Walter G. Wasser

AI-Powered Health Risk Assessment: A New Frontier

“AI-Powered Forecasting” recently made headlines on the cover of Science, highlighting a groundbreaking deep learning model for faster and more accurate weather predictions. Known as GraphCast, it outperformed the gold-standard system, achieving an astounding 99.7% accuracy for tropospheric predictions. This advancement promises better warnings for extreme weather events like hurricanes and cyclones, potentially saving countless lives. But what if we could apply similar predictive precision to medicine, particularly in forecasting individual health risks to prevent diseases or severe acute events? This vision may soon become a reality.

Current Challenges in Health Risk Prediction

Consider the example of cancer screening. Nearly 90% of women will never develop breast cancer, yet current guidelines recommend annual or biennial mammograms for all women aged 45 and older. This age-based approach results in diagnosing only 14% of all cancers in the United States. Similarly, in a large clinical trial of a multicancer early detection test, only 1.4% of participants showed a cancer signal, with a mere 0.5% being true positives. Now, the Alzheimer’s Association suggests blood tests for healthy individuals to detect biomarkers indicative of early-stage Alzheimer’s. These simplistic, singular measures can lead to false positives, unnecessary anxiety, and wasted resources.

The Potential of Multimodal AI in Medicine

With multimodal AI, there’s an extraordinary opportunity to revolutionize medicine, much like GraphCast has for weather forecasting. Take Alzheimer’s disease, for example. Blood biomarkers are just one data layer. Integrating additional genomic data, such as the APOε4 allele and polygenic risk scores, along with electronic health records, retinal imaging, and wearable biosensor data, could significantly enhance risk assessment. A recent study using machine learning predicted Alzheimer’s up to seven years before diagnosis by analyzing electronic health records. Other predictive factors include gut microbiome diversity and environmental exposures.

Advancing Cancer Prevention

Cancer prevention is another area poised for transformation. The National Human Genome Research Institute’s eMERGE Network recently validated polygenic risk scores for clinical use. These scores, combined with genomic data on cancer predisposition genes, can pinpoint high-risk individuals more accurately. For instance, the Mass General Brigham health system identified a significant portion of their patient population at high risk for breast, colon, and prostate cancers using these scores.

Electronic health records also hold untapped potential for predicting hard-to-diagnose cancers like pancreatic cancer. AI models integrating numerous features from these records have shown promise in identifying high-risk individuals. For example, AI outperformed radiologists in detecting pancreatic abnormalities on CT scans, highlighting the power of integrating multiple data layers for risk assessment.

Moving Beyond Traditional Screening

Instead of relying solely on multicancer early detection (MCED) blood tests, integrating additional risk data can enhance their utility. High-risk individuals could undergo regular surveillance, potentially leading to earlier cancer detection and improved outcomes. However, this approach requires rigorous validation through randomized clinical trials.

The Path Forward

Our ability to integrate diverse data into multimodal AI models is still evolving. Current efforts are primarily focused on electronic health records and genomics, with continuous data from wearable sensors yet to be fully incorporated. Tackling these analytical challenges could pave the way for more accurate and actionable medical forecasting.

Recent Advances in Medical Forecasting

This week, several new reports have highlighted different steps towards accurate medical forecasting:

  1. Kidney Disease Prediction: Using multiple omics (DNA methylations, proteins, micro-RNAs, gene variants) for predicting kidney disease and failure in people with Type 1 diabetes compared to standard clinical models.
  2. Multi-Cancer Risk Prediction: Using national health resources (Denmark) to predict multi-cancer risk over three years.
  3. Type 1 Diabetes Prediction in Children: Enhancing accuracy and timing of prediction.

Summing Up

Accurate forecasting of high-risk individuals for major medical conditions holds great promise for prevention. By incorporating multiple data layers and using multimodal AI, we can improve risk stratification and timing predictions, leading to more effective interventions.

Historically, medical practice has focused on reactive care rather than proactive prevention. Efforts to advance highly accurate medical forecasting offer a chance to shift this paradigm, achieving primary prevention of cardiovascular, metabolic, and neurodegenerative diseases, as well as early cancer detection. This transformation will require robust validation to ensure these models effectively change the natural history of diseases and improve patient outcomes.


Lam, R., Sanchez-Gonzalez, A., Willson, M., Wirnsberger, P., Fortunato, M., Alet, F., Ravuri, S., Ewalds, T., Eaton-Rosen, Z., […], & Battaglia, P. (2023). Learning skillful medium-range global weather forecasting. Science, 382(6677), 1416-1421. DOI: 10.1126/science.adi2336.

Topol, E. J. (2024). Medical forecasting. Science, 384(6698). DOI: 10.1126/science.adp7977.

Topol, E. J. (2024, May 23). Medical Forecasting: The exciting opportunity to provide highly accurate risk and timing assessment at the individual level for actionable conditions to potentiate prevention. Ground Truths. Retrieved from

About the Author
The author is a specialist in nephrology and internal medicine and lives with his wife and family in Jerusalem.