About

Andrea Wright is a Lecturer on Anthropology at Harvard University and serves as the Allston Burr Resident Dean of Eliot House as well as an Assistant Dean of Harvard College. Her research interests encompass migration, intimate labor, feminist anthropology, care, gender, activist anthropology, race, engaged scholarship, and feminist pedagogy. She is affiliated with the Tozzer Anthropology Building and the Peabody Museum in Cambridge, MA. Her work focuses on understanding social dynamics related to migration and gender, emphasizing engaged scholarship and feminist approaches within anthropology.

Research topics

• Artificial Intelligence
• Medicine
• Computer Science
• Family medicine
• Internal medicine
• Machine Learning
• Nursing
• Intensive care medicine
• Data science
• Psychology
• Virology
• Pathology

Selected publications

• Institutional Trends in Penicillin Allergy: A New Era of Active Penicillin Allergy Delabeling

  Clinical & Experimental Allergy · 2026-04-21

  articleOpen access

  Summary Paediatric penicillin allergy prevalence declined from 7.57% to 6.65% (2018–2024), especially in youngest children. An overall decrease in new PAL and an increase in institutional penicillin allergy delabelling was observed.

  PublisherDOI

• Comparison of machine learning methods for prediction of venous thromboembolism among hospitalized adults

  Journal of Hospital Medicine · 2026-05-14

  articleOpen access

  BACKGROUND: Hospital-acquired venous thromboembolism (HA-VTE) is a significant cause of morbidity and mortality among hospitalized adults. Accurate prediction of HA-VTE is crucial for timely intervention and prevention. While logistic regression is widely used for the development of clinical prediction models, there is ongoing interest in the potential for machine learning methods to enhance prediction performance. OBJECTIVES: This study aimed to identify the most practical method for the prediction of HA-VTE based on electronic health record (EHR) data. METHODS: We evaluated and compared the performance of prognostic models developed using logistic regression, random forest, extreme gradient boosting (XGBoost), deep neural networks, and an ensemble method to predict HA-VTE using EHR data from a large academic medical center. Models were evaluated in a temporal external validation cohort based on discrimination and calibration metrics, overall and among key patient subgroups. RESULTS: All models demonstrated similarly high discrimination, with C statistics ranging from 0.886 to 0.900. Logistic regression, random forest, and XGBoost showed excellent calibration, whereas deep neural networks exhibited poorer calibration. Additionally, prediction accuracy at a probability cut-off of 0.02 and correlations between models indicated that logistic regression performed comparably well, if not better, compared with machine learning methods. Subgroup analyses further confirmed that the logistic regression model demonstrated strong discrimination and calibration among patient subpopulations. CONCLUSIONS: Our findings suggest that logistic regression is a highly effective tool for EHR-based prediction of HA-VTE.

  PublisherDOI

• Patient‐reported outcomes for monitoring substance use treatment: A systematic review of single‐item measures

  Addiction · 2026-04-22

  articleOpen access

  BACKGROUND AND AIMS: Measurement-based care (MBC) is a structured approach using standardized, repeated assessments to monitor treatment progress and guide clinical decision-making. MBC improves outcomes for substance use treatment but can be time consuming due in part to lengthy assessment tools. Single-item, patient-reported outcome measures (PROMs) offer a more acceptable alternative for routine monitoring, yet their psychometric properties have not been systematically evaluated. We sought to identify constructs assessed by single-item PROMs in substance use treatment and critically appraise their validity, reliability and overall quality using standardized criteria. METHODS: We conducted a systematic review following COSMIN and PRISMA guidelines. MEDLINE, Embase and PsycINFO were searched from January 2005 to August 2025 for studies evaluating single-item PROMs in adults with substance use. We assessed psychometric properties, including content validity, test-retest reliability, construct validity, responsiveness and predictive validity using COSMIN criteria. Quality of evidence was assessed using a modified GRADE approach. RESULTS: Of 4722 records screened, 35 studies met inclusion criteria, evaluating 68 single-item PROMs across 9 clinical constructs for more than 50 000 participants. Fifteen studies achieved an overall rating of sufficient measure properties and moderate-or-above level of evidence rating across domains. Test-retest reliability ranged approximately from Intraclass Correlation Coefficient = 0.60-0.85; construct validity correlations approximately ranged r = 0.11-0.98. Predictive validity was strong for several measures, with odds ratios up to 7.3 for treatment readiness. Measures assessing craving, treatment readiness and self-efficacy demonstrated the most robust evidence and, in some cases, outperformed multi-item scales. However, over half of measures lacked empirically validated thresholds and responsiveness to change analyses, limiting clinical interpretability and treatment monitoring. CONCLUSIONS: Single-item patient-reported outcome measures (PROMs) are pragmatic tools for implementing measurement-based care in substance use treatment, offering strong implementation feasibility and, in some cases, predictive performance comparable to longer instruments. PROMs lacking validated thresholds or responsiveness may be best used as complementary tools, whereas those with strong evidence and thresholds can support primary monitoring.

  PublisherDOI

• Integrating rule-based NLP and large language models for statin information extraction from clinical notes

  International Journal of Medical Informatics · 2025-09-06 · 8 citations

  articleOpen accessSenior author
  PublisherOA PDFDOI

• Real-time automated billing for tobacco treatment: developing and validating a scalable machine learning approach

  JAMIA Open · 2025-05-02

  articleOpen accessSenior author

  Objectives: To develop CigStopper, a real-time, automated medical billing prototype designed to identify eligible tobacco cessation care codes, thereby reducing administrative workload while improving billing accuracy. Materials and Methods: ChatGPT prompt engineering generated a synthetic corpus of physician-style clinical notes categorized for CPT codes 99406/99407. Practicing clinicians annotated the dataset to train multiple machine learning (ML) models focused on accurately predicting billing code eligibility. Results: Decision tree and random forest models performed best. Mean performance across all models: PRC AUC = 0.857, F1 score = 0.835. Generalizability testing on deidentified notes confirmed that tree-based models performed best. Discussion: CigStopper shows promise for streamlining manual billing inefficiencies that hinder tobacco cessation care. ML methods lay the groundwork for clinical implementation based on good performance using synthetic data. Automating high-volume, low-value tasks simplify complexities in a multi-payer system and promote financial sustainability for healthcare practices. Conclusion: CigStopper validates foundational methods for automating the discernment of appropriate billing codes for eligible smoking cessation counseling care.

  PublisherOA PDFDOI

• Impact of PREVENT Cardiovascular Risk Equations on Statin Eligibility by Subgroup and Risk Thresholds: A Cross-Sectional Study

  Journal of General Internal Medicine · 2025-09-25

  articleOpen accessSenior author

  BACKGROUND: Replacing the pooled cohort equations (PCEs) with the Predicting Risk of Cardiovascular Disease EVENTs (PREVENT) equations for atherosclerotic cardiovascular disease (ASCVD) risk is projected to reduce statin eligibility, prompting discussion of lowering the risk threshold used with PREVENT. The potential impact on statin eligibility in different subgroups is unknown. OBJECTIVE: To assess the impact of replacing PCEs with PREVENT equations on statin eligibility for a real-world population of primary care patients, incorporating social deprivation index (SDI) and lowering the ASCVD risk thresholds for statin eligibility. DESIGN: Cross-sectional analysis comparing 10-year ASCVD risk scores and statin eligibility using the PCEs and PREVENT equations within a primary care population. Subgroup analyses were conducted by age, sex, and race. Risk thresholds for statin eligibility were varied to assess the effect on eligibility. PARTICIPANTS: Adult patients who visited a Vanderbilt primary care clinic in 2023. MAIN MEASURES: Estimated 10-year ASCVD risk and proportion of patients eligible for statin therapy using the PCEs vs. PREVENT equations. KEY RESULTS: In 50,123 patients, the mean 10-year ASCVD risk was significantly lower with PREVENT compared to the PCEs (3.6 vs. 7.5, p < 0.0001). In 36,430 patients not on statins, PREVENT reduced statin eligibility by 78.2%, with the largest reductions in women (82.6%), patients aged 40-49 (97.8%), and Black patients (81.2%). Lowering the statin eligibility threshold from 7.5 to 3% led to a 27.5% overall increase in eligibility rather than 78.2% reduction. However, gaps between subgroups expanded, and younger and Black patients retained relative reductions in eligibility (e.g., 4.7% decrease in statin eligibility among Black patients compared to a 32.7% increase among White patients). CONCLUSIONS: In a real-world primary care population, replacing the PCEs with the PREVENT equations would significantly reduce statin eligibility at the 7.5% threshold. Lowering the risk threshold would increase overall eligibility but disproportionately affect eligibility within certain subgroups.

  PublisherOA PDFDOI

• The Pragmatic Removal of Penicillin Allergy Electronic Health Record Labels (PROPEL) Trial: A Randomized Clinical Trial

  The Journal of Allergy and Clinical Immunology In Practice · 2025-07-16 · 3 citations

  articleOpen access
  PublisherOA PDFDOI

• Applying an Empirical Taxonomy to Alert Malfunctions in a Pragmatic Trial for Hypertension Management in Chronic Kidney Disease

  Applied Clinical Informatics · 2025-10-01

  article

  Clinical decision support (CDS) systems have been widely adopted across clinical settings to promote evidence-based practice for clinicians. CDS malfunctions often affect the user experience and indirectly or directly interfere with patient care. To enhance optimal performance, it is critical to constantly monitor the performance of the tool and react promptly when malfunctions are identified.This study aimed to describe malfunctions identified in the development and implementation of a CDS alert as well as lessons learned.A pragmatic randomized controlled trial of a CDS alert for primary care patients with chronic kidney disease and uncontrolled blood pressure was conducted. The alert included prechecked default orders for medication initiation or titration, basic metabolic panel, and nephrology electronic consult. Alert monitoring involved retrospective chart review and review of alert firing reports.Eight CDS malfunctions were identified. The most common causes of malfunctions were due to conceptualization and build errors. Provider feedback and retrospective chart review were the primary methods of identifying the root cause of malfunctions.Our findings highlight the need for CDS interventions to be continuously monitored through chart review, alert firing reports, and opportunities for provider feedback. Lessons learned from CDS malfunctions can be implemented to improve provider trust in automated electronic health record-based alerts, reduce administrative burden, and prevent inappropriate alert recommendations that can negatively affect patient outcomes. This study is registered with Clinivaltrials.gov (identifier: NCT03679247).

  PublisherDOI

• Abstract 4370328: Hybrid Rule-Based and Large Language Model Framework Extracts Statin-Related Information from Clinical Notes

  Circulation · 2025-11-03

  articleSenior author

  Background: Atherosclerotic cardiovascular disease is a leading cause of morbidity and mortality; statin therapy reduces risk but adherence is suboptimal. Clinical notes contain details on statin intolerance, contraindications, and patient deferral that structured data miss, yet manual extraction is time-consuming. Hypothesis: A hybrid AI framework combining rule-based NLP and LLM-based methods can accurately extract statin-related information from clinical notes to inform clinical decision support. Methods: We developed a three-component framework: (1) a rule-based NLP filter to exclude irrelevant notes, (2) an LLM-based refinement filter to identify notes likely containing relevant information, and (3) an LLM-based multicategory classifier to categorize records into intolerance, contraindications, and deferral. Dataset A (2,000 notes; July 1–August 1, 2024) from adult primary care visits at Vanderbilt University Medical Center (VUMC) was split into training (n = 1,200) and testing (n = 800) subsets for development and evaluation. Dataset B (197,761 notes; August 1–September 1, 2024) was used for retrospective evaluation. Performance metrics included precision, recall, F1, accuracy, and filter-out rate. Patient-level prevalence for each category was measured in Dataset B. Results: In Dataset A, the rule-based NLP filter excluded 81% of notes while retaining all relevant ones (precision = 1.00). The LLM-based refinement filter achieved precision = 0.973, recall = 0.947, F1 = 0.960, accuracy = 0.996, and a filter-out rate of 95.4% on the testing subset. The multicategory classifier attained F1 scores of 0.99 (intolerance), 0.81 (contraindications), and 0.86 (deferral). In Dataset B, after sequential filtering, 45,253 of 197,761 notes remained; the classifier identified 3,027 patients (6.4%) with documented intolerance, 310 (0.7%) with contraindications, and 1,391 (2.9%) who deferred therapy. Conclusions: The hybrid AI framework efficiently processes clinical notes, filtering out over 90% of irrelevant records while maintaining high precision for relevant content. This scalable approach enables extraction of actionable statin-related information and has potential to enhance clinical decision support by integrating patient-level insights to optimize statin therapy.

  PublisherDOI

• A standard-based taxonomy of features that affect user response to clinical decision support alerts

  BMC Medical Informatics and Decision Making · 2025-10-21 · 1 citations

  articleOpen accessSenior author

  OBJECTIVE: The objective of this work is to develop a standard-based taxonomy of features that might affect user response to alerts using evidence from literature and public alert logic repositories. METHODS: We developed a taxonomy of features using multiple sources: (1) the Agency for Healthcare Research and Quality (AHRQ) CDS Connect Repository, (2) alert logic from commercial electronic health record (EHR) customers, and (3) published literature. Three categories (patient, provider, environment/context) were used a priori to develop the taxonomy. The final taxonomy was mapped to the Fast Healthcare Interoperability Resources (FHIR) standard for development of standardized CDS services. RESULTS: Aggregating potential features extracted from three data sources, we identified 95 unique features, which we then mapped to the FHIR standard, encompassing 24 FHIR resources. The common features differed depending on the knowledge source. In the AHRQ public alert repository, frequently occurring features were observations in flowsheets, procedures, diagnoses, medications, and patient age. On the other hand, the commercial EHR customers primarily presented features such as diagnosis type, patient age, diagnosis grouper, diagnosis, medication value set. Literature-based insights revealed that provider type, medication, patient age, alert severity, and medication dose were the most common features. CONCLUSION: This study demonstrated a standard-based taxonomy of features that could impact user responses to CDS alerts, bridging insights from academic studies and industry practices. The taxonomy stands as a foundational tool, guiding the CDS development, implementation, and evaluation, with the overarching goal of improving user acceptance and healthcare quality.

  PublisherOA PDFDOI

Recent grants

• Improving Quality by Maintaining Accurate Problem Lists in the EHR (IQ-MAPLE)

  NIH · $2.2M · 2014–2019

• Safety Promotion through Early Event Detection in the Elderly (SPEEDe)

  NIH · $802k · 2019–2020

• Improving clinical decision support reliability using anomaly detection methods

  NIH · $2.4M · 2014–2020

• Safety Promotion through Early Event Detection in the Elderly (SPEEDe)

  NIH · $2.7M · 2020–2025

Frequent coauthors

• David W. Bates

  Brigham and Women's Hospital

  255 shared

• Dean F. Sittig

  The University of Texas Health Science Center at Houston

  231 shared

• Blackford Middleton

  165 shared

• Allison B. McCoy

  Vanderbilt University Medical Center

  96 shared

• Gian Vincenzo Zuccotti

  University of Milan

  88 shared

• Lipika Samal

  Brigham and Women's Hospital

  84 shared

• Gordon D. Schiff

  Harvard University

  72 shared

• Joan S. Ash

  Oregon Health & Science University

  66 shared