Stanley Xu

BACKGROUND Longitudinal electronic health record clinical notes contain rich information for artificial intelligence (AI) applications, but they often exceed practical language model context windows and impose substantial computational burden. Clinically informed context reduction may improve scalability by retaining information-rich note sections while removing less relevant text. OBJECTIVE To develop and evaluate a scalable context-reduction framework that uses a fine-tuned open-source language model to extract clinically relevant information from longitudinal clinical notes for downstream AI applications. METHODS We conducted a multiphase study within Kaiser Permanente Southern California using clinical notes from adults with mild asthma. Manual annotation was first performed on a training dataset of 1,500 notes and an independent validation dataset of 250 notes. A Flan-T5 large model was then fine-tuned on the training notes to extract visit reason, chief complaint, and history of present illness. Model outputs from the validation dataset were compared with manually extracted reference standards using contextual precision, recall, relevancy, and faithfulness, with GPT-OSS-20B model serving as the evaluation judge. The fine-tuned model was then applied to 3,815,537 implementation notes from 198,667 patients to extract targeted clinically relevant information at the note level and aggregate it to the patient level. RESULTS The study included 3,817,287 clinical notes from 198,669 patients. Fine-tuning substantially improved extraction performance compared with zero-shot inference, increasing contextual precision from 0.342 to 1.000, recall from 0.305 to 0.958, relevancy from 0.329 to 0.991, and faithfulness from 0.128 to 0.984. In the implementation dataset, mean note length decreased from 213.1 to 27.7 tokens, and mean patient-level token burden decreased from 4,093.6 to 518.1 tokens, corresponding to a 7.90-fold reduction. Context-window feasibility also improved substantially. The proportion of patients with 500 tokens or fewer increased from 4.9% to 65.6%, while the proportion with more than 5,000 tokens decreased from 23.3% to 0.2%. CONCLUSIONS A fine-tuned open-source language model accurately retained targeted clinical information while substantially reducing longitudinal clinical text burden at population scale. Clinically informed context reduction may support more efficient and scalable use of electronic health record notes in downstream AI applications.

Background: The Vaccine Safety Datalink (VSD) detected a statistical signal for ischemic events (ischemic stroke or transient ischemic attack) following bivalent mRNA COVID-19 vaccination through prospective surveillance during 2022-2023. Although multiple studies from other surveillance systems and countries reported no increased risk, important methodological limitations remained. This U.S. study addressed those limitations by evaluating the ischemic stroke risk following bivalent mRNA COVID-19 vaccination, influenza vaccination, and their same-day coadministration using event-dependent self-controlled case series (SCCS) design. Methods: Study outcomes included first-ever ischemic stroke (primary outcome), first-in-1-year ischemic stroke (secondary outcome), and ischemic events (exploratory outcomes), identified using ICD-10-CM codes in inpatient and emergency department settings during September 1, 2022–March 31, 2023, among individuals aged ≥12 years across eight VSD sites. Analyses were conducted separately for Pfizer-BioNTech and Moderna bivalent vaccines, with relative incidences (RI) and 95% confidence intervals (CI) estimated for 1–21-day and 1–42-day risk intervals, using person-time outside these intervals as the control period. Subgroup analyses were performed by age group (12–64, >65 years) and history of documented SARS-CoV-2 infection. Results: A total of 6,510 first-ever ischemic strokes were identified among more than 6.8 million participants. Among recipients of Pfizer-BioNTech bivalent COVID-19 and influenza vaccines, no statistically significant increased risk of first-ever ischemic stroke was observed following bivalent COVID-19 vaccination (RI = 0.94; 95% CI: 0.63–1.41), influenza vaccination (RI = 0.95; 95% CI: 0.82–1.10), or same-day coadministration (RI = 1.15; 95% CI: 0.88–1.49) within 1– 21-day risk intervals; findings were similar for 1–42-day intervals. Comparable null results were observed for Moderna vaccines and across all subgroups, secondary, and exploratory outcomes. Conclusion: No increased risk of ischemic stroke was found following bivalent mRNA COVID-19 vaccination, influenza vaccination, or their coadministration in this multi-site SCCS study. These findings are consistent with previous studies and underscore the importance of continued vaccine safety monitoring. Highlights: We evaluated ischemic stroke risk after bivalent mRNA COVID-19 vaccination, influenza vaccination, and their coadministration. No increased risk of ischemic stroke was observed within 1–21 or 1–42 days following these vaccination exposures. Findings were consistent across vaccine type, age group, and prior SARS-CoV-2 infection status. These results support the safety of bivalent mRNA COVID-19 vaccines with respect to ischemic stroke.

Background: Deep learning models for clinical risk prediction are commonly developed using large numbers of input features and long patient histories, increasing computational cost and implementation complexity. However, there is limited evidence on how reducing feature dimensionality and sequence length affects predictive performance and clinically relevant model behavior in longitudinal electronic health record (EHR)-based prediction. Methods: Using transformer-based models and longitudinal EHR data from adults with mild asthma within an integrated healthcare system, we systematically evaluated feature-level and sequence-level dimensionality reduction. Reduced-feature models were constructed using attribution-based (Integrated Gradients), univariate performance-based, and clinically guided feature selection strategies. Maximum sequence length was varied using percentile-based truncation of longitudinal histories. Model performance was assessed using discrimination, calibration, risk stratification, threshold-based event capture, and computational efficiency across six-fold cross-validation defined by geographically distinct Medical Service Areas. Findings: Models using substantially fewer input features achieved discrimination comparable to the high-dimensional reference model. Moderate sequence-length truncation reduced training time by more than 50% while preserving discrimination. Although reduced-dimensional models showed attenuated predicted risk at the highest risk levels, they identified similar high-risk groups and captured comparable proportions of observed asthma exacerbation events at clinically relevant decision thresholds. Interpretation: Transformer-based clinical risk prediction can be achieved using simpler and more computationally efficient input representations than commonly assumed. While dimensionality reduction may affect calibration at the upper tail of predicted risk, discrimination and threshold-based clinical performance remain largely preserved. Treating input dimensionality and sequence length as explicit, tunable design parameters may facilitate the development of scalable and resource-efficient clinical prediction models.

Rituximab (RTX) dosed every 6 months is an affordable and highly effective treatment for relapsing multiple sclerosis (RMS). However, higher cumulative doses are associated with greater risks of infections. Whether extending dosing intervals at a low dose (500 mg) increases the risk of return of disease activity, particularly in those with only 1 year of stability or with significant return of B cells, is uncertain. Our objective was to compare whether extending RTX dosing to every 12 months (q12mo) vs continuing every 6 months (q6mo) dosing in patients with RMS increased the risk of return of disease activity. We emulated a target trial comparing the effectiveness of RTX 500 mg q12mo to q6mo starting in year 2 (baseline) among patients with RMS who had active disease at RTX start and no evidence of disease activity (NEDA) during the first year on RTX. Eligible patients were identified from the retrospective cohort of patients with MS treated with RTX between 2008 and 2023 in Kaiser Permanente Southern California. The absence of relapses, MRI disease activity, and disability progression (NEDA-3) were analyzed by intention-to-treat (ITT) and per-protocol (PP) with inverse probability of treatment weighting of receiving q6mo or q12mo dosing. We identified 140 patients (mean age 40.4 years, 73.6% female) who received q12mo dosing and 468 patients (mean age 38.4, 70.9% female) on q6mo dosing after baseline with a median ITT follow-up of 2.6 (interquartile range [IQR] 1.9-3.3) and 3.9 years (IQR 2.4-4.8), respectively. Relapses and/or MRI disease activity (3.5%) and failing NEDA-3 (3.8%) up to 4 years later were uncommon. Extending dosing to q12mo was associated with significant B-cell repopulation (70.5% with ≥80 cells/μL PP, 75.7% ITT) compared with q6mo intervals (23.3% with ≥80 cells/μL PP, 60.2% ITT) but not with an increased risk of failing NEDA-3 up to 4 years after baseline in propensity score-adjusted ITT (hazard ratio [HR] 1.60, 95% CI 0.69-3.68) or PP analyses (HR 1.44, 95% CI 0.45-4.61). Extending low-dose RTX treatment intervals after 1 year of stability to q12mo was highly effective and did not result in an increased risk of disease activity despite high B-cell counts compared with continuing q6mo dosing over 2 or more years of follow-up. This study provides Class IV evidence that in persons with RMS with NEDA-3 in the 12 months following RTX, extending the dosing interval to 12 months did not increase the risk of disease activity compared with continuing q6mo dosing.