Rescuing a Failed Phase III Trial: Finding the Patients Who Respond to Oral Insulin

After a Phase III trial missed its primary endpoint, PhaseV’s causal machine learning identified the patient subgroups that responded to Oramed’s investigational oral insulin, and Oramed re-initiated a Phase III trial in the U.S. under an amended protocol.
At a glance

Phase III Oral Insulin, Type 2 Diabetes

Item
Detail
Trial
Phase III, investigational oral insulin, Type 2 diabetes
Original result
Primary endpoint missed (A1C reduction vs. placebo at 26 weeks)
Responders identified by
Body mass index (BMI), baseline HbA1c and age
Outcome
Phase III re-initiated in the U.S. under an amended protocol
01

The challenge

When a Phase III trial misses its main goal, the whole program is at risk. Oramed’s oral insulin candidate did not reduce A1C (a measure of long-term blood sugar) more than placebo after 26 weeks. But an average result can hide the fact that some patients respond well and others don’t. Without knowing who might benefit, Oramed had no clear path forward.

02

What we did

PhaseV applied its Response Optimizer pipeline, which uses causal machine learning to study how treatment effects differ from patient to patient (known as heterogeneity of treatment effect, or HTE).

  • Analyzed the Phase III data to find patient subgroups with a stronger response.
  • Added Phase II data and validated the findings against multiple external Phase III studies.
  • Guarded against false discoveries. Rigorous sensitivity testing and false discovery analysis helped ensure the patterns were real and not chance findings.
03

The results

  • Subgroups characterized by BMI, baseline HbA1c and age responded significantly to oral insulin
  • Based on this analysis, Oramed re-initiated a Phase III trial of its oral insulin candidate in the U.S. under an amended protocol
Why it matters

What it changes, team by team.

  • Executives and investors: A missed endpoint doesn’t have to end a program. Existing data can show where the value is.
  • Clinical development: A more targeted trial, enrolling the patients most likely to benefit.
  • Biostatisticians: Subgroup findings stress-tested with external validation and false discovery controls.
  • Patients: A path to testing a therapy in the people it is most likely to help.
FAQ

Questions, answered.

Can a failed Phase III trial be rescued?

Sometimes. If the drug works well in some patients but not others, the average can mask it. Causal machine learning can find those patients, and a new trial can then target them.

How do you know the subgroups aren’t a statistical fluke?

Findings were validated against Phase II data and several external Phase III studies, and tested with sensitivity and false discovery analyses.

What is heterogeneity of treatment effect (HTE)?

The idea that the same treatment helps different patients by different amounts.

Source

Related publication: “Machine Learning-Based Optimization of Patient Eligibility for Oral Insulin Treatment: A Post Hoc Analysis of Phase III Trial Results,” Diabetes.

More case studies

All case studies →
How Smarter Site Selection Saved a Phase 3 Lung Cancer Trial 9 Months and $30M

How Smarter Site Selection Saved a Phase 3 Lung Cancer Trial 9 Months and $30M

A Phase 3 lung cancer trial was set to miss its 24-month goal. Smarter sites and design cut 9 months, 33 sites and 95 patients.
$30Mestimated cost saved
How Causal ML Site Selection Nearly Halved Enrollment Time in Two Global COPD Trials

How Causal ML Site Selection Nearly Halved Enrollment Time in Two Global COPD Trials

Two global Phase 3 COPD trials faced up to 3.8 years of enrollment. A causal ML site list cut the estimate to 1.95 years with 25 fewer sites.
1.95years to enroll, down from 3.8
How AI-Optimized Adaptive Design Cut a Phase 3 Infectious Disease Trial by 6 Months

How AI-Optimized Adaptive Design Cut a Phase 3 Infectious Disease Trial by 6 Months

Simulating millions of designs found a Phase 3 adaptive trial that is 25% shorter with 9% fewer patients, saving $6.05M at about 90% power.
25%shorter trial
How Adaptive Enrichment Gave a Phase 2 Rheumatic Disease Trial 25% More Statistical Power

How Adaptive Enrichment Gave a Phase 2 Rheumatic Disease Trial 25% More Statistical Power

Simulation designed a Phase 2 rheumatic disease trial that focuses enrollment on likely responders, with a 25% gain in statistical power.
25%more statistical power