Scientists in Europe have unveiled Delphi-2M, an artificial-intelligence system that can forecast an individual’s risk of developing more than 1,000 diseases—and estimate when they might occur—up to 20 years in advance. The model, detailed this week in Nature, was trained on anonymized health records from about 400,000 participants in the UK Biobank and then externally validated—without retraining—on 1.9 million patient records from Denmark’s National Patient Registry. Performance was comparable to leading single-disease tools for common conditions such as cardiovascular disease and diabetes, the authors report.
Speaking about the work, Dr. Tomas Fitzgerald, a staff scientist at EMBL-EBI and a co-author, joined media briefings to explain that Delphi-2M treats each event in a patient’s history (diagnoses, prescriptions, demographics, lifestyle factors) like “tokens” in a language model, allowing it to learn disease trajectories and the interplay of multiple conditions over time. The approach aims to move beyond one-condition risk calculators toward a holistic, longitudinal forecast of health.
According to EMBL-EBI, the tool is not a diagnostic and isn’t ready for routine clinical use; instead, it could help researchers and health systems prioritize prevention and early interventions, with future versions likely to incorporate genetic and proteomic data to refine predictions.
Independent coverage notes strengths and caveats: the model achieved average AUCs around the mid-0.7s across many endpoints, but performance varied by population and may reflect dataset biases (e.g., healthier, wealthier cohorts in biobanks). Experts say careful validation, fairness checks, and privacy safeguards are essential before deployment.
Bottom line: Delphi-2M represents a significant step toward lifecourse risk prediction at scale—offering “weather-style” health forecasts that could guide personalized prevention—while underscoring the need for rigorous oversight before clinical rollout
