Biomedical Artificial Intelligence · global
Teaching AI When to Abstain: NetraAI Searches Clinical Data for Patient Subgroups
A retrospective study across three disease areas shows that identifying a small number of key variables first can improve the predictive performance of existing models. But the system classifies only some patients, and the striking results still require validation before they can inform treatment selection.
Patients with the same disease may respond very differently to treatment; when everyone is included in a single analysis, important clues about a minority of patients may be lost. NetraAI, developed by NetraMark, attempts to identify patient subgroups with shared characteristics in clinical data and abstain when evidence is insufficient. A newly published retrospective study provides preliminary support for this analytical strategy while also revealing its limitations.
The peer-reviewed paper was published on September 4, 2026, in MDPI’s journal *AI*. NetraMark announced the news on September 10 and added the study to its publication list. The study used three existing datasets: CATIE for schizophrenia, CAN-BIND for depression, and COMPASS for pancreatic cancer. These involved, respectively, comparisons of antipsychotic drugs, treatment response to the antidepressant escitalopram, and associations between different chemotherapy regimens and response.
NetraAI’s task is to find compact combinations of variables that describe specific patient groups, then allow other models to use these clues. For cases it cannot classify consistently, the system can return “No Call,” indicating that it makes no determination. This design allows researchers to distinguish which features may carry useful information and to which patients that information applies.
The company’s announcement emphasized that all eight methods tested across the three datasets showed improved predictive performance after using variables identified by NetraAI. The paper further explained that these results came from an internal evaluation: with all patients retained and difficult-to-classify cases included, the overall range of AUC, a measure of model discrimination, rose from 0.46–0.62 in analyses using the original variables to 0.54–0.78. This supports the utility of variable selection, but does not yet allow an inference that it can improve the success rate of new trials.
When the analysis was further restricted to subgroups the system could classify consistently, the effects varied by dataset. All eight methods improved again in the depression and pancreatic cancer datasets, while only one improved in the schizophrenia dataset. The patients the system could actually classify accounted for approximately 27.7%–40.4% of the total in each dataset. High performance scores within these subgroups must therefore be considered alongside the fact that most patients received no determination.
The pancreatic cancer analysis was particularly striking. According to the company’s announcement, NetraAI narrowed approximately 25,000 genomic variables down to a three-variable feature combination. The paper described these as three single-nucleotide variants associated with the ranking of chemotherapy regimen–related responses among classifiable patients. However, COMPASS is an observational dataset, and this association cannot directly establish which treatment is more effective for a particular patient. Both the company and the paper characterized the findings as exploratory and requiring external validation.
The NetraMark team participated in the study, whose current purpose is to generate testable hypotheses about patient subgroups. Using these findings to decide whom to enroll in the next trial, or even to help guide clinical treatment selection, would require independent data and prospectively planned studies to confirm whether these features can be reproduced and to assess the situation of patients who remain unclassified. The real test is whether clues found in historical data continue to hold in the next group of patients.