Preprocessing of Physician Notes by LLMs Improves Clinical Concept Extraction Without Information Loss

Abstract

Clinician notes are a rich source of patient information, but often contain inconsistencies due to varied writing styles, abbreviations, medical jargon, grammatical errors, and non-standard formatting. These inconsistencies hinder their direct use in patient care and degrade the performance of downstream computational applications that rely on these notes as input, such as quality improvement, population health analytics, precision medicine, clinical decision support, and research. We present a large-language-model (LLM) approach to the preprocessing of 1618 neurology notes. The LLM corrected spelling and grammatical errors, expanded acronyms, and standardized terminology and formatting, without altering clinical content. Expert review of randomly sampled notes confirmed that no significant information was lost. To evaluate downstream impact, we applied an ontology-based NLP pipeline (Doc2Hpo) to extract biomedical concepts from the notes before and after editing. F1 scores for Human Phenotype Ontology extraction improved from 0.40 to 0.61, confirming our hypothesis that better inputs yielded better outputs. We conclude that LLM-based preprocessing is an effective error correction strategy that improves data quality at the level of free text in clinical notes. This approach may enhance the performance of a broad class of downstream applications that derive their input from unstructured clinical documentation.

1. Introduction

Electronic Health Records (EHRs) have transformed healthcare documentation by improving the legibility and availability of patient data [1]. They have solved the “availability problem”—where paper charts were often inaccessible—and the “legibility problem” of handwritten notes [2,3,4]. Nonetheless, EHRs face major challenges (

Table 1. Challenges for Medical Record Documentation.

Challenge	Comments
Legibility	Former issue with handwritten notes; Solved with EHRs
Availability	Paper chart inaccessibility; resolved through digitization
Physician Burden	Ongoing crisis; contributes to dissatisfaction and burnout
Note Quality	Persistent problem; bloat, errors, abbreviations, jargon
Unstructured Free Text	Inability to transfer data to downstream applications

) that include physician dissatisfaction with the EHR, poor note quality, and difficulties in converting unstructured free text into a computable format for downstream applications.

Physician dissatisfaction with EHRs remains high, driven by poor interface design, excessive documentation burden, clerical overload, and limited perceived benefit for patient care [5,6,7,8,9,10,11,12].

Note quality in the EHR remains a persistent problem. Notes may suffer from inconsistent structure, misspellings, excessive abbreviations, slang, and ambiguous terminology [13,14,15,16,17,18]. Up to 20% of note content may be abbreviations or acronyms [19,20,21,22,23,24,25,26]. Additionally, notes are often poorly formatted and unnecessarily verbose. One study of redundancy in EHR notes showed that 78% of the content in sign-out notes and 54% of the content of progress notes was carried over from prior notes [27].

The electronic health record (EHR) serves as the official medical-legal document memorializing the care provided to a patient and the reasoning behind clinical decisions. It also supports billing and reimbursement for services rendered. However, data within the EHR is increasingly being reconfigured and repurposed for a range of secondary uses (

Table 2. Secondary Uses of EHR Data.

Use	Comments
Decision Support	Early stages; requires structured data
Precision Medicine	Needs structured data
Population Health	Underutilized due to fragmented and unstructured data
Data Exchange	FHIR adoption is underway, but not yet universal
Clinical Research	Hindered by a lack of structured data
Quality Improvement	Untapped potential, needs structured data

), including clinical decision support, precision medicine, quality improvement, population health management, data exchange, and clinical research [28,29,30,31,32,33,34].

1.1. Secondary Uses of the EHR Require Structured Data as Inputs

Secondary applications of electronic health records (EHRs)—such as precision medicine, population health management, clinical research, decision support, and quality improvement—depend on structured data to function effectively. However, a large portion of clinically meaningful information in the EHR is unstructured narrative text, such as physician notes, discharge summaries, and radiology reports. These free-text data are not readily usable by computational systems. It is estimated that up to 80% of information in the EHR exists in such unstructured text [35,36].

To make this information accessible for secondary use, two main strategies have emerged. The first approach involves transforming textual data into high-dimensional vector representations using language models. These embeddings capture latent semantic relationships and support tasks such as clustering, classification, and semantic similarity search. The second approach extracts biomedical entities from text and maps them to standardized terminologies or ontologies, including SNOMED CT, LOINC, HPO, ICD-10, and RxNorm [35,36,37,38,39,40,41]. While the embedding approach excels at capturing meanings, ontology-based mapping provides superior interpretability. Embeddings and ontology mappings are often used in tandem to leverage their respective strengths.

Nonetheless, the inability to consistently extract and ontologize clinically relevant information from free text remains a significant barrier to realizing the full potential of secondary EHR applications. Without reliable transformation of narrative content into a computable form, downstream systems are constrained by incomplete inputs (Table 2).

1.2. Current Approaches to Reducing Physician Burden and Improving Note Quality

Efforts to reduce documentation burden and improve note quality have included a range of strategies, each with its own limitations (

Table 3. Current Approaches to Reducing Physician Burden and Improving Note Quality.

Approach	Comments
Dictation	Costly; requires editing; complex workflow.
Ambient AI	Promising; hallucinations; workflow issues.
Smart Phrases or Copy-Paste	Adds redundancy; note bloat.
Spelling and Grammar Checkers	Piecemeal; limited context awareness.
Templates	Rigid; low physician satisfaction.
Documentation training	Poor physician acceptance; limited efficacy
LLM as a documentation coach	Promising emerging strategy
Scribes	Costly; workflow disruption; authorship concerns.

). Dictation and use of medical scribes remain costly and logistically complex [42,43]. Templates, smart phrases, and copy-paste functionality—originally intended to streamline documentation—have instead contributed to note bloat and redundant content [44,45]. Spelling and grammar checkers offer piecemeal solutions and may err due to lack of clinical context [18,46,47]. Educational interventions such as documentation training have had limited impact and are often poorly received by physicians [48]. Among emerging innovations, two LLM-based approaches stand out: (1) real-time coaching to guide clinicians toward better documentation [49], and (2) automated note generation via ambient listening [50,51,52]. A recent comparative review by Avendaño et al. [53] emphasizes the growing interest in AI-based documentation tools, including digital scribes and real-time assistants, as scalable solutions. These strategies create new opportunities to reduce clinician documentation burden and to improve the quality and computability of clinical notes (Table 3).

1.3. Our Proposed Approach to Improving Note Quality

Since the release of ChatGPT in late 2022, interest in large language models (LLMs) has surged across healthcare domains [54,55,56,57,58]. LLMs have been explored for a range of clinical applications, including decision support, diagnostic explanation, text summarization, concept extraction, ontology mapping, and document restructuring [59,60,61,62,63,64]. We propose that LLMs can also play a focused role in improving physician documentation through preprocessing.

For this submission, we define preprocessing as the enhancement of a clinical note’s structural and linguistic integrity. This includes correcting spelling and grammatical errors [65], expanding abbreviations (e.g., OU →both eyes), replacing colloquial expressions with standardized medical terminology (e.g., upgoing toe→ Babinski sign) [66], and enforcing consistent formatting—all without introducing new content or altering clinical meaning. In contrast to generative documentation systems, preprocessing seeks to improve clarity and consistency while preserving the clinician’s original narrative.

Preprocessing is not intended to reduce physicians’ documentation burden. However, we hypothesize that it offers two main benefits:

• Improved note quality, reflected in fewer spelling, grammar, and formatting errors.
• Greater retrieval of clinical concepts, as measured by Doc2Hpo.

2. Methods

This study evaluates the ability of GPT-4o 2024-11-20 to correct 1618 de-identified clinical notes from a neurology clinic. We hypothesize that preprocessing will find and correct spelling and grammatical errors, correct non-standard terminology, and expand acronyms and abbreviations. We further hypothesize that these corrections will enhance the yield of HPO concepts from the Doc2Hpo pipeline.

The evaluation framework (Figure 1) assesses the extraction and mapping of Human Phenotype Ontology (HPO) concepts under three conditions: (1) from raw notes, (2) from preprocessed notes, and (3) from GPT-extracted ground truth phrases. We define concept extraction as the identification of relevant clinical terms or phrases within free text. Concept mapping refers to the assignment of each extracted term to a standardized HPO concept and its corresponding identifier.

Data Acquisition: Institutional Review Board (IRB) approval was obtained from the University of Illinois to use de-identified clinical notes from the neurology clinic. All notes were de-identified using REDCap [67]. Selected notes met the following inclusion criteria:

• A neuroimmunological diagnosis, including multiple sclerosis, myasthenia gravis, neuromyelitis optica, and Guillain–Barré syndrome.
• Patient seen between 2016 and 2022
• Outpatient visit in the neurology clinic
• Examined by a neurology resident or attending physician
• Note length of at least 2000 characters

From an initial pool of 21,028 notes, the final dataset comprised 1618 typed notes exported as ASCII text and stored in a JSON-compatible format (Figure 2).

Preprocessing: GPT-4o (Open AI) was prompted to preprocess neurology notes (Appendix A and Appendix B). GPT-4o produced a detailed list of all modifications for each note (Appendix C). Character count, word count, sentence count, spelling and grammar corrections, abbreviation expansion, and word substitutions were calculated for each note.

Concept Extraction Using Doc2Hpo: We used the open-source Python implementation of Doc2Hpo [68], executed locally with the Aho-Corasick string-matching algorithm and a custom negation detection module (‘NegationDetector‘) that applies a windowed keyword search within sentence boundaries. We used the HPO release dated 3 March 2025. No modifications were made to the Doc2Hpo source code. Doc2Hpo was applied under three conditions:

• raw notes
• preprocessed notes
• ground truth phrases

To obtain the ground truth terms, GPT-4o was prompted to identify all candidate HPO-eligible phrases in each note (Appendix D). These phrases were combined into a single synthetic sentence for Doc2Hpo input (e.g., “The patient has ataxia, weakness, aphasia, and hyperreflexia.”). Doc2Hpo, which returned a list of matched terms and corresponding HPO IDs. For each of the 1,618 notes, we compiled three term sets based on the output of Doc2Hpo:

• ground truth terms
• raw-note terms
• preprocessed-note terms

Metric Computation: For each extracted term, we classified its match status as follows:

• True Positive (TP): Term present in both the extracted set and the ground truth set.
• False Negative (FN): Ground truth term not captured in extraction set.
• False Positive (FP): Term extracted but not part of ground truth set.
• True Negative (TN): Term present in the note but not successfully mapped to HPO by Doc2Hpo.

Performance metrics (precision, recall, accuracy, and F1 score) were calculated by the method of Powers [69].

3. Results and Discussion

3.1. Preprocessing Finds and Corrects Errors

We used GPT-4o to preprocess 1,618 de-identified clinical notes from a neurology clinic. Each note was converted from plain text to a structured JSON format with high-level section headers (e.g., History, Examination, Impression, Plan). GPT-4o was prompted to preprocess each note by correcting spelling and grammatical errors, expanding acronyms and abbreviations, and replacing non-standard phrasing (Appendix E).

On average, GPT-4o corrected 4.9 ± 1.8 grammatical errors, 3.3 ± 5.2 spelling errors, and expanded 12.6 ± 8.1 abbreviations or acronyms per note (Figure 3). Additionally, it replaced 3.1 ± 3.0 instances of non-standard language or jargon with standardized clinical terminology. A manual review of 100 randomly selected notes confirmed preservation of clinical meaning, along with improvements in clarity, structure, and readability. No hallucinated or spurious content was identified by expert reviewers.

3.2. Preprocessing Improves HPO Term Extraction by Doc2Hpo

To evaluate whether preprocessing improves phenotype term extraction, we compared three parallel data flows:

• Raw Notes → Doc2Hpo → Raw Terms
• Preprocessed Notes → Doc2Hpo → Preprocessed Terms
• GPT-4o Extracted Phrases → Doc2Hpo → Ground Truth Terms

All extracted terms were mapped to the Human Phenotype Ontology (HPO) using Doc2Hpo. The ground truth set was defined as the union of HPO terms extracted by Doc2Hpo from GPT-4o-supplied candidate phrases. For each note, term sets from the raw and preprocessed notes were compared against the ground truth using standard classification metrics: precision, recall, accuracy, and F1 score.

Preprocessed notes yielded higher precision and accuracy compared to raw notes. Recall also improved, rising from 0.61 to 0.68. Although modest in absolute terms, this gain was statistically significant due to the large sample size (n = 1618) and resulted in a meaningful improvement in F1 score (Figure 4). The concurrent improvement in both precision and recall suggests that GPT-4o neither introduced spurious concepts nor omitted clinically important terms. These results support the hypothesis that preprocessing enhances the performance of NLP pipelines performing ontology-based concept extraction.

Major academic medical centers are already exploring applications of LLMs to free-text EHR data—both to power predictive analytics [70] and to improve documentation quality among house officers and early-career clinicians [71].

4. Limitations

Preprocessing of physician notes does not directly reduce documentation burden, which is fundamentally tied to EHR design, clinical workflows, and organizational factors [10,72,73,74]. Technologies such as ambient AI and real-time documentation coaching may provide more comprehensive burden relief [75]. Encouragingly, recent surveys suggest that AI tools are improving physician satisfaction with EHRs [76].

This study was conducted using 1618 de-identified neurology notes, most of which were from patients with multiple sclerosis. Broader validation across multiple specialties, note types, and care settings is needed. Ground truth terms were generated using GPT-4o phrase extraction followed by human review. While this approach enabled scalable benchmarking, future studies may benefit from full manual annotation by domain experts.

We did not formally assess computational cost or latency. In our setup, processing time was approximately 10 s per note and cost averaged $5 per 1000 notes. Actual costs will vary by vendor and system configuration. Despite these costs, many institutions are actively piloting or adopting LLM-based tools in production environments [71,77].

Although notes were de-identified, we did not obtain explicit consent from patients or physicians. Real-time deployment in clinical settings may require additional legal or regulatory safeguards. US HIPAA-compliant LLM offerings are available through major cloud providers, including Microsoft, Amazon, and Google.

This cross-sectional study does not capture two key dynamics: (1) model performance drift—as clinical language evolves, periodic re-tuning may be needed to maintain accuracy; and (2) behavioral adaptation—as physicians grow accustomed to GPT-based editing, their writing may change in ways that affect both the editor’s output and downstream data use. Longitudinal studies should examine these phenomena using periodic F1 audits, user surveys, and dynamic prompts.

5. Conclusions

We demonstrate that LLM-based preprocessing—focused on correcting errors, expanding abbreviations, and standardizing terminology—enhances the clarity of clinical notes without altering their meaning or introducing hallucinated content. Across 1618 de-identified neurology notes, this approach yielded a statistically significant improvement in F1 score for HPO concept extraction, rising from 0.40 to 0.61.

Preprocessing is a lightweight, non-intrusive intervention that improves note quality without requiring any workflow changes from clinicians. Editing can be implemented transparently, and the resulting improvements strengthen the EHR as input for downstream applications, including precision medicine, decision support, quality improvement, and clinical research.

Looking ahead, LLM-powered preprocessing may serve as a practical on-ramp for broader reuse of EHR free text in real-time applications—such as quality monitoring, error correction, ontology mapping, cross-lingual translation, and generation of FHIR-compatible structured data.