Integrating structured and unstructured data for predicting emergency severity: an association and predictive study using transformer-based natural language processing models

Abstract

Background

Efficient triage in emergency departments (EDs) is critical for timely and appropriate care. Traditional triage systems primarily rely on structured data, but the increasing availability of unstructured data, such as clinical notes, presents an opportunity to enhance predictive models for assessing emergency severity and to explore associations between patient characteristics and severity outcomes. This study aimed to evaluate the effectiveness of combining structured and unstructured data to predict emergency severity more accurately.

Methods

Data from the 2021 National Hospital Ambulatory Medical Care Survey (NHAMCS) for adult ED patients were used. Emergency severity was categorized into urgent (scores 1–3) and non-urgent (scores 4–5) based on the Emergency Severity Index. Unstructured data, including chief complaints and reasons for visit, were processed using a Bidirectional Encoder Representations from Transformers (BERT) model. Structured data included patient demographics and clinical information. Four machine learning models—Logistic Regression, Random Forest, Gradient Boosting, and Extreme Gradient Boosting—were applied to three data configurations: structured data only, unstructured data only, and combined data. A mean probability model was also created by averaging the predicted probabilities from the structured and unstructured models.

Results

The study included 8,716 adult patients, of whom 74.6% were classified as urgent. Association analysis revealed significant predictors of emergency severity, including older age (OR = 2.13 for patients 65 +), higher heart rate (OR = 1.56 for heart rates > 90 bpm), and specific chronic conditions such as chronic kidney disease (OR = 2.28) and coronary artery disease (OR = 2.55). Gradient Boosting with combined data demonstrated the highest performance, achieving an area under the curve (AUC) of 0.789, an accuracy of 0.726, and a precision of 0.892. The mean probability model also showed improvements over structured-only models.

Conclusions

Combining structured and unstructured data improved the prediction of emergency severity in ED patients, highlighting the potential for enhanced triage systems. Integrating text data into predictive models can provide more accurate and nuanced severity assessments, improving resource allocation and patient outcomes. Further research should focus on real-time application and validation in diverse clinical settings.

Introduction

Emergency departments (EDs) are critical points of care that manage a diverse array of medical conditions with varying degrees of severity [1, 2]. Efficient triage and resource allocation are vital to ensuring timely and appropriate care for patients [3]. Traditionally, triage systems have relied on structured data such as patient demographics, vital signs, and medical history to assess the urgency of cases [4, 5]. Modern ED triage systems typically involve a nurse performing patient assessment and using an algorithm to determine triage acuity. The nurse incorporates subjective information obtained from the patient with structured data from the EHR into their triage decision. While triage systems are effective in predicting resource use and likelihood hospitalization [6], current systems rely heavily on the training and experience of the nurse to gather necessary data and correctly apply the algorithm. The wealth of unstructured data now available in EHRs presents an opportunity to refine and augment severity predictions by incorporating this rich source of information [7,8,9].

Natural Language Processing (NLP) techniques have shown promise in extracting valuable insights from unstructured text data. Specifically, Transformer-based models like Bidirectional Encoder Representations from Transformers (BERT) have revolutionized the field of NLP by enabling deep bidirectional understanding of text [10, 11]. BERT is uniquely suited for processing complex medical narratives because it captures the context of words in relation to all other words in a sentence. This ability sets BERT apart from traditional NLP techniques, such as Bag of Words or Term Frequency-Inverse Document Frequency (TF-IDF), which are limited by their reliance on isolated word frequencies and their inability to account for word dependencies and nuanced meanings in clinical language [12]. The use of BERT allows the models to understand the deeper semantics of clinical text, which is crucial in an ED setting where word context can dramatically alter the interpretation of patient symptoms [13]. Moreover, BERT is pre-trained on vast datasets and fine-tuned on specific tasks, making it effective even when applied to smaller datasets. This advantage of BERT has made it a key tool for integrating unstructured and structured data to enhance predictive tasks in healthcare [14, 15].

In recent years, NLP models have been applied to various clinical tasks, including analyzing clinical notes, predicting patient outcomes, and assisting in care prioritization [7,8,9, 16]. Appling NLP for ED triage has also grown significantly. For instance, Stewart et al. [17] reviewed various applications of NLP at ED triage, highlighting the potential of these techniques to enhance triage accuracy and efficiency. The advancements underscore the promising contribution of NLP to transforming triage practices in EDs. Despite these advances, the application of novel approaches to integrate structured and unstructured clinical data for predicting emergency severity in ED settings remains underexplored [17].

This study aimed to fill this gap by developing and comparing models that predict the emergency severity score, a critical triage indicator, using both structured and unstructured data from the National Hospital Ambulatory Medical Care Survey—Emergency Department (NHAMCS-ED) for the year 2021 [18]. We hypothesized that a combined approach integrating structured data with unstructured text data processed through a Transformer-based model would outperform models utilizing structured data alone or unstructured data alone. By leveraging the comprehensive NHAMCS-ED dataset, our goal was to harness the full spectrum of available information, enhancing the predictive power and clinical utility of our models.

Method

Data source

The data for this study was obtained from the NHAMCS-ED dataset for the year 2021. Only adult patients (age > 18 years) were included in the study, and patients with missing emergency severity scores were excluded. This dataset provides comprehensive information on patient visits to emergency departments across the United States, including demographic details, reasons for visit, and diagnostic and treatment data.

Study outcome

The emergency severity score used in this study is based on the Emergency Severity Index (ESI), a widely used triage tool in EDs. The ESI assigns scores from 1 to 5 to prioritize patients based on the urgency of their condition. A score of 1 represents the most critical cases requiring immediate life-saving interventions, while a score of 5 indicates non-urgent cases that can safely wait for care. These scores are typically assigned by a triage nurse during the initial evaluation, using a combination of objective measures (such as vital signs) and clinical judgment [19, 20]. ESI scores directly inform decisions about the urgency of treatment. For example, patients with a score of 1 need immediate intervention to prevent death, while those with a score of 2 are high-risk and must be seen quickly to avoid deterioration. Patients with a score of 3, though stable, still require timely care but can wait longer than those with more urgent scores. Meanwhile, scores of 4 and 5 represent minor conditions that can safely wait for extended periods [21, 22]. While the ESI is an ordinal scale, the differences in urgency between consecutive scores are not evenly spaced. The difference in urgency between a score of 1 and 2 is much greater than between scores 3 and 4. For this study, we grouped ESI scores into two categories: urgent (scores 1–3) and non-urgent (scores 4–5). This binary categorization reflects common clinical practice, where the main concern is whether a patient requires urgent intervention. Although this approach reduces some granularity, it aligns with the critical decision-making process in EDs, prioritizing the need for urgent care [23].

Structured predictors

The structured data extracted from the dataset included a variety of variables related to patient demographics, visit characteristics, and clinical information. Specifically, the structured data encompassed patient demographics such as age, sex, and race/ethnicity. Visit characteristics included arrival time, mode of arrival, day of the week, and whether the patient arrived by ambulance. Clinical information comprised vital signs (temperature, heart rate, diastolic blood pressure, systolic blood pressure, pulse oximetry, respiratory rate), pain level, and medical history (conditions such as Alzheimer's disease/dementia, asthma, cancer, cerebrovascular disease, chronic kidney disease, chronic obstructive pulmonary disease, congestive heart failure, coronary artery disease, depression, diabetes mellitus types I and II, end-stage renal disease, pulmonary embolism, HIV infection/AIDS, hyperlipidemia, hypertension, obesity, obstructive sleep apnea, osteoporosis, and substance abuse or dependence). Additional factors considered were the type of residence (private residence, nursing home, homeless, or other), insurance type, whether the visit was a follow-up or within the last 72 h, and the nature of any injury or trauma, overdose/poisoning, or adverse effect of medical/surgical treatment. Missing values in the structured data were handled using median imputation, and the data were standardized using StandardScaler [24].

Unstructured data and BERT model

Unstructured data consisted of the chief complaints and reasons for the injury presented at the ED visits. To ensure the quality and consistency of the input data, a structured text cleaning process was applied. This involved converting all text to lowercase for uniformity, removing punctuation and numbers, and filtering out common stopwords (e.g., "and," "the") that do not contribute meaningfully to clinical interpretation. These steps ensured that the text data retained only relevant clinical information.

These cleaned text fields were tokenized using the BERT tokenizer from the HuggingFace library [25], preparing the text data for input into a BERT-based model. The BERT model represents a significant advancement in natural language processing by enabling deep bidirectional understanding of text [26, 27]. Unlike traditional models that read text either left-to-right or right-to-left, BERT processes text in both directions simultaneously, allowing it to understand the context of a word based on all surrounding words. This bidirectional approach enables BERT to capture the nuanced meanings of words and phrases in their specific contexts. BERT's architecture is based on transformers, a type of deep learning model that relies on self-attention mechanisms to weigh the importance of different words in a sentence. This allows BERT to excel at tasks that require understanding the relationships between words and the overall meaning of sentences. Pre-trained on a vast corpus of text data, including books and Wikipedia articles, BERT can be fine-tuned on specific tasks such as classification, question answering, and named entity recognition.

To prepare the unstructured text data for analysis, we used the BERT tokenizer. This process converts the clinical text into a structured format that BERT can interpret, ensuring that important contextual information is preserved. The tokenizer breaks down sentences into smaller units, allowing BERT to understand the relationships between words in a given sentence. Following tokenization, the text was passed through the BERT model to generate numerical embeddings—dense vectors that represent the semantic meaning of the text. These embeddings capture the context and meaning of the text, allowing the model to utilize the full depth of clinical narratives. The embeddings were then combined with the structured data, integrating both textual and numerical information to enhance the predictive capability of the model.

Predictive model development

For this study, four different machine learning models were applied: Logistic Regression (LR), Random Forest (RFM) [28], Gradient Boosting (GB) [29], and Extreme Gradient Boosting (XGB) [30]. We implemented four machine learning models using Python’s scikit-learn and xgboost libraries to evaluate the predictive performance of structured data in classifying emergency severity. For each model, key parameters were configured, while all other parameters were set to their default values. The LogisticRegression function from sklearn.linear_model was set with a maximum iteration limit of 1000 (max_iter = 1000) to ensure convergence. The RandomForestClassifier function from sklearn.ensemble was employed with 500 estimators (n_estimators = 500), balancing accuracy and computational efficiency. The GradientBoostingClassifier, also from sklearn.ensemble, was applied with default settings, allowing the model to iteratively adjust for errors made by prior trees, thereby focusing subsequent trees on misclassified instances to improve predictive precision. For XGBoost, we utilized XGBClassifier from the xgboost library, configuring it with logloss as the evaluation metric to prioritize probability calibration and classification accuracy. Each of these models was trained and evaluated using four distinct approaches: structured data, unstructured data, combined data, and a mean probability model.

The first approach used only structured data, including patient demographics, clinical information, and visit characteristics from the NHAMCS-ED dataset. Logistic Regression, Random Forest, Gradient Boosting, and XGBoost models were trained on this data. The second approach focused solely on unstructured data, processed using the BERT model to generate feature vectors. These vectors were used as input for the same machine learning models. The third strategy combined both structured and unstructured data, merging quantitative information with BERT-extracted features to provide a comprehensive input for the models. The final method employed a mean probability model, which averaged the predicted probabilities from the structured and unstructured models. This technique combined the strengths of both data types without retraining. All approaches were evaluated using fivefold cross-validation.

Evaluation metrics

The evaluation of all models involved calculating ROC AUC, accuracy, F1 score, precision, recall, sensitivity, and specificity. The models' predictive probabilities and true labels were recorded, and ROC curves were plotted to visualize the performance of each model. The cutoff points for classification were determined by finding the thresholds closest to the top-left corner of the ROC curve [31, 32]. The ROC AUC quantifies the model’s ability to differentiate between these two categories, with higher values (closer to 1.0) indicating better discrimination across various threshold values. Accuracy reflects the proportion of correct classifications (both urgent and non-urgent) out of the total predictions; however, its utility may be limited when class distribution is imbalanced. Precision measures the proportion of true positives (correctly classified urgent cases) among all instances predicted as urgent, making it particularly useful when minimizing false positives is important. Sensitivity (Recall), on the other hand, evaluates the model’s ability to correctly identify all urgent cases, which is crucial in emergency department settings where missing urgent cases could have serious consequences. Specificity assesses the model's ability to correctly classify non-urgent cases, thereby avoiding over-triage, where non-urgent patients are incorrectly labeled as urgent. Finally, the F1 score, which is the harmonic mean of precision and recall, offers a balanced evaluation of the model’s handling of both false positives and false negatives, especially valuable in scenarios with uneven class distributions. ROC curves were plotted for each model to compare their performance. Additionally, visualizations such as forest plots of odds ratios and word clouds of unstructured variables were generated to illustrate the significance and frequency of different variables in the dataset.

Results

Among the 8,716 patients included in the study, 25.4% were categorized as non-urgent or semi-urgent, while 74.6% were classified as urgent, emergent, or immediate. Table 1 and Supplement Table 1 present the baseline characteristics of U.S. patients presenting to the ED, stratified by emergency severity score. Significant differences were observed between the two groups in terms of gender, with a higher proportion of females in the urgent category (55.1%) compared to the non-urgent group (52.0%, p = 0.0096). Age also varied significantly, with older patients more likely to be in the urgent category (p < 0.0001). Specifically, 27.5% of patients aged 65 and above were in the urgent group, compared to 15.7% in the non-urgent group. Race/ethnicity did not show significant differences between groups (p = 0.0603). However, differences were noted in residence type (p < 0.0001), with a higher percentage of urgent patients residing in nursing homes (2.8% vs. 1.0%) and a greater proportion of non-urgent patients living in private residences (95.8% vs. 94.2%). Insurance type also showed significant differences (p < 0.0001), with a higher percentage of urgent patients covered by Medicare (29.7% vs. 19.0%) and a higher percentage of non-urgent patients being uninsured (10.4% vs. 8.0%). Arrival by ambulance was significantly more common in the urgent group (25.4% vs. 8.2%, p < 0.0001). Follow-up visits were slightly more frequent in the non-urgent group (8.9% vs. 7.4%, p = 0.0243). Pain levels, temperature, heart rate, diastolic blood pressure, systolic blood pressure, pulse oximetry, and respiratory rate all showed significant differences between the two groups. In terms of medical history, conditions such as cancer, cerebrovascular disease, chronic kidney disease, chronic obstructive pulmonary disease, congestive heart failure, coronary artery disease, diabetes mellitus type II, end-stage renal disease, pulmonary embolism, hyperlipidemia, hypertension, obesity, obstructive sleep apnea, osteoporosis, and substance abuse were more common in the urgent group.

Table 1 Baseline characteristics of U.S. patients presenting to the ED, stratified by Emergency Severity Score, NHAMCS 2021

Figure 1a and Fig. 1b display forest plots of odds ratios with 95% confidence intervals for the various structured variables used in the study. These figures illustrate the significant predictors of emergency severity, highlighting the relative importance of different factors. In Fig. 1a, demographic and visit characteristics are detailed. Female patients had higher odds of being classified as urgent (OR = 1.15, 95% CI: 1.05–1.26). Age was a significant predictor, with patients aged 40–65 having higher odds of urgency (OR = 1.32, 95% CI: 1.21–1.44) compared to those aged 18–39. Patients aged 65 and above had even higher odds (OR = 2.13, 95% CI: 1.88–2.41). Arrival by ambulance markedly increased the odds of being urgent (OR = 3.65, 95% CI: 3.05–4.36). Medicare coverage was associated with higher odds of urgency (OR = 1.79, 95% CI: 1.58–2.02), while being uninsured was associated with lower odds (OR = 0.75, 95% CI: 0.61–0.91). Patients from nursing homes had higher odds of being classified as urgent (OR = 2.80, 95% CI: 1.73–4.54). Figure 1b focuses on clinical information and medical history. Heart rate was a significant predictor, with patients having heart rates over 90 bpm showing higher odds of being urgent (OR = 1.56, 95% CI: 1.42–1.72). Blood pressure was also significant; diastolic blood pressure less than 60 mm Hg was associated with higher odds of urgency (OR = 1.53, 95% CI: 1.20–1.95), and DBP greater than 80 mm Hg showed increased odds (OR = 1.29, 95% CI: 1.17–1.42). Systolic blood pressure greater than 120 mm Hg was associated with higher urgency (OR = 1.10, 95% CI: 1.00–1.22). Several medical conditions significantly increased the odds of being classified as urgent, including cancer (OR = 2.54, 95% CI: 1.91–3.37), chronic kidney disease (OR = 2.28, 95% CI: 1.71–3.03), chronic obstructive pulmonary disease (OR = 1.78, 95% CI: 1.46–2.17), congestive heart failure (OR = 2.45, 95% CI: 1.88–3.18), coronary artery disease (OR = 2.55, 95% CI: 2.03–3.20), end-stage renal disease (OR = 3.37, 95% CI: 1.88–6.05), diabetes mellitus type II (OR = 1.90, 95% CI: 1.57–2.30), hyperlipidemia (OR = 1.63, 95% CI: 1.40–1.90), hypertension (OR = 1.73, 95% CI: 1.55–1.93), and obesity (OR = 1.25, 95% CI: 1.08–1.44).

Fig. 1

a Forest Plot of Odds Ratios with 95% CI (Log Scale). b Forest Plot of Odds Ratios with 95% CI (Log Scale). The odds ratios (ORs) presented in Figure. 1 were derived from a logistic regression model where all variables were mutually adjusted. This means that the ORs account for the influence of all other variables included in the model. For example, the OR for age reflects the effect of age on emergency severity while controlling for other factors such as gender, vital signs, and medical history. This mutual adjustment allows for a more accurate estimation of the individual contribution of each variable to the prediction of emergency severity, minimizing potential confounding effects

Figure 2 presents the frequency and word cloud of the words in the unstructured variables, providing a visual representation of the most common terms found in the chief complaints and reasons for the injury presented at the ED visits. Table 2 and Fig. 3 summarize the performance metrics for the different models. The results demonstrate that integrating structured and unstructured data leads to improved model performance across all classifiers. Logistic Regression showed significant improvements when combining both data types, achieving an AUC of 0.784, an accuracy of 0.717, and a high precision of 0.894. Random Forest and Gradient Boosting models similarly benefited from the combination, with Random Forest achieving an AUC of 0.766 and Gradient Boosting reaching 0.789. In particular, Gradient Boosting demonstrated strong predictive capabilities with a precision of 0.892 and an F1 score of 0.797. Extreme Gradient Boosting, although slightly weaker with structured data alone, showed notable gains when unstructured data was included, with a combined AUC of 0.779 and a precision of 0.886.

Fig. 2

Frequency and the word cloud of the word in the unstructured variables

Table 2 Performance metrics for different data models

Fig. 3

Receiver Operating Characteristic (ROC) curves for the four models evaluated in the study. Each model include the structured data model, which uses only structured data such as patient demographics, visit characteristics, vital signs, and medical history; the unstructured data model, a BERT-based natural language processing (NLP) model that uses only unstructured data, including chief complaints and reasons for injury; the combined input model, a machine learning classification model that integrates both structured data and BERT-extracted features from the unstructured data; and the mean probability model, which averages the predicted probabilities from the structured data model and the unstructured data model

Discussion

Our study demonstrated that combining structured and unstructured data significantly improved the prediction of emergency severity in an ED setting. By integrating clinical narratives with traditional patient demographics, vital signs, and medical history, we were able to capture a more comprehensive representation of the patient's condition. The results showed that models incorporating both data types outperformed those relying solely on structured or unstructured data. This finding highlights the potential of leveraging advanced NLP techniques, such as BERT, in conjunction with structured clinical data to enhance decision-making in emergency care. While the BERT model effectively captured the contextual nuances in clinical notes, the combined approach proved most robust, supporting the idea that integrating diverse data sources can yield more accurate and actionable predictions in complex medical environments like the ED.

Association analysis and clinical implications

While the association analysis identified several statistically significant predictors of emergency severity, it is crucial to differentiate between statistical significance and clinical relevance. Chronic conditions such as coronary artery disease, chronic kidney disease, and chronic obstructive pulmonary disease were significant predictors in our model. However, the practical relevance of these findings for real-time decision-making in ED settings should be critically examined. For instance, while the presence of chronic conditions may inform long-term risk stratification, their immediate impact on triage decisions may be limited unless the condition is actively contributing to the acute presentation. Thus, although these conditions were associated with increased urgency, further research is needed to explore their practical role in ED triage processes.

In addition, older age and higher heart rate emerged as significant predictors, aligning with clinical expectations that elderly patients and those with abnormal vital signs require urgent attention. However, it is important to interpret these findings with caution, particularly in the context of retrospective analysis. While our model can identify factors associated with higher acuity, it does not substitute clinical judgment, which remains critical in real-time decision-making. The inclusion of unstructured data, particularly chief complaints, offers a way to incorporate nuanced patient information that is often missing in structured data, thus improving the predictive accuracy of the models.

A notable finding was the association between insurance status and emergency severity. Patients covered by Medicare had higher odds of being classified as urgent, while uninsured patients were less likely to be classified as urgent. This result raises important questions regarding access to care and its influence on triage outcomes. One potential explanation is that uninsured patients may delay seeking care due to financial concerns, leading to underrepresentation in our dataset or potentially presenting with less acute conditions. Alternatively, these findings may reflect broader disparities in healthcare access and utilization, where insurance status influences not only access to primary care but also ED triage decisions [33,34,35]. The association between Medicare coverage and higher urgency might reflect the higher baseline health risks of the elderly population, who are more likely to suffer from multiple comorbidities. Further exploration of how insurance status interacts with other social determinants of health, such as socioeconomic status and healthcare access, is warranted. Future studies should aim to validate these findings and examine whether controlling for other factors, such as pre-existing health conditions, changes the relationship between insurance status and triage classification. Moreover, this finding highlights the need for ED policies that address potential biases in triage based on insurance status and other social determinants of health.

Model performance and clinical application

Our results demonstrated that integrating structured and unstructured data improves the performance of predictive models, particularly in complex cases where traditional triage systems may fall short. The Gradient Boosting and Extreme Gradient Boosting models achieved the highest performance, with AUCs of 0.789 and 0.779, respectively, when both data types were combined. The strong performance of these models underscores the value of using machine learning techniques that can account for non-linear interactions and complex relationships between variables, which are often present in clinical data.

Our findings can be compared to the findings of Brouns et al. (2019) [36] and Veldhuis et al. (2022) [37]. Brouns et al. evaluated the Manchester Triage System in older emergency department patients, reporting an AUC of 0.74 for predicting hospital admissions, a result similar to the AUCs achieved by our combined data models. However, their study noted that MTS had a lower AUC of 0.71 for predicting in-hospital mortality, highlighting the limitations of relying solely on structured triage systems in medically complex populations. Our results demonstrate that combining structured and unstructured data can address some of these limitations by improving predictive accuracy, particularly in more complex cases. Similarly, Veldhuis et al. compared clinical judgment to early warning scores and found that clinical judgment outperformed risk stratification models, with AUCs between 0.70 and 0.89, especially for ICU admissions and severe adverse events [37]. While our models performed similarly, this emphasizes the need to integrate machine learning with clinical judgment. Our models, combining structured and unstructured data, outperformed single-source models, aligning with Veldhuis et al.'s suggestion that clinical tools combined with automated systems yield the best results.

In comparison with traditional triage systems, such as the Manchester Triage System [36], our models show promise in enhancing predictive accuracy by leveraging a broader range of patient data, particularly unstructured clinical narratives. However, it is essential to emphasize that clinical judgment remains a critical component of ED decision-making. Predictive models, while valuable, should complement—not replace—the expertise of healthcare providers, who are best equipped to make nuanced decisions in real-time clinical settings.

Limitations and future directions

There are several limitations to this study. First, the study is retrospective and relies on the accuracy and completeness of the NHAMCS-ED dataset. Any missing or inaccurately recorded data could impact the model's performance. Although the proportion of missing data was relatively low (< 10%), the method of imputation (median) might not capture the true underlying values in all cases, and different imputation techniques could lead to slightly different results. Second, the study focuses on data from a single year (2021), which may limit the generalizability of the findings to other years or different hospital settings. Emergency presentations can vary over time due to factors such as seasonal changes, pandemics, or other public health events. Future studies should validate these findings with data from multiple years and diverse clinical environments to ensure the robustness and applicability of the models across varying contexts. Third, while BERT proved highly effective in processing unstructured clinical text, it is computationally intensive compared to simpler models such as TF-IDF or logistic regression. The complexity and resource demands of BERT may limit its use in real-time ED settings, particularly in resource-constrained environments. For real-time applications, it may be beneficial to explore lighter models like DistilBERT [38] or other simplified NLP approaches that balance computational efficiency with performance. Additionally, another important limitation involves the potential bias inherent in machine learning models [39]. Bias can emerge from the data used to train the model, particularly if the dataset reflects existing disparities in healthcare access, treatment, or outcomes. For instance, the underrepresentation of uninsured patients in the dataset may skew the model's ability to predict outcomes for this group, potentially reinforcing inequities in healthcare delivery. Furthermore, models trained on past data may perpetuate historical biases in clinical decision-making, such as differences in treatment recommendations based on race, gender, or insurance status. Addressing this issue will require careful evaluation of the model’s performance across diverse patient populations and the implementation of fairness-enhancing techniques, such as bias mitigation algorithms, to ensure that the model does not exacerbate existing healthcare disparities.

Future research should explore the integration of additional data sources, such as imaging and laboratory results, to further enhance predictive models [40]. These data sources could provide additional valuable information that can improve the accuracy of severity predictions. Prospective studies are also needed to validate the performance of these models in real-time clinical settings. Implementing these models in actual ED workflows and assessing their impact on clinical outcomes and operational efficiency will provide crucial insights [41]. Exploring the use of other advanced NLP models and techniques could yield further improvements in handling unstructured data. For example, models that incorporate contextual embeddings or use transfer learning from larger clinical datasets could enhance the performance of text-based predictions. Additionally, investigating the specific contributions of different types of unstructured data to the model's performance can provide valuable insights for improving triage protocols and decision-making processes in EDs.

Conclusion

The integration of structured and unstructured data shows promise in enhancing the prediction of emergency severity in ED settings. By leveraging advanced NLP techniques and comprehensive data sources, healthcare providers may improve the accuracy of severity predictions, potentially leading to more informed resource allocation and better patient outcomes. However, while these findings are encouraging, further validation in diverse clinical environments is necessary before definitive claims can be made about the model's broader applicability. Prospective studies involving multi-year datasets and real-world implementation will be essential to confirm the impact on clinical decision-making. This study underscores the potential of combining diverse data types to support predictive modeling, but caution is required when interpreting the results, given the need for additional validation. The findings contribute to the growing body of work on data integration in healthcare and suggest that this approach has the potential to support clinical decision-making in emergency settings, pending further research.