Two people can share a diagnosis and have different disease drivers, histories, and responses to treatment. Precision medicine tries to make those differences useful for a specific decision.
AI can help connect molecular measurements with clinical records and outcomes. But finding a pattern is only the start. The harder question is whether that pattern identifies a treatment choice that improves outcomes for the people who will use it.
A useful patient model needs a clearly defined job. Predicting who will do poorly is different from predicting who will benefit more from a particular treatment.
The Patient Stratification Challenge
Patient stratification means defining groups using characteristics relevant to a research question. These might include a tumor alteration, a laboratory measurement, disease stage, or previous treatment.
Some groups are defined before a study using an established biological rationale. Others emerge from exploratory analysis. The second approach can generate useful hypotheses, but searching many possible groups also creates opportunities for chance findings.
A subgroup should therefore have a reproducible definition. What was measured, with which assay, at what time, and using which threshold? A category that changes when the data are cleaned differently will be difficult to apply consistently.
Prognosis and Treatment Benefit
The FDA-NIH BEST resource distinguishes prognostic biomarkers, which indicate the likelihood of a clinical outcome, from predictive biomarkers, which help identify differences in response to an exposure such as treatment.
A treated group can have a favorable outcome because its members were likely to do well anyway. To understand treatment benefit, researchers need an appropriate comparison. Outcomes within the treated group alone cannot supply it.
Difference: +20 percentage points
Difference: +20 percentage points
The illustration deliberately omits sampling uncertainty. In a real study, subgroup estimates can be imprecise, particularly when only a few participants have the biomarker. Researchers need a planned comparison of treatment effects, rather than assuming that significance in one group and nonsignificance in another proves a difference between groups.
AI-Powered Stratification
Models can help analyze complex molecular profiles, identify candidate markers, or combine several measurements into a score. Those tasks become more useful when they start with an explicit intended use: finding a trial population, estimating prognosis, or supporting a treatment comparison.
Keep the timing clear. A model intended for use before treatment cannot rely on a laboratory result collected afterward. Missingness can also carry clues about a hospital’s workflow instead of a patient’s biology.
At AgentCures, considering patient differences during target selection and trial planning is part of our approach. A proposed group should remain connected to the evidence and assumptions used to define it.
What It Means for Patients
HER2 illustrates how a molecular characteristic can help define a treatment-development strategy. A 2001 randomized trial studied chemotherapy with or without trastuzumab in metastatic breast cancer that overexpressed HER2 and demonstrated clinical benefit from adding trastuzumab in that setting.
That result came from a defined test, treatment, population, and comparison. It was not established merely by clustering patients into groups, and it does not imply that every molecular distinction is actionable.
The practical chain matters: a reliable measurement identifies a reproducible group, and clinical evidence supports a decision for that group.
The Clinical Trial Advantage
Enriching a trial for a biologically relevant group may make its question more precise. It can also make recruitment harder and limit what the result says about people outside the group.
Teams need to examine the marker’s prevalence, the availability of testing, and how many otherwise eligible people may be excluded. A highly specific hypothesis is of limited practical use if the study cannot find participants or the necessary assay cannot be deployed reliably.
Scaling Precision Medicine
More records create more opportunities to learn and more ways to accidentally overstate performance. Repeated visits from the same person can appear very similar. If some are used for training and others for testing, a model may appear better at handling new patients than it really is.
Visit 1 · Visit 2 · Visit 3
Visit 1 · Visit 2
Visit 1 · Visit 2 · Visit 3
No visit from person C enters model fitting or tuning. This is one evaluation requirement, not sufficient evidence for clinical use.
A 2024 study of data leakage in clinical machine learning demonstrates how evaluation choices can inflate prediction performance. The evaluation must match the proposed use, including who is new to the model and what information is available at the decision time.
Also inspect performance across relevant patient groups. A good average can hide poor calibration or a high error rate in an underrepresented group. If a model reports probabilities, compare predicted probabilities with observed outcomes rather than relying only on how well it ranks people.
From Bench to Bedside
Precision medicine becomes useful through a chain of evidence, not a single accurate-looking model. The measurement needs to work, the patient group needs to be reproducible, and the treatment decision needs support in the intended setting.
AI can help investigate each link. Its value grows when it makes the limits visible and helps a team design the next study needed to decide whether a patient difference should change care.