Background: Limitations of Disease Prediction Models Confined to Fragmented Diagnoses

Electronic Health Records (EHRs), which patients generate during hospital visits, are important clues that show changes in an individual's health status. However, until now, they have only recorded fragmented diagnostic results at a specific point in time. Existing disease risk prediction models also have limitations. Models such as the Pooled Cohort Equations (PCE) for assessing cardiovascular disease risk or the Gail model for calculating breast cancer incidence typically calculate the incidence of only a single disease independently. However, in the human body, multiple diseases are interconnected and tend to change dynamically over time. Research has continued to combine this with Polygenic Risk Scores (PRS), which represent a patient's innate genetic risk, to predict long-term disease trends, but it has not been easy to organically connect complex clinical data and genetic information.

Key Findings: Dynamic Disease Prediction AI Model Designed with Data from 680,000 Individuals

The research team from Harvard Medical School, including Massachusetts General Hospital (MGH) and Dana-Farber Cancer Institute, addressed these challenges by introducing a new dynamic Bayesian generative framework called 'ALADYNOULLI'. Inspired by the name of Aladdin's magic lamp and the mathematician Jacob Bernoulli, this Artificial Intelligence (AI) model precisely simulates lifetime disease risk changes by combining longitudinal EHRs and PRS. The key is the introduction of Gaussian process priors, which mathematically smoothly connect disease incidence patterns over time. The research team validated the model using large-scale, multi-ethnic data from the UK Biobank, MGB, and the 'All of Us' cohort of the US National Institutes of Health, which included data from 683,000 individuals. The analysis revealed that ALADYNOULLI exhibits more sophisticated short- and long-term risk discrimination performance than existing standard risk scores, such as PCE, QRISK3, and PREVENT, in over 300 disease areas, including cancer, cardiovascular, and mental disorders. Furthermore, the framework has successfully identified new genetic regions that may have been overlooked in existing Genome-Wide Association Studies (GWAS), which focus on single diseases. This is a noteworthy point. The framework also succeeded in distinguishing biologically distinct sub-types within a group of patients diagnosed with the same disease, raising the possibility of precision diagnostics to the next level.

Significance and Prospects: Challenges to be Addressed to Complete a Lifetime Disease Map

This research has opened the way for clinicians to understand patients not only by the presence or absence of a specific disease but also by the organic disease development trajectory throughout their lives. This patient-centered precision medicine is expected to be a major driving force in upgrading the level of preventive medicine. However, there are also clear challenges to be addressed at this time. Since most large-scale biobanks are biased towards European populations, it is uncertain whether AI trained on data with limited ethnic diversity will perform equally well in Asians and other ethnic groups. Therefore, additional validation with diverse multi-ethnic cohorts is now required. Furthermore, it is necessary to establish a system infrastructure that can quickly refine complex hospital records entered in real-time in the medical field and link them to the AI model while protecting personal information.