AIM-AHEAD Research Spotlight Series: Showcasing Innovative Research Across the Consortium

Clinical Validation of an Artificial Intelligence Platform for Detection of Pulmonary Nodules

Kalyani Narra, MD
JPS Health Network
South Central Hub

Pulmonary nodules are commonly identified during chest computed tomography (CT) imaging, with more than 95% of detected nodules being benign. However, a small percentage may represent malignancy, making timely identification and appropriate follow-up critical. Previous studies have shown that many incidental pulmonary nodules do not receive recommended follow-up due to workflow and resource limitations. Through participation in the AIM-AHEAD CLINAQ Fellowship Program, Kalyani Narra, MD, of JPS Health Network, evaluated the clinical performance of an artificial intelligence (AI)-based natural language processing (NLP) tool designed to identify pulmonary nodule findings within radiology reports and support more consistent follow-up processes.

JPS Health Network serves a large and medically complex patient population in Tarrant County, Texas, providing care to more than 261,000 uninsured individuals between the ages of 18 and 65 and supporting more than 1.5 million patient encounters annually. With patients speaking more than 80 languages and many patients relying on financial assistance programs or government-sponsored insurance, JPS continues to explore scalable approaches that support efficient and accessible healthcare delivery.

JPS began low-dose CT lung cancer screening in 2016 and updated screening criteria in 2022 following expanded recommendations from the U.S. Preventive Services Task Force and Centers for Medicare & Medicaid Services. While standardized assessment categories and management recommendations exist for pulmonary nodules identified through formal lung cancer screening programs, similar processes are not consistently available for nodules identified incidentally through other imaging pathways. This project evaluated whether an AI-enabled approach could help address this gap by improving identification of pulmonary nodules requiring clinical follow-up.

Study Design and Methods

The study evaluated a commercial AI tool that uses NLP to identify pulmonary nodule-related findings within dictated radiology reports. Researchers developed a Pulmonary Nodule Risk Score (PNRS) system to classify findings based on established guideline recommendations and assessed the AI tool’s ability to identify pulmonary nodules compared with independent manual review.

Researchers selected a sample of adult patients who underwent chest CT imaging at JPS Health Network between March and December 2024. Radiology reports were reviewed to determine whether pulmonary nodules were present, and findings identified by the AI tool were compared against manual review performed by independent reviewers. The team also evaluated changes in pulmonary clinic referrals before and after implementation of the AI tool.

The NLP platform was implemented into clinical workflows in April 2024. Between April and December 2024, the system reviewed 23,913 chest CT examinations and identified pulmonary nodule findings in 3,531 studies (15%). Among identified findings, 75% did not require dedicated follow-up, while 879 patients had actionable findings classified within PNRS categories 2–4B. A total of 227 patients received recommendations for referral.

Key Findings

The validation study demonstrated strong performance of the AI tool in identifying pulmonary nodules from radiology reports. Among 693 patients included in the study population, 19% had pulmonary nodules identified through the reference standard review. The patient population had a median age of 55 years, with 47% female participants, 36% receiving self-pay or charity care, and 30% reporting current tobacco use.

The AI tool demonstrated a sensitivity of 92% and specificity of 98% for identifying pulmonary nodule findings. Five false-negative findings and two false-positive findings were identified during validation. Clinical verification confirmed that no clinically significant pulmonary nodule findings were missed by the AI tool, and patients with false-negative findings did not experience adverse outcomes related to missed findings.

Following implementation of the AI tool, referrals for higher-risk pulmonary nodules increased. Referral rates increased from 38% to 66% for PNRS category 4A findings and from 77% to 92% for PNRS category 4B findings. These findings suggest that AI-supported screening of radiology reports may help healthcare teams identify patients who require additional evaluation and improve follow-up processes.

Implications for Research and Clinical Practice

This work demonstrates the potential of AI-enabled tools to support clinical workflows by identifying important findings within large volumes of radiology reports. For healthcare systems managing high volumes of imaging data, automated identification of pulmonary nodules may help improve consistency in follow-up processes and support timely referrals for patients with higher-risk findings.

The project also highlights the importance of validating AI tools within real-world clinical environments before broader implementation. By evaluating performance, assessing clinical outcomes associated with missed findings, and measuring workflow changes, researchers provided important insight into the potential benefits and limitations of AI-assisted approaches in pulmonary nodule management.

More broadly, this work demonstrates how AI/ML technologies can be integrated into healthcare operations to support clinical decision-making and improve care coordination. The findings may inform future efforts to apply similar approaches across additional clinical areas and healthcare settings.

Recognition and Dissemination

The project’s findings have been shared through multiple research and innovation forums, including the 2025 AIM-AHEAD Annual Meeting in Dallas, Texas; the JPS Quality Fair 2025, where the project received first place in the Innovation category; VITAL2026 in Minneapolis, Minnesota; the AACR Annual Meeting in San Diego in April 2026; and TxSCO 2025, where the work was selected as one of four oral presentations.

In April 2026, Dr. Narra was promoted to Associate Professor at the Burnett School of Medicine and was nominated to Epic Cosmos on behalf of JPS Health Network, supporting opportunities for continued collaboration and clinical research advancement.

Next Steps

Building on this work, the team plans to continue evaluating opportunities to expand AI applications across additional disease processes and clinical workflows. Future efforts will focus on sharing findings through peer-reviewed publications, collaborating with other safety-net healthcare organizations to evaluate broader implementation opportunities, and pursuing additional funding and leadership opportunities to advance AI-enabled innovation at JPS Health Network.

As radiology reporting processes continue to evolve and pulmonary nodule findings become more consistently documented, AI-supported approaches may provide additional opportunities to strengthen identification, referral, and follow-up pathways. Through continued evaluation and collaboration, this work contributes to the growing use of AI/ML tools to support more efficient and patient-centered healthcare delivery.

Hypertension Research Through AI/ML-Driven Innovation: Building the Foundation for Predictive Modeling

Dhruvangi Sharma, PhD
Georgia State University
Southeast-Morehouse Hub

Hypertension remains one of the leading modifiable risk factors for cardiovascular disease and affects nearly half of adults in the United States. Despite the availability of effective treatments, achieving long-term blood pressure control remains a significant challenge for many patients. Through participation in the AIM-AHEAD CLINAQ Fellowship Program, Dhruvangi Sharma, PhD, of Georgia State University explored how artificial intelligence and machine learning (AI/ML) can be used to better understand the factors associated with successful hypertension outcomes and establish the foundation for future predictive modeling.

The project's primary objective was to develop a supervised AI/ML model capable of predicting successful hypertension outcomes, defined as the absence of clinical and subclinical cardiovascular disease among individuals with hypertension. Recognizing that hypertension is influenced by a combination of clinical, behavioral, psychosocial, demographic, and non-medical health factors, the project focused on identifying and organizing the data needed to support the development of a robust predictive model while documenting a reproducible AI/ML workflow that could be applied to similar datasets.

Study Design and Methods

The study used a retrospective analysis of two large cardiovascular research datasets: the Jackson Heart Study (JHS) and the Coronary Artery Risk Development in Young Adults (CARDIA) study. After obtaining the necessary institutional approvals and data use agreements, researchers evaluated both datasets to determine their suitability for developing a predictive AI model.

Participants eligible for analysis included adults aged 18 years or older with hypertension at baseline and no history of stroke, myocardial infarction, coronary artery disease, or chronic kidney disease. Candidate predictor variables were organized into five categories: demographic characteristics; biophysiological measures, including blood pressure, cholesterol, body mass index, and laboratory values; behavioral factors such as smoking, alcohol use, and physical activity; psychosocial factors including stress and depression; and non-medical health factors, including income, education, neighborhood characteristics, and insurance status.

The planned AI/ML workflow included data preparation, feature selection, preprocessing, model development, training, validation, and performance evaluation using supervised learning approaches, including linear regression, random forest, and XGBoost algorithms.

Preparing Research Data for AI

An important component of the project involved evaluating the readiness of large clinical research datasets for AI/ML applications. Researchers assessed four JHS cohorts containing 440 datasets and explored approximately 18,700 variables to identify those most relevant to predicting successful hypertension outcomes over a 10- to 12-year follow-up period. The CARDIA dataset included two cohorts with 644 datasets and more than 18,700 available variables, which were similarly reviewed to identify candidate predictors.

During the project, the research team encountered several technical challenges related to data access and preparation. Because the datasets were not directly compatible with the Seven Bridges PIC-SURE platform, data files required manual retrieval, decompression, and evaluation before analysis. Differences in file structures and data dictionary formats also required alternative approaches for variable assessment. Despite these challenges, the team successfully identified relevant study variables, documented the data preparation process, and established a workflow that can inform future AI/ML research using large clinical datasets.

Key Findings

Baseline analyses demonstrated that the Jackson Heart Study provides a strong foundation for investigating hypertension outcomes. Among 3,883 participants, 57.4% had hypertension at baseline, and more than half reported taking blood pressure medication. After applying inclusion and exclusion criteria, researchers identified an eligible cohort of participants with hypertension and no baseline cardiovascular disease for future predictive modeling.

The project also demonstrated that the Jackson Heart Study is well suited for examining hypertension and long-term cardiovascular outcomes, while the CARDIA dataset contains many relevant variables despite technical limitations that affected data extraction. Together, these findings support the feasibility of using large longitudinal datasets to develop AI/ML models that examine factors associated with successful hypertension management.

Implications for Research and Clinical Practice

This project highlights the importance of preparing high-quality research data before developing AI/ML models. Establishing standardized approaches for dataset evaluation, variable selection, and data preparation helps ensure that future predictive models are built on reliable and clinically meaningful information.

By documenting each step of the AI/ML development process, this work provides a framework that can be adapted for future studies using similar datasets. As predictive models continue to evolve, they may help researchers and healthcare organizations better understand the complex factors associated with successful hypertension management and support more informed clinical decision-making.

Recognition and Next Steps

During the fellowship, Dr. Sharma advanced this work through several research and professional achievements, including submission of an AIM-AHEAD Hub Specific Project grant proposal, participation in Georgia State University's Presidential Strategic AI Initiative, and recognition by the Georgia State University Office of the Vice President for Research for research funding achievements in 2025.

Building on this foundation, future efforts will focus on completing development and validation of the predictive AI model, expanding analyses across additional datasets, pursuing external funding opportunities, and evaluating how AI/ML approaches can be applied to other chronic disease research. Through continued collaboration and refinement of AI-ready research infrastructure, this work contributes to the growing use of AI/ML methods to support data-driven approaches for improving hypertension research and cardiovascular health.

Related Links

• AIM-AHEAD CLINAQ Fellowship Program