From Prediction to Enrichment
Prediction is useful when it changes what happens next.
A predictive model does not have to make the final clinical decision. In a trial setting, it can help determine who should move to the next, more expensive stage of screening.
The enrichment workflow
Move expensive confirmation downstream
The model sits upstream of protocol-defined confirmatory testing. Its role is to help trial teams decide how aggressively to narrow the candidate pool.
01
Candidate pool
Start with the population available for screening rather than requiring every candidate to undergo confirmatory testing.
02
Risk stratification
Apply a frozen model to available clinical and cognitive data and estimate the probability of the target endpoint.
03
Prioritization
Use the trial's chosen operating point to identify candidates for the next stage of screening.
04
Confirmatory testing
Reserve PET, CSF, or other protocol-defined confirmation for the candidates selected by the enrichment strategy.
Alzheimer's disease case
A+T+ biomarker status as an enrichment example
In one current research case, Clinovia uses age, sex, MMSE, RAVLT immediate recall, and LDELTOTAL to estimate the likelihood of A+T+ CSF biomarker status.
Current model
Clinical + cognitive A+T+ classifier
Five inputs, one predicted probability, and a set of frozen operating points chosen around the trial team's tolerance for missed candidates.
Choosing an operating point
Different trials can tolerate different tradeoffs
Instead of treating one threshold as universally correct, the model can be operated at different sensitivity targets. Higher sensitivity captures more potential A+T+ candidates while increasing the number sent forward for confirmation.
| Target sensitivity | Threshold | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|
| 90% | 0.322 | 90.3% | 32.5% | 51.0% | 81.1% |
| 95% | 0.268 | 95.4% | 24.3% | 49.5% | 87.1% |
| 98% | 0.145 | 98.3% | 10.2% | 46.0% | 88.6% |
ADNI development cohort, N=542. Thresholds were selected from out-of-fold predictions at each target sensitivity, maximizing specificity subject to the target.
What this demonstrates
The model becomes part of a decision system
The model does not replace protocol-defined biomarker confirmation.
The threshold can be selected according to the trial's tolerance for missed candidates.
The output can be used to prioritize candidates rather than make a diagnostic determination.
The same framework can be evaluated for other endpoints and trial populations.
Validation status: this A+T+ classifier has been evaluated internally in ADNI but has not yet undergone external validation. The operating points shown above are therefore research results, not validated clinical screening thresholds.
Beyond the AD case
One application of a broader approach
Alzheimer's disease is the current research case. The underlying approach is broader: use available data to identify predictive signal, validate it in the target population, and determine whether that signal can improve how candidates move through a clinical trial.
Discuss a trial applicationFor Research Use Only. The models and examples described on this page are not diagnostic devices and do not replace protocol-defined clinical or biomarker assessment.