Medicine's prediction problem is becoming personal
A single blood methylation profile stratifies ten-year risk across 21 common diseases, a new model estimates treatment effects for individual patients from health records, and a bedside tool predicts immunotherapy response in lung cancer. The shared move: from population averages to one patient.
- Methylation risk scores for eight diseases reached a holdout C-index of at least 0.80, in a preprint. Self-reported m:…6304
- The highest risk tertile of methylation risk scores contained 73 to 90% of the ten-year events for the top five diseases, in a preprint. Self-reported m:…6304
- The ITE-Transformer improves discrimination and calibration of treatment assignment and outcome models compared with tabular machine learning and non-pre-trained transformers, in a preprint. Self-reported m:…6303
- LCatlas predicts immunotherapy outcomes in non-small cell lung cancer with 82% accuracy in a 28-patient validation cohort, in a preprint. Self-reported m:…9058
One blood profile, twenty-one diseases
The most striking of the three papers starts with a single blood draw. A preprint reports that one blood methylation profile stratifies ten-year risk across 21 common diseases, with methylation risk scores for eight diseases reaching a holdout C-index of at least 0.80. A C-index of 0.80 means the model ranks a person who will develop the disease as higher risk than a person who will not about four times out of five, which is strong for a single test applied across many conditions at once. m:…6304
The scores also beat the established alternatives. Methylation risk scores outperformed published polygenic and clinical covariate models for seven of eight top-tier diseases. And the concentration of risk is sharp: the highest risk tertile of methylation risk scores contained 73 to 90% of the ten-year events for the top five diseases. In plain terms, a third of the population, as ranked by the test, holds roughly three-quarters to nine-tenths of the events. That is the shape of a screening tool: most of the risk sits in a group small enough to do something about. m:…6304
From risk to treatment choice
Knowing who is at risk is only half the problem. The harder half is knowing which treatment helps which patient, and a second preprint takes that on directly. The ITE-Transformer is a pre-trained foundation model for predicting individualized treatment effects from electronic health records, and it improves discrimination and calibration of treatment assignment and outcome models compared with standard tabular machine learning and non-pre-trained transformer baselines. m:…6303
The framework enables the study of counterfactual questions regarding individual patient outcomes under alternative treatment strategies: what would have happened to this patient on the other drug. And the model identifies substantial heterogeneity in predicted treatment effects associated with patient diagnoses, medication histories and comorbidities. That heterogeneity is the point. If every patient responded the same way, a population average would be enough. The model is useful exactly to the extent that responses differ, and it finds that they differ a lot. m:…6303
A tool at the bedside
The third paper is the most concrete: a point-of-care tool called LCatlas that predicts immunotherapy outcomes in non-small cell lung cancer using nine clinical and molecular variables. It achieved an AUC of 79%, with a range of 72 to 85%, in cross-validation, and demonstrated 82% prediction accuracy in a 28-patient validation cohort. Nine variables is a small enough number to collect in a clinic, which is the design goal: a tool a clinician can actually use, rather than a model that needs data nobody has. m:…9058
The validation cohort of 28 patients is small, and the authors' own framing is that further prospective validation is required. That caveat applies to all three papers, which are preprints. But the direction is consistent across them: prediction is moving from the population to the patient, and from risk alone to risk plus treatment response. A methylation score that concentrates risk, a model that estimates individual treatment effects, and a bedside tool that predicts response are three versions of the same idea. m:…6304 m:…6303 m:…9058
What should slow adoption down
The numbers that look strongest also carry the clearest limits. A C-index of 0.80 on holdout data is a model result, not a clinical result, and a 28-patient validation cohort is a pilot, not proof. Methylation profiles change over time, which is both their promise as dynamic biomarkers and a reason to ask how stable a single measurement is. And individualized treatment effect models are only as good as the records they learn from, which carry the biases and gaps of the care that produced them. m:…6304 m:…9058 m:…6303
For health system leaders, the practical question is what to do with a prediction. A risk score that concentrates events in a third of the population is useful only if there is a follow-up: screening, prevention, earlier treatment. A treatment-effect model is useful only if clinicians trust it enough to change a prescription. The papers show the prediction step is advancing quickly. The follow-up step is where the value will be decided. m:…6304 m:…6303
What is still unresolved
- Whether methylation risk scores hold up in prospective cohorts outside the study population.
- Whether ITE-Transformer treatment-effect estimates change real prescribing decisions or outcomes.
- Whether LCatlas accuracy holds in a larger prospective validation cohort.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.