic stroke is a major cause of death and disability in adulthood worldwide. Because it has highly heterogeneous phenotypes, phenotyping of ischemic stroke is an essential task for medical research and clinical prognostication. However, this task is not a trivial one when the study population is large. Phenotyping of ischemic stroke depends primarily on manual annotation of medical records in previous studies. This article evaluated various strategies for automated phenotyping of ischemic stroke into the four subtypes of the Oxfordshire Community Stroke Project classification based on structured and unstructured data from electronical medical records (EMRs). A total of 4640 adult patients who were hospitalized for acute ischemic stroke in a teaching hospital were included. In addition to the structured items in the National Institutes of Health Stroke Scale, unstructured clinical narratives were preprocessed using MetaMap to identify medical concepts, which were then encoded into feature vectors. Various supervised machine learning algorithms were used to build classifiers. The study results indicate that textual information from EMRs could facilitate phenotyping of ischemic stroke when this information was combined with structured information. Furthermore, decomposition of this multi-class problem into binary classification tasks followed by aggregation of classification results could improve the performance.