AI models trained on electronic health record (EHR) data have already proven their worth by helping researchers identify patterns that enhance disease prediction and personalize patient care. The ongoing exploration into integrating additional health data modalities presents an exciting opportunity for a more holistic understanding of human health.
Breaking New Ground in Health Data
A notable development in this area comes from Verily Health, a health technology and platform company. In partnership with NVIDIA and the National Institutes of Health (NIH) All of Us Research Program, they have created the first multimodal foundation model that merges EHR data with genomic information. This project exemplifies how combining different health data types can significantly advance precision medicine.
Integrating Diverse Health Dimensions
While EHRs provide critical insights into a patient’s health journey, they represent only a portion of the entire narrative. Jonathan Amar, senior manager of Verily Data Science, emphasizes the need for AI to learn from multiple aspects of health, including both genetic and experiential factors. By merging EHR data with genomic insights, researchers can uncover inherited disease risks that may not be evident until later in life.
The challenge of connecting “nature” (genetics) and “nurture” (lifestyle and clinical history) has been a longstanding hurdle in healthcare AI. Successfully integrating these dimensions can enhance our understanding of disease risk and management.
Testing the Multimodal Approach
Verily’s research aimed to evaluate the effectiveness of combining EHR and genomic data into a single foundation model. Foundation models are adept at identifying complex disease risk patterns because they can learn from extensive and varied datasets before being fine-tuned for specific predictions. As larger multimodal datasets become accessible, these models will help reveal connections that might be missed when relying on a single data source.
However, the integration process was not without difficulties. Amar pointed out the complexities of merging static genetic data with dynamic clinical histories, as they differ fundamentally in structure. The computational requirements of processing both types of data together often exceed the capabilities of conventional CPU-based workflows.
Access to genomic data also poses challenges, as it is frequently unavailable for a significant number of patients. Nation-level initiatives like the All of Us Research Program, which collects multimodal data including genomic information, play a crucial role in overcoming these barriers and enabling the development of more comprehensive health models.
Achievements in Disease Prediction
Verily’s innovative approach allowed for the combination of genetic risk scores—metrics that gauge an individual’s likelihood of developing diseases based on their DNA—with EHR data in a singular foundation model, utilizing NVIDIA GPUs. The results were promising.
Incorporating genetic risk information markedly enhanced the model’s ability to identify individuals at high risk for Type 2 diabetes. The model not only increased the detection of true cases but also minimized false alerts, resulting in a substantial performance boost. Moreover, it demonstrated an ability to project risk over clinically relevant timeframes of five to ten years, rather than focusing solely on short-term predictions.
On the computational front, the model achieved a threefold increase in pre-training efficiency by employing NVIDIA NeMo AutoModel, significantly reducing the time required for insights in a complex research environment.
Opening Doors for Future Research
This groundbreaking model, now available to researchers through GitHub as Forecast™ 1.0, provides global research teams with a secure platform to expedite their disease risk discovery processes.
The improvement in Type 2 diabetes prediction through the combination of EHR and genomic data is just the beginning. The implications of this work extend beyond single diseases or even genomics. It lays the foundation for future multimodal models that integrate various health dimensions, including data from wearables, clinical notes, medical imaging, and patient-reported outcomes, which are vital for precision medicine.
Continuous Improvement and Future Prospects
Verily is committed to enhancing its predictive and generative models through ongoing research and development. As the datasets grow richer and more diverse—including unstructured clinician notes—Forecast™ models will evolve, capturing a more detailed and dynamic view of patient health.
By merging typically disparate data types, this initiative showcases the potential of multimodal AI in improving disease prediction and sets the stage for future advancements in areas like pharmaceutical biomarker discovery and health system precision medicine programs.
In conclusion, the integration of multimodal data into foundation models represents a significant leap toward a more precise understanding of health. The journey ahead is filled with potential, as ongoing innovations promise to unlock deeper insights into human health and disease management.
- Key Takeaways:
- Multimodal foundation models integrate EHR and genomic data for enhanced disease prediction.
- Combining various health data modalities provides a holistic view of patient health.
- Ongoing research and larger datasets will continue to improve predictive models and insights.
Read more → www.biopharmadive.com