The critical role of standards for AI in digital pathology

Journal of Pathology Informatics · 2026-02-04 · 3 citations

Background: The field of pathology has not yet fully realized the potential of artificial intelligence (AI) and digital pathology. Adoption must be driven by demonstrable utility, and successful implementation depends on interoperability and sustainability, which require established standards. We address the imminent challenges facing the field of AI in digital pathology, which currently suffers from a lack of coordinated and adopted standards. Methods: We conducted several roundtable discussions with key opinion leaders from multiple sectors across the healthcare ecosystem. Based on how standards are used, we distinguish different areas of practice (relevance, endorsement, and utility) and emphasize the importance of standards. Results: Our roundtable discussion centered on one key theme: successfully implementing AI in digital pathology depends on achieving a certain level of uniformity across practices. We derive an approach to describe the critical role of standards consisting of seven interdependent areas of practice: value recognition, existing standards, dependencies for AI, failures, management of standards, trends, and a roadmap for accelerated and sustainable adoption. The promise of standards and our approach can be understood as the interconnection of these areas. We address imminent challenges surrounding the field of digital pathology by providing an approach for interoperable, coordinated, and sustainable use of standards across diverse practice settings. Conclusion: With the concepts and frameworks outlined in this article, we highlight the importance of standards in pathology and their crucial role in driving computational advances and enabling AI solutions to enhance patient care.

Development of a national pathology training system using digital pathology and SNOMED-CT

Journal of Pathology Informatics · 2025-06-18 · 1 citations

Introduction: Digital pathology is an important resource in modern pathology education, with many examples of large databases of educational cases now available online. However, there remains a lack of standardization in retrieval and categorization of cases for training within and between institutions. Methods: Over 1600 teaching terms applicable to histopathology were developed and mapped to corresponding SNOMED CT terms to create a large directory of teaching cases. This was then integrated as a pre-defined list into a clinical PACS system for a national digital pathology project covering multiple hospitals and pathology training programs. Results: This resource allows easy allocation of teaching term labels to cases with educational value. A substantial catalog of educational cases has been generated already, with ongoing efforts to expand this with cases from routine clinical practice. The catalog is fully searchable by term for use in training and examinations. Conclusions: A large directory of digital pathology teaching cases was developed with associated corresponding SNOMED CT terms. The directory of terms will be shared with other healthcare providers to expand its use and utility.

Future of Artificial Intelligence—Machine Learning Trends in Pathology and Medicine

Modern Pathology · 2025-01-04 · 136 citations

Advancements in Interoperability: Achieving Anatomic Pathology Reports That Adhere to International Standards and Are Both Human-Readable and Readily Computable

JCO Clinical Cancer Informatics · 2025-02-01 · 4 citations

PURPOSE: Over the past 50 years, multiple pathology organizations worldwide have evolved in cancer histopathology reporting from subjective, narrative assessments to structured, synoptic formats using controlled vocabulary. These reporting protocols include the required data elements that represent the minimum set of evidence-based, clinically actionable parameters necessary to convey the diagnostic, prognostic, and predictive information essential for patient care. Despite these advances, the synoptic reporting protocols were not harmonized across the various pathology organizations. Cancer pathology continues to be widely reported and stored in free-text format, or without encoded data such that it is neither computable nor interoperable across organizations. METHODS: In 2020, SNOMED International created the Cancer Synoptic Reporting Working Group (CSRWG). This resulted in international collaboration across multiple pathology organizations. CCRWG's mission was to use SNOMED Clinical Terms (CT) concepts to represent the required content within the College of American Pathologists (CAP) and International Collaboration on Cancer Reporting (ICCR) published pathology reporting protocols. RESULTS: In late 2023, the CSRWG published over 1,300 new or revised SNOMED CT concepts to represent all required pathology cancer data elements for adult and pediatric solid tumors in both CAP and ICCR using the semantic principles of the SNOMED-CT concept model. Thus, computability and interoperability would be broadly established. CONCLUSION: This work brings to fruition the longstanding desire for an international, interoperable, human- and machine-readable cancer pathology report for use in patient care, health care quality improvement, population health, public health surveillance, and translational and clinical trial research. The following report describes the project, its methods, and applications in the stated use cases.

Rapid slide-level cytopathology diagnosis enabled by a multicamera array scanner (MCAS)

2025-03-19

Developing digital pathology solutions for thick cytology specimens such as FNA smears is challenging because of their large area and 3D nature. Additionally, traditional deep learning methods for computational pathology require manual annotation of whole slide images, which is labor-intensive and time-consuming. To image thick FNA smears quickly, we have developed a multi-camera array scanner (MCAS) that digitizes entire cytology slides in 3D within 4 minutes. Here, we present a multiple-instance learning-based weakly-supervised deep learning method that predicts the presence of adenocarcinoma from the whole-slide images obtained by the MCAS using only slide-level diagnosis label for training, Our method also locates the regions with the highest contribution to the algorithm decision, thus enhancing interpretability. Given the speed of scanning and the interpretability that our method offers, it can serve as a very helpful screening tool for pathologists, increasing the speed of decision-making.

Evaluating Large Language Models and Retrieval-Augmented Generation Enhancement for Delivering Guideline-Adherent Nutrition Information for Cardiovascular Disease Prevention: Cross-Sectional Study

Journal of Medical Internet Research · 2025-10-07 · 4 citations

BACKGROUND: Cardiovascular disease (CVD) remains the leading cause of death worldwide, yet many web-based sources on cardiovascular (CV) health are inaccessible. Large language models (LLMs) are increasingly used for health-related inquiries and offer an opportunity to produce accessible and scalable CV health information. However, because these models are trained on heterogeneous data, including unverified user-generated content, the quality and reliability of food and nutrition information on CVD prevention remain uncertain. Recent studies have examined LLM use in various health care applications, but their effectiveness for providing nutrition information remains understudied. Although retrieval-augmented generation (RAG) frameworks have been shown to enhance LLM consistency and accuracy, their use in delivering nutrition information for CVD prevention requires further evaluation. OBJECTIVE: To evaluate the effectiveness of off-the-shelf and RAG-enhanced LLMs in delivering guideline-adherent nutrition information for CVD prevention, we assessed 3 off-the-shelf models (ChatGPT-4o, Perplexity, and Llama 3-70B) and a Llama 3-70B+RAG model. METHODS: We curated 30 nutrition questions that comprehensively addressed CVD prevention. These were approved by a registered dietitian providing preventive cardiology services at an academic medical center and were posed 3 times to each model. We developed a 15,074-word knowledge bank incorporating the American Heart Association's 2021 dietary guidelines and related website content to enhance Meta's Llama 3-70B model using RAG. The model received this and a few-shot prompt as context, included citations in a Context Source section, and used vector similarity to align responses with guideline content, with the temperature parameter set to 0.5 to enhance consistency. Model responses were evaluated by 3 expert reviewers against benchmark CV guidelines for appropriateness, reliability, readability, harm, and guideline adherence. Mean scores were compared using ANOVA, with statistical significance set at P<.05. Interrater agreement was measured using the Cohen κ coefficient, and readability was estimated using the Flesch-Kincaid readability score. RESULTS: The Llama 3+RAG model scored higher than the Perplexity, GPT-4o, and Llama 3 models on reliability, appropriateness, guideline adherence, and readability and showed no harm. The Cohen κ coefficient (κ>70%; P<.001) indicated high reviewer agreement. CONCLUSIONS: The Llama 3+RAG model outperformed the off-the-shelf models across all measures with no evidence of harm, although the responses were less readable due to technical language. The off-the-shelf models scored lower on all measures and produced some harmful responses. These findings highlight the limitations of off-the-shelf models and demonstrate that RAG system integration can enhance LLM performance in delivering evidence-based dietary information.

Evaluating Large Language Models and Retrieval-Augmented Generation Enhancement for Delivering Guideline-Adherent Nutrition Information for Cardiovascular Disease Prevention: Cross-Sectional Study (Preprint)

2025-06-06

<sec> <title>BACKGROUND</title> Cardiovascular disease (CVD) remains the leading cause of death worldwide, yet many web-based sources on cardiovascular (CV) health are inaccessible. Large language models (LLMs) are increasingly used for health-related inquiries and offer an opportunity to produce accessible and scalable CV health information. However, because these models are trained on heterogeneous data, including unverified user-generated content, the quality and reliability of food and nutrition information on CVD prevention remain uncertain. Recent studies have examined LLM use in various health care applications, but their effectiveness for providing nutrition information remains understudied. Although retrieval-augmented generation (RAG) frameworks have been shown to enhance LLM consistency and accuracy, their use in delivering nutrition information for CVD prevention requires further evaluation. </sec> <sec> <title>OBJECTIVE</title> To evaluate the effectiveness of off-the-shelf and RAG-enhanced LLMs in delivering guideline-adherent nutrition information for CVD prevention, we assessed 3 off-the-shelf models (ChatGPT-4o, Perplexity, and Llama 3-70B) and a Llama 3-70B+RAG model. </sec> <sec> <title>METHODS</title> We curated 30 nutrition questions that comprehensively addressed CVD prevention. These were approved by a registered dietitian providing preventive cardiology services at an academic medical center and were posed 3 times to each model. We developed a 15,074-word knowledge bank incorporating the American Heart Association’s 2021 dietary guidelines and related website content to enhance Meta’s Llama 3-70B model using RAG. The model received this and a few-shot prompt as context, included citations in a &lt;i&gt;Context Source&lt;/i&gt; section, and used vector similarity to align responses with guideline content, with the temperature parameter set to 0.5 to enhance consistency. Model responses were evaluated by 3 expert reviewers against benchmark CV guidelines for appropriateness, reliability, readability, harm, and guideline adherence. Mean scores were compared using ANOVA, with statistical significance set at &lt;i&gt;P&lt;/i&gt;&lt;.05. Interrater agreement was measured using the Cohen κ coefficient, and readability was estimated using the Flesch-Kincaid readability score. </sec> <sec> <title>RESULTS</title> The Llama 3+RAG model scored higher than the Perplexity, GPT-4o, and Llama 3 models on reliability, appropriateness, guideline adherence, and readability and showed no harm. The Cohen κ coefficient (κ&gt;70%; &lt;i&gt;P&lt;/i&gt;&lt;.001) indicated high reviewer agreement. </sec> <sec> <title>CONCLUSIONS</title> The Llama 3+RAG model outperformed the off-the-shelf models across all measures with no evidence of harm, although the responses were less readable due to technical language. The off-the-shelf models scored lower on all measures and produced some harmful responses. These findings highlight the limitations of off-the-shelf models and demonstrate that RAG system integration can enhance LLM performance in delivering evidence-based dietary information. </sec>

Elevated Lipoprotein(a) Independently Increases Risk for Short-Term Atherosclerotic Cardiovascular Events in Machine Learning Predictive Models

JACC Advances · 2025-10-25

BACKGROUND: Lipoprotein(a) [Lp(a)] is underutilized in short-term atherosclerotic cardiovascular disease (ASCVD) risk prediction. OBJECTIVES: This study investigates Lp(a) contribution to short-term ASCVD event prediction using contemporary real-world data and machine learning (ML). METHODS: A cohort of 731,983 individuals from a claims database was used to investigate the association of Lp(a) with incident ASCVD and all-cause mortality using Cox proportional hazards models. Novel ML models were developed to predict incident ASCVD events at 1, 2, and 3 years after Lp(a) testing. The models were validated in an independent cohort of 53,930 patients. RESULTS: An increase of 50 nmol/L in Lp(a) was independently associated with incident ASCVD events (HR: 1.072; 95% CI: 1.059-1.084) and all-cause mortality (HR: 1.041; 95% CI: 1.015-1.068) after adjustment for age, sex, and race/ethnicity. Novel ML models featuring Lp(a) predicted incident ASCVD events at 1, 2, and 3 years with robust discrimination (C-statistic: 0.83-0.84) in both the derivation and validation cohorts. Modest underestimation of risk was observed in the validation cohort for the 1-year model (calibration slope 1.25). Lp(a) contributed more to 1-year ASCVD prediction than smoking, diabetes, and other lipid parameters. Inclusion of Lp(a) in the 1-year model led to an integrated discrimination improvement of 0.03 and an optimal net reclassification improvement of 10% at a risk threshold of 26%. CONCLUSIONS: Lp(a) is a significant predictor of short-term ASCVD risk. Assessing Lp(a) and imminent ASCVD risk may assist in identifying patients who may benefit from escalation of preventative therapies.

Impact of Initial Cardiology Telemedicine Evaluation on Follow-Up Visits for Common Conditions: Quasi-Experimental Study

Journal of Medical Internet Research · 2025-08-05

Background: Telemedicine use has increased significantly in cardiology clinics, but the impact of initial telemedicine evaluation on total visit usage is unknown. Objective: This study aimed to determine the effect of initial telemedicine evaluation on the number of follow-up visits within 6 months for common cardiovascular conditions at an academic health system. Methods: Electronic health records data were extracted for general cardiology visits. New patient visits (NPVs) were included occurring from June 1, 2020, to May 31, 2023, for 10 common cardiovascular conditions-atrial fibrillation or flutter, chest pain, coronary artery disease, dyslipidemia, dyspnea, heart failure, hypertension, palpitations, preoperative evaluation, and syncope or dizziness. The effect of initial telemedicine versus in-person evaluation on follow-up visits within 6 months was assessed using a 2-stage least squares instrumental variable model with the proportion of clinician telemedicine use as the instrument and adjustment for patient and visit characteristics. Results: There were 5528 NPVs conducted by 40 general cardiology clinicians during the study period. The average patient age was 56 (SD 17.5) years, 54.2% (2998/5528) were female, 43.2% (2389/5528) were non-Hispanic White, 24.7% (1368/5528) were Asian, 13.8% (761/5528) were Hispanic, 34.4% (1904/5528) were on Medicare, and 13.2% (729/5528) were on Medicaid. Of the NPVs, 53.5% (2959/5528) were conducted via telemedicine (2814/5528, 50.9% via video and 145/5528, 2.6% via phone). Telemedicine use for NPVs ranged from 0% to 100% (N=40) across individual clinicians. The average number of follow-up visits was 57 visits per 100 patients within 6 months across all diagnosis groups. Patients receiving telemedicine NPVs were more likely to have telemedicine follow-up visits than those receiving in-person NPVs (1354/1619, 83.6% vs 680/1533, 44.4%). In the instrumental variable analysis, the impact of initial telemedicine evaluation differed by presenting condition. There was an increase in follow-up visits for patients with syncope or dizziness (29.8 visits/100 patients, 95% CI 6.4-53.1), palpitations (34.9 visits/100 patients, 95% CI 18.6-51.1), chest pain (36.9 visits/100 patients, 95% CI 18.5-55.2), and dyspnea (37.0 visits/100 patients, 95% CI 11.8-62.0). There was a decrease in follow-up visits for patients with coronary artery disease (-29.5 visits/100 patients, 95% CI -50.3 to -8.6) and dyslipidemia (-24.5 visits/100 patients, 95% CI -40.2 to -8.8). There was no significant effect for patients presenting for atrial fibrillation or flutter, heart failure, hypertension, and preoperative evaluation. Conclusions: The effect of initial telemedicine evaluation on follow-up visits varied significantly by presenting condition in this cardiology practice. Telemedicine use resulted in increased follow-up visits for patients presenting with symptomatic complaints, while for those presenting with chronic conditions, there was no significant effect or a decrease in visits. Future studies should assess strategies to target initial care modalities to appropriate patients in cardiology clinics with early in-person evaluation for symptomatic patients.

Standards in digital pathology

Elsevier eBooks · 2024-11-29