TLDR: This paper introduces an innovative Electronic Health Record (EHR) foundation model that integrates Polygenic Risk Scores (PRS) as a foundational data modality, moving beyond traditional EHR-only approaches to build more holistic health profiles. Leveraging the extensive and diverse data from the All of Us (AoU) Research Program, this multimodal framework aims to learn complex relationships between clinical data and genetic predispositions. Evaluation on AoU data demonstrates the model’s predictive value for the onset of various conditions, particularly Type 2 Diabetes (T2D), and illustrates the interplay between PRS and EHR data, paving the way for more personalized and actionable healthcare strategies.
In a significant stride towards truly personalized healthcare, a new research paper introduces an innovative Electronic Health Record (EHR) foundation model that moves beyond traditional EHR-only approaches by integrating Polygenic Risk Scores (PRS) as a fundamental data type. This groundbreaking work, leveraging the vast and diverse data from the All of Us (AoU) Research Program, aims to build more holistic health profiles by understanding the intricate connections between clinical data and an individual’s genetic predispositions.
The core idea behind this research is to enhance the predictive capabilities and interpretability of health models. While existing EHR foundation models have shown great promise in predicting clinical outcomes from longitudinal health records, they often fall short due to their reliance on a single data modality. Real-world health is complex, influenced by a myriad of factors including genetics, lifestyle, and social determinants. By incorporating additional data types like genomics, the new multimodal framework seeks to provide a more complete picture.
The researchers from Verily Life Sciences, Nvidia, and Google have adapted advancements in generative AI to the EHR foundation model space. They propose two main architectural frameworks for integrating diverse data modalities: cross-attention and adapter modules. Cross-attention mechanisms, inspired by encoder-decoder paradigms, are used to align and integrate representations from different modalities with tokenized clinical histories. Adapter modules, on the other hand, insert vectorized modality embeddings directly into health trajectories, allowing for static modalities (like genomics) to be prepended and dynamic modalities (like survey data) to be inserted mid-sequence, preserving temporal context.
Polygenic Risk Scores (PRS) are a powerful tool in personalized medicine, quantifying an individual’s genetic likelihood of developing complex diseases by aggregating the effects of multiple genetic variants. Integrating PRS into clinical practice can enable earlier identification of at-risk individuals, leading to proactive interventions such as targeted screenings or personalized preventive strategies. This genetic information, combined with traditional risk factors, offers a more comprehensive view of a person’s health trajectory.
The study utilized data from the All of Us Research Program, a landmark initiative by the U.S. National Institutes of Health (NIH) designed to build one of the most diverse and comprehensive health databases. With over 500,000 participants, the AoU dataset includes structured clinical events, genomic profiles, wearable data, and survey-derived indicators, making it an ideal resource for developing and evaluating multimodal models.
Evaluation on AoU data demonstrated the model’s significant predictive value for the onset of various conditions, particularly Type 2 Diabetes (T2D). The multimodal EHR+PRS model consistently outperformed the EHR-only model for T2D prediction, showing statistically significant improvements in metrics like AUPRC and AUROC. Interestingly, the incremental benefit of PRS was most pronounced when less EHR history was available, suggesting that genetic predispositions play a larger role in early risk assessment before extensive clinical data accumulates.
A novel aspect of this research is the introduction of a ‘path-computing probability’ for generated health trajectories. Unlike traditional Monte Carlo sampling which yields bucketed prediction scores, this new method provides much more nuanced and continuous risk estimates. This allows for finer-grained risk stratification and more granular operational decision-making, as it extracts more meaningful signals from the generated trajectories.
The researchers also validated that the model’s predictions align with established PRS, confirming that the self-supervised learning process effectively learned meaningful biological signals across modalities. For instance, the model’s T2D risk predictions were positively correlated with known T2D-associated PRS.
Beyond direct prediction, the paper explores transfer learning, showcasing the architecture’s versatility and efficiency. The pre-trained multimodal representations can be adapted for custom classification tasks, such as 10-year stroke prediction or Chronic Obstructive Pulmonary Disease (COPD) onset, either through direct fine-tuning or feature extraction. This flexibility means the model can be applied to a wide range of downstream prediction problems.
This approach is pivotal for unlocking new insights into disease prediction, proactive health management, risk stratification, and personalized treatment strategies. It lays the groundwork for more personalized, equitable, and actionable real-world evidence generation in healthcare. The ultimate vision is a ‘digital twin’ – a high-fidelity virtual replica of an individual’s health that can simulate disease progression and test therapeutic strategies virtually, providing continuous, personalized recommendations for predictive and preventative healthcare. To delve deeper into the technical details, you can read the full paper here.
Also Read:
- Improving Clinical Predictions: Benchmarking Imputation Techniques for ICU Data
- Assessing AI’s Prowess in Clinical Text Analysis: A Deep Dive into NER Performance
Future Directions and Limitations
While promising, the study acknowledges several limitations. The generalizability of PRS, often derived from predominantly European ancestry cohorts, remains a concern for diverse populations. Future work will focus on mitigating bias by up-sampling under-represented ancestries. The study also highlights the need to incorporate other valuable modalities like detailed survey data, wearable device measurements, medical images, and clinical notes. Further research will explore more contemporary transformer models and advanced training techniques to optimize predictive power and achieve benchmark-leading performance, ultimately moving towards truly intelligent, personalized, and preventative healthcare solutions.
