Dr. Joseph Geraci reports

PEER-REVIEWED STUDY PUBLISHED IN MDPI'S AI JOURNAL SHOWS NETRAAI UNCOVERS CLINICAL TRIAL SIGNALS MISSED BY CONVENTIONAL AI AND FOUNDATION MODELS

Netramark Holdings Inc. has published new research in MDPI's peer-reviewed AI (artificial intelligence) journal.

The study, "Interpretable Subgroup Discovery with Abstention in Small, Heterogeneous Clinical Trials: A Retrospective Multi-Dataset Study," evaluated NetraMark's proprietary NetraAI technology across clinical data sets in schizophrenia, major depressive disorder and pancreatic cancer.

• NetraAI's biological and clinical signatures successfully boosted the accuracy of all eight evaluated algorithms across every single data set tested.
• The study showed improved performance across conventional machine learning, neural-network and foundation-model (including Fable 5) approaches in schizophrenia, major depressive disorder and pancreatic cancer.
• Findings support what the company believes is the potential of an unbiased discovery-first approach to uncover clinically relevant structure in complex trial data and inform future patient-stratification and trial-enrichment strategies.

The study benchmarked NetraAI against eight industry-standard predictive models, ranging from traditional biostatistical methods and deep neural networks to major AI foundation models. When provided with the compact variables discovered by NetraAI, every model saw an immediate, significant increase in predictive power across all three clinical data sets tested. The findings provide what the company believes is important evidence of NetraAI's fundamental differentiation from conventional machine learning, neural networks and newer foundation-model approaches.

Across all three data sets, researchers tested eight different AI and machine-learning methods using the original clinical data. Performance was generally weak to modest.

When those same AI methods were instead provided with the variables discovered by NetraAI, every method improved on every dataset tested.

"This study shows what makes NetraAI different," said Dr. Joseph Geraci, PhD, founder, chief scientific and technical officer of NetraMark. "NetraAI is not another prediction model. It is designed to find clinically meaningful patient subpopulations that other analytical methods can miss. In this study, every AI and machine learning method tested performed better after being given the variables discovered by NetraAI. That is the role we believe NetraAI can play in clinical development: helping companies find the patients, variables and signals that are otherwise hidden in complex trial data."

A different kind of AI for clinical trials

Most AI systems are built to learn from enormous data sets. Clinical trials present a very different challenge: relatively few patients, but often thousands of clinical, genomic and biological measurements.

NetraAI was built for this problem. Rather than trying to predict an answer for every patient, NetraAI searches for small, explainable combinations of variables that identify meaningful patient subpopulations. When the data do not support a reliable assignment, NetraAI can make a "no call" instead of forcing a prediction.

To test whether general-purpose AI models could natively discover these same patient subpopulations without NetraAI's specialized approach, the study also evaluated a pretrained foundation model. On the original clinical trial data, its performance was near chance 0.53 to 0.59 area under curve (AUC). When it was given the variables discovered by NetraAI, its performance improved (achieving a near perfect AUC in the pancreatic cancer analysis) for a specific explainable subpopulation discovered by NetraAI.

"We do not see foundation models as competing with NetraAI," Dr. Geraci continued. "The study suggests that NetraAI can do something complementary by discovering the clinical structure first. Other AI systems can then make better use of the data. For pharmaceutical companies, that could mean a clearer path to characterize the patients most likely to drive a successful clinical trial. The power of explainability makes all of this possible. "

One platform across multiple diseases

In this paper, the same NetraAI architecture was evaluated across schizophrenia, major depressive disorder and pancreatic cancer, using data sets ranging from clinical measurements to high-dimensional genomic information.

In the pancreatic cancer analysis, for example, NetraAI reduced a search space containing approximately 25,000 genomic variables to a compact three-variable signature associated with observed regimen-associated response. While none of the eight other industry-standard AI and machine learning models tested could independently discover this compact signature on their own, they were all able to leverage NetraAI's variables to achieve near-perfect predictive accuracy within the called patient subgroup. The finding remains exploratory and requires external validation.

NetraMark believes the ability to identify such compact, explainable patient subgroups could ultimately help pharmaceutical sponsors better understand drug response, comparator, or placebo response and patient heterogeneity, and use those insights to inform future trial design and patient-enrichment strategies.

"Pharmaceutical companies spend enormous amounts of capital and years generating clinical trial data," said George Achilleos, chief executive officer of Netramark "Our commercial thesis is straightforward: valuable information may remain hidden inside those data sets after conventional analysis is complete. NetraAI is designed to find it. If those discoveries can help sponsors make better decisions about the next clinical trial, we believe that represents a substantial opportunity for Netramark."

About NetraAI

NetraAI is engineered to include focus mechanisms that separate small data sets into explainable and unexplainable subsets. Unexplainable subsets are collections of patients that can lead to suboptimal overfit models and inaccurate insights due to poor correlations with the variables involved. NetraAI uses explainable subsets to derive insights and hypotheses (including factors that influence treatment and placebo responses and adverse events), potentially increasing the likelihood of a clinical trial's success. Many other AI methods lack these focus mechanisms and assign every patient to a class, often leading to overfitting, which drowns out critical information that could have been used to improve a trial's chance of success.

About Netramark Holdings Inc.

Netramark is focused on being a leader in the development of generative artificial intelligence (Gen AI)/machine learning (ML) solutions that are targeted at the biotechnology and pharmaceutical industries. The company's product offering uses a novel topology-based algorithm that has the ability to parse patient data sets into subsets of people that are strongly related according to several variables simultaneously. This allows Netramark to use a variety of ML methods, depending on the character and size of the data, to transform the data into powerfully intelligent data that activates traditional AI/ML methods. The result is that Netramark can work with much smaller data sets often prevalent in clinical trials, and accurately segment diseases into different types, as well as help to classify patients for sensitivity to drugs and/or efficacy of treatment.

We seek Safe Harbor.