← Articles & Insights

Science & Medicine

AI Can Shorten Clinical Trials. Can It Fast-Forward Ageing?

AI Can Shorten Clinical Trials. Can It Fast-Forward Ageing? A digital patient can be copied, examined and subjected to hypothetical treatments without complaining about parking. That makes the idea attractive: collect someone’s medical history, genetics, blood results and wearable data, then simulate their future against an enormous reference population. Could years of clinical research become months? Sometimes, AI can shorten the route to an answer. But three different shortcuts are frequently bundled together: recruiting fewer people, predicting outcomes more precisely, and spending less time observing what actually happens. Only the last one fast-forwards the biological question. The most credible breakthrough sounds disappointingly administrative: prognostic covariate adjustment. A model predicts a participant’s outcome from information recorded before treatment. Researchers then incorporate that forecast into their analysis of a randomised trial, reducing background variation that might otherwise obscure a treatment’s effect. Human participants still receive treatments and provide observed outcomes. The forecast helps interpret the experiment; it does not become an immortal placebo volunteer. [1] The European Medicines Agency has issued a qualification opinion for PROCOVA, a procedure implementing this approach. That is meaningful progress, but it is not blanket permission to replace human controls with computer-generated patients. Nor does statistical adjustment automatically establish how a drug works differently across sex, age or disease subgroups. Representation, subgroup performance and treatment interactions still require examination.[2] An illustrative calculation explains the appeal. Under a simplified continuous-outcome model, a prognostic score explaining 25% of outcome variance could reduce a 400-person requirement to roughly 300, with comparable statistical power. Actual savings depend on assumptions and predictive performance in the relevant population. Even then, a two-year endpoint remains a two-year endpoint. The recruitment queue may shrink; the calendar does not necessarily surrender. [3] There is also a surprisingly revealing distinction between billions of model parameters and billions of patients. The published DT-GPT study used datasets containing 35,131 intensive-care patients, 16,496 lung-cancer patients and 1,140 Alzheimer’s participants. Its underlying model had seven billion parameters. Those parameters are internal numerical settings, not seven billion extensively documented human beings. [4] More strikingly, this capable forecasting system struggled to identify critically low haemoglobin in its lung- cancer task: its ROC AUC was 0.506, approximately chance-level discrimination. A model can forecast ordinary trajectories reasonably well while missing the unusual event that matters most. The digital patient may look healthy precisely when the physical patient needs urgent attention. [4] That finding changes the questions consumers should ask. An impressive average accuracy score is insufficient. A preventive-health or performance platform should disclose which outcome it predicts, over what period, in which population, and how well it detects important adverse events. The FDA–EMA principles published in January 2026 emphasise a clear context of use, suitable data and risk-based performance assessment. “AI- powered” supplies none of those answers.[5] Longevity introduces a second trap: changing a measurement is not necessarily changing the fate it predicts. A treatment might improve a molecular ageing score without establishing longer life or fewer years of disability. The CALERIE trial illustrates the distinction. In a post-hoc analysis of a two-year randomised calorie-restriction trial, researchers found a roughly 2–3% slowing on DunedinPACE, a DNA-methylation measure of ageing pace. They did not find significant changes on the PhenoAge and GrimAge biological-age measures. The experiment did not demonstrate that participants lived longer.[6] The clocks were measuring different constructs. An attractive number on one dashboard cannot settle every biological question, however enthusiastically a clinic’s marketing department may attempt the conversion. Shorter studies can legitimately use surrogate endpoints: measurements that stand in for outcomes such as illness or death. But the FDA stresses that acceptance requires evidence linking the surrogate to clinical benefit in its intended context. A marker associated with mortality does not automatically become a validated substitute for survival when a treatment changes it.[7] For readers considering anti-ageing services, the practical implication is concrete: before purchasing testing or treatment, establish what decision the result would change. Does the provider offer evidence of better health outcomes, or merely movement in its preferred score? Keeping dated clinical results, medication changes and relevant training records could support future personalised modelling, but collecting every available measurement is not proof that a useful digital twin exists. Performance enhancement raises the same evidentiary problem. An athlete considering hormones or peptides needs answers about benefits, harms and interactions in people resembling that athlete. A model trained on hospital populations does not earn that authority through a muscular avatar. Supporting access to enhancement is compatible with demanding stronger evidence; moral disapproval and scientific uncertainty are separate arguments. A plausible future is faster, better-targeted human research: simulations helping select promising questions, forecasts improving trial precision, and continuing observation checking whether predictions survive contact with biology. That is an inference from current capabilities, not an established timetable. The immediate opportunity is less glamorous than simulated immortality: distinguishing a useful forecast from a sales pitch. A digital twin should help medicine discover uncertainty—not give uncertainty a lab coat and a subscription fee.

References

[1] Schuler et al. Increasing the efficiency of randomized trial estimates via linear adjustment for a prognostic score. [2] EMA. Qualification opinion for Prognostic Covariate Adjustment . [3] On the relevance of prognostic information for clinical trials: A theoretical quantification . The 400-to-300 example is an illustrative calculation, not a reported trial result. [4] Makarov et al. (2025). Large language models forecast patient health trajectories enabling digital twins . [5] FDA–EMA (2026). Guiding Principles of Good AI Practice in Drug Development . [6] Waziry et al. (2023). Effect of long-term caloric restriction on DNA methylation measures of biological aging in healthy adults from the CALERIE trial. [7] FDA. Surrogate Endpoint Resources for Drug and Biologic Development . 3