Berry Consultants

A clinical trial might enroll 400 patients, and we learn an amazing amount from them. Around 100 times as many patients are being treated in health systems, and we don't learn from them in any formal way.

On this week's "In the Interim…", Dr. Scott Berry describes what happens instead: "One doctor gets anecdotal information from three patients. Another doctor gets anecdotal information from four patients. They may be misled."

Large observational datasets help, but causal inference from them is incredibly hard. In Scott's definition, a learning healthcare system is randomized, embedded in routine care, multifactorial, and adaptive, so everyday treatment decisions produce randomized evidence.

When one weighs the impact of 40,000 patients in a randomized learning healthcare system against a 400-patient trial, the results could be enormous.

Hear the full discussion: berryconsultants.com/resource/84-ma…

A listener to "In the Interim..." wrote in to ask what "astronomical infrastructure investment and coordination" a learning healthcare system would need.

Dr. Scott Berry gives his answer: "Astronomical? It's actually the opposite."

A health system already collects data on its patients, and clinicians already make many decisions without high-level evidence. Randomize those decisions through the order sheet, draw outcomes from the data the system already captures, and the added cost is minimal.

As Scott explains, "you don't build a healthcare system to run a learning healthcare system. You import the learning technology into the healthcare system that already exists."

REMAP Periop at UPMC shows what that looks like. Fluids before surgery, perioperative pain management, and antiemetic management are all open questions, and a system that performs tens of thousands of elective surgeries a year could learn from every one of them.

Hear the full discussion: berryconsultants.com/resource/84-ma…

On this week's "In the Interim…", Dr. Scott Berry opens the mailbag with two listener questions, then weighs in on the pace of innovation in clinical trial science.

Highlights from the episode:
✅️ What we mean by a learning healthcare system, through the REMAP framework
✅️ Why embedding trials in routine care lowers cost
✅️ Learning from tens of thousands of surgeries a year with REMAP Periop
✅️ A Bayesian view of multiplicities: hierarchical models over alpha adjustment
✅️ AI, digital twins, and ARPA-H's SURPASS call for adaptive platform trials
✅️ The contrast with the EMA platform trial reflection paper and ICH E20

Have a question for a future episode? Send it to scott@berryconsultants.com or post a comment below.

The full episode is live now: berryconsultants.com/resource/84-ma…

Two-stage randomization makes platform trials easier for patients. A patient is first randomized to a cohort, such as drug B, and then to active B or its matching placebo. Only one therapy has to be explained at consent, and each placebo can match its drug's mode of administration.

The trade-off is several placebo groups, and the question of whether they can be pooled.

On this week's "In the Interim…", Dr. Scott Berry and Dr. Kert Viele point out that this is not an untestable assumption. Patients eligible for several cohorts are randomized among them, so the trial produces a randomized comparison of the placebo groups. If the assumption is a concern, enrolling more controls gives that comparison more power.

In our experience, we have not seen a case where it mattered, and the HEALEY ALS Platform Trial has released a poster showing the placebo groups were similar. Scott and Kert note that it may depend on the indication, with pain and PTSD as areas where it deserves a closer look.

Hear the full discussion: berryconsultants.com/resource/83-re…

On this week's "In the Interim…", Dr. Scott Berry calls non-concurrent controls "the least risky thing and incredibly valuable." Consider an ALS platform trial. Setting aside earlier controls means continuing to enroll more patients on placebo in a deadly disease, rather than fitting a simple model that adjusts for time and making room to test more arms.

Scott's question for trial sponsors: would you rather use non-concurrent controls and run five arms, or use only concurrent controls and run three? Being conservative has consequences.

Dr. Kert Viele adds that the adjustment is not exotic. You can bin enrollment by time and add a covariate to the model. "We have this technology."

The European Medicines Agency's draft reflection paper gives a strong preference to concurrent controls. A strong preference is not a hard no, and we expect to make the case for non-concurrent controls where the science supports them.

Hear the full discussion: berryconsultants.com/resource/83-re…