NetraMark 's AI Significantly Outperforms ChatGPT, DeepSeek and Traditional Machine Learning in Clinical Trial Subpopulation Discovery, Offers New Path for Trial Success
TORONTO, June 23, 2025 (GLOBE NEWSWIRE) -- NetraMark Holdings Inc. (the “Company” or “NetraMark”) (CSE: AIAI) (OTCQB: AINMF) (Frankfurt: PF0)a premier artificial intelligence (AI) company that is transforming clinical trials has released a new preprint on arXiv demonstrating that its flagship AI platform, NetraAI, has substantially outperformed leading large language models, DeepSeek and ChatGPT. NetraAI also outperformed traditional machine learning techniques in identifying clinically meaningful patient subgroups from real-world clinical trial data.
The study, titled “Integrating Dynamical Systems Learning with Foundational Models: A Meta- Evolutionary AI Framework for Clinical Trials,” showcases NetraAI’s ability to deliver clear, interpretable, and statistically significant insights from clinical trial data - capabilities that DeepSeek and ChatGPT failed to achieve when prompted with the same tasks.
“The future of AI will depend on a variety of AI agents working in concert—and NetraAI brings something fundamentally distinct to that collaboration,” said Dr. Joseph Geraci, Founder, CSO, and CTO of NetraMark. “Foundational models like DeepSeek and ChatGPT struggled to uncover anything clinically actionable in these datasets. NetraAI not only identified high-impact patient subgroups but delivered clear and clinically meaningful explanations. This is a new class of AI, designed to complement and extend what other less specialized systems cannot achieve alone.”
Study Overview
The study put NetraAI, DeepSeek, ChatGPT, and traditional machine learning models to the test using three, complex clinical trial datasets: CATIE (focused on schizophrenia), CAN-BIND (focused on depression), and COMPASS (focused on pancreatic cancer chemotherapy). These datasets are notoriously difficult—filled with noisy, messy real-world, multi-variable patient data where most AI models stumble. This is the type of data that pharmaceutical companies must work with in real trials.
Summary of Results:
Across both trials, NetraAI was the only AI system that could extract statistically valid, clinically actionable insights from real-world trial data. It revealed interpretable patient subgroups, boosted predictive model performance to upwards of 100% accuracy, and delivered clarity where others generated noise.
- NetraAI: Produced clear subgroups with minimal variables, enabling smarter, faster, and more successful trial strategies.
- DeepSeek & ChatGPT: Failed to generate meaningful outputs. Despite advanced architecture, they were incapable of effectively dealing with these clinical data.
- Traditional ML: Plateaued on noisy data and provided no interpretability—insufficient for real clinical applications.
- The utilization of NetraAI’s insights actually improved all the models generated by the other methods.
Head-to-Head Detailed Trial Comparison
1. CATIE Trial Data (Schizophrenia)
- NetraAI:
- Successfully identified high-value “personas”—specific patient subgroups with shared clinical characteristics linked to treatment response.
- Used only 2–4 clinical variables per subgroup, making results easily actionable and explainable to clinicians and regulators.
- Boosted predictive model performance from 55–66% to over 85%.
- Provided insights ready for direct application in trial design and stratification.
- DeepSeek & ChatGPT:
- Failed to extract a single statistically valid patient subgroup.
- Despite extensive expert prompting and structured input, these models returned:
- Generic summaries lacking clinical relevance.
- Random or extreme patient outliers with no value for decision-making.
- Fundamentally, could not discover robust structures in the data, as the methods it employs are based on known ML methods, and patient heterogeneity is a real problem
- These foundational models are not designed to deal with the complexity of clinical trial data, even after data is structured and simplified
- In practical terms: delivered zero usable insights.
- Traditional ML:
- Struggled with signal-to-noise issues, plateauing at 55–66% accuracy.
- Could not segment the patient population meaningfully.
- Produced black-box predictions with no explanatory pathways—a red flag for regulatory submission.
2. CAN-BIND Trial Data (Depression)
- NetraAI:
- Identified clinically interpretable subgroups tied to treatment success using a handful of variables.
- Enabled models to achieve 91–100% accuracy, with Area Under the Curve (AUCs) nearing 0.99, (AUC - Area Under the Curve is a standard metric used to evaluate the accuracy of predictive models. A higher AUC means the model is better at distinguishing between different patient outcomes, making it a key indicator of clinical relevance and model performance) an almost unheard-of benchmark in real-world mental health data.
- Can be used to empower trial designs that target the most responsive patients, cutting waste and risk.
- DeepSeek & ChatGPT:
- Delivered no actionable insights.
- Could not identify any subgrouping, even when patient numbers were reduced for simplicity.
- Their inability to handle multi-variable relationships in tabular data rendered them non-functional in this context.
- For clinical discovery, these models were dead ends.
- Traditional ML:
- Reached a ceiling around 66% accuracy.
- Could not differentiate which patients were predictable or explain why outcomes varied.
3. COMPASS Trial Data (Pancreatic Cancer Chemotherapy)
- NetraAI:
- Identified clinically relevant patient personas tied to divergent survival outcomes using just 2–3 variables per subgroup.
- Achieved model accuracy of 90–95% on stratified subgroups, significantly outperforming traditional machine learning baselines.
- Persona-specific analyses revealed clear survival curve separations, indicating potential for guiding treatment selection.
- Provided interpretable feature sets and subgroup patterns reflective of real-world clinical complexity in pancreatic oncology.
- DeepSeek & ChatGPT:
- Produced no clinically interpretable patient subgroups.
- Returned output lacking statistical grounding or alignment with survival outcomes.
- Were unable to differentiate between high-risk and low-risk patient profiles.
- Performance was on par with random segmentation, highlighting their unsuitability for precision oncology applications.
- Traditional ML:
- Displayed predictive accuracies less than 60% for the best models, with inconsistent results across cross-validation runs.
- Lacked subgroup resolution and generated opaque predictions without clinical utility.
ChatGPT and DeepSeek: Inadequate for Clinical Discovery
Despite their scale and popularity, ChatGPT and DeepSeek were incapable of extracting actionable patient subgroups from structured clinical datasets. Even when tasked with just 50 patients and extensive expert prompting, both models:
- Failed to identify statistically valid clusters
- Produced only generic summaries or extreme outliers
- Lacked internal structure to interpret trial variables meaningfully
“ChatGPT and DeepSeek are incredibly powerful technologies,” said Dr. Geraci. “But when it comes to uncovering patterns in structured clinical trial data about patient populations, they fall short. NetraAI not only succeeds where they cannot — it delivers validated, regulator-ready insights.”
How NetraAI Works Differently
NetraAI is built upon a unique and differentiated mathematical foundation drawn from dynamical systems theory, evolutionary computation, and information geometry. At its core is a long range memory mechanism for discovering hard to find bundles of outcome-aligned clinical variables and their corresponding patient subtypes, that we call personas.
Each discovered persona is:
- Defined by just a few interpretable variables for feasibility (2–4)
- Supported by statistical significance (p < 0.05)
- Actionable in real-world decision-making
NetraAI also includes a knowledge-layer strategist—nicknamed Dr. Netra—which integrates scientific literature and past experience using a separate LLM layer. Unlike ChatGPT and DeepSeek, which operate as generalists, NetraAI is purpose-built for discovery in clinical trial contexts.
Explainable, Honest, and High-Impact AI
Whereas most machine learning tools “always predict” regardless of confidence, NetraAI explicitly identifies which patients it can predict well—and which it cannot, avoiding overfitting and enabling targeted insight. It explains its reasoning in human-readable terms, helping scientists, sponsors, and regulators align on the implications. It seeks to pinpoint the patient groups likely to benefit from a drug and transforms these insights into practical enrichment criteria.
This makes it ideal for:
- Trial enrichment and design optimization
- Precision medicine strategy
- Regulatory submission support
- Reducing placebo effects and heterogeneous trial failure
Why This Matters
Pharmaceutical companies spend billions on trials, in many cases trials fail not because the drugs don’t work but often because the wrong patients were enrolled. NetraAI addresses that problem. It finds the right patients, reveals actionable hidden treatment signals that can alter the trajectory of a clinical trial, and helps sponsors design trials that are more likely to succeed.
About the Preprint
The full preprint, “Integrating Dynamical Systems Learning with Foundational Models: A Meta- Evolutionary AI Framework for Clinical Trials,” is now available on arXiv - link here. The study represents a defining moment in the shift away from generic AI tools like ChatGPT and DeepSeek toward multi-agent precision-built AI systems designed for medicine.
About NetraMark
NetraMark is a company focused on being a leader in the development of Generative Artificial Intelligence (Gen AI)/Machine Learning (ML) solutions targeted at the Pharmaceutical industry. Its 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 datasets and accurately segment diseases into different types, as well as accurately classify patients for sensitivity to drugs and/or efficacy of treatment.
For further details on the Company please see the Company’s publicly available documents filed on the System for Electronic Document Analysis and Retrieval+ (SEDAR+).
Forward-Looking Statements
This press release contains "forward-looking information " within the meaning of applicable Canadian securities legislation including statements regarding the potential use of NetraMark’s AI solutions to drive intelligent, accurate patient-centric clinical trial optimization, the optimization of clinical trials by uncovering personas, the production of validated, regulatory-ready insights, the potential of generalist or more traditional AI and machine learning tools to provide meaningful patient sub groups from clinical trial data, ,the ability of NetraAI to predict patients, the benefits of NetraAIthe integration ofNetraMark’s AI as a dedicated solution to advance clinical trial success, which are based upon NetraMark’s current internal expectations, estimates, projections, assumptions and beliefs, and views of future events. Forward-looking information can be identified by the use of forward-looking terminology such as “expect”, “likely”, “may”, “will”, “should”, “intend”, “anticipate”, “potential”, “proposed”, “estimate” and other similar words, including negative and grammatical variations thereof, or statements that certain events or conditions “may”, “would” or “will” happen, or by discussions of strategy. Forward-looking information includes estimates, plans, expectations, opinions, forecasts, projections, targets, guidance, or other statements that are not statements of fact. The forward-looking statements are expectations only and are subject to known and unknown risks, uncertainties and other important factors that could cause actual results of the Company or industry results to differ materially from future results, performance or achievements. Any forward-looking information speaks only as of the date on which it is made, and, except as required by law, NetraMark does not undertake any obligation to update or revise any forward-looking information, whether as a result of new information, future events, or otherwise. New factors emerge from time to time, and it is not possible for NetraMark to predict all such factors.
When considering these forward-looking statements, readers should keep in mind the risk factors and other cautionary statements as set out in the materials we file with applicable Canadian securities regulatory authorities on SEDAR+ at www.sedarplus.ca including our Management’s Discussion and Analysis for the year ended September 30, 2024. These risk factors and other factors could cause actual events or results to differ materially from those described in any forward-looking information.
The CSE does not accept responsibility for the adequacy or accuracy of this release.
Contact Information:
Swapan Kakumanu - CFO | swapan@netramark.com | 403-681-2549
© 2025 GlobeNewswire, Inc. All Rights Reserved.
