Abstract: Our simulations using drug efficacy show sample sizes are infeasible, even for potentially unethical placebo-control trials. We highlight the strength of simulations with observational data to inform the null hypotheses and assess feasibility of future trials of EA treatment.
Introduction
Designing clinical trials for acute brain injury is notoriously difficult, often resulting in "underpowered" studies that fail to yield statistically significant results. When a patient suffers an acute brain injury, such as an aneurysmal subarachnoid hemorrhage (aSAH), their brain frequently enters a state of "electrical storms" known as epileptiform activity (EA). While clinicians routinely prescribe anti-seizure medications (ASMs) like levetiracetam or anesthetics like propofol to suppress this activity, there is a surprising lack of high-quality evidence confirming that these treatments improve long-term functional outcomes.
The gold standard for evidence is the Randomized Controlled Trial (RCT). However, RCTs in neurology face a persistent "Goldilocks" problem: if a trial is too small, it lacks the statistical power to detect a benefit; if it is large enough to be definitive, it often becomes prohibitively expensive and slow. This study utilizes a machine-learning-powered simulator built from real-world observational data to determine the sample sizes required to design a trial that can actually detect treatment efficacy.
Methodology
The researchers developed a "pharmacological model-guided simulator" using a sophisticated two-step approach to create a "digital twin" environment for virtual clinical trials.
1. Mechanistic PK/PD Modeling
Rather than relying on a simple "black box" statistical model, the authors employed a pharmacokinetic-pharmacodynamic (PK/PD) framework. This models how drugs move through the body (PK) and how they specifically suppress electrical brain activity (PD) at the individual patient level. This approach accounts for the high variability in how different patients respond to identical medication dosages.
2. Double Machine Learning (DML)
To predict long-term outcomes—measured by the modified Rankin Scale (mRS) at hospital discharge—the study utilized Double Machine Learning. This technique allows researchers to estimate the causal effect of a treatment while adjusting for a wide array of "confounders," such as baseline age and injury severity, which might otherwise bias the results in observational data.
By combining these methods, the researchers could run thousands of simulated RCTs with varying dosages and sample sizes to perform a rigorous power analysis.
Data Source
The study utilized a retrospective dataset from Massachusetts General Hospital (2012–2017). While the initial database included 136 patients with aSAH, the final cohort for the simulation consisted of 48 adult patients who met strict criteria: at least 24 hours of continuous EEG monitoring and the presence of epileptiform activity. The data included high-resolution EEG recordings, minute-by-minute medication dosing, and detailed clinical demographics.
Key Insights and Results
The "Power Gap"
The most striking finding is the massive discrepancy between current trial sizes and the requirements suggested by the simulation. While many existing neurology trials enroll dozens or a few hundred patients, this study suggests that thousands may be required to prove efficacy.
To achieve 80% statistical power, the required sample sizes were:
- Levetiracetam (7 mg/kg vs. Placebo): 500 patients per arm.
- Levetiracetam (15 mg/kg vs. 7 mg/kg): >4,000 patients per arm.
- Propofol (1 mg/kg/hr vs. Placebo): 1,200 patients per arm.
- Propofol (Dose Comparisons): Simulations comparing different active doses failed to reach 80% power even with more than 1,200 participants per arm.
The Ethical Paradox
The simulations revealed that placebo-controlled trials require the fewest patients to reach statistical significance. However, because EA is known to be harmful, withholding treatment in a placebo group is often considered unethical. Trials comparing two different active doses—the more ethical approach—require sample sizes that are often logistically impossible to achieve.
Non-Linear Benefits and Dosing
For levetiracetam, the model suggested a specific "sweet spot." Interestingly, increasing the dose from 7mg/kg to 15mg/kg actually worsened predicted outcomes. This suggests that the side effects of high-dose ASMs, such as heavy sedation, may outweigh the clinical benefits of seizure suppression.
Discussion and Future Directions
These findings suggest that many previous "negative" trials in the field of brain injury may not have failed because the drugs were ineffective, but because the trials were simply too small to detect the effect.
Limitations
The authors acknowledge that the model was built on data from a single center and a relatively small training set of 48 patients. Furthermore, the study focused on outcomes at hospital discharge; looking at 6-month or 12-month outcomes might reveal larger treatment effects, potentially reducing the required sample size.
Conclusion
The authors propose this simulation framework as a general "recipe" for future trial design. By testing hypotheses in silico before launching expensive clinical trials, researchers can better assess feasibility and optimize dosing strategies. Future work may involve "multi-tiered" trials that combine ASMs with interventions to improve cerebral blood flow, as suppressing seizures alone may be insufficient for full neurological recovery.
Summary updated