Multivariate time-series forecasting of liver biomarkers from longitudinal lifestyle data for nonalcoholic steatohepatitis detection

Abstract Objectives To develop a machine learning method that estimates future liver biomarkers’ values from longitudinal lifestyle (diet, activity) data for early detection of nonalcoholic steatohepatitis (NASH). Materials and Methods The method in this study is developed utilizing the nonalcoholic fatty liver disease adult dataset, by National Institute of Diabetes and Digestive and Kidney Diseases, a real-world dataset representative of common electronic health records in the United States. We have developed time-series Machine Learning/Deep Learning and tree-based models to forecast future values for liver biomarkers, identified the minimum requirement of initial data points for optimal forecasting performance, and developed time-series classifier models for detecting NASH from longitudinal lifestyle data and initial biomarker values. Results Our experiments show that lifestyle-informed forecasting models, such as Attention-long short-term memory and TimeSeriesForestRegressor accurately predict future biomarker trajectories with as few as 2 observed timepoints (prediction error as low as 0.62), and NASH classifiers trained on these Forecasting liver Biomarkers (FoBi) estimated biomarkers achieve performance (accuracy 86%) comparable to or exceeding existing biopsy-aligned methods. Discussion The proposed approach, FoBi, is the first method to forecast liver biomarker trajectories from lifestyle data and demonstrate that both observed and model-estimated biomarkers can support effective NASH detection in real-world clinical settings. Conclusion Lifestyle-driven biomarker forecasting offers a promising, minimally invasive foundation for early NASH detection and long-term disease management, reducing dependence on frequent laboratory testing and biopsy-aligned measurements.