Statistical analysis is the bedrock of all clinical research and drug approvals. However, this was not always the case.

Clinical trials designed using statistical principles became the norm in the US following World War II. They helped developers determine the efficacy of a drug in the broader population, but failed to consider variabilities in treatment response among individuals.

To account for those variabilities, subgroup analysis was used. Subgroup analysis was considered an extension of the traditional statistical analysis. It required teams to pre-select variables they thought would cause outcome variations then analyzed the data using mean, median, standard deviations, correlations, etc. to prove or disprove their prediction.

This type of traditional Subgroup analyses are easy to conduct and understand, however, the results can be limited. Because relying on predetermined sets of variables will not uncover novel traits that impact treatment outcomes that are not part of the original analysis plan. For example, if the team assumes that age or gender have a significant impact on outcomes, they will not discover variables related to genetics, biomarkers, or comorbidities.

Traditional statistical methods also cannot identify patterns that involve the impact of multiple variables – e.g. obese patients over the age of 50, or patients that have a specific biomarker, and failed on a prior treatment.

However, recent advances in analytics capabilities, using artificial intelligence and machine learning (AI/ML), are bringing new agility to subgroup analysis methods. With faster processing and expanded memory capacity, analysts can conduct more robust subgroup analyses, asking broader questions and relying on algorithms to uncover meaningful trends.

The most advanced platforms, including IQVIA’s Subpopulation Optimization and Modeling Solution (SOMS) solution, use algorithms to rapidly analyze broad data sets and identify traits or combinations of traits that are linked to treatment outcomes. The advanced AI/ML and augmented data visualization tools automate the analysis process, uncovering variables that bring additional precision and insight to the analysis process.

This new approach allows for:

• Broader group definitions. Algorithms can determine more precise boundaries for subgroup populations. For example, if a team wants to understand how time since diagnosis affects treatment outcomes, with traditional analysis methods they might pre-group patients based on whether they were diagnosed in the prior 6, 12, and 24 months. Whereas using the AI/ML platform the algorithm can look at outcomes among all patients, then determine which time-frame triggers significant variations. For example, patients diagnosed in prior 18 months have better treatment outcome. This allows analysts to link diagnosis time more precisely to outcomes, which could open trial recruiting to larger populations, and support arguments for expanded label claims.
• Obscure combinations of traits. Algorithms can be trained to identify obscure combinations of traits that would otherwise go undiscovered. For example, IQVIA recently used SOMS to help a pharma company evaluate the safety and efficacy of a neurological drug in a broad patient population. The platform identified 30 primary and derived candidate biomarkers that impacted treatment response and reflected various demographics and disease characteristics. Several of the subpopulations identified were counter-intuitive. The data played a key role in developing inclusion/exclusion criteria for the phase 3 trial and aided in supporting broader label claims.
• Clarify patients at risk. AI/ML subgroup analysis platforms can more precisely identify subpopulations at risk of adverse events. For example, if a percentage of patients in a study experienced high blood pressure after taking the drug, and all are over 65, it would be easy to assume all older patients face this risk. However, the algorithm might discover that all of those patients also had a shared biomarker, or comorbidity, thus narrowing the defined population at risk.  This could allow more people to safely benefit from a drug rather than excluding everyone over age 65.
• Reduce human error. AI driven algorithms aren’t burdened by bias, which avoids Type-1 error rates – in which analysts come to the wrong conclusions when they find data that supports their hypothesis. IQVIA’s SOMs platform is specifically trained to control for such errors, reducing the risk of false assumptions.

A more precise future

This ‘subgroup analysis on steroids’ brings new flexibility that wasn’t previously possible. Analysts can use it to tackle new research questions related to genomics and precision medicine, and better understand the interplay between multiple biomarkers.

These innovations do come with some new complications. Such high dimensional analysis can bring a lack of clarity and deliver results that can’t be easily explained. This risk can be mitigated by working with experts who have proficiency in using AI/ML based analysis methodologies, though such talent can be hard to find.

The lack of available skills is slowing adoption of these technologies and can affect whether regulatory agencies and providers embrace the results.

Protagonist biostatisticians and managers can address this caution by using AI/ML platforms to generate new hypotheses, then validating them using traditional statistical methods. This allows developers to leverage the power of AI/ML for cutting edge of analysis, while providing an extra layer of traditional analysis to validate the results.

As the industry becomes more familiar with this technology it will bring new power and agility to the subgroup analysis process. Eventually we will be able to determine exactly when and why patients will respond to a treatment, bringing greater safety, efficacy, and market performance for new drugs coming to market.