Leena Choi

Purpose: This study aimed to develop and evaluate machine‐learning models for estimating gestational age at delivery using EHR data from two health systems. Methods: We trained random forest, gradient boosting, and ensemble models using EHR data from mother–infant dyads at Vanderbilt University Medical Center (VUMC) and independently reimplemented and retrained the modeling pipeline at the University of Michigan (UMich). We further analyzed EHR predictors of gestational age, assessed temporal drift in EHR data elements, and evaluated model performance stratified by gestational‐age category at delivery. Results: The study included pregnancies corresponding to 54 344 and 34 345 mother–infant dyads at VUMC (2005–2025) and UMich (2012–2024), respectively. The gestational age predictions of the ensemble models showed high agreement with the reference standard on the VUMC dataset (±1 week: 85.1%, ±2 weeks: 94.3%, MAE: 4.4 days). When the same modeling pipeline was independently implemented and retrained on the UMich dataset, the ensemble models achieved higher agreement with the reference standard (±1 week: 93.1%, ±2 weeks: 97.8%, MAE: 2.8 days). Further, performance was better among pregnancies delivered in more recent years, and among full‐ and late‐term deliveries compared with preterm deliveries. Discussion: The results indicate that supervised machine learning methods leveraging linked mother–infant EHRs can accurately estimate gestational age at delivery overall, although performance was substantially lower for preterm deliveries. Reproducibility of the modeling approach and portability of the analytic workflow were demonstrated across the two health systems. Conclusion: This study presents a robust machine learning framework to estimate gestational age at delivery, whose development process was reproducible across two health systems. The framework has the potential to support large‐scale real‐world maternal and neonatal health studies requiring retrospective estimation of gestational age.

Bayesian borrowing of external information can improve trial efficiency, particularly in pediatric and rare disease settings where patient populations are limited, but may introduce bias and inflate the Type I error rate when the trial differs from external studies. Recent U.S. Food and Drug Administration (FDA) draft Bayesian guidance emphasizes careful evaluation of external information, prior specification, and assessment of operating characteristics. This paper compares three meta-analytic-predictive (MAP)-based methods for Bayesian borrowing: the MAP prior, robust MAP (RMAP) prior, and self-adapting mixture (SAM) prior. An adaptive platform trial design in Rett syndrome is used as a case study. Simulation studies evaluate frequentist operating characteristics under varying prior–data conflict, between-study heterogeneity, treatment effects, and clinically significant differences (CSDs) for the SAM prior. The MAP prior achieved the greatest efficiency when external and current data were compatible but exhibited the largest bias under substantial prior–data conflict. The RMAP priors improved robustness through fixed robust-component weights, whereas the SAM prior adaptively adjusted borrowing and was less sensitive to prior–data conflict while retaining efficiency gains when the data were compatible. Although the CSD influenced the degree of adaptive borrowing, as reflected by effective sample size, it had only a modest impact on frequentist operating characteristics. Sensitivity analyses using a skeptical robust component yielded similar qualitative conclusions, while accentuating the differences between the MAP and RMAP priors. These findings provide guidance for evaluating and selecting MAP-based borrowing strategies before trial implementation, particularly in rare disease settings, consistent with current FDA recommendations.

Objective: To develop a transformer-based natural language processing (NLP) system for detecting adverse drug events (ADEs) from clinical notes in electronic health records (EHRs). Materials and Methods: We fine-tuned BERT Short-Formers and Clinical-Longformer using the processed dataset from the 2018 National NLP Clinical Challenges (n2c2) shared task Track 2. Two data processing methods, window-based and split-based approaches, were compared to identify the optimal processing method. Model generalizability was evaluated on a dataset extracted from Vanderbilt University Medical Center (VUMC) EHRs. Results: On the n2c2 dataset, the best 5-fold cross-validation AUPRC, micro F1, and macro F1 scores were 0.840 (Clinical-Longformer, 4-chunk split), 0.965 (BioBERT, 15-word window), and 0.852 (Clinical-Longformer, 10-chunk split). On the VUMC dataset, the best AUPRC, micro F1, and macro F1 scores were 0.536 (Clinical-Longformer, 6-chunk split), 0.966 (BERT-base-uncased, 6-chunk split), and 0.762 (Clinical-Longformer, 4-chunk split). Discussion: Transformer-based models demonstrated strong performance for ADE detection, with split-based processing generally outperforming window-based methods. Clinical-Longformer combined with a practical split-based approach showed promise for real-world implementation. Beyond token limits, chunk size substantially influenced model performance, even when text length remained within limits. Conclusion: Our findings provide practical guidance for developing transformer-based ADE detection systems from clinical notes. The selection of both text preprocessing strategies and model architectures should be guided by note characteristics and practical considerations such as annotation burden.

Purpose: Patients being initiated on treatment with select immunomodulators should be screened for tuberculosis (TB) and hepatitis B virus (HBV). Previous studies found disparities in TB/HBV screening before therapy initiation. Health-system specialty pharmacies (HSSPs) have access to electronic health records and can evaluate infection screening completion. Summary: A multisite, prospective, observational cohort study was conducted at 12 HSSPs. Patients 18 years of age or older prescribed an immunomodulator between March 1 and May 31, 2023, were included. Patients prescribed medications associated with a risk of TB or HBV reactivation within the previous year or referred from a non–health-system provider were excluded. Patients were followed for 6 months. Screening completion was summarized as percentages. Univariate logistic regression models were used to evaluate the association between the number of HSSP services provided and completion of safety screening, and safety screening–related pharmacist interventions. Of 1,597 pharmacy referrals, 737 patients were included. Patients had rheumatological (55.5%; n = 409), dermatological (29%; n = 214), or gastroenterological conditions (15.5%; n = 114). In total, 95.1% of patients (n = 701) had any screening for TB and HBV, and 81.5% (n = 601) completed screenings within 2 years before or 60 days after referral. Screening was completed within 1 year before or 60 days after referral for 75.3% of patients (n = 555). More HSSP services were associated with increased likelihood of completed safety screening (odds ratio [OR], 1.23; P = 0.012) and pharmacist interventions (OR, 2.68; P = 0.04). Conclusion: High rates of screening for TB and HBV before immunomodulator therapy initiation were observed in patients referred to HSSPs.

Objective: This study aimed to train and evaluate supervised machine learning walgorithms using electronic health record (EHR) data to accurately estimate gestational age at delivery. Materials and Methods: We trained random forest, gradient boosting, and ensemble models on EHR data of mother-infant dyads from Vanderbilt University Medical Center(VUMC) and replicated the analyses at University of Michigan (UMich). We further analyzed EHR predictors of gestational age, assessed temporal drift in EHR data elements, and evaluated model performance stratified by delivery status. Results: The study included pregnancies corresponding to 54,344 and 34,345 mother-infant dyads at VUMC (2005-2025) and UMich (2012-2024), respectively. The gestational age predictions of the ensemble models achieved the highest agreement with the reference standard on the VUMC dataset (±1 week: 85.2%, ±2 weeks: 94.3%, MAE: 4.4 days) and demonstrated stronger generalization on the UMich dataset (±1 week: 93.1%, ±2 weeks: 97.8%, MAE: 2.8 days). Further, performance was better among pregnancies delivered in more recent years, and among full- and late-term deliveries compared with preterm deliveries. Discussion: The results indicate that supervised machine learning methods leveraging linked mother-infant EHRs can accurately estimate gestational age at delivery, while demonstrating the generalizability of the modeling approach and the portability of the analytic workflow across healthcare sites. Conclusion: This study presents a robust and generalizable machine learning framework to estimate gestational age at delivery. The framework can be reliably used to impute gestational age in large-scale, real-world clinical studies to support maternal and neonatal health research, in which accurate estimation of pregnancy onset is critical.

Purpose: High-dose methotrexate is an effective treatment for adult and pediatric patients with acute lymphoblastic leukemia, osteosarcoma, and lymphoma. However, its clearance is highly variable, and delayed clearance can lead to significant toxicity. This study aimed to identify a population pharmacokinetic model that accurately characterizes high-dose methotrexate clearance in an adult population. Methods: We developed a new population pharmacokinetic model using a training dataset derived from a cohort of 208 adult patients who received high-dose methotrexate at Vanderbilt University Medical Center. To assess predictive performance, we evaluated both our model and several externally developed models using an independent test dataset. Results: The final model was a three-compartment model incorporating body surface area as a covariate on all pharmacokinetic parameters, and serum creatinine and sex as covariates on clearance. Our newly developed model predicted with the most accuracy and precision at the first concentration measurement, taken at 24 h. Our model, along with five external models, was selected based on predictive performance for further assessment with maximum a posteriori Bayesian forecasting to compare predictions at the individual-level. The model by Hui et al. was more accurate at 48 h, while our model performed similarly at 72 h. Conclusion: These findings suggest an optimal strategy for therapeutic drug monitoring and dosing decisions: use our model for prediction at the 24 h mark when no prior drug levels are available, followed by Bayesian forecasting using our new model supplemented by the Taylor model.

Purpose: To examine interventions, recommendations, and direct cost avoidance from integrated outpatient specialty neurology pharmacists in the non–multiple sclerosis (MS) setting. Methods: A single-center, retrospective cohort study was performed, including patients prescribed specialty medications from non-MS neurology clinics with at least one pharmacist intervention from January to March 2023. Outcomes included pharmacist intervention types, intervention recommendations to patients and providers, and direct cost avoidance from interventions. Cost avoidance was calculated by summing the total costs (average wholesale price minus 20%) of one fill of medications that were discontinued or had a dose change as a result of an intervention or refill deferral. Patients who used excess supply on hand from a previous fill were accounted for during the calculation. The cost per unit of each medication/dosage was multiplied by the quantity of the last medication fill to calculate the cost avoided. Results: The study included 741 patients with a median age of 60 (interquartile range, 22-72) years; 83% were White; 55% were male; and 69% used the integrated health-system specialty pharmacy (IHSSP) for medication fulfillment. Pharmacists performed 2,244 interventions, generating 392 recommendations, with a 98% acceptance rate. The most common recommendation outcomes were medication changes (34%, n = 129) and financial assistance referrals (17%, n = 64). During the 3-month study period, discontinuations and dose changes resulted in $172,757 in direct cost avoidance for IHSSP patients. Conclusion: IHSSP neurology pharmacists in the non-MS setting frequently perform interventions that result in therapy changes and direct cost avoidance. IHSSP pharmacist interventions benefit patients, providers, the healthcare system, and insurers.

Purpose: This study evaluated the effectiveness of electronic early oral anticancer medication monitoring questionnaires in identifying adverse effects (AEs) requiring pharmacist intervention among patients initiated on oral anticancer therapy. Methods: A prospective randomized study was conducted within an integrated health system specialty pharmacy. Patients were excluded if they never started treatment, discontinued oral anticancer therapy within 14 days, were not new to therapy, or had a hematologic condition. Stratified randomization by age (≥60 years versus <60 years) and sex was performed to assign patients to intervention or usual care. Intervention patients received an electronic questionnaire 7 to 14 days after initial counseling. Outcomes included time to first AE-related pharmacist intervention and intervention frequency within 45 days and dose modifications, healthcare utilization, and treatment modifications within 90 days. Per-protocol analysis compared outcomes among questionnaire responders and patients assigned to usual care. Cox proportional hazard (time to intervention) and ordinal logistic regression (intervention number) models adjusted for age, sex, race, diagnosis, Charlson comorbidity index score, and medication type. Results: Of the 446 patients randomized, 388 were included in the analysis (183 in the intervention cohort and 205 in the usual care cohort). The questionnaire response rate was 47%. In the intent-to-treat analysis, no significant differences were observed in time to intervention (hazard ratio [HR], 1.2; P = 0.360) or intervention frequency (odds ratio [OR], 1.2; P = 0.409). Per-protocol analysis (n = 291) showed responders had significantly faster time to intervention (HR, 2.4; P < 0.001), more interventions (OR, 2.7; P < 0.001) and lower healthcare utilization (OR, 0.4; 95% CI, 0.2-0.9; P = 0.023). Conclusion: Electronic monitoring resulted in earlier and more frequent interventions and less healthcare utilization among questionnaire responders.