NMES

138 patients: sizing a trial for electrical stimulation in the PICU

Biostatistics consultingLocal hospitalConsulting analysis / presentation deliveredPower analysis, G*Power, two-sample t-test sample-size formula

Trial sized at 138 patients — achievable in under a year, not the assumed twoCaught a tautological framing in the original question before running any numbers

Context

Critically ill children in the ICU can lose significant muscle mass in a short time. The hospital's clinical team had already run a pilot observational study in the PICU and found a mean 25% reduction in quadriceps muscle cross-sectional area, present in 57% of patients — roughly 2.7% muscle loss per day. The open question was whether Neuromuscular Electrical Stimulation (NMES) could reduce or prevent that atrophy, and before running that trial, the team needed to know how many patients it would take.

Question

Not whether NMES helps at all, but by how much — and how confidently we needed to know it.

As originally posed, the question — does NMES reduce or prevent atrophy in critically ill pediatric patients? — assumes its own answer. It isn't really testable as worded, since any observed change technically satisfies it. Before I could size a trial, I had to fix what we were actually asking.

Approach & rationale

A sample-size calculation for a randomized controlled trial depends on three levers: power (the probability of detecting a real effect, commonly 0.80–0.90), alpha (the false-positive rate, usually 0.05), and the Minimal Clinically Important Difference (MCID) — the smallest change clinicians actually care about. I converted the clinician's MCID into Cohen's d, a standardized effect size, so it could plug into the sample-size formula. Set the MCID too large and the trial may miss a modest but real benefit; too small, and the trial needs far more participants to detect an effect that may not even be clinically meaningful.

G*Power software interface for the sample-size calculation — The G*Power interface used for the power analysis.

Methodology & iteration

Both are valid: the tradeoff is that the broad question needs a larger N and a longer timeline (δ ≈ 10–15%), while the narrow question needs a smaller N and a sharper focus (a fixed ≥75% threshold).

Power for the observed effect

TriedUsing the clinician's pilot data (25% mean CSA loss vs. a ≤10% target in the treatment arm), powering for a 15-percentage-point between-group difference — σ = 0.28, δ = 0.15, α = 0.05, 80% power — gives n ≈ 55 per group in the two-sample t-test formula.

Held upAccounting for the ~15–21% attrition seen in the pilot scales that to 65 per arm, 138 total, matching the clinician's own original estimate and feasible given the hospital's admission volume.

Reframe the question instead

TriedAn alternative: ask a binary, threshold-based question — in PICU patients receiving NMES, what proportion achieve ≥75% CSA preservation over 15 days compared to standard care?

Held upA binary endpoint is simpler statistically and often needs a smaller sample, answering a clinically actionable threshold directly rather than an abstract percent-change average.

In short

Powered at α = 0.05 and 80% power for a 15-percentage-point difference in CSA loss, the trial needs 138 total patients (65 per arm) — comfortably within the hospital's patient volume.

Results

Powered at α = 0.05 and 80% power for a 15-percentage-point difference in CSA loss, the trial needs 138 total patients (65 per arm) — comfortably within the hospital's patient volume. At roughly 730 eligible admissions per year, that's achievable in well under a year of active recruitment, not the two years originally assumed.

Worked two-sample t-test sample-size formula — The full worked formula, arriving at n ≈ 55 per group before attrition.

Significance

Delivered to the hospital's team as a consulting analysis.

Reflection

This project was a different kind of contribution than my other case studies — advising someone else's study design with statistics, rather than running my own bench experiment. The real work wasn't the formula; it was translating between what a clinician cares about (will this actually help my patients) and what a trial needs to prove that (a specific, falsifiable, adequately powered question). Fixing the framing before touching the numbers turned out to matter more than the calculation itself.

BackAll work NextLamins →