SHIELD: A Diverse Clinical Note Dataset and Distilled Small Language Models for Enterprise-Scale De-identification

De-identification of clinical text remains essential for secondary use of electronic health records (EHRs), yet public benchmarks such as i2b2 2006/2014 are over a decade old and lack the semantic and demographic diversity of modern narratives. While Large Language Models (LLMs) achieve state-of-the-art zero-shot extraction, enterprise deployment is hindered by compute costs and governance restricting Protected Health Information (PHI) from cloud APIs. We introduce SHIELD (Synthetic Human-annotated Identifier-replaced Entries for Learning and De-identification), a diverse dataset of 1,394 notes with 10,505 gold-standard PHI spans across 9 categories, built via set-cover diversity sampling with human-in-the-loop adjudication. We evaluate four LLMs (two proprietary, two open-weight) to establish a performance ceiling, then distill these capabilities into locally deployable Small Language Models (SLMs). Distributional analysis using Frechet Text Distance and Jensen-Shannon Divergence confirms SHIELD occupies a distinct region of biomedical embedding and vocabulary space versus legacy benchmarks. Our best distilled model matches its teacher on structured PHI categories (DATE, DOCTOR, ID, PATIENT, PHONE) and achieves micro-averaged span-level precision of 0.88 and recall of 0.86 on standard workstation hardware. Cross-dataset evaluation shows diversity-trained models generalize well on universal structured PHI, while institution-specific entities remain hard to transfer, suggesting optimal deployment combines broad-coverage models with specialized models for high-volume notes. We publicly release the SHIELD dataset and the distilled DeBERTa v3 model.

Comparison of deep learning and conventional methods for disease onset prediction

Background: Conventional prediction methods such as logistic regression and gradient boosting have been widely utilized for disease onset prediction for their reliability and interpretability. Deep learning methods promise enhanced prediction performance by extracting complex patterns from clinical data, but face challenges like data sparsity and high dimensionality. Methods: This study compares conventional and deep learning approaches to predict lung cancer, dementia, and bipolar disorder using observational data from eleven databases from North America, Europe, and Asia. Models were developed using logistic regression, gradient boosting, ResNet, and Transformer, and validated both internally and externally across the data sources. Discrimination performance was assessed using AUROC, and calibration was evaluated using Eavg. Findings: Across 11 datasets, conventional methods generally outperformed deep learning methods in terms of discrimination performance, particularly during external validation, highlighting their better transportability. Learning curves suggest that deep learning models require substantially larger datasets to reach the same performance levels as conventional methods. Calibration performance was also better for conventional methods, with ResNet showing the poorest calibration. Interpretation: Despite the potential of deep learning models to capture complex patterns in structured observational healthcare data, conventional models remain highly competitive for disease onset prediction, especially in scenarios involving smaller datasets and if lengthy training times need to be avoided. The study underscores the need for future research focused on optimizing deep learning models to handle the sparsity, high dimensionality, and heterogeneity inherent in healthcare datasets, and find new strategies to exploit the full capabilities of deep learning methods.

A study linking patient EHR data to external death data at Stanford Medicine

This manuscript explores linking real-world patient data with external death data in the context of research Clinical Data Warehouses (r-CDWs). We specifically present the linking of Electronic Health Records (EHR) data for Stanford Health Care (SHC) patients and data from the Social Security Administration (SSA) Limited Access Death Master File (LADMF) made available by the US Department of Commerce's National Technical Information Service (NTIS). The data analysis framework presented in this manuscript extends prior approaches and is generalizable to linking any two cross-organizational real-world patient data sources. Electronic Health Record (EHR) data and NTIS LADMF are heavily used resources at other medical centers and we expect that the methods and learnings presented here will be valuable to others. Our findings suggest that strong linkages are incomplete and weak linkages are noisy i.e., there is no good linkage rule that provides coverage and accuracy. Furthermore, the best linkage rule for any two datasets is different from the best linkage rule for two other datasets i.e., there is no generalization of linkage rules. Finally, LADMF, a commonly used external death data resource for r-CDWs, has a significant gap in death data making it necessary for r-CDWs to seek out more than one external death data source. We anticipate that presentation of multiple linkages will make it hard to present the linkage outcome to the end user. This manuscript is a resource in support of Stanford Medicine STARR (STAnford medicine Research data Repository) r-CDWs. The data are stored and analyzed as PHI in our HIPAA-compliant data center and are used under research and development (R&D) activities of STARR IRB.