Can we trust an AI agent with a cancer treatment decision?
AI agents can already draft a treatment plan for a cancer patient. The harder question is whether their reasoning holds up well enough for a clinician to rely on it. My thesis developed a framework for evaluating LLM agents that support treatment decisions in advanced lung cancer.
Evaluating Clinical Agentic AI Systems in Oncology
The question
When an AI agent recommends a lung cancer treatment, how do we know whether to trust it? A correct answer is not enough if the reasoning behind it is weak, inconsistent or unsupported.
Why it matters
Treatment decisions in advanced lung cancer weigh many clinical factors. Before agents like these can support oncologists, we need rigorous ways to test them, much as we test any other clinical tool.
The approach
Instead of only checking the final recommendation, the framework looks at the whole path to it, assessing the agent's reasoning and its plan along several dimensions. It was built together with oncologists at AI-On-Lab.
The thesis was defended at Politecnico di Milano in October 2026, and a co-authored abstract was presented as a poster at the ESMO Congress 2026. The methods and results are not public yet.
Papers and abstracts
Older papers are published under the spelling “Alipour”.
-
2026
124P Multidimensional evaluation of AI agents: Reasoning quality and clinical effectiveness in real-world I line decision-making for NSCLC
ESMO Real World Data and Digital Oncology, 14, 100940 (abstract). Presented as a poster at the ESMO Congress 2026
Published Read -
2021
Countermovement jump variables and estimated peak power in female children with dynamic knee valgus
ISBS Proceedings Archive, Vol. 39
Published Read -
2020
Correlation between tibial external rotation and lower extremity biomechanical characteristics during gait in female juniors with dynamic knee valgus
International Conference on Motion Analysis, Tehran
Published Read -
2018
Assessing wear factor of the elbow joint in shoulder amputation by biceps performance analysis
World Congress of Biomechanics (WCB)
Accepted
Where I want to go deeper
Evaluation and safety of clinical AI
Methods that tell clinicians when an AI system's reasoning can be trusted, and when it can't.
LLM agents and reasoning in medicine
Multi-agent systems and retrieval for clinical decision support.
Medical imaging and biosignals
Deep learning on CT, MRI, fMRI and EEG, with explainability built in.