A Seamless + Adaptive + Sequential Clinical Trial Design

1. Overall Trial Framework

• Indication: Metastatic solid tumor
• Investigational drug: Novel targeted small molecule (monotherapy)
• Traditional development: Phase II (dose-finding) → Phase III (confirmatory)
• Proposed design: Phase II/III seamless adaptive group-sequential trial

This single trial simultaneously incorporates:

• Seamless design (Phase II → Phase III),
• Adaptive design (dose and population selection), and
• Sequential design (interim analyses with early stopping).

---

2. Seamless Design Component

Stage 1 (Exploratory Stage – “Phase II–like”)

• Arms:
• 3 active doses (Low / Medium / High)
• 1 control arm
• Sample size: ~200 patients
• Objectives:
• Characterize dose–response
• Assess preliminary efficacy (PFS)
• Evaluate safety

Seamless Transition

• Enrollment is not paused
• A pre-planned interim analysis is conducted
• One dose is selected to continue
• Data from patients on the selected dose in Stage 1 are retained and contribute to the final confirmatory analysis

➡️ This data reuse across stages defines the seamless design.

---

3. Adaptive Design Component

Interim Analysis for Adaptation (before any sequential look)

• Conducted by an independent DMC
• Uses unblinded data

Adaptive decisions (pre-specified in the protocol/SAP):

1. Dose selection
• Eliminate doses with unacceptable safety
• Select the dose with the most favorable benefit–risk profile
2. Population adaptation
• Exploratory analyses suggest stronger efficacy in a biomarker-positive subgroup (e.g., PD-L1 high)
• Decision rule triggers population enrichment for Stage 2

Resulting adaptations:

• Only the selected dose proceeds
• Stage 2 enrollment may be restricted to the biomarker-positive population
• Statistical adjustments are pre-planned to preserve type I error

➡️ These data-driven, pre-specified modifications constitute the adaptive design.

---

4. Sequential Design Component

Group Sequential Testing Strategy

• Total of three formal analyses:
• One or two interim looks
• One final analysis
• Uses a group sequential design with an alpha-spending function (e.g., O’Brien–Fleming)

Sequential looks (after adaptation is complete):

• Look 1 (~60% information fraction):
• May stop early for efficacy or futility
• Look 2 / Final (100% information):
• Final confirmatory test

Key features:

• No further design changes after the first sequential look
• Alpha is spent only at formal hypothesis-testing looks

➡️ This constitutes the sequential design.

---

5. Statistical Control

• Type I error control is ensured via:
• Pre-specified alpha allocation
• Combination testing (e.g., inverse normal combination test)
• Independent DMC oversight
• Stage 1 data are used for:
• Adaptation (decision-making)
• Final inference only through valid combination methods
• No unplanned data-driven decisions are allowed

---

6. Role of the DMC

The DMC:

• Is the only body with access to unblinded interim data
• Executes adaptive decisions
• Evaluates stopping boundaries
• Provides recommendations without operational bias

---

7. One-Sentence Summary

This trial seamlessly combines exploratory and confirmatory phases, adaptively selects the optimal dose and population based on interim data, and uses a group sequential framework to allow early stopping for efficacy or futility—while maintaining strict control of type I error.

Combination testing Example

Pooled Active Doses vs Control → p₁, p₂, and Final Combined p-value

Endpoint: Progression-Free Survival (PFS)
Model: Cox proportional hazards (log-rank equivalent)
Direction: Treatment benefit (HR < 1) → one-sided tests
Combination method: Inverse Normal Combination Test
Information fractions:

• Stage 1: 40%
• Stage 2: 60%

---

0) Hypothetical Stage 1 Data (same dataset throughout)

Four arms: Control (C) + three active doses (D1, D2, D3).
Cox model with dose as a categorical variable (control as reference) yields:

Comparison	HR	β = log(HR)	SE(β)
D1 vs C	0.95	−0.0513	0.28
D2 vs C	0.75	−0.2877	0.20
D3 vs C	0.70	−0.3567	0.22

---

A) Stage 1: Compute p₁

(Pooled Active Doses vs Control)

We treat all active doses as a single “Active” strategy and compare it with control.

---

Step A1: Inverse-variance weights

• 
• 
• 

---

Step A2: Pooled log-hazard ratio

---

Step A3: Standard error of pooled estimate

---

Step A4: Wald Z statistic

---

Step A5: One-sided p₁

Alternative:

✅ Stage 1 result:

---

B) Stage 2: Compute p₂

(Selected Dose vs Control)

Assume the DMC selects Dose D3 for confirmatory testing.

Stage 2 Cox model (new data only) gives:

• 
• 

---

Step B1: Wald Z statistic

---

Step B2: One-sided p₂

✅ Stage 2 result:

---

C) Final Analysis: Inverse Normal Combination Test

Step C0: Define weights

Using information fractions:

---

Step C1: Convert p-values to normal scores

• 
• 

---

Step C2: Weighted combination

---

Step C3: Final combined p-value

---

✅ Final Results Summary

• Stage 1 (pooled active doses vs control):

• Stage 2 (selected dose vs control):

• Final combined p-value (inverse normal):

---

Key Takeaway

This example shows how:

• Stage 1 provides a global, selection-independent signal,
• Stage 2 provides confirmatory evidence for the selected dose, and
• The inverse normal combination test validly integrates both stages into a single confirmatory inference while preserving type I error.