Integrating multimodal clinical data to predict intravenous (IV) fluid utilization: a comparative analysis of natural language processing techniques

Introduction

Intravenous (IV) fluid therapy (Hoste et al., 2014; Myburgh & Mythen, 2013) is commonly administered in emergency departments (EDs), playing a critical role in managing conditions ranging from mild dehydration to severe, life-threatening illnesses requiring urgent fluid resuscitation. Despite its routine use, variability and uncertainty persist regarding IV fluid administration decisions, potentially leading to inconsistent patient care, unnecessary resource utilization, and adverse patient outcomes (Pines, Griffey & Cone, 2015). Optimizing the decision-making process for IV fluid use can substantially enhance clinical effectiveness, patient safety, and operational efficiency (Kawamoto et al., 2005).

Advancements in medical informatics, particularly the widespread implementation of electronic health records (EHRs), have facilitated access to extensive datasets comprising both structured clinical variables and unstructured patient narratives (Hripcsak & Albers, 2013; Jensen, Jensen & Brunak, 2012). Structured data typically include quantifiable metrics such as demographic information, vital signs, medical history, and clinical severity indicators. While structured data offer easily interpretable and standardized measures, they often fail to capture the complete clinical context required for precise patient management decisions. In contrast, unstructured data, such as narrative text documented by clinicians during patient visits, provide richer context and nuanced details that are often absent from structured data (Meystre et al., 2017; Xiao, Choi & Sun, 2018).

Recent developments in natural language processing (NLP), especially transformer-based models such as Generative Pre-trained Transformer 2 (GPT-2), have revolutionized the extraction and interpretation of critical medical insights from unstructured textual data (Kreimeyer et al., 2017; Lee et al., 2020; Radford et al., 2019). These advanced NLP methods enable sophisticated understanding and quantification of patient narratives, which can significantly augment clinical predictions derived from structured data alone (Esteva et al., 2019; Wu et al., 2020). Indeed, prior research has consistently demonstrated the value of combining these data modalities for a range of clinical predictions, from inpatient mortality and 30-day readmissions (Rajkomar et al., 2018) to predicting patient disposition from the ED (Hong, Haimovich & Taylor, 2018). These studies affirm that unstructured narratives contain diagnostic and prognostic information often absent from structured fields alone. However, while the general benefit of data fusion is recognized, its application to resource-specific predictions like IV fluid utilization remains less explored. Furthermore, which NLP techniques are most effective for extracting signals from the short, keyword-driven texts common in the ED is an area requiring further investigation.

This study aimed to address this gap by developing and evaluating predictive models of IV fluid utilization in emergency departments using structured clinical data and NLP-derived insights from unstructured patient narratives. The primary objective was to assess whether an integrated predictive modeling approach, combining structured and unstructured data, would yield superior performance compared to models relying exclusively on either data type. Such models could inform clinical decisions, enhance patient care management, and optimize resource allocation in emergency medical practice.

Method

The overall pipeline of our study is illustrated in Fig. 1. The process begins with data acquisition from the NHAMCS-ED dataset, followed by cleaning and a train/test split. The data is then separated into two streams: a structured data stream and an unstructured text stream. For the text data, we extracted features using three distinct NLP techniques: CountVectorizer, Word2Vec, and pre-trained GPT-2 embeddings. Finally, both single-modality (structured-only, unstructured-only) and integrated (combined via early fusion) datasets were used to train and evaluate Logistic Regression (LR) and Gradient Boosting Classifier (GBC) models for predicting IV fluid administration.

Results

A total of 13,115 adult patients were included in this analysis, of which 4,193 (32.0%) received intravenous (IV) fluids during their ED visits (Table 1). Patients who received IV fluids were more frequently female (56.1% vs. 53.0%, p = 0.0008), older (28.8% aged ≥65 years vs. 19.6%, p < 0.0001), and predominantly White (65.1% vs. 56.1%, p < 0.0001). They were also more likely to reside in nursing homes (3.4% vs. 1.5%, p < 0.0001) and had higher proportions of Medicare insurance coverage (33.9% vs. 23.7%, p < 0.0001). Clinically, IV fluid recipients showed greater severity indicators, such as higher rates of ambulance arrival (29.2% vs. 17.3%, p < 0.0001), severe or unbearable pain levels (31.1% vs. 28.1%, p = 0.0187), elevated body temperature above 38°C (3.6% vs. 1.1%, p < 0.0001), increased heart rates greater than 90 bpm (44.7% vs. 36.0%, p < 0.0001), lower oxygen saturation levels (15.2% vs. 8.2% with pulse oximetry ≤94%, p < 0.0001), and elevated respiratory rates greater than 20 breaths per minute (14.1% vs. 6.6%, p < 0.0001). Additionally, IV fluid recipients more often presented with conditions related to overdose, poisoning, adverse medical effects, or trauma (21.6% vs. 36.2%, p < 0.0001) and had higher Emergency Severity Index scores categorized as immediate or emergency (28.4% vs. 14.3%, p < 0.0001). Patients receiving IV fluids also had notably higher prevalence of chronic conditions, including coronary artery disease (12.4% vs. 5.7%, p < 0.0001), diabetes mellitus type II (12.6% vs. 7.9%, p < 0.0001), hypertension (41.8% vs. 28.5%, p < 0.0001), chronic kidney disease (7.7% vs. 3.5%, p < 0.0001), and obesity (10.3% vs. 7.5%, p < 0.0001).

Figure 2 presents a heatmap visualizing the percentage point differences in demographic and clinical characteristics between patients who received IV fluids and those who did not. This visualization highlights the most significant differentiating factors. Notably, patients with an ‘Urgent’ or ‘Semi-urgent’ Emergency Severity Index score, those with a history of hypertension or diabetes, and those arriving by ambulance showed a markedly higher prevalence in the IV fluid group. Conversely, patients presenting with trauma or injury were substantially less likely to receive IV fluids, likely because their treatments were focused on other interventions. This graphic representation complements the detailed statistics provided in Table 1.

Figure 3 presents a forest plot illustrating odds ratios (ORs) and corresponding 95% confidence intervals (CIs) for significant predictors of IV fluid utilization derived from multivariable logistic regression analysis. The plot visually depicts the relative influence of various patient demographics, clinical characteristics, and medical histories on the likelihood of receiving IV fluids in emergency department visits. Key findings include notably higher odds for IV fluid administration among patients arriving by ambulance (OR = 1.99, 95% CI [1.82–2.17]), those aged ≥65 years (OR = 2.05, 95% CI [1.86–2.25]), patients with elevated temperatures (>38 °C; OR = 3.27, 95% CI [2.52–4.25]), and elevated respiratory rates (>20 breaths per minute; OR = 2.41, 95% CI [2.13–2.73]). Additionally, patients categorized with higher Emergency Severity Index (immediate or emergency) and those experiencing injury or adverse medical effects also exhibited significantly increased odds. Chronic conditions such as congestive heart failure, coronary artery disease, chronic kidney disease, diabetes, cancer, and hypertension significantly contributed to the likelihood of IV fluid administration. Conversely, factors associated with reduced likelihood included non-White race/ethnicity, residence in nursing homes or homelessness, and patients on Medicaid or uninsured. This detailed representation emphasizes clinically relevant predictors influencing IV fluid use, highlighting critical factors for clinical decision-making and resource optimization in emergency settings.

Figure 4 reports results for the LR models and Fig. 5 for GBC models, each using three NLP representations of the unstructured narratives—CountVectorizer, Word2Vec, and pre-trained GPT-2. Across both model families, combined models that integrated structured and unstructured data consistently outperformed models using either data type alone, regardless of the NLP technique. Among the NLP methods, clear performance differences emerged: in the GBC analyses (Fig. 3), the combined CountVectorizer model achieved the best AUC (0.786), followed by Word2Vec (0.785) and GPT-2 (0.772); the unstructured-only models showed the same ordering (CountVectorizer 0.735, Word2Vec 0.725, GPT-2 0.698). For LR (Fig. 2), the combined models followed the identical pattern (CountVectorizer 0.781, Word2Vec 0.775, GPT-2 0.766). Taken together, these results indicate that while incorporating patient narratives is critical, simpler keyword-based representations are more effective than a complex, context-based pre-trained model like GPT-2 for extracting predictive signal from the short emergency department chief-complaint texts.

Discussion

This study evaluated the predictive capability of integrating structured clinical variables with unstructured patient narratives to forecast IV fluid administration in ED settings. The primary finding confirms that an integrated, multimodal approach consistently outperforms models that rely on either structured or unstructured data alone. This aligns with emerging evidence supporting the fusion of diverse data sources to enhance the accuracy and efficiency of clinical decision-making. The improved performance underscores that patient narratives contain crucial, clinically relevant information not fully captured by traditional structured metrics (Su et al., 2022; Zhang et al., 2019, 2024).

A key and unexpected finding of this study, however, emerged from our comparative analysis of different NLP techniques. Contrary to the prevailing trend of leveraging large, complex models, we found that traditional NLP methods, specifically the frequency-based CountVectorizer model, yielded superior predictive performance compared to the advanced, pre-trained GPT-2 transformer model (Shaik et al., 2023). This suggests that for the specific task of predicting IV fluid utilization based on short ED chief complaints, the choice of NLP methodology is not trivial and “more complex” does not necessarily mean “more effective”.

Our primary finding that an early fusion of features provides a significant performance boost is consistent with previous work in the ED setting. For example, Hong, Haimovich & Taylor (2018) found that incorporating chief complaint text features with structured triage data significantly improved the prediction of hospital admissions. Our results, however, provide a crucial, nuanced counterpoint to the narrative that newer, more complex NLP models are always superior. While unexpected, this finding is not without precedent. In a comparison of transformer models and bag-of-words for classifying radiology reports, Lu, Ehwerhemuepha & Rakovski (2022) noted that the performance gain from a sophisticated Clinical-BERT model over simpler baselines was modest, highlighting that model complexity must be justified by the task. Our study extends this concept, demonstrating that for the concise, keyword-driven nature of ED chief complaints, the efficient keyword detection of a CountVectorizer model is more effective than the deep contextual understanding of a general-purpose transformer.

We hypothesize several reasons for this outcome. First, the nature of clinical text in the ED is typically short, factual, and keyword-driven (e.g., “chest pain,” “shortness of breath”). “Bag-of-words” models like CountVectorizer excel at capturing the presence and frequency of these critical keywords, which are themselves powerful predictive signals. Second, the strength of models like GPT-2 lies in understanding nuanced syntax and long-range context, which are largely absent in these brief patient complaints. Attempting to model deep context in short, factual text may introduce noise rather than improving signal detection. Finally, the GPT-2 embeddings used were from a general-purpose model, not fine-tuned on medical text. In contrast, CountVectorizer builds its vocabulary directly from the task-specific data, making its features highly relevant and optimized for this particular clinical dataset. Our findings thus highlight the critical importance of aligning model choice with data characteristics.

Despite the variations in NLP model performance, the key clinical predictors identified by our best-performing models remained consistent and clinically coherent. Factors such as arrival by ambulance, older age, a history of coronary artery disease, and elevated heart rate prominently emerged as significant predictors of IV fluid administration. The identification of these factors, which logically align with higher patient acuity, reinforces the clinical validity and applicability of the integrated modeling approach. This demonstrates that such models can reliably capture the key indicators that influence real-world clinical decision-making in emergency care.

The implications of this study for healthcare informatics are twofold. It reinforces the value of leveraging unstructured narrative data, but it also serves as a crucial reminder that the optimal NLP approach is task-dependent. For clinical prediction tasks involving short, keyword-rich text, simpler, more interpretable, and computationally efficient models may be superior to large, general-purpose language models. This methodological insight is valuable, suggesting that future research should focus not just on applying the newest models, but on rigorously comparing different techniques to find the best fit for specific clinical problems. This approach could be expanded to enhance predictive models for various other clinical outcomes and resource allocation challenges in healthcare.

Conclusions

In conclusion, this study demonstrates that integrating structured clinical data with unstructured patient narratives substantially improves the accuracy of predictive models for IV fluid utilization in emergency departments. Our research further reveals that for the short clinical texts characteristic of emergency chief complaints, traditional NLP techniques like CountVectorizer are more effective at extracting predictive features than advanced, pre-trained GPT-2 models. This multimodal approach, combining the interpretability of logistic regression and the high performance of Gradient Boosting, offers a robust framework for enhancing clinical decision-making, streamlining resource allocation, and elevating patient care.

Supplemental Information