Artificial Intelligence in Medicine

Background Hepatocellular carcinoma (HCC) treatment selection demands nuanced integration of heterogeneous patient data, yet prevailing predictive models rely on restricted data modalities and oversimplified therapeutic frameworks, compromising clinical translation. Objective We developed and validated a multimodal artificial intelligence framework to guide optimal treatment strategy selection across the full spectrum of HCC interventions. Methods This retrospective study comprised 1,043 HCC patients (development cohort, January 2017-December 2023) and 55 external validation patients (2023) from Wuxi Peoples Hospital. We engineered Embedding-Augmented Extra Trees (ET-Emb), a novel model fusing structured clinical variables with contextual text embeddings derived from medical histories and radiology reports. ET-Emb quantifies probabilities for five primary treatments: open/laparoscopic resection, transarterial chemoembolization, radiofrequency ablation (RFA), and chemotherapy. Model performance was rigorously assessed via 10-fold cross-validation and external validation using ROC-AUC and PR-AUC metrics. Results ET-Emb demonstrated robust performance in the development cohort (ROC-AUC: 0.84 {+/-} 0.04; PR-AUC: 0.55 {+/-} 0.06), significantly outperforming established benchmarks. This generalizability was preserved in external validation (ROC-AUC: 0.77 {+/-} 0.02; PR-AUC: 0.47 {+/-} 0.03). SHAP analysis identified textual clinical narratives and socioeconomic determinants as critical predictive drivers. Conclusions By unifying structured and unstructured data modalities, ET-Emb delivers accurate, multi-treatment strategy prediction for HCC. Its clinical validity and the demonstrated significance of textual features establish multimodal AI as an essential paradigm for simulating complex oncological decision-making, positioning ET-Emb as a transformative tool for precision HCC management.

Connected speech is essential for everyday communication, but clinical constraints and patient fatigue limit detailed evaluation in acute stroke (<1-week post-stroke). Bedside assessments may sample discourse but rarely quantify language impairment (LI) in connected speech, leaving patient communication poorly characterized. We analyzed brief story retellings from 86 patients with left-hemisphere stroke (~4 days post-stroke; 63 classified with LI using composite clinical and naming criteria). From transcripts generated with automatic speech recognition, we derived discrete linguistic features and embeddings with Large Language Models (LLMs). An ensemble of embedding-based classifiers distinguished patients with and without LI with 90% balanced accuracy (79% sensitivity, 100% specificity), outperforming independent embedding and discrete-linguistic-based classifiers, showing distinct LLMs contributed complementary information. Adding the discrete-linguistic-based classifier to the ensemble did not improve balanced accuracy but modestly increased sensitivity at the expense of specificity. We provide proof of concept for a fast, largely automated discourse screener of acute LI.

Background: Free-text notes in electronic health records (EHRs) contain fine-grained psychiatric information that is essential for psychiatric research and clinical care, and often absent or under-recorded in structured codes alone. Clinical natural language processing (cNLP) can support extraction of this information from EHR notes, yet Spanish-language cNLP remains under-developed. Moreover, broad evaluations comparing multiple encoder-based language models across extensive, fine-grained psychiatric concept sets remain scarce, and it remains unclear how these models compare with traditional NLP (tNLP) systems and much larger generative large language models (LLMs). In addition, cross-site performance of fine-tuned models is rarely tested, and limited annotated training data remains a major challenge, especially for rare symptoms. Objectives: We aimed to advance scalable, global psychiatric cNLP by fine-tuning multiple encoder-based models with differing architectures and pre-training strategies for detecting fine-grained psychiatric concepts in Spanish EHRs. We further evaluated the impact of augmenting the fine-tuning data with precision-weighted weak labels for less-frequent concepts, and compared the performance of the encoder-based models to that of tNLP and a fine-tuned generative LLM trained on the same data. Finally, we evaluated model cross-site generalizability on an external EHR dataset. Methods: Three encoder-based models (BETO, XLM-RoBERTa-large, and bsc-bio-ehr-es) were fine-tuned on 1,642 clinician-annotated EHR documents from Colombia to detect 110 psychiatric concepts in Spanish text. To address the limited annotated examples available for less-frequent concepts, 12,000 additional documents were weakly-labeled for less-frequent concepts using tNLP, and incorporated into the fine-tuning data with labels weighted by pattern precision. Models were compared with tNLP and a generative LLM, and evaluated on an external EHR dataset from another psychiatric hospital in Colombia. Results: Encoder model performance varied substantially, with macro-F1 ranging from 0.64 to 0.81. BETO achieved the highest macro-F1 (0.81; median F1=0.88 [IQR=0.77-0.96]). Adding precision-weighted weak labels for less-frequent concepts improved BETO's overall macro-F1 to 0.83 and increased mean F1 for the 55 augmented concepts from 0.82 to 0.86. Under matched fine-tuning conditions, fine-tuned BETO and the tNLP method were equivalent in F1, whereas the LLM significantly outperformed BETO in F1. After weak-label augmentation, BETO significantly outperformed tNLP in F1 (PFDR<.001) and narrowed the performance gap with the LLM, although equivalence was not established. Lastly, fine-tuned BETO maintained reasonably strong performance on data from an external hospital not used for model fine-tuning (out-of-domain macro-F1=0.78). Conclusions: General-purpose pre-trained encoders had strong performance for psychiatric concept extraction from Spanish EHRs. Weak-label augmentation improved BETO's performance and strengthened results relative to a tNLP baseline, while reducing, but not eliminating, the performance gap with a much larger fine-tuned generative LLM. These findings highlight the utility of these relatively lightweight models for scalable, accurate and reproducible detection of psychiatric concepts in Spanish-language EHRs.

The choice of FHIR-to-text serialisation format significantly impacts clinical LLM quality (Kruskal-Wallis H=163.86, p<10^-33, delta=0.24 on a 5-point scale), yet remains unstudied as a clinical deployment variable. We present FHIRBench-UK, evaluating five large language models across six serialisation formats and three clinical tasks on 100 UK Core FHIR patient bundles (18,000 scored prompts across clean and perturbed cohorts). Our findings converge with independent work on open-weight models (Pator, 2026). The optimal format is context-dependent: raw_json dominates for clinical QA, hybrid_adaptive for clinical reasoning, and structured_markdown for summarisation. In 58% of model-task-complexity scenarios, raw_json is suboptimal. Model capability moderates format sensitivity: Claude Sonnet 4.5 shows 0.10-point sensitivity versus Llama 3.3's 0.39, making adaptive serialisation most valuable for budget-constrained deployments using mid-tier models. All findings replicate under clinically realistic data perturbation. The study additionally confirms a complete ranking inversion between token-level F1 and clinical quality (rho=-0.90), replicating US findings across UK Core profiles. We recommend task-aware serialisation routing as a zero-cost quality intervention for NHS FHIR-based LLM deployments.

Self-repairs, in-utterance revisions in which a speaker abandons and reformulates their speech, are a promising interpretable marker for speech-based cognitive screening. Detecting them automatically is difficult because a self-repair is defined by its relationship to surrounding speech rather than by fixed lexical cues. On the DementiaBank ADReSS corpus, we compared the capability of generative LLMs under a five-condition prompt ablation against a fine-tuned DistilBERT token classifier at detecting self-repairs. GPT-5 performed best (test F1 = 0.73) and was largely insensitive to prompt design, whereas the LLaMA (open-weight alternative) was both weaker and far more prompt-sensitive (test F1 = 0.47). DistilBERT, nearly 100 times smaller, matched the open-weight LLM at a fraction of the computational cost. These results suggest that a locally deployable encoder, given sufficient in-domain annotation, is a more plausible route to clinical self-repair detection than scaling model size or prompt complexity.

Background Naturalistic audiovisual recordings of caregiver-child interactions contain rich developmental signals. However, extracting interpretable clinical measures requires resource-intensive manual coding. To address this bottleneck, we evaluated natural-language queries for retrieving specific behavioral moments from these recordings, applying multimodal embeddings as an automated evidence-selection layer. Methods We compared three embedding models (Jina Embeddings v5 Omni, LanguageBind, and Wave7B) for natural-language retrieval directly from audio and video streams, bypassing transcript text. We assessed performance across 27 behavioral targets in 277 caregiver-child recordings (14, 24, and 36 months of age) from the Early Head Start Talkbank corpus, yielding 7,479 recording-target queries. Results Jina Embeddings v5 Omni achieved the highest top-10 retrieval success (text-to-audio 38.3%; text-to-video 36.4%), ahead of LanguageBind (37.0%; 34.5%) and Wave7B (36.1%; 35.0%). Across models, retrieval was substantially more successful for common targets than for rare vocal and gestural behaviors, such as pointing and babbling. By analyzing the spoken words within the retrieved audio clips, we found that Jina accurately ranked the children by their relative vocabulary size at each age (Spearman = 0.68, 0.82, and 0.90 at 14, 24, and 36 months). However, the model severely underestimated the total number of unique words each child used throughout the full session. Conclusion Multimodal embeddings can successfully pinpoint important developmental behaviors and speech patterns within lengthy caregiver-child recordings. However, these systems still struggle to locate rare events. Additionally, while they can accurately rank children by relative vocabulary size, they fail to measure a child's complete vocabulary. We conclude that these models are currently best suited for automated evidence-selection to prioritize relevant segments for expert interpretation rather than acting as an independent replacement for manual behavioral coding or language assessment. Improving the detection of infrequent behaviors and validating these models across external datasets are essential next steps before real-world clinical deployment.

Extracting clinical information from Dutch free-text medical notes requires language-specific annotation resources, yet Dutch primary care lacks a reusable event-annotation framework for infections, post-acute infection syndromes (PAIS), and related symptoms. We adapted the COVID-19 Annotated Clinical Text (CACT) framework to Dutch and applied it to GP notes for PAIS event extraction. The framework has three annotation layers: a DiagnosticExpression typology covering acute infections, post-acute syndromes, and relevant comorbidities; an eleven-subtype Evidence inventory grounded in Dutch primary-care testing practice; and explicit decision rules for the SOEP structure of Dutch general practitioner (GP) notes (Subjective, Objective, Evaluation, Plan), including the distinction between clinician hedging and patient-side hypotheticals. On a 200-note pilot, span-level F1 under the Lybarger criterion reached 0.51 [95% CI: 0.47, 0.55] across six core entities; restricted to spans both annotators noticed, conditional F1 reached 0.78 [0.75, 0.80], indicating that most disagreement stems from annotation coverage rather than label assignment. The adaptation illustrates how an English event-based clinical annotation framework can be extended to a new language and clinical setting, yielding a reusable resource for Dutch clinical NLP; which steps generalise beyond this case (CACT to Dutch primary care) and which are specific to Dutch or PAIS remain to be tested.

Motivation: Graph Neural Networks (GNNs) have gained increasing interest in the biomedical domain, as the integration of prior knowledge and deep neural networks has the potential to enhance insights into molecular processes and disease mechanisms. However, a comprehensive and systematic assessment of model architectures, data modalities, graph structures, and their performance for graph signal classification in the biomedical domain is yet to be performed. In order to close this gap, we conducted a benchmarking study on multiple GNNs on a Protein-Protein Interaction (PPI) network for Kidney Renal Clear Cell Carcinoma and Breast cancer subtype prediction, performing an in-depth investigation of architectures, incorporating skip connections and various data modalities. Results: While none of the GNNs outperforms the structure-agnostic Multi-Layer Perceptron baseline, all of them can handle bimodal data (gene methylation and expression) and offer the ability to gain explainability based on PPIs. We offer practical guidelines for applying GNNs to graph signal processing tasks specifically for cancer classification. Depending on the underlying dataset and PPI structure employed, models on different data modalities outperform others. Overall, we suggest using ChebNet, which tends to outperform the Graph Convolutional Network and the Graph Attention Network in cancer subtype prediction. We recommend using GNN architectures that employ a simple flattening readout layer, as they provide better classification performance and faster training time than those with global average pooling. Additionally, we tested residual connections, but they had only an insignificant impact on classification performance.

Background: People affected by cancer often face difficulty finding relevant clinical, psychological, and practical support services. AI-powered navigation assistants may improve access to these resources, but their retrieval performance must be reliable. Objective: To develop a single-pass retrieval-augmented generation assistant for cancer-care navigation and compare the retrieval strategies, including their robustness to reworded questions. Methods: We created a database of 853 cancer-support resources reviewed by librarians, clinicians, researchers, and patient partners. We evaluated the system using 100 questions derived from questions submitted by patients. We compared keyword-based, semantic, and hybrid retrieval using Precision@K, Hit@K, and nDCG@K. The best-performing configuration was then tested using semantically equivalent rewordings of the original questions. Results: Keyword-based retrieval performed poorly, achieving a P@1 of 25.0% and Hit@5 of 43.0%. Semantic retrieval improved these results to 58.0% and 86.0%, respectively. The best hybrid configuration achieved a P@1 of 64.0%, Hit@5 of 90.0%, and nDCG@5 of 51.0%. Performance remained relatively stable when the questions were reworded, with a P@1 of 61.0%, Hit@5 of 88.0%, and nDCG@5 of 46.1%. Conclusions: Hybrid retrieval performed best and remained relatively stable when questions were reworded. However, its limited ability to rank a relevant resource first highlights the limitations of single-pass retrieval for patient-facing cancer navigation. Future work will explore metadata filtering and a multi-agent architecture to improve retrieval reliability.

Background: Large language models (LLMs) that use retrieval-augmented generation (RAG) are increasingly used to answer clinical questions, although the evaluation of these systems remains limited. Building on previous studies conducted by our team, this case report aimed to improve upon this knowledge gap by applying a reusable methodology to compare the performance of eight LLMs that utilize RAG techniques for evidence synthesis. Case Presentation: Eight commercially available RAG LLM tools (OpenEvidence, Undermind, Consensus, SciSpace, Elicit, MediSearch, EvidenceHunt, and Scite) were evaluated using twelve ChatGPT-generated clinical questions on the topics of treatment, etiology, and prognosis. To enable comparison, we prompted ChatGPT to identify all key unique medical concepts from the full set of LLM responses to each question. Concepts were categorized as critical ("must-have") or non-critical ("nice-to-have") for answering the clinical question. Experienced information scientists were consulted at each step for their expertise. Descriptive statistics and Kruskal-Wallis tests were used to compare performance across tools and question categories. No significant differences were found among the eight RAG LLMs in their coverage of "must-have" (p=0.95) or "nice-to-have" (p=0.16) key unique medical concepts, and no single tool consistently captured all identified concepts. Conclusions: These findings suggest that RAG LLMs may be supplementary tools for evidence retrieval and synthesis but cannot, at this time, fully replace comprehensive expert review of the medical literature. The evaluation framework presented here may be a useful model for future comparative assessments of rapidly evolving AI evidence synthesis tools.

While Directed Acyclic Graphs (DAGs) are essential for causal inference, their construction often relies on expert heuristics, which bypasses systematic evidence synthesis and creates a critical "evidence retrieval gap" in causal modeling. This study introduces EpiKG2DAG, a framework that supports evidence-anchored candidate DAG generation by transforming unstructured biomedical abstracts into structured epidemiological associations. We utilized DeepSeek-V3 to extract exposure-outcome association triplets from 189,266 abstracts and employed SapBERT for semantic normalization against UMLS concepts. The resulting Epidemiological Knowledge Graph (EpiKG) enables the automated identification of candidate confounders, mediators, and colliders based on graph-theoretic motifs and literature-derived evidence. A case study on COVID-19 and AKI demonstrates that the framework uncovers non-obvious confounders, such as air pollution, while ensuring evidence traceability. This work contributes to the field by mitigating the knowledge-acquisition bottleneck and providing a transparent, reproducible foundation for evidence-based causal modeling.

Conventional electrocardiography is highly effective for waveform and rhythm diagnosis, but it is less suited to showing how the internal shape of hundreds or thousands of consecutive heartbeats changes over time. We introduce FOXTAIL, a complementary view that represents each cardiac cycle as an ordered sequence of changes in signal direction. Overlaying these sequences in a fixed visual field makes beat-to-beat organization visible and allows the density, size, stability, and scale persistence of those changes to be measured. We evaluated the representation in recordings containing normal sinus rhythm, paroxysmal atrial fibrillation, severe heart failure, ventricular tachyarrhythmia, and controlled electrode-motion noise. Paired recordings showed that FOXTAIL descriptors can reveal within-person state changes that are not conveyed by a single average beat. The noise and pre-fibrillation analyses also showed that a dense event pattern is not automatically equivalent to physiological complexity, measurement artifact, or impending disease. FOXTAIL is therefore not proposed as a replacement for the diagnostic ECG or as a new classifier, but as an observation and measurement domain for asking a more basic question: how is the electrical organization of the heart changing from one beat to the next, and which of those changes persist across scale?

Large language models perform well on medical examinations, but users routinely challenge their answers and invoke professional roles, and it is unclear what a system does when a medical credential and a stated task-specific accuracy point in opposite directions. In a factorial experiment on 480 items from four Polish specialty examination sets and three consumer large language model systems (ChatGPT, Claude, Gemini), each item and system received eleven independent conversations. Conditions crossed attributed source role (medical student, experienced specialist), stated prior accuracy on similar questions (2/10, 8/10) and suggestion correctness. The primary outcome was adoption of a prespecified incorrect option when the baseline answer matched the official key, comparing a specialist described as 2/10 with a student described as 8/10. Baseline agreement with the key was 87.2% across 15,683 analyzable conversations. The incorrect option was adopted more often from the specialist described as 2/10 than from the student described as 8/10 (10.2% vs. 7.6%; adjusted risk difference +2.82 percentage points, 95% CI +0.65 to +4.99). Estimates varied across the three systems and only one system-specific interval excluded zero. In a prespecified exploratory analysis with a shared eligibility rule, correct suggestions were adopted far more often than incorrect ones (risk difference +35.7 percentage points, 95% CI +30.8 to +40.7), indicating selective rather than indiscriminate compliance. An incorrect suggestion from a specialist with low stated accuracy was therefore slightly more influential than the same suggestion from a student with high stated accuracy, although the difference was modest and varied across systems. Agreement reached only after a user has disclosed a preferred answer should not automatically be treated as an independent second opinion, and medical large language model systems should be evaluated on how they revise answers after such disclosure, not solely on initial accuracy.

Background: Adverse event (AE) coding is essential for safety monitoring in oncology clinical trials, particularly in acute myeloid leukemia (AML), where intensive therapies are associated with frequent and heterogeneous toxicities requiring standardized MedDRA (Medical Dictionary for Regulatory Activities) coding. However, manual Low-Level Term (LLT) assignment remains labor-intensive, subjective, and difficult to scale. Although large language models (LLMs) have emerged as promising decision-support tools for automated coding, unguided zero-shot generation remains insufficient for reliable fine-grained MedDRA coding. Objective: To develop and evaluate a retrieval-augmented reasoning pipeline for clinically aligned LLT-level MedDRA coding of free-text adverse events from prospective AML clinical trials. Methods: We implemented a retrieval-augmented reasoning pipeline inspired by the retrieval-augmented generation (RAG) paradigm using LLaMA-3.3-70B-Instruct as the primary backbone and benchmarked the framework across multiple open instruction-tuned LLMs. Dense semantic retrieval first generated a constrained top-100 LLT candidate set for each AE, followed by structured LLM reasoning to select a single best-matching LLT and deterministic mapping to Preferred Term (PT) and System Organ Class (SOC) levels. The pipeline was evaluated retrospectively on AE datasets from three prospective AML clinical trials (MOSAIC, DELTA, and DaunoDouble) with automated LLT/PT/SOC metrics and expert-assessed Clinical Correctness Rate (CCR). Results: Clinical expert review showed high clinical acceptability of the RAG pipeline across datasets (91-97%). Under automated evaluation, the pipeline achieved LLT exact accuracy of 50-58%, PT accuracy of 78-85%, and SOC accuracy of 90-93%. Zero-shot generation and random candidate selection performed substantially worse. Semantic retrieval more often included the coder-assigned LLT among the candidate terms available to the model than retrieval based on lexical similarity. Multi-model benchmarking showed that backbone choice mainly affected LLT exact agreement, whereas PT and SOC performance remained comparatively stable. Conclusions: Retrieval-augmented reasoning supports clinically aligned MedDRA coding of free-text adverse events under realistic candidate constraints in AML clinical trials. Evaluation across three AML clinical trials showed that strict LLT-level string agreement underestimated clinical ap-propriateness, highlighting the importance of combining hierarchical evaluation metrics with clini-cal expert validation for AI-assisted MedDRA coding in hematology trials.

Machine learning workflows frequently incorporate data preprocessing to enhance predictive performance. However, the need for Super Learner ensembles made up only of preprocessing-invariant tree-based algorithms remains unexplored. Using three benchmark clinical classification datasets, this study examined how preprocessing affected the Super Learner's prediction performance, learner weight distribution, and oracle behavior. The Heart Disease (207 observations), Indian Liver Patient Dataset (583 observations), and Pima Indians Diabetes (768 observations) datasets were used to create a Super Learner ensemble model that included Classification and Regression Trees (CART), Random Forest, Ranger, and Extreme Gradient Boosting (XGBoost). Models were evaluated under raw and preprocessed data conditions using repeated cross-validation. Predictive performance was assessed using the area under the receiver operating characteristic curve (AUC), Matthews correlation coefficient (MCC), and Brier score. Learner weight allocation and Oracle Gap were compared using paired Wilcoxon signed-rank tests with Benjamini-Hochberg adjustment. Preprocessing produced negligible changes in predictive performance for the Heart Disease and Pima datasets. For the ILPD dataset, preprocessing significantly improved AUC (0.746 to 0.752; adjusted p = 0.0017) and reduced the Brier score (0.177 to 0.175; adjusted p < 0.001). Learner weights remained largely stable, although Random Forest replaced Ranger as the dominant learner for the Heart Disease dataset. Oracle Gaps remained extremely small (<0.002) across all datasets and did not differ significantly between preprocessing conditions. Preprocessing provides limited benefit for Super Learner ensembles composed of preprocessing-invariant learners and does not materially alter their oracle behavior. Preprocessing decisions should therefore be guided by dataset characteristics rather than adopted as a universal modelling practice.

Symbolic clinical natural language processing (NLP) systems remain widely used for extracting clinical concepts from electronic health record (EHR) narratives, but maintaining rule resources requires extensive manual error analysis and rule refinement. This study investigates whether large language models (LLMs) can assist in identifying extraction errors and generating candidate rules to improve symbolic clinical NLP systems. Using error reports derived from a multi-site evaluation of a previously validated symbolic model for cognitive and neuropsychiatric-related clinical concepts, we developed a human-in-the-loop framework, REFINE. The framework first uses LLMs to classify extraction errors and generate explanatory reasoning, which can then be incorporated into prompts for rule generation. Three LLMs (GPT-5.2, GPT-4o, GPT-4o-mini) were evaluated under four prompting conditions. LLM-generated rule sets improved performance compared with the baseline NLP-CAM system, increasing F1-score from 0.37 to 0.58. These findings suggest that LLMs can support scalable rule refinement for symbolic clinical NLP systems.

Uncertainty quantification is proposed as a safeguard for machine-learning systems in health-related signal analysis, but an uncertainty score is useful only if it behaves as a reliability signal. Free-living wearable electrocardiogram (ECG) signal-quality assessment provides a test bed because ambiguity, artifact, and acquisition shift can alter the relationship between confidence and correctness. This study evaluates predictive uncertainty under ambiguity, controlled corruption, and external distribution shift. 32,224 non-overlapping 10-s windows of synchronised single-lead ECG and three-axis accelerometry from 15 subjects in the Brno University of Technology ECG Quality Database were analysed. Two model families were compared: multinomial logistic regression and Classification and Regression Tree (CART), each progressing from a point estimate to a fixed-structure posterior and then a structure posterior. Expected conditional entropy and mutual information were evaluated as designated aleatoric and epistemic uncertainty measures, with max-softmax uncertainty as a confidence baseline. Validation covered error ranking, selective prediction, behavioural probes, posterior structural diversity, recorded-noise stress testing, and zero-shot external transfer. The logistic structure posterior retained an expected 8.5 of nine features and concentrated on near-complete masks, yielding little additional predictive diversity. Bayesian CART produced 221 distinct complete topologies among 238 retained draws and stronger score-dependent selective-risk behaviour. Conditional entropy increased with local class overlap, whereas mutual information increased when training information was reduced, although both showed cross-sensitivity. Under recorded noise, predicted quality severity changed more consistently than uncertainty, while external transfer preserved ordinal severity more reliably than uncertainty ordering. These findings show that posterior richness alone does not establish reliable uncertainty. Model-derived uncertainty should therefore be validated against prespecified ambiguity, information, and shift probes before supporting abstention, reacquisition, or downstream decisions.

Automatic coding from clinical notes has been studied extensively for International Classification of Diseases (ICD) codes, yet broad Current Procedural Terminology (CPT) and Healthcare Common Procedure Coding System (HCPCS) recommendation remains comparatively underexplored. Existing studies often focus on one specialty, a limited code vocabulary, or a single model family, leaving it unclear how different artificial intelligence (AI) paradigms perform under a common, clinically meaningful evaluation. We formulate CPT and HCPCS coding as an AI-assisted recommendation task in which a physician or professional coder reviews a short, ranked list of candidate codes supported by the clinical note. Using operative notes from Vanderbilt University Medical Center (VUMC) and discharge summaries from Medical Information Mart for Intensive Care IV (MIMIC-IV), we compare lexical retrieval, Clinical-Longformer, GPT-5.6-Sol, MedGemma-27B, and an inspectable agentic-style retrieve-and-verify system under a controlled review budget. Micro-averaged recall within a fixed number of recommendations measures whether reference codes reach the reviewable list; micro-F1 is reported only where reference labels are sufficiently complete. Zero-shot GPT-5.6-Sol achieves the highest recall within five and ten candidates: 0.717 and 0.800 on VUMC and lower-bound values of 0.689 and 0.738 on MIMIC-IV. The retrieve-and-verify system reaches 0.695 and 0.784 on VUMC and lower-bound values of 0.575 and 0.657 on MIMIC-IV, with a candidate-linked evidence window attached to each retained recommendation. Diagnostic analyses reveal distinct failure sources, including output-length underfilling, confusion among closely related codes, out-of-knowledge-base generation, and incomplete evidence support. These findings establish a systematic evaluation framework for procedure-code recommendation and identify practical requirements for future systems that are accurate, review-efficient, and grounded in clinical evidence.

The performance of an EHR-based deep learning model trained on a small sample can be improved if more data is collected. Instead of collecting more data, the model can be trained on additional data from an analogous external source. However, this risks the model learning patterns in the external data that do not generalize to the target sample. Furthermore, data use agreements often prohibit combining datasets with medical records of different sources. We consider utilizing pre-existing methods in continual learning, namely the elastic weight consolidation (EWC) loss function and variational continual learning (VCL), both of which are regularization-based methods that we use to borrow external data and incorporate parameters from a model on external data into local model training. To investigate the utility of this modeling framework, we consider two binary classification tasks: (1) predicting which children will be diagnosed with autism spectrum disorder (ASD) from medical claims up to 18 months, and (2) predicting which patients with end-stage renal disease (ESRD) will be re-hospitalized within 30 days. Target datasets were derived from Duke University's EHR warehouse, and external datasets were sourced from either NC Medicaid claims for the ASD prediction task, or the United States Renal Data System (USRDS) for the rehospitalization prediction task. For both of these tasks, borrowing models - using either the EWC loss function or VCL - performed similarly to that of a model trained only on the full external data, when the sample size of target data used to train the model was small. That is, while a model that does not borrow using our methods performed poorly in low data regimes, the borrowing model instead matched the performance of a model trained on external data even when sample size of target data was small. In addition, an analysis of model predictions showed that models with small samples are better calibrated and more functionally similar to a model trained only on external data when the sample size is small.

Background: Large language models (LLMs) have shown increasing capability in medical knowledge tasks, yet how they perform in extracting structured clinical information from real-world clinical documentation remains uncertain. We evaluated the performance of LLMs relative to medical professionals in extracting SNOMED-coded clinical information from openly available Ear, Nose and Throat (ENT) EHRs from MTSamples, examining both reliability and accuracy metrics. Methods: We evaluated the performance of seven LLMs (including GPT-4o, Claude 3.5, Gemini 1.5 Pro, Gemma 3 and three LLAMA variants) against annotations from fourteen medical professionals who served as both study authors and data annotators. Each annotator independently extracted seven categories of clinical information from 98 publicly available ENT clinical documents: socio-demographics, symptoms, signs, diagnoses, treatments, risk factors, and test results. Standardised medical terminology was enforced through SNOMED-CT code assignment, enabling standardised comparison through Cohen's Kappa. We employed Bayesian hierarchical modelling to test non-inferiority of medic-LLM agreement compared to medic-medic agreement, using Beta distributed likelihood functions with weakly informative priors. Non-inferiority margins of 0.05, 0.10, and 0.15 were assessed with 95% posterior probability thresholds. Results: Cohen's Kappa for inter-rater reliability was 0.752 (95% CI: 0.710 - 0.794) between medical professionals and 0.391 (95% CI: 0.362-0.420) between LLMs and medical professionals. Bayesian analysis showed medic-medic agreement (posterior mean 0.813, 95% CI: 0.755-0.860) exceeded medic-LLM agreement (0.659, 95% CI: 0.633-0.684) by 0.154 (95% CI: 0.091-0.209). Non-inferiority was rejected at all tested margins (delta = 0.05, 0.10, 0.15). Agreement varied by clinical category, with smallest differences for test results and largest for diagnoses. GPT-4o achieved 97.0% precision and 84.9% recall, with a 7.5% false positive rate. Conclusions: Current LLMs do not achieve inter-rater reliability levels comparable to medical professionals in clinical information extraction from ENT documentation. These findings provide evidence-based guidance for LLM deployment in clinical documentation workflows, suggesting they are best suited for initial extraction with human verification rather than autonomous operation.