AI Revolutionizes Medicine Through Data-Driven Disease Prediction

Data reveals the future. This statement aptly describes the artificial intelligence era we now inhabit. The medical field is no exception—as genomic data, personal medical records, and lifestyle information on smoking, exercise, and sleep accumulate, AI is opening an era where it can simultaneously interpret this complex information to more accurately identify disease risks and treatment directions.

First, AI based on medical records has begun predicting disease trajectories. A prime example is the Delphi-2M model, a modified transformer architecture. Published in Nature several months ago, this research trained on UK Biobank data from 400,000 individuals and demonstrated its capability on 1.9 million Danish subjects. The model simultaneously predicts incidence rates for over 1,000 diseases based on individuals' past diagnostic histories and lifestyle factors, generates future health trajectories, and estimates disease burden over 10 to 20-year horizons. Rather than examining diseases in isolation, it learns and predicts which conditions are likely to follow others in temporal sequence.

As the speed and accuracy of genomic interpretation have dramatically improved, identifying variants from DNA sequences and predicting diseases including genetic disorders has become standard practice. The most challenging aspect of genomic testing is determining whether a discovered variant actually causes disease. AlphaMissense, developed by Google researchers, learned protein structural context and evolutionary conservation to predict at scale whether missense variants—single amino acid changes—would eliminate protein function. This helps reduce variants of uncertain significance, narrow candidate causes in rare disease diagnosis, and prioritize targetable mutations in cancer genomics.

Many disease-causing variants reside not in protein-coding regions but in non-coding DNA that affects gene expression levels without directly altering proteins. Models like Enformer use transformer-based deep learning to predict how much a specific gene will be expressed in particular cell types from DNA sequence alone, enabling more detailed tracking of how non-coding variants increase disease risk. For splicing variants—changes affecting how RNA is cut and joined—that are easily missed in genomic testing, models like SpliceAI predict both the disruption of normal splice sites and the creation of cryptic new ones. Reflecting on my own research from over 20 years ago, when extensive experiments were required to discover that specific cytokine splicing variants affected cancer patients, I am struck by AI's tremendous power.

Using AI for protein structure prediction and cellular response prediction to guide experimental design has become essential for researchers in the field. By learning single-cell gene expression data and gene perturbation information, AI can predict how cells will change when specific genes are manipulated. This proves valuable for identifying drug targets and designing combination therapy strategies.

Systematically utilizing individual lifestyle data for health management and disease treatment is critically important. Beyond smoking and drinking status and frequency, sleep patterns, and dietary habits, activity metrics including exercise are crucial indicators. Activity levels measured by wearables such as smartwatches or fitness bands with wrist accelerometers are not merely exercise metrics—they serve as important indicators correlated with long-term risk for hundreds of diseases. Research utilizing UK Biobank wrist accelerometer data systematically evaluated associations between moderate-intensity physical activity and incidence of 697 diseases, clearly demonstrating that lifestyle is a core element of disease prevention. When such lifestyle data combines with medical records and genomic analysis, AI will be able to further personalize prevention strategies for individual patients.

One limitation is that many current findings are centered on Western populations. When data skews toward specific population groups, predictions can become biased. Additionally, privacy protection and explainability—understanding why AI makes certain predictions—must be addressed. Nevertheless, recent trends are clear. AI is enabling a new system that narrows disease candidate causes based on genomic and other biological information, predicts future risks from medical records, and designs prevention and lifestyle interventions from behavioral data. This is the medical revolution that AI has brought.