Clinical notes contain a vast amount of potentially useful information about adverse drug event (ADE) signals that never reach pharmacovigilance databases. Traditional rule-based or sentence-level models often miss subtle causal cues and generate excess false positives. To build a large language model (LLM) pipeline that reads entire electronic health record (EHR) notes, identifies drug-event pairs with a “reasonable possibility” of causation, and infers other important properties of each ADR, such whether the event is “serious” or “unlabeled”. We adopted a two-pass workflow using model “OpenAI o1”: Pass 1 screens each note for ADEs; Pass 2 adds 20 structured fields. A diverse sample of 372 deidentified notes from physicians and pharmacists at the University of California, San Francisco (UCSF), from 31 specialty/setting cells, yielded 191 ADEs. One medical expert reviewed each ADE for validity, seriousness, and label status. Another expert created a gold standard manually curated ADE set on 100 of the 372 ADEs to give us a percent “recall” estimate. A third expert met with the first expert to arrive at a consensus on the validity of LLM ADEs validated by the first expert but not found in the gold standard ADEs, giving us a estimate of “accuracy”. Of 191 ADEs, 180 were true positives (94.2

Background and Aims:Outpatient clinical notes are a rich source of information regarding drug safety. However, data in these notes are currently underutilized for pharmacovigilance due to methodological limitations in text mining. Large language models (LLM) like BERT have shown progress in a range of natural language processing tasks but have not yet been evaluated on adverse event detection. Methods:We adapted a new clinical LLM, UCSF BERT, to identify serious adverse events (SAEs) occurring after treatment with a non-steroid immunosuppressant for inflammatory bowel disease (IBD). We compared this model to other language models that have previously been applied to AE detection. Results:We annotated 928 outpatient IBD notes corresponding to 928 individual IBD patients for all SAE-associated hospitalizations occurring after treatment with a non-steroid immunosuppressant. These notes contained 703 SAEs in total, the most common of which was failure of intended efficacy. Out of 8 candidate models, UCSF BERT achieved the highest numerical performance on identifying drug-SAE pairs from this corpus (accuracy 88-92%, macro F1 61-68%), with 5-10% greater accuracy than previously published models. UCSF BERT was significantly superior at identifying hospitalization events emergent to medication use (p < 0.01). Conclusions:LLMs like UCSF BERT achieve numerically superior accuracy on the challenging task of SAE detection from clinical notes compared to prior methods. Future work is needed to adapt this methodology to improve model performance and evaluation using multi-center data and newer architectures like GPT. Our findings support the potential value of using large language models to enhance pharmacovigilance.

Several biomedical language models have already been developed for clinical language inference. However, these models typically utilize general vocabularies and are trained on relatively small clinical corpora. We sought to evaluate the impact of using a domain-specific vocabulary and a large clinical training corpus on the performance of these language models in clinical language inference. We trained a Bidirectional Encoder Decoder from Transformers (BERT) model using a diverse, deidentified corpus of 75 million deidentified clinical notes authored at the University of California, San Francisco (UCSF). We evaluated this model on several clinical language inference benchmark tasks: clinical and temporal concept recognition, relation extraction and medical language inference. We also evaluated our model on two tasks using discharge summaries from UCSF: diagnostic code assignment and therapeutic class inference. Our model performs at par with the best publicly available biomedical language models of comparable sizes on the public benchmark tasks, and is significantly better than these models in a within-system evaluation on the two tasks using UCSF data. The use of in-domain vocabulary appears to improve the encoding of longer documents. The use of large clinical corpora appears to enhance document encoding and inferential accuracy. However, further research is needed to improve abbreviation resolution, and numerical, temporal, and implicitly causal inference.

There is a great and growing need to ascertain what exactly is the state of a patient, in terms of disease progression, actual care practices, pathology, adverse events, and much more, beyond the paucity of data available in structured medical record data. Ascertaining these harder-to-reach data elements is now critical for the accurate phenotyping of complex traits, detection of adverse outcomes, efficacy of off-label drug use, and longitudinal patient surveillance. Clinical notes often contain the most detailed and relevant digital information about individual patients, the nuances of their diseases, the treatment strategies selected by physicians, and the resulting outcomes. However, notes remain largely unused for research because they contain Protected Health Information (PHI), which is synonymous with individually identifying data. Previous clinical note de-identification approaches have been rigid and still too inaccurate to see any substantial real-world use, primarily because they have been trained with too small medical text corpora. To build a new de-identification tool, we created the largest manually annotated clinical note corpus for PHI and develop a customizable open-source de-identification software called Philter ("Protected Health Information filter"). Here we describe the design and evaluation of Philter, and show how it offers substantial real-world improvements over prior methods.

Predictive modeling with electronic health record (EHR) data is anticipated to drive personalized medicine and improve healthcare quality. Constructing predictive statistical models typically requires extraction of curated predictor variables from normalized EHR data, a labor-intensive process that discards the vast majority of information in each patient's record. We propose a representation of patients' entire raw EHR records based on the Fast Healthcare Interoperability Resources (FHIR) format. We demonstrate that deep learning methods using this representation are capable of accurately predicting multiple medical events from multiple centers without site-specific data harmonization. We validated our approach using de-identified EHR data from two US academic medical centers with 216,221 adult patients hospitalized for at least 24 h. In the sequential format we propose, this volume of EHR data unrolled into a total of 46,864,534,945 data points, including clinical notes. Deep learning models achieved high accuracy for tasks such as predicting: in-hospital mortality (area under the receiver operator curve [AUROC] across sites 0.93-0.94), 30-day unplanned readmission (AUROC 0.75-0.76), prolonged length of stay (AUROC 0.85-0.86), and all of a patient's final discharge diagnoses (frequency-weighted AUROC 0.90). These models outperformed traditional, clinically-used predictive models in all cases. We believe that this approach can be used to create accurate and scalable predictions for a variety of clinical scenarios. In a case study of a particular prediction, we demonstrate that neural networks can be used to identify relevant information from the patient's chart.

Hepatitis B virus (HBV) reactivation in the setting of rituximab use is a potentially fatal but preventable safety event. The rate of HBV screening and proportion of patients at risk who receive antiviral prophylaxis in patients initiating rituximab is unknown.We analyzed electronic health record (EHR) data from 2 health systems, a university center and a safety net health system, including diagnosis grouper codes, problem lists, medications, laboratory results, procedures codes, clinical encounter notes, and scanned documents. We identified all patients who received rituximab between 6/1/2012 and 1/1/2016. We calculated the proportion of rituximab users with inadequate screening for HBV according to the Centers for Disease Control guidelines for detecting latent HBV infection before their first rituximab infusion during the study period. We also assessed the proportion of patients with positive hepatitis B screening tests who were prescribed antiviral prophylaxis. Finally, we characterized safety failures and adverse events.We included 926 patients from the university and 132 patients from the safety net health system. Sixty-one percent of patients from the university had adequate screening for HBV compared with 90% from the safety net. Among patients at risk for reactivation based on results of HBV testing, 66% and 92% received antiviral prophylaxis at the university and safety net, respectively.We found wide variations in hepatitis B screening practices among patients receiving rituximab, resulting in unnecessary risks to patients. Interventions should be developed to improve patient safety procedures in this high-risk patient population.