Artificial Intelligence–Assisted Current Procedural Terminology Coding in Pediatric Orthopaedic Surgery: Promise or Premature?

Abstract

Background

Accurate assignment of Current Procedural Terminology (CPT) codes is essential for reimbursement, compliance, and quality reporting in pediatric orthopaedic surgery. Manual coding remains resource-intensive and prone to error and variability. Large language models (LLMs) represent a potential solution for automating CPT assignment directly from operative documentation. This study evaluated the performance of ChatGPT-4o and ChatGPT-5 in generating accurate CPT codes from deidentified pediatric orthopaedic operative notes.

Methods

A retrospective observational study was performed using 20 deidentified pediatric orthopaedic operative notes collected between January and April 2025. Notes were drawn from cases performed by 7 attending pediatric orthopaedic surgeons, representing a diverse academic case mix to expose the models to variability in operative dictation style and documentation standards. Each case was categorized as single-code (1 CPT) or multicode (≥2 CPT). Each operative note was entered into ChatGPT-4o and ChatGPT-5 using a standardized zero-shot prompt without priming or fine-tuning, and model-generated CPT codes were recorded. CPT codes assigned by the institutional billing department served as the gold standard reference. Model outputs were adjudicated as true positives, false positives, or false negatives. Performance metrics included F1 score, precision, recall, exact-match accuracy, and rates of overcoding and undercoding.

Results

ChatGPT-4o achieved an overall F1 score of 0.48 (precision: 0.45, recall: 0.56), with higher accuracy in single-code cases (F1 score: 0.53) than in multicode cases (F1 score: 0.41). Overcoding occurred in 70% of cases and undercoding occurred in 60%, and exact-match accuracy was 30%. ChatGPT-5 outperformed ChatGPT-4o across all measures, achieving an overall F1 score of 0.69 (precision: 0.67, recall: 0.74), with higher accuracy in single-code cases (F1 score: 0.76) than in multicode cases (F1 score: 0.61). Overcoding occurred in 25% of cases and undercoding occurred in 30%, and exact-match accuracy improved to 60%.

Conclusions

ChatGPT-5 demonstrated markedly improved performance over ChatGPT-4o in CPT code prediction for pediatric orthopaedic operative notes, particularly in single-code procedures. While accuracy gains were substantial, error rates remain above acceptable thresholds for unsupervised deployment. These findings suggest that LLMs may serve as augmentative tools to streamline coding workflows and reduce administrative burden, with potential earlier applicability in single-code pediatric orthopaedic cases. Larger-scale validation and integration studies are warranted to support broader clinical implementation.

Key Concepts

• (1)ChatGPT-5 substantially outperformed ChatGPT-4o in Current Procedural Terminology (CPT) code prediction from pediatric orthopaedic operative notes, achieving an F1 score of 0.69 versus 0.48 and an exact-match accuracy of 60% versus 30%.
• (2)Both models performed better on single-code procedures than on multicode procedures, with ChatGPT-5 achieving an F1 score of 0.76 in single-code cases compared to 0.61 in multicode cases.
• (3)Overcoding was the dominant error pattern for ChatGPT-4o (70% of cases), while ChatGPT-5 substantially reduced both overcoding (25%) and undercoding (30%), suggesting improved contextual bundling recognition.
• (4)Current error rates remain above acceptable thresholds for unsupervised clinical deployment; however, supervised integration of large language models (LLMs) as a first-pass coding assistant may reduce administrative burden and variability in routine pediatric orthopaedic cases.
• (5)Larger multi-institutional validation studies with domain-specific fine-tuning and electronic health record integration will be necessary before LLM-assisted CPT coding can be responsibly implemented in clinical billing workflows.

Level of Evidence

III, Retrospective Comparative Study

Introduction

Accurate assignment of Current Procedural Terminology (CPT) codes is essential for reimbursement, institutional budgeting, and quality reporting in pediatric orthopaedic surgery [[1], [2], [3]]. These codes form the foundation of healthcare analytics, directly influencing hospital revenue, surgeon productivity metrics, and participation in value-based care programs [4,5]. Despite their importance, CPT coding remains a predominantly manual process dependent on coder expertise and documentation clarity, resulting in considerable interobserver variability and administrative inefficiency [[6], [7]].

In orthopaedic surgery, coding inaccuracies are common, with documented error rates varying widely across institutions and procedure types [[6], [8],]. Multicenter audits have identified significant rates of procedure miscoding, with financial and compliance implications that are well established in the orthopaedic literature [[8], [9]]. Even subtle documentation inconsistencies, such as omitted modifiers, ambiguous procedural language, or variable terminology, can lead to underpayment, overbilling, or compliance risk [[7], [10]]. Pediatric orthopaedics presents unique challenges in this regard as procedures often integrate several orthopaedic modalities like fracture fixation, deformity correction, and soft-tissue reconstruction within a single operative session; moreover, many pediatric orthopaedic procedures lack well-established CPT codes, contributing to considerable variability in how surgeons code for the same procedure [[2], [3], [6], [11]]. This variability further amplifies coding inaccuracies and increases the likelihood of misclassification compared to adult subspecialties.

Traditional solutions, including coder retraining, standardized operative templates, and post hoc auditing, can improve consistency but are resource-intensive and difficult to scale across large health systems [[7], [9]]. As documentation volume continues to expand, there is growing interest in using artificial intelligence (AI) to augment coding efficiency. Large language models (LLMs), such as OpenAI’s ChatGPT series, have proven an ability to interpret unstructured text and generate structured outputs, including diagnostic and procedural codes [[12], [13], [14]]. Early studies across neurosurgery, craniofacial surgery, and anesthesiology suggest that LLMs can achieve moderate coding accuracy, improving with each successive model generation [[15], [16], [17], [18], [19]].

However, the application of LLMs to pediatric orthopaedic surgery has not been systematically evaluated. Given the specialty’s high procedural diversity and frequent multicode cases, pediatric orthopaedics provides a rigorous test of model interpretability and contextual reasoning. This study aimed to evaluate the performance of two contemporary LLMs, ChatGPT-4o and ChatGPT-5, in generating CPT codes directly from deidentified pediatric orthopaedic operative notes. We hypothesized that ChatGPT-5 would outperform ChatGPT-4o in accuracy, precision, and recall, particularly in single-code cases, and that qualitative analysis of coding errors would find specific documentation patterns limiting model performance.

Methods

Study design

This retrospective observational study evaluated the accuracy of LLMs in generating CPT codes from pediatric orthopaedic operative notes (see Fig. 1). The study was conducted from January to August 2025. A total of 20 deidentified pediatric orthopaedic operative notes were selected from a larger data set of over 400 procedures performed between January and April 2025. Each note represented a distinct pediatric surgical case performed by attending pediatric orthopaedic surgeons at our institution.

Figure 1.

Study design and evaluation workflow. Flow diagram illustrating case selection, deidentification, model input, generation of CPT code predictions by ChatGPT-4o and ChatGPT-5, comparison with institutional gold standard codes, and calculation of performance metrics. Created in BioRender. Manem, K. (2026) https://BioRender.com/gho22id. CPT, Current Procedural Terminology.

The study aimed to assess whether general-purpose LLMs could replicate or approximate institutional billing department accuracy in CPT assignment. We hypothesized that LLMs would demonstrate high precision for single-code procedures and lower accuracy for multicode procedures due to greater contextual ambiguity.

Data source and collection

An a priori power analysis confirmed that 20 operative notes provided sufficient precision to estimate a moderate effect size (Cohen’s d = 0.6) between models at 80% power and alpha = 0.05 for this exploratory comparison. Consistent with the design, formal inferential testing was not the primary analytic goal; a Wilcoxon signed-rank test was subsequently performed on per-case F1 scores as a paired nonparametric comparison. Operative notes were extracted from the institutional electronic health record (EHR) and stored in Research Electronic Data Capture (REDCap) for structured analysis. The 20 operative notes represented cases performed by 7 attending pediatric orthopaedic surgeons, with individual contributions ranging from 2 to 5 cases per surgeon, ensuring exposure to variability in operative dictation style, terminology, and documentation format. Only notes that contained clearly delineated procedure details were included. For model input, three standardized fields were extracted from each operative note: the postoperative diagnosis, procedure name, and description of procedure. Adult cases (occasionally performed by pediatric ortho faculty) and incomplete notes were excluded.

Each operative note was linked to its gold standard CPT code(s), as determined by our institutional billing department. Discrepancies were adjudicated by a panel of three independent reviewers until consensus was achieved. Cases were classified using a standardized framework including coding complexity (single-code, 1 CPT code, or multicode, 2+ CPT codes) and primary procedure type (fracture, soft tissue, spine, or deformity/reconstruction and other). This framework ensured consistent stratification and reproducibility in subsequent analyses. A complete breakdown of all 20 cases by procedure type, gold standard CPT codes, and complexity classification is provided in Supplementary Table S1.

Model setup and query structure

Two LLMs were tested: ChatGPT-4o and ChatGPT-5 (OpenAI, San Francisco, CA, USA). ChatGPT was selected based on a review of the existing surgical coding literature, in which it demonstrated the highest CPT coding accuracy among publicly available LLMs across multiple surgical subspecialties, including a direct five-model comparison by Isch et al. [15] in craniofacial surgery. The within-family comparison of ChatGPT-4o and ChatGPT-5 was designed to isolate the effect of model generation advances on coding performance while controlling for prompt structure and interface. Each model was accessed via a secure institutional interface using identical prompt structures. Models were neither primed with prior examples or external datasets nor evaluated against authoritative coding references such as the American Medical Association (AMA) CPT manual, reflecting how a clinician or trainee would be most likely to encounter and use these tools in practice.

Each operative note was entered using the following standardized prompt:

“You are an experienced medical coder. Read the following operative note and output the appropriate 2025 CPT codes with modifiers. Give a 1 sentence rationale for your selection.”

The prompt structure was informed by similar zero-shot prompting approaches used in prior LLM-based medical coding studies [14,18]. All outputs were manually copied into REDCap for evaluation. The models were blinded to the gold standard codes throughout the process.

Adjudication and scoring

The institutional billing department’s finalized CPT codes served as the gold standard reference. The coders designated as the gold standard each possessed at least 10 years of professional coding experience, with a combined total exceeding 40 years, predominantly in orthopaedic coding. Three independent reviewers compared each LLM output to the gold standard codes, classifying results as a true positive, a false positive, or a false negative. Disagreements were discussed and adjudicated by consensus with attending oversight. Model-generated rationales accompanying CPT code selections were recorded and reviewed by adjudicators during scoring, informing the qualitative error analysis; however, they were not formally scored as an independent outcome as adjudication was based solely on code-level accuracy.

Evaluation metrics

For each model and case, precision, recall, and F1 score were calculated using the adjudication categories (see Table 1). An exact match percentage was also recorded as a proportion of cases in which all CPT codes exactly matched the gold standard. Definitions and formulas are outlined in Table 2. Additionally, the proportion of overcoding and undercoding events was analyzed to identify systematic trends. Performance was further stratified by case complexity (simple vs complex).

Table 1.

Adjudication framework for model output evaluation.

Category	Definition
True positive (TP)	Model-predicted CPT code matches the gold standard code assigned by the institutional billing department.
False positive (FP)	Model generated a CPT code not present in the gold standard (ie, an extra or incorrect code).
False negative (FN)	A gold standard CPT code was missing from the model output.

CPT, Current Procedural Terminology.

Definitions used to classify model-generated CPT code outputs relative to gold standard institutional billing codes.

Table 2.

Performance metrics for model evaluation.

Metric	Formula	Interpretation
Precision (PPV)	TP/(TP + FP)	Proportion of predicted codes that were correct.
Recall (sensitivity)	TP/(TP + FN)	Proportion of true codes the model retrieved.
F1 score	2 × (precision × recall)/(precision + recall)	Harmonic mean of precision and recall.
Exact-match accuracy	Notes with all CPT codes correct/total notes	Case-level complete agreement.
Overcoding rate	Cases with FP/total cases	Frequency of extra code prediction.
Undercoding rate	Cases with FN/total cases	Frequency of missed code prediction.

CPT, Current Procedural Terminology; FN, false negative; FP, false positive; PPV, positive predictive value; TP, true positive.

Definitions and formulas for precision, recall, F1 score, exact-match accuracy, overcoding rate, and undercoding rate used to assess model outputs relative to gold standard CPT codes.

Statistical analysis

All statistical analyses were performed using R version 4.4.0 (R Foundation for Statistical Computing, Vienna, Austria). Descriptive statistics were used to summarize model performance. A Wilcoxon signed-rank test was used to compare per-case F1 scores between models, selected for its suitability with paired, non-normally distributed continuous data at this sample size. Visualizations, including accuracy and F1-score comparisons, were generated using the ggplot2 package in R.

Data management and reproducibility

All deidentified data, prompts, and model outputs were stored in REDCap and analyzed locally. Statistical code and analysis scripts are available upon reasonable request. No model fine-tuning, retraining, or external data ingestion was performed to preserve reproducibility and ensure clinical applicability.

Results

Quantitative performance

Twenty deidentified pediatric orthopaedic operative notes were analyzed across two LLMs: ChatGPT-4o and ChatGPT-5. Eleven notes (55%) represented single-code procedures, and nine represented multicode procedures. Both models were assessed for precision, recall, F1 score, and exact-match accuracy relative to institutional gold standard CPT codes. Overall, ChatGPT-5 substantially outperformed ChatGPT-4o across all performance metrics, with the greatest improvement observed in single-code procedures.

ChatGPT-4o achieved an overall F1 score of 0.48 (precision: 0.45, recall: 0.56) with an exact-match accuracy of 30% (see Figure 2, Figure 3). Overcoding occurred in 70% of cases (mean: 1.25 excess codes per note) and undercoding in 60% (mean: 0.95 missed codes per note). ChatGPT-5 demonstrated consistent improvement across all measures, achieving an overall F1 score of 0.69 (precision: 0.67, recall: 0.74), with an exact-match accuracy of 60%. Overcoding was reduced to 25% of cases (mean: 0.3 extra codes per note) and undercoding to 30% (mean: 0.45 missed codes). Formal statistical comparison confirmed significantly higher per-case F1 scores for ChatGPT-5 than for ChatGPT-4o (Wilcoxon signed-rank test: W = 56, P = .020).

Figure 2.

Overall CPT prediction performance by model. Comparison of overall precision, recall, F1 score, and exact-match accuracy for ChatGPT-4o and ChatGPT-5 in CPT code prediction from pediatric orthopaedic operative notes. CPT, Current Procedural Terminology.

Figure 3.

Comparison of F1 scores for ChatGPT-4o and ChatGPT-5 in single-code versus multicode cases (left), demonstrating improved performance and a larger effect size for ChatGPT-5 in single-code procedures. Comparison of model error rates (right).

When stratified by coding complexity, both models demonstrated higher accuracy in single-code cases than in multicode cases. For single-code procedures, ChatGPT-4o achieved an F1 score of 0.53, compared with 0.76 for ChatGPT-5. For multicode procedures, ChatGPT-4o achieved an F1 score of 0.41, compared with 0.61 for ChatGPT-5. This pattern indicates that ChatGPT-5 was better able to parse contextual and procedural cues, though multicode documentation remained more challenging due to embedded and overlapping procedural language.

Further subgroup analysis by procedure classification revealed variable model performance across case types (Table 3). ChatGPT-5 demonstrated the greatest relative improvement in fracture cases (ΔF1 = +0.381), followed by spine (ΔF1 = +0.267) and soft tissue (ΔF1 = +0.222). Notably, ChatGPT-5 achieved a recall of 1.000 in soft-tissue procedures, correctly identifying all gold standard codes in that category, though a precision of 0.800 reflects residual overcoding. The Deformity/Reconstruction and Other category demonstrated the smallest improvement between models (ΔF1 = +0.038), driven by consistently poor performance on complex reconstructive cases, which involve unlisted or infrequently used CPT codes unlikely to be well represented in the model’s training data. Simpler procedures within this category (for example, hardware removal and hip arthrotomy) were correctly coded by both models.

Table 3.

Model performance by procedural classification.

Classification	N	ChatGPT-4o			ChatGPT-5			ΔF1
		Precision	Recall	F1 score	Precision	Recall	F1 score
Fracture	7	0.214	0.429	0.286	0.625	0.714	0.667	+0.381
Soft tissue	4	0.600	0.750	0.667	0.800	1.000	0.889	+0.222
Spine	3	0.429	0.429	0.429	0.667	0.727	0.696	+0.267
Deformity/reconstruction and other	6	0.357	0.500	0.417	0.417	0.500	0.455	+0.038

Precision, recall, and F1 scores for ChatGPT-4o and ChatGPT-5 across fracture, soft tissue, spine, deformity/reconstruction and other procedures, with ΔF1 representing the difference in F1 scores between models for each category.

These findings suggest that ChatGPT-5’s largest gains occurred in categories with more standardized CPT code sets and operative language, whereas complex reconstructive procedures involving rare or unlisted codes remained the most challenging for both models.

Qualitative error analysis

A qualitative review of model outputs identified recurring error patterns that differentiated ChatGPT-5 from ChatGPT-4o. While quantitative metrics captured overall performance, this analysis highlighted the linguistic and contextual factors underlying model behavior. Contextual comprehension improved in ChatGPT-5. The ChatGPT-5 model more consistently linked operative verbs and anatomical references to their corresponding CPT concepts (for example, correctly associating “open reduction and fixation” with fracture repair codes), reducing semantic misclassification.

Redundant and bundled predictions also decreased. ChatGPT-5 more reliably recognized when secondary procedures, such as wound closure or hardware fixation, were already encompassed within the primary CPT code. This improvement lowered false-positive and overcoding rates. Despite these improvements, multicode ambiguity persisted. Both models struggled when several discrete operations (for example, osteotomies and hardware removal) were documented within a single paragraph in addition to the primary procedure, often omitting one or more applicable codes. These errors likely reflect inherent challenges in parsing long, unstructured operative narratives that lack clear procedural segmentation. These patterns support the quantitative findings and illustrate how ChatGPT-5’s improved contextual reasoning reduces certain error types while leaving multiprocedure narratives particularly vulnerable to omission.

Discussion

This study demonstrates that LLMs can meaningfully interpret operative documentation to generate CPT codes in pediatric orthopaedic surgery. ChatGPT-5 substantially outperformed its predecessor, achieving higher precision, recall, F1 scores, and exact-match accuracy, with particularly strong performance in single-code cases. These findings indicate that iterative improvements in model architecture and contextual reasoning directly translate to more accurate surgical text interpretation. However, this gain likely reflects broader improvements in general-purpose reasoning rather than coding-specific competency since neither model was evaluated against an authoritative coding reference. Both models therefore remained substantially less accurate than the institutional coders against whom they were compared.

Previous investigations have shown similar trends across surgical disciplines. Isch et al. [15] reported that ChatGPT-4 achieved the highest accuracy among public LLMs for craniofacial CPT coding, particularly in complex multicode cases. Zaidat et al. [20] demonstrated that ChatGPT-4 accuracy in spine surgery improved substantially following exposure to prior operative examples, underscoring the benefit of contextual priming. Our findings extend this literature to pediatric orthopaedics, a specialty with distinct linguistic and procedural variability, showing that ChatGPT-5 achieved an overall F1 score of 0.69 and 60% exact-match accuracy without any model training or prompt priming. Together, these results suggest that newer-generation LLMs possess sufficient contextual understanding to generalize across surgical subspecialties, even in complex procedural environments.

From a clinical operations perspective, partial automation of CPT generation could substantially reduce administrative workload and variability in pediatric orthopaedic billing. Manual coding remains labor-intensive and error-prone, often delaying claims and introducing compliance risk. Integrating LLMs into a supervised coding workflow—where the model provides preliminary CPT suggestions reviewed by a trained coder—could streamline billing review and enhance accuracy in routine, single-code cases. A significant advantage is real-time query capability, allowing LLMs to deliver immediate feedback and reduce the delays typically encountered in human-driven coding queries. Additionally, these tools may serve as educational resources, offering immediate feedback to trainees and clinicians on how operative documentation influences CPT assignment.

The viability threshold for LLM-assisted CPT coding depends on the intended deployment model. For autonomous unsupervised use, the American Health Information Management Association establishes a professional coding accuracy benchmark of 95% or above [[1], [19], [21]]—a standard that ChatGPT-5’s current precision of 67% and exact-match accuracy of 60% fall well short of, consistent with our conclusion that unsupervised deployment remains premature. For supervised integration, where the LLM serves as a first-pass screening tool reviewed by a credentialed coder, the threshold is considerably lower and governed by workflow efficiency rather than a fixed accuracy number. In this model, output that is correct 60% to 70% of the time still reduces cognitive load and narrows the coder’s review to exceptions rather than requiring full de novo code selection. For single-code procedures, where ChatGPT-5 achieved an F1 score of 0.76, supervised integration may already represent a practical efficiency gain in routine clinical settings. Achieving thresholds suitable for broader supervised use will likely require model fine-tuning on specialty-specific operative documentation, structured templating of operative notes, and prospective validation across institutions and procedure volumes.

Despite these encouraging results, several limitations must be acknowledged. First, the sample size of 20 operative notes limits statistical generalizability. Larger, multi-institutional studies will be necessary to validate these findings across diverse documentation styles. Second, model performance was assessed using static versions of ChatGPT-4o and ChatGPT-5; ongoing model updates may yield variable outcomes. Third, both models were tested without priming or fine-tuning, which preserved reproducibility but likely constrained maximal achievable accuracy. Because authoritative CPT descriptors and coding conventions are not well represented in public pretraining corpora, these results should be read as a floor for LLM-assisted coding performance rather than a ceiling and should not be generalized to systems explicitly constrained to such references. Finally, LLMs continue to exhibit difficulty parsing multiprocedure notes containing several distinct operations within a single narrative, a limitation inherent to free-text operative reporting that warrants further study. Furthermore, prompt sensitivity was not formally evaluated; results reflect the single standardized prompt used, and performance may differ with alternative phrasings. Additionally, while the prompt requested CPT codes with modifiers, modifier accuracy was not formally evaluated in this study as modifier assignment often depends on payer-specific rules, procedural context beyond the operative note, and factors outside the scope of this initial evaluation. While the institutional billing department’s codes served as the gold standard, it should be noted that this study was primarily designed to evaluate whether LLMs can approximate existing human-level coding practice rather than to adjudicate absolute coding accuracy, a distinction that contextualizes the use of institutional coders as the reference standard.

Future directions include expanding the dataset to several hundred operative notes encompassing a broader spectrum of pediatric orthopaedic procedures, mapping specific documentation patterns associated with misclassification, and developing fine-tuned domain-specific models trained on operative syntax. At the time of this study, constraining a model to the full CPT code set was not feasible through standard interfaces, given the length and licensing of AMA CPT content as well as institutional restrictions. A retrieval-augmented implementation surfacing only relevant code descriptors represents a more practical path and would allow direct benchmarking against the baseline established here. Integration with EHRs could allow real-time model-assisted CPT suggestions paired with coder verification, creating a hybrid human-AI workflow that improves efficiency while maintaining oversight.

In summary, this study provides early evidence that LLMs can augment CPT coding in pediatric orthopaedic surgery. While unsupervised use remains premature, ChatGPT-5 demonstrated clear quantitative and qualitative improvements over prior models, suggesting rapid progress toward clinically viable performance. Continued refinement, validation, and ethical implementation will determine how these systems ultimately integrate into surgical documentation and billing workflows.

Conclusion

LLMs demonstrate growing potential as assistive tools in pediatric orthopaedic billing and documentation. In this initial pilot evaluation, ChatGPT-5 substantially outperformed ChatGPT-4o in generating CPT codes from operative notes, achieving higher accuracy, precision, and recall, particularly for single-code cases. These findings suggest that contemporary LLMs are beginning to capture the procedural syntax and contextual nuances of operative documentation.

Although general-purpose LLMs without coding-specific training are not sufficient for autonomous CPT coding, supervised integration of such models could still reduce administrative burden, enhance billing accuracy, and provide educational feedback for trainees and coders. Continued validation using larger, multi-institutional data sets, combined with domain-specific model training and real-time workflow integration, will be essential for establishing their reliability and safety in clinical practice. Ultimately, human-AI collaboration may offer a scalable path toward more efficient, standardized, and transparent coding processes in pediatric orthopaedic surgery.

Additional links

• •AAOS: Coding and Reimbursement

Author contributions

Krishi S. Manem: Writing – review & editing, Writing – original draft, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Gloria R. Gogola: Writing – review & editing, Supervision, Software, Resources, Data curation. Rohini M. Vanodia: Writing – review & editing, Supervision. Adarsh Suresh: Writing – review & editing, Writing – original draft, Investigation, Data curation. Surya N. Mundluru: Writing – review & editing, Supervision.

Ethics approval and consent

This study was reviewed and approved as exempt research by the institutional review board. All operative notes were fully deidentified prior to model testing. No patient identifiers, demographic data, or protected health information was included in model inputs. Provider and institutional identifiers were removed from all operative notes. Because the project involved analysis of deidentified text only, it did not constitute human participant research under federal guidelines (45 CFR 46). All analyses used institutionally licensed, secure deployments of ChatGPT-4o and ChatGPT-5.

The authors declare that no patient consent was necessary as no images or identifying information is included in the article.

Funding

This research received no external funding from any source, including the National Institutes of Health, Wellcome Trust, Howard Hughes Medical Institute, or any other organization.

Declaration of competing interests

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Supplementary material

The following are the Supplementary data to this article: