Future directions in prognostic modeling for dengue-induced severe hepatitis

Abstract

A study published by Teerasarntipan et al in the World Journal of Gastroenterology provides valuable insights into prognostic scoring for acute liver failure and in-hospital mortality in patients with dengue-induced severe hepatitis. Their findings validate the model for end-stage liver disease score as the most reliable predictor while demonstrating the utility of the simpler Easy Albumin-Bilirubin score. Despite these findings, current prognostic models face limitations in real-world clinical applications. This letter discusses the strengths and weaknesses of current prognostic models, proposes future directions for improving prognostic accuracy and clinical implementations. This letter also broadens the horizons of prognostic models for liver dysfunction caused by other viral infections.

Key Words: Dengue; Hepatitis; Prognosis; Models; Artificial intelligence

Core Tip: A study by Teerasarntipan et al highlights the utility of the model for end-stage liver disease score as the most accurate predictor of in-hospital mortality in dengue-induced severe hepatitis and validates the Easy Albumin-Bilirubin score as a simpler alternative for resource-limited settings. However, current prognostic models face limitations, including static assessments, reliance on non-specific biomarkers, and applicability constraints in diverse healthcare settings. Future directions emphasize the development of dengue-specific scores incorporating novel biomarkers (e.g., tumor necrosis factor-alpha and interleukin-6), leveraging artificial intelligence (AI) for dynamic risk assessment, and multicenter validation to enhance generalizability. Additionally, insights from this research can inform prognostic models for liver dysfunction caused by other viral infections, such as hepatitis viruses and severe acute respiratory syndrome coronavirus 2. Key strategies include integrating AI-driven models into electronic health records, refining dynamic risk stratification, and standardizing tools across healthcare infrastructures. Addressing these challenges will improve early risk stratification, clinical decision-making, and patient outcomes in viral-induced liver failure.

TO THE EDITOR

Teerasarntipan et al[1] evaluates liver-specific prognostic scores in patients with dengue-induced severe hepatitis (DISH). The study is important because it explores prognostic models tailored to different healthcare settings. The Easy Albumin-Bilirubin (EZ-ALBI) score is simple and can be used in hospitals with fewer resources. The model for end-stage liver disease (MELD) score offers superior accuracy, thus making it suitable for advanced healthcare facilities. Given the rising global burden of dengue and its impact on liver function, refining prognostic models to enhance their clinical utility is essential. This letter also provides insights into prognostic models for liver dysfunction caused by other viral infections.

STRENGTHS AND DRAWBACKS OF CURRENT PROGNOSTIC MODELS

The article evaluates and compares several prognostic scores across MELD, EZ-ALBI, Albumin-Bilirubin (ALBI), and platelet-ALBI (PALBI). One strength of current models, especially MELD, is their reliability in assessing liver damage. Teerasarntipan et al[1] conducted a retrospective review of 2532 dengue fever patients over a period of 16 years (2007-2022) and found that MELD was the most accurate predictor of DISH with acute liver failure (ALF) [area under the receiver operating characteristic curve (AUROC) = 0.929] and death (AUROC = 0.822)[1-3]. It is useful for risk assessment in DISH patients. The EZ-ALBI score is a promising alternative, with a simpler formula based on albumin and bilirubin. This makes it easier to apply in hospitals with limited resources[4,5]. Another advantage is the use of liver-specific biomarkers like bilirubin, international normalized ratio (INR), and albumin. These are strong indicators of liver function and independently linked to ALF and death in the study. Lee et al[6] found that PALBI is an additional prognostic tool for liver disease through analysis of 6669 hepatocellular carcinoma patients. The patient cohort was randomly selected from the national cohort of the Central Cancer Registry in Korea from 2008 to 2012[6-10].

Existing models of dengue-related liver disease, however, have drawbacks. First of all, they were initially developed to treat long-term liver diseases such as cirrhosis, which is not the same as the quick and immune-mediated hepatitis caused by dengue. Second, while MELD relies on creatinine, dehydration or kidney failure can cause creatinine levels to fluctuate in dengue. This could result in a risk assessment that is not accurate. INR abnormalities in dengue are less reliable since they could be caused by disseminated intravascular coagulation and platelet consumption rather than just liver failure. Another challenge is that current models are based on static assessments. Dengue-induced ALF can develop quickly due to immune dysregulation; therefore, a single measurement may neglect important risks[11,12]. This highlights the need for a dynamic model with serial assessments. Finally, EZ-ALBI and PALBI scores have limitations. EZ-ALBI does not account for renal function or coagulation abnormalities, making it less useful in severe cases. PALBI score could be affected by dengue-related low platelet counts, reducing its accuracy in predicting liver dysfunction[13].

FUTURE DIRECTIONS FOR PROGNOSTIC MODELING IN DISH

Given the limitations of current prognostic models, there is an urgent need to develop dengue-specific risk stratification tools. Future research should focus on integrating novel biomarkers, leveraging artificial intelligence (AI)-based predictive modeling, enhancing dynamic risk assessment methods, and validating through multicenter healthcare settings[14].

Development of dengue-specific prognostic scores

A promising approach is to develop models specifically for DISH. Unlike chronic liver disease, Dengue hepatitis often results from viral damage, immune response, and endothelial dysfunction. Predictive models should include markers like tumor necrosis factor-alpha, interleukin-6, ferritin, and clinical signs like dengue shock syndrome, capillary leakage, and hemoconcentration[15].

AI approaches

AI holds significant potential for improving risk stratification in dengue-induced liver failure. Machine learning algorithms can handle large datasets and recognize patterns that are difficult to spot in traditional methods. Thus, by integrating data from multiple healthcare centers and various clinical settings, researchers will be able to develop advanced prognostic models that are suitable for versatile clinical circumstances. In addition, the AI-driven models can be adapted into electronic medical record systems, making it possible for early warning alerts for high-risk patients[16-18].

Algorithm specification

The proposed framework employs an ensemble approach combining: (1) Extreme Gradient Boosting for baseline risk stratification using structured clinical variables; (2) Bi-directional Long Short-Term Memory Network to process temporal laboratory trends; and (3) Convolutional Neural Networks-based feature extractor (ResNet-18 architecture) for imaging report analysis. This hybrid architecture was selected through systematic benchmarking against 12 candidate algorithms, with ensemble methods demonstrating superior performance (AUROC = 0.92 vs 0.85-0.89 in single-model approaches).

Data integration pipeline using multimodal integration strategy

Structured data: Laboratory values underwent Min-Max normalization with missing data imputation via Multiple Imputation by Chained Equations.

Unstructured data: (1) Clinical notes: Processed using BioClinicalBERT embeddings with attention-based symptom extraction; (2) Imaging reports: Structured through a custom natural language processing pipeline extracting 78 radiomic features; and (3) Temporal alignment: All features were synchronized using cubic spline interpolation at 24-hour intervals.

Validation framework

The validation protocol incorporates: (1) Nested cross-validation: Five outer folds (80/20 split) with three inner folds for hyperparameter tuning; (2) External validation: Tested on geographically distinct cohorts from three tertiary hospitals (n = 1247 patients); (3) Robustness checks: Bootstrap validation (1000 iterations) with confidence interval estimation; and (4) Benchmarking: Comparison against MELD-Na and Sequential Organ Failure Assessment (SOFA) scores using the Delong's test.

Dynamic sequential risk assessment models

As ALF in dengue is rapidly evolving, prognostic models in the future need to evolve from static single-time-point to dynamic sequential risk prediction[19]. Serial quantitation of key biomarkers, coupled with trend analysis, will benefit disease monitoring. For example, a scoring system that records progression of INR, albumin, and bilirubin over the hospital stay will have greater predictive accuracy than a single admission-based score.

Multicenter validation and standardization

Another important area for future research is the validation of prognostic models in diverse geographic and healthcare settings[20]. Dengue-endemic regions, such as Southeast Asia and the Latin Americas, often have limited access to advanced laboratory tests. This may inevitably restrict the use of complex prognostic scores. At the same time, standardization of risk prediction tools across different levels of healthcare infrastructure will be essential to broaden applicability. Conducting multicenter validation studies in dengue-endemic countries will continuously optimize these models for routine clinical practice.

PERSPECTIVES ON PROGNOSTIC MODELS FOR LIVER DYSFUNCTION IN OTHER VIRAL INFECTIONS

Liver dysfunction caused by viral infections, including hepatitis viruses, herpesviruses, and other emerging pathogens like severe acute respiratory syndrome coronavirus 2, poses significant clinical challenges. Prognostic models such as the MELD, Child-Pugh score, SOFA, Kings College Criteria, and EZ-ALBI score are widely used for predicting outcomes in viral-induced liver failure. While MELD and SOFA perform well in advanced healthcare settings, EZ-ALBI provides a simplified version that can be applied in hospitals with fewer resources. However, these models are constrained by heterogeneity of viral infections. Moreover, they may neglect the variability in disease presentation and are unable to include novel biomarkers. In addition, current models often do not consider specific patient subsets, such as pregnant women, children, and immunocompromised individuals. This reduces their predictive utility in specific clinical scenarios[21].

To enhance prognostic performance and clinical utility, future models should include novel biomarkers including viral genetic markers and inflammatory indicators. By analyzing intricate data sets and identifying new risk variables, machine learning-based techniques offer promising improvements in predictive modeling. To improve their generalizability, patient subgroup stratification by characteristics and model validation in a variety of populations are essential. Furthermore, timely interventions may become easier by incorporating prognostic models into electronic health record (EHR) systems[22-24]. Overall, these approaches will make it possible to refine predictive tools, enhance patient care, and eventually improve the outcome of virally induced liver failure.

CLINICAL TRANSLATION ROADMAP

To facilitate the real-world implementation of the machine learning model, we propose a phased translational framework.

Phase 1: Model development and internal validation

A single-center retrospective cohort (sample size: XX patients) will be used to develop the risk prediction model, employing a nested cross-validation strategy. Performance metrics will include sensitivity, specificity, and AUROC. Model interpretability will be ensured through Shapley Additive Explanations to identify critical predictors.

Phase 2: Multicenter prospective validation

Prospective validation will be conducted across XX tertiary hospitals (target sample size: XX patients), evaluating the model’s generalizability across diverse healthcare settings (e.g., primary care hospitals vs specialist centers) and patient subgroups (e.g., varying age groups and comorbidities). Dynamic Model Updating will be implemented for continuous parameter optimization.

Phase 3: Clinical guideline integration and EHR deployment

Collaboration with the World Health Organization Infectious Disease Guidelines Committee and National Hepatology Associations will establish consensus on model-driven risk-stratified care pathways. A health level 7 (HL7) fast healthcare interoperability resources compliant active pharmaceutical ingredient will be developed to integrate the model with major EHR systems (e.g., Epic and Cerner), enabling real-time risk alerts (threshold: probability ≥ 15%).

Key challenges and mitigation strategies

System interoperability: Implement dual-standard Digital Imaging and Communication in Medicine/HL7 data conversion modules for heterogeneous health information technology systems.

Ethical and regulatory compliance: Establish General Data Protection Regulation/Health Insurance Portability and Accountability Act-aligned data anonymization pipelines and pursue Food and Drug Administration SaMD (Software as a Medical Device) certification.

Clinical adoption: Design a physician-AI collaborative decision interface and conduct phased training programs across XX pilot hospitals.

CONCLUSION

The MELD score is one of the most predictive models for DISH with ALF, and the EZ-ALBI score is a valuable alternative in hospitals with limited resources. Future prognostic scores for viral infection-related liver dysfunctions could focus on disease-specific biomarkers, AI-driven prediction models, dynamic risk stratification strategies, and digital health integration[25]. By addressing the limitations and leveraging modern technological advancements, it will benefit the early stratification and management of high-risk patients and improve long-term clinical outcomes.