What is the AI in your clinical trial not telling you?
After working across multiple clinical programmes, we identified
6 critical points where AI consistently fails to surface the right signals - and most teams never see it coming.
Scroll down⌄
Clinical Intelligence
A Phase II that misses its primary endpoint is not necessarily a dead programme. In our experience, the signal is often still in the data - buried in subgroup heterogeneity, obscured by the wrong endpoint, or invisible to standard statistical approaches not designed for small, complex cohorts.
Below are the six questions we ask every time a biotech team brings us a dataset that didn't deliver the answer they were looking for.
1
Is your population too heterogeneous to detect a real effect?
Standard analysis treats the trial population as uniform. In precision medicine and complex indications, responders and non-responders are often fundamentally different patients. A null result at the population level can mask a strong signal in a subgroup that was never defined at the outset.
Have you looked for latent subgroups in your data using unsupervised approaches - independent of your pre-specified endpoints?
Have you looked for latent subgroups in your data using unsupervised approaches - independent of your pre-specified endpoints?
2
Was the dose right for the patients who actually responded?
Patient-level covariates - weight, comorbidities, baseline severity, concomitant medications - interact with dose in ways aggregate analyses cannot capture. A failed dose-ranging study may contain a clear dose-response relationship hidden within a subset of patients.
Has the relationship between dose, patient characteristics, and outcome been modelled at the individual patient level - not just the arm level?
Has the relationship between dose, patient characteristics, and outcome been modelled at the individual patient level - not just the arm level?
3
Did you measure the right thing at the right time?
Primary endpoints are chosen before the trial. Biological response rarely follows the timeline the protocol assumes. A surrogate biomarker measured at week 12 may tell you more than the clinical endpoint at week 24 - if you know where to look.
Have your secondary and exploratory endpoints - biomarkers, safety signals, PK/PD data - been fully interrogated for surrogate value?
Have your secondary and exploratory endpoints - biomarkers, safety signals, PK/PD data - been fully interrogated for surrogate value?
4
Is there a safety signal that reframes the benefit-risk?
Safety data is routinely analysed for harm. It is rarely analysed for information. Unexpected adverse event patterns or differential tolerability across subgroups can reveal mechanistic insights - and sometimes identify the population in which the drug works best.
Have safety and tolerability patterns been cross-referenced with efficacy data at the patient level - not just reported as aggregate rates?
Have safety and tolerability patterns been cross-referenced with efficacy data at the patient level — not just reported as aggregate rates?
5
Does the sequence of treatment matter?
In indications with complex treatment histories: prior lines of therapy, combination regimens, rescue medications - the order in which treatments are received can fundamentally alter response. This is almost never modelled in standard analysis, yet it can be the difference between a null result and a clear mechanistic story.
Has treatment sequencing and prior therapy history been incorporated as a variable in your outcome analysis?
Has treatment sequencing and prior therapy history been incorporated as a variable in your outcome analysis?
6
What does your data tell you about mechanism of action?
A failed trial is a dataset about your drug's biology, not just its clinical performance. Multi-dimensional patient data - genomics, proteomics, imaging, clinical covariates - often contains patterns that clarify or challenge the assumed MoA. That knowledge has direct value for the next trial design, a partnership conversation, or a regulatory submission.
Has the full breadth of biomarker and molecular data collected in the trial been used to build or refine your mechanistic hypothesis?
Has the full breadth of biomarker and molecular data collected in the trial been used to build or refine your mechanistic hypothesis?