AI System for Disease Detection and Risk Prediction from Longitudinal Multi-Modal Medical Data

Background & Unmet Need

• Much of biomedical research and healthcare is based on characterizing and responding to relevant changes in data that have been collected over time, such as imaging, laboratory results, pathology, free-text clinical notes, and wearable signals
• Many artificial intelligence (AI)-based diagnostic tools rely primarily on single-timepoint data—such as a single imaging scan, a one-off lab reading, or both—when assessing a patient’s health status
• This can result in suboptimal early detection or less accurate predictions of disease course, since meaningful patterns often emerge over multiple patient visits
• Classical risk models for disease prediction rely on relatively few variables and do not exploit the richness of longitudinal image data or other repeated clinical information
• Unmet Need: Methods for integrating multimodal longitudinal data to support clinical decisions

Technology Overview

• The Technology: A time-aware AI architecture which integrates multimodal longitudinal data to provide improved clinical decision support
• The system accommodates virtually any data type including imaging data, lab values, pathology, clinical notes, genomic data, and wearable signals
• The system learns temporal relationships across a patient’s entire clinical history to perform a variety of tasks (e.g., disease classification, regression, multi-label tasks, survival/risk estimation)
• PoC Data: The inventors developed a longitudinal model for mammogram risk prediction with superior predictive value (AUC 0.92 at 1-year follow up vs. 0.9 and 0.83 for comparator models)
• An additional model for AD progression in normal vs. MCI patients based on multimodal inputs (cognitive assessments, MRI, genetic, and clinical data) with an AUC curve improved by up to 15% over state-of-the-art models

Technology Applications

• Clinical decision support (e.g., diagnosis, detection, risk stratification, therapy selection, monitoring)
• Population health management (e.g., patient outreach and screening, proactive interventions, healthcare policy and budget planning)
• Research and clinical trials (e.g., longitudinal outcome analysis, biomarker discovery, adaptive trial design)

Technology Advantages

• The architecture is flexible and can incorporate virtually any data modality or set of data modalities
• The architecture design enables robust handling missing or asynchronous data without severely degrading performance
• The system is generalizable; it scales to large datasets and can be applied to numerous clinical domains (e.g., oncology, neurology, and cardiology)

Figure 1: The LSM-AD model integrates clinical and MRI biomarkers over a series of patient visits to predict control vs. MCI-to-AD progression and outperforms prior state-of-the art longitudinal models.