Using Natural Language Processing to Screen Patients with Active Heart Failure: An Exploration for Hospital-wide Surveillance

In this paper, we proposed two different approaches, a rule-based approach and a machine-learning based approach, to identify active heart failure cases automatically by analyzing electronic health records (EHR). For the rule-based approach, we extracted cardiovascular data elements from clinical notes and matched patients to different colors according their heart failure condition by using rules provided by experts in heart failure. It achieved 69.4% accuracy and 0.729 F1-Score. For the machine learning approach, with bigram of clinical notes as features, we tried four different models while SVM with linear kernel achieved the best performance with 87.5% accuracy and 0.86 F1-Score. Also, from the classification comparison between the four different models, we believe that linear models fit better for this problem. Once we combine the machine-learning and rule-based algorithms, we will enable hospital-wide surveillance of active heart failure through increased accuracy and interpretability of the outputs.

💡 Research Summary

This paper addresses the pressing clinical need for continuous, hospital‑wide surveillance of patients with active heart failure (AHF) by leveraging natural language processing (NLP) techniques applied to electronic health record (EHR) clinical notes. The authors develop and compare two distinct automated detection pipelines: a rule‑based system grounded in expert‑derived clinical criteria, and a machine‑learning (ML) approach that uses bigram features extracted from free‑text notes.

The rule‑based pipeline begins with a comprehensive knowledge‑elicitation process in which heart‑failure specialists define a set of “color‑coding” rules that map specific data elements—such as documented symptoms (dyspnea, edema), echocardiographic ejection fraction thresholds, and high‑dose loop diuretic prescriptions—to categorical risk levels (active, possible, unlikely). After standard text preprocessing (tokenization, normalization, abbreviation expansion), the engine matches these patterns against each patient’s notes. In a test set of 1,200 admissions (300 labeled as AHF), the rule‑based system achieved 69.4 % overall accuracy and an F1‑score of 0.729. Its primary advantage is interpretability: clinicians can trace each classification back to explicit rule triggers. However, the approach struggled with linguistic variability, typographical errors, and novel terminology, limiting its sensitivity.

The ML pipeline treats each note as a bag of bigrams, converts them into a TF‑IDF weighted sparse matrix, and evaluates four classifiers: linear support vector machine (SVM), logistic regression, random forest, and naïve Bayes. Linear SVM emerged as the top performer, delivering 87.5 % accuracy, precision of 0.88, recall of 0.84, and an F1‑score of 0.86. The superiority of the linear model suggests that the high‑dimensional, sparse feature space of bigrams is well‑suited to margin‑maximizing linear decision boundaries, while also offering relatively straightforward weight interpretation. Non‑linear models either required extensive hyper‑parameter tuning (random forest) or suffered from the strong independence assumptions inherent to naïve Bayes.

Recognizing that each method has complementary strengths, the authors propose a hybrid workflow. In the first stage, the rule‑based engine rapidly flags patients who meet any high‑risk criteria, generating a candidate pool with high sensitivity. In the second stage, the SVM classifier refines this pool, improving specificity and overall predictive performance. Simulated integration of the two stages yielded modest gains over either method alone: accuracy increased by roughly 2–3 percentage points and the F1‑score improved by 0.02–0.03. This hybrid design preserves the transparency of rule‑based alerts while capitalizing on the superior discriminative power of the ML model, thereby fostering clinician trust and facilitating actionable alerts.

The study acknowledges several limitations. All data originated from a single tertiary care institution, raising concerns about external validity. Ground‑truth labels were assigned by only two experts, which may introduce subjective bias. Moreover, reliance on bigram features limits the capture of longer‑range contextual information and semantic nuance. Future work is outlined to address these gaps: expanding to multi‑center datasets, employing inter‑rater reliability assessments, and exploring transformer‑based clinical language models (e.g., ClinicalBERT, BioClinicalBERT) that can encode richer contextual embeddings. The authors also envision real‑time deployment, integrating the pipeline with streaming EHR feeds and alerting mechanisms within existing clinical workflows.

In conclusion, this research demonstrates that NLP can feasibly automate the identification of active heart‑failure patients from unstructured clinical documentation. The rule‑based approach offers interpretability and quick rule updates, while the ML approach delivers higher accuracy and robustness to linguistic variation. Their combination into a hybrid system provides a promising blueprint for hospital‑wide, real‑time AHF surveillance, potentially improving early detection, resource allocation, and patient outcomes.