Definition
Ratio of patient characteristic variance in later enrollment waves to variance in the first wave. A falling ratio means AI matching is progressively enrolling a narrower population than the trial began with.
The Measurement Gap
An AI patient matching platform accelerated enrollment for a Phase III oncology trial from 18 to 11 months. Enrollment targets met 7 months ahead of schedule. But a biostatistician reviewing trial data notices the enrolled population's profile changed over the enrollment period. Early enrollees had broad distribution: ages 38-74, BMI 19-38, multiple comorbidity profiles. Late enrollees cluster remarkably: ages 52-63, BMI 23-29, mostly one prior treatment line.
The AI is learning from its own successes. Patients who enrolled early and stayed resemble each other. The model updates preferences accordingly. With each wave, the ideal patient profile becomes more specific and the enrolled population more homogeneous.
This is invisible in enrollment metrics. Targets met, dropout rate low, screen failure rate low. The consequence appears in the data: the trial population no longer represents the intended patient population. The trial was designed for a broad label, but the enrolled population skews toward a narrow subgroup. The FDA may restrict the label to this subgroup or require an additional trial in excluded populations.
No current metric tracks whether AI matching is progressively narrowing enrolled population diversity across enrollment waves.
Formula
PMHS = Var(characteristics_wave_N) / Var(characteristics_wave_1)
Where Var = multivariate variance of patient characteristics for the enrollment wave
Worked Example:
Phase III trial, 743 patients over 11 months, analyzed in quarterly waves:
Wave 1 (months 1-3, 162 patients):
- Age variance: 198.7 (SD 14.1 years)
- BMI variance: 31.4 (SD 5.6)
- Prior treatment lines variance: 1.23 (SD 1.11)
- Comorbidity count variance: 2.87 (SD 1.69)
- Composite multivariate variance: 58.6
Wave 4 (months 10-11, 207 patients):
- Age variance: 71.3 (SD 8.4 years)
- BMI variance: 14.8 (SD 3.8)
- Prior treatment lines variance: 0.42 (SD 0.65)
- Comorbidity count variance: 0.93 (SD 0.96)
- Composite multivariate variance: 21.9
PMHS = 21.9 / 58.6 = 0.374.
The final wave has 63% less population diversity than the first. The AI converged on: age 55-62, BMI 24-28, one prior treatment line, 0-1 comorbidities. The screened patient pool maintained variance of 61.2 throughout. The narrowing is in the AI's selections, not the available patients.
Proposed formula, pending validation pilots.