A Scalable Deep Mixed-Effects Learning Framework for High-Dimensional and Multi-Modal Biomarker Integration and Complex Disease Outcome Prediction

The integration of high-dimensional and multi-modal biomarker data remains a central challenge in precision medicine, hindered by noise, weak individual signals, and heterogeneous data structures. We propose AdaMixNet, an adaptive mixed-effects deep learning framework that unifies nonlinear fixed-effects modeling with kernel-based random-effects estimation to robustly and accurately predict complex disease outcomes. By leveraging feature screening to distinguish sparse, high-impact biomarkers from dense, low-signal features, AdaMixNet captures both strong and subtle biological effects across diverse data modalities. Through comprehensive simulations and applications to two large-scale real cohorts, i.e., METABRIC (breast cancer) and ADNI (Alzheimer’s disease), AdaMixNet shows robust overall performance, often outperforming strong machine-learning and statistical baselines in simulation and achieving competitive or best results for several real-data outcomes while maintaining good performance across sample sizes from 1000 to 20,000. AdaMixNet offers a generalizable and interpretable framework for integrating high-dimensional and multi-modal omics profiles with low-dimensional clinical data, accelerating the translation of molecular insights into clinical applications. Modern medicine collects large amounts of biological data, such as gene-expression profiles, genetic variants, and medical test results. However, combining these different types of data to better predict disease remains difficult. The data are often noisy, and important signals can be weak or hidden among many measurements. In this study, we developed a method called AdaMixNet. It is a computer-based tool that learns from both strong and subtle biological signals while also using standard clinical information. This helps improve the accuracy of predicting disease outcomes. We tested AdaMixNet using simulated data and two large real-world studies of breast cancer and Alzheimer’s disease. Our method reduced prediction errors compared to existing approaches. This work may help researchers and doctors better use complex biological data to support more personalized healthcare decisions in the future. Dai et al. develop AdaMixNet, an adaptive mixed-effects deep learning framework that integrates high-dimensional omics data with clinical variables for disease outcome prediction. Across simulations and the METABRIC and ADNI cohorts, AdaMixNet reduces prediction error by up to 25% and outperforms existing approaches.