Emerging applications of natural language processing for the identification of steatotic liver disease

INTRODUCTION Steatotic liver disease (SLD) encompasses a broad clinical spectrum, ranging from simple fat accumulation in the liver to hepatic inflammation (steatohepatitis) and fibrosis.1 Metabolic dysfunction–associated steatotic liver disease (MASLD) and alcohol-associated liver disease account for the majority of SLD cases. MASLD is the leading cause of chronic liver disease worldwide, and metabolic dysfunction–associated steatohepatitis (MASH) cirrhosis is now the second most common indication for liver transplant in the United States.2 Despite the increasing global prevalence of SLD and its potential to lead to advanced liver disease, findings of steatosis, particularly on imaging studies, remain overlooked in clinical practice.3 A multidisciplinary consensus group of experts drafted a research agenda calling attention to specific target areas to confront the global epidemic of MASLD, including identification, prevention, and management of SLD.3 The strategic framework emphasizes the need for improved quantification of the burden of SLD and investigation into how artificial intelligence can be used to improve SLD diagnosis. Under-recognition of high-risk SLD leads to missed opportunities for early disease-modifying interventions. Compounding this issue is the limited understanding of individual-level risk factors that drive progression from steatosis to steatohepatitis, advanced fibrosis, and HCC. Therefore, there is an urgent need for innovative population-level strategies to improve the early identification of SLD and promote accurate risk stratification of individuals with SLD. Natural language processing (NLP) is a type of artificial intelligence that enables computers to understand and learn from large amounts of text by extracting context and meaning from human language.4 In hepatology, NLP has emerged as a strategy for identifying patients with SLD through the analysis of clinical data within electronic health records (EHRs). The application of NLP in this manner has opened the door to large-scale screening and identification of patients with SLD. Herein, we explore 2 recent studies published in Hepatology Communications (Torgersen et al5 and Sherman et al6) that highlight the opportunities for NLP algorithms to reshape the landscape of early SLD detection and management. Additionally, we outline advancements in the use of NLP and examine how emerging artificial intelligence technologies are transforming SLD care. NLP FOR THE DETECTION OF SLD Structured data (ie, laboratory values and diagnostic codes) is easy to extract from EHRs. In contrast, unstructured data—narrative text from notes and test reports—is an untapped, rich source of clinical information.4 NLP algorithms facilitate the automated identification of important clinical details from unstructured EHR text that would otherwise require extensive manual review.4 Despite the widespread availability of unstructured data in radiology imaging reports that identify the presence of SLD, linking these findings to a clinical diagnosis remains a challenge. Large-scale applications of NLP to identify patients with SLD are increasing, providing opportunities for advancing patient care through improved detection, risk stratification, and linkage to care. INNOVATIONS IN NLP To address this gap, Torgersen et al5 conducted a cross-sectional study investigating the performance characteristics of an NLP algorithm to identify the presence of hepatic steatosis in patients with and without HIV using available imaging reports. The algorithm was designed to capture key terms such as “fatty,” “steatosis,” “steatotic,” or “steatohepatitis” from free text in imaging reports from the EHR. Validation was achieved by comparing the model’s performance in identifying the presence or absence of SLD within imaging reports to a gold-standard manual case review. Following validation, the NLP algorithm was applied to ultrasound, CT, and MRI studies (n=26,706) for Veteran patients with and without HIV. Concordance between the NLP algorithm and manual review was 94.3%, and the model identified SLD with 100% sensitivity, 89.7% specificity, 88.5% positive predictive value (PPV), and 100% negative predictive value (NPV). Importantly, the NLP algorithm had superior sensitivity, NPV, and agreement with manual review for the diagnosis of SLD compared to using NAFLD ICD-9/ICD-10 codes, which had an NPV of 33.0%. This study demonstrates the impactful use of NLP algorithms to improve the detection of SLD. This approach uncovered a substantial number of previously unrecognized cases of SLD with high accuracy and without the need for exhaustive manual chart review, and the NLP algorithm outperformed traditional ICD code-based methods of identifying SLD. Elsewhere, Sherman et al evaluated the accuracy of an NLP algorithm in identifying histologic features of MASLD within liver biopsy reports obtained from 1987 to 2021 in a large healthcare system. The researchers developed and validated an NLP algorithm that extracted histologic features of MASLD from text contained in liver biopsy pathology reports using key MASH terms, such as steatosis, lobular inflammation, ballooning, and steatohepatitis. Additionally, they evaluated the ability of the NLP algorithm to correctly identify the fibrosis score from the pathology report. The algorithm was validated following review of liver biopsy pathology reports (n=1224) that were manually reviewed and annotated by a group of expert clinicians. Agreement between the NLP algorithm and manual review had PPVs and NPVs ranging from 93.5% to 100% for all histologic concepts and fibrosis stages. Subsequently, the authors used the validated NLP algorithm to establish a very large MASLD cohort to examine the incidence of long-term MASLD-related outcomes, including progression to cirrhosis and hepatic decompensation. In this MASLD cohort, the incidence of cirrhosis and decompensation was significantly associated with increasing NLP-extracted MASH/fibrosis stagings on liver biopsy. Overall, the researchers demonstrated that an NLP algorithm can accurately identify steatosis and fibrosis features from free text contained within liver biopsy reports and can facilitate the creation of a cohort capable of longitudinal risk assessments. Taken together, these studies suggest that the performance of NLP algorithms to identify SLD using readily available unstructured data is excellent and can be leveraged to evaluate long-term clinical outcomes. INTEGRATING NLP ALGORITHMS INTO ROUTINE CLINICAL CARE In routine clinical practice, hepatic steatosis is often detected incidentally on abdominal imaging studies, but the clinical importance of this finding is too often overlooked, resulting in the underuse of risk reduction interventions.7 As illustrated above, NLP algorithms can bridge this gap by using broad EHR-based screening to accurately identify individuals with SLD needing further evaluation (eg, fibrosis testing) or who may be eligible for targeted treatments. For example, there is an opportunity to leverage NLP [or large language models (LLMs), discussed later] to identify patients with MASLD and evidence of stage F2–F3 fibrosis on biopsy or elastography who may be eligible for treatments and linkage to specialty care.1 In this manner, NLP may serve as an automated surveillance tool to streamline care delivery, increase treatment access, and improve clinical outcomes in SLD. LIMITATIONS OF NLP ALGORITHMS NLP algorithms perform very well in assisting in the identification of patients with SLD and MASLD. However, the reliability of these algorithms depends on the presence and quality of EHR documentation, which may have variability that can impact the algorithm’s performance. Furthermore, because these algorithms have frequently been trained on datasets from single healthcare systems, an investigation is needed to assess the generalizability of these algorithms across diverse health systems. Finally, integration of NLP algorithms within routine clinical care remains an unmet need hindered by several factors. First, scaling these interventions for broad uptake and use requires prioritizing interoperability early so that these tools can be used across diverse health systems and EHRs. Second, these tools must be created with the involvement of key stakeholders and a strong understanding of how they would be incorporated into existing workflows. For example, when these algorithms identify actionable clinical insights, the alerts to the end-user should adhere to the 5 rights of clinical decision support tools—right information, right person, right channel, right format, and right time in the workflow—to best achieve the intended outcomes.8 Despite these limitations, NLP algorithms present a valuable advancement in leveraging unstructured clinical data to identify persons with SLD. EMERGENCE OF LLMS Building on advancements in NLP technology, the field of artificial intelligence has entered a new era with the emergence of LLMs, which offer enhanced capabilities in contextual understanding and inference of human language,4 making them powerful tools for interpreting complex, structured and unstructured clinical data. Recent studies have begun exploring how these models can be applied to real-world clinical data, including imaging reports, to support disease identification and decision-making. Vong et al9 studied the use of LLMs in detecting hepatic steatosis from abdominal CT imaging reports performed in the emergency department using 3 Open AI LLMs (ChatGPT-3.5, ChatGPT-4, and ChatGPT-4o). The LLMs reviewed text from imaging reports and were able to accurately identify the presence of hepatic steatosis when prompted. Each LLM tested was found to have high sensitivity, specificity, and interrater reliability in identifying hepatic steatosis from imaging reports. In another study, Panzeri et al10 evaluated the diagnostic accuracy of ChatGPT-4 in assessing liver fibrosis stages in patients with MASH using digital biopsy images. When provided with tissue biopsy images selected by a senior pathologist, ChatGPT-4 achieved a diagnostic accuracy rate of 81% in the assessment of fibrosis stage in comparison to expert liver pathologists reads. However, its accuracy dropped to 54% when analyzing randomly selected tissue biopsy image fields of view, indicating a reliance on curated input. Notably, the model’s performance was lower in identifying stage F4, with an accuracy of 40%, compared to higher accuracy in stages F2–F3. These findings suggest that while LLMs show promise in the histopathological evaluation of liver fibrosis, further refinement is needed before they can be reliably integrated into clinical diagnostic workflows. CONCLUSIONS NLP algorithms are effective for early detection, classification, and risk stratification in SLD. These algorithms have the potential to bridge the gap between under-recognized key clinical data in the EHR and the diagnosis and management of SLD. Future integrations of advanced artificial intelligence applications will continue to improve care across the spectrum of SLD.