Causal AI Tool Withholds Ovarian Cancer Treatment Recommendations When Evidence Insufficient
Posted on 18 Nov 2026
Ovarian cancer treatment often requires individualized decisions, particularly when available patient data do not clearly favor one therapeutic approach. However, artificial intelligence (AI) systems may still generate recommendations despite insufficient evidence, raising concerns about their reliability in high-stakes clinical decisions. To improve decision-making, researchers have developed a causal AI approach that can withhold treatment recommendations when the available evidence is inconclusive.
Researchers at Constructor University are investigating how causal artificial intelligence (AI) and large language models (LLMs) can support decision-making in cancer treatment. Their approach combines causal machine learning, uncertainty assessment, and AI-generated explanations to distinguish between predicting treatment outcomes and determining whether the available evidence is sufficiently reliable to guide individual clinical decisions.
The study used real-world ovarian cancer patient data with simulated treatment assignments and outcomes to evaluate the system’s ability to generate reliable recommendations. Rather than providing a treatment direction when evidence was insufficient, the model was designed to abstain from making a recommendation. At the selected confidence threshold, it withheld recommendations in 99.4% of cases, demonstrating that improved predictive performance alone does not necessarily translate into reliable individualized treatment guidance.
The work builds on earlier causal AI research in supply-chain decision-making conducted with BMW Group and published in Supply Chain Analytics. That prior project used causal models and large language models to examine possible cause-and-effect pathways in supplier risk, including delivery delays, backlogs, and escalation levels. The cancer-focused research, titled “When Individualized Causal Predictions Are Not Actionable: Failure Analysis, Abstention and a Runtime-Checked Narrative Agent on Ovarian-Cancer Covariates,” will be presented at NeurIPS 2026 on Dec. 11 in Sydney, Australia.
“If we can use causal AI to understand complex industrial systems and make better decisions, why couldn't we explore similar ideas in medicine, where the consequences of those decisions are far more significant?” said Ishansh Gupta, Constructor University Lecturer and Postdoctoral Researcher.
“Our cancer research is still at an early stage. The next steps would involve working closely with medical researchers and clinicians, validating the models against appropriate clinical data and, eventually, investigating prospective clinical evaluations and trials, where justified and approved. The ambition is not to replace doctors with AI, but to develop tools that could eventually help them understand complex treatment decisions with greater confidence,” added Gupta.
Related Links
Constructor University