Editing Openai/695f02d3-11f8-800e-993a-ad71db6d42aa (section)

=== Assistant: This is an important point for cancer vaccines and immunotherapies like GPS, and it’s where a lot of investors and even… ===
This is an important point for cancer vaccines and immunotherapies like GPS, and it’s where a lot of investors (and even clinicians) get tripped up.

I’ll walk through it step-by-step and then tie it directly back to why medians can mislead you.

==== 1️⃣ What “median-based intuition” assumes ====

When people think:

: “Treatment median OS is 12.6 months vs control 8 months — that’s a big win”

they are implicitly assuming:
* The treatment effect starts immediately
* The hazard reduction is constant over time
* The survival curves separate early and stay separated
* The hazard ratio is roughly constant (proportional hazards)

That mental model is basically an exponential survival world.

This is often wrong for immunotherapy.

==== 2️⃣ What delayed immunotherapy effects actually look like ====

For cancer vaccines and immune-based therapies:

===== Early period (first months) =====
* Immune system is still being primed
* Patients may progress or die at similar or even higher rates than control
* Survival curves often overlap or cross

===== Later period =====
* A subset of patients mount a durable immune response
* Their risk of death drops substantially
* A long tail appears in the survival curve

This produces:
* Non-proportional hazards
* A hazard ratio that changes over time

==== 3️⃣ Why medians fail in this situation ====

===== 🔴 Median OS only reflects the 50th percentile =====

The median is the time when half the patients have died.

If:
* Many early events happen before the immune benefit kicks in
* The benefit is concentrated in a subset of responders

Then:
* The median may move very little
* Even if long-term survival is meaningfully improved

====== Example ======
* Control median OS = 8 months
* Treatment median OS = 9 months

That looks unimpressive.

But…
* 2-year OS: - Control = 5% - Treatment = 20%

That is a huge clinical benefit — completely invisible at the median.

==== 4️⃣ How this distorts “median → HR” intuition ====

Recall the shortcut:

HR≈mCmTHR \approx \frac{m_C}{m_T}HR≈mTmC
This assumes:
* Constant hazard ratio
* No delayed effect

But with immunotherapy:
* Early HR ≈ 1.0 or >1.0
* Late HR ≪ 1.0
* Overall HR is a weighted average across time

So:
* Median OS might barely change
* Yet the true HR (from Cox/log-rank) can still be favorable
* Or the opposite can happen: median improves, but HR misses significance

==== 5️⃣ Visual intuition (very important) ====

===== Chemotherapy-like curves =====

<syntaxhighlight>Survival
|
|\
| \
|  \      Control
|   \____
|        \____ Treatment
+---------------------- Time

</syntaxhighlight>
* Early separation
* Median shifts nicely
* HR ≈ median ratio

===== Immunotherapy-like curves =====

<syntaxhighlight>Survival
|
|\
| \_____ Control
|   \
|    \______
|           \_____ Treatment (long tail)
+------------------------- Time

</syntaxhighlight>
* Curves overlap early
* Separate late
* Median barely moves
* Tail drives benefit

==== 6️⃣ Why this matters for statistical testing ====

===== Cox model & log-rank test =====
* Assume proportional hazards
* Lose power when hazards are delayed
* Can underestimate benefit for immunotherapy

This is why immunotherapy trials increasingly look at:
* Landmark survival (e.g., 12-, 18-, 24-month OS)
* Restricted mean survival time (RMST)
* Piecewise hazard ratios
* Milestone survival percentages

==== 7️⃣ REGAL / GPS relevance (without speculating on results) ====

For a WT1-targeted cancer vaccine like GPS:
* Mechanistically, benefit would not be immediate
* Immune priming + memory formation take time
* A subset of patients may derive durable benefit

Therefore:
* Median OS may be a blunt instrument
* HR thresholds derived from medians (like 0.636) may misrepresent true benefit
* Long-term OS and tail separation are often more informative

This is exactly the pattern historically seen with:
* Sipuleucel-T
* Checkpoint inhibitors (early days)
* Several cancer vaccine programs

==== 8️⃣ One-sentence takeaway ====

: Delayed immunotherapy effects shift benefit into the survival tail, where medians and median-derived intuition break down — making long-term survival metrics far more informative than early median OS.

If you want, next we can:
* Quantify how much tail benefit is needed to “beat” a weak median
* Compare HR vs RMST numerically
* Walk through a toy dataset showing median ≈ unchanged but OS benefit is real

Just say the word.