From supply chains to cancer research: how causal AI is opening new research paths at Constructor University

How can automotive supply-chain research inform a new line of inquiry in cancer treatment research? For Constructor University Lecturer and Postdoctoral Researcher Ishansh Gupta, the answer is Causal AI—a form of artificial intelligence that goes beyond simple pattern recognition to analyze why certain patterns exist, and help predict outcomes of changes. Gupta's research into AI applications in the automotive sector has produced multiple publications—including his latest study on supplier risk and evaluation published in the journal Supply Chain Analytics. Now, Gupta is exploring how this same technology could potentially support health professionals with decision-making during cancer treatment, research that has been accepted for presentation next month at NeurIPS, one of the world's foremost AI research conferences.

The connection between the two projects was sparked in 2025 when Gupta attended the Lindau Nobel Laureate Meeting, where he met Nobel Prize-winning researchers who challenged him to think beyond the traditional bounds of his field. When asked which field outside of economics had most influenced his life, the answer was clear and immediate: medicine, the profession of both his parents. Gupta found himself challenged to consider whether he could apply his work in industrial AI and the automotive sector to other fields like human health.

“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 Gupta.

Causal AI and cancer treatment

Together with Prof. Dr.-Ing Hendro Wicaksono, who leads Constructor University's Data-Driven Collaborative Decision Making in Complex Industrial Systems research group, Gupta began a new research project to investigate how Causal AI and LLMs could be leveraged to support decision-making during cancer treatment. The study examined one of the key challenges to using AI to support individualized treatment decisions: a model may produce recommendations for every patient, even when the available evidence is not strong enough to support it. Gupta and his colleagues wanted to examine when an AI system could and should provide a treatment direction, versus when it should acknowledge that the available evidence is uncertain and abstain from making a recommendation.

“In business, a poor decision can cost millions. In medicine, the consequences can be measured in human lives,” Gupta said. “The future of AI should not simply be about producing more accurate predictions. It should be about helping human professionals make better, more transparent and more trustworthy decisions.”  

Gupta and his colleagues used real-world ovarian cancer data and simulated treatment assignments and outcomes to test an approach that combined causal machine learning with methods designed to measure uncertainty and give the system access to LLM-generated explanations. The system was also allowed to abstain from giving treatment directions if the model concluded that the available evidence was insufficient.

The results showed that improving a model’s predictions did not necessarily make its individual recommendations reliable. Under the conditions tested, the system abstained from recommending treatment in 99.4% of cases at the selected confidence target. The findings highlight a key distinction in AI-assisted decision-making: producing a prediction is not the same as producing evidence strong enough to guide a decision. This suggests that causal AI should not be judged solely on its ability to predict outcomes and recommend actions, but also on its ability to assess the likelihood of the outcomes, including knowing when to refrain from giving an answer at all.

"This is precisely what makes our research so interesting," reflected Gupta. "We are looking for ways to combine causal models, LLMs and human expertise to build systems that can reliably optimize decision-making, including allowing the reasoning behind each recommendation to be examined, questions and validated."

Gupta and Prof. Wicaksono will present the research, “When Individualized Causal Predictions Are Not Actionable: Failure Analysis, Abstention and a Runtime-Checked Narrative Agent on Ovarian-Cancer Covariates,” at the NeurlPS 2026 conference on December 11 in Sydney, Australia.

"NeurIPS is among the most prestigious conferences in artificial intelligence, so having our work accepted at one of its workshops is personally a very meaningful milestone," said Gupta. "I’m excited to exchange ideas with leading AI researchers, challenge our assumptions and hopefully establish collaborations that help take this work further. For Constructor, it demonstrates our ability to build on established strengths in industrial AI and move into new interdisciplinary fields like medical AI. Hopefully, this is the beginning of a much larger research direction, rather than a single publication."

Reimagining supply chain decision-making

The industrial research that would eventually inspire the cancer treatment project began with a question about how AI can support decisions in complex supply chains. In "a Causal Analytics Approach for Interpretable Supply Chain Risk and Supplier Performance Evaluation," Gupta, Prof. Wicaksono et al. collaborated with industry partner BMW Group to explore how causal AI could help analyze possible cause-and-effect relationships among problems such as incorrect deliveries, missing parts, backlogs and supplier escalations. The approach aimed to not only identify risks, but to show why certain risks may develop and what could potentially be changed.

"Imagine a strike or natural disaster disrupting a supplier," explained Gupta. "A traditional predictive model might flag that supplier as high-risk. Our approach aimed to go further. It connected external events with internal operational indicators, allowing managers to examine possible causal pathways, such as how a disruption might increase delivery delays, create backlogs and eventually escalate supplier risk.

The researchers worked with 18 domain experts representing over 300 years of combined professional experience, as well as actual internal supplier-performance data. LLMs helped organize information from external events such as strikes, natural disasters and supplier disruptions into draft causal models, with were reviewed and validated by the domain experts rather than accepted automatically.

The simulations run using these models indicated that reducing supplier backlogs by 50% to 70% consistently reduced predicted escalation levels. While it does not confirm that the same outcome would occur in a live supply chain, the study illustrates how causal AI can help decision-makers explore possible interventions before taking action. "The broader idea is to move from firefighting after disruptions occur to understanding possible consequences and evaluating preventive actions beforehand," said Gupta.

From simulation to reality

While both projects were simulation-based, Gupta and his colleagues in Prof. Wicaksono's research group are committed to pursuing research outcomes with real-world application and impact.

“For me, this is one of the most important aspects of our work," said Gupta. "A scientifically interesting model is not automatically a trustworthy decision-making system. In our industrial research, domain expertise helped us validate causal relationships and challenge what the algorithms discovered. We want to carry the same philosophy into medicine, but with even greater scientific and ethical rigor.”

" 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."

Follow these links to learn more about AI study programs and Prof. Wicaksono's research group at Constructor University.