- Industry
- Health and life sciences
- Berkan Sesen’s role
- Researcher and lead author
The business need
Lung cancer multidisciplinary team meetings, where treatment is chosen, put substantial time pressure on their members, and survival prediction and treatment selection carry high uncertainty.
The same question now applies to language models: does giving a model access to clinical guidelines make its answers more correct, more specific to the patient and clearer?
What Berkan Sesen did
- During an Oxford DPhil, Berkan evaluated Bayesian networks for survival prediction and treatment selection, using the English Lung Cancer Database (LUCADA).
- Berkan compared a network structure elicited from clinicians with structures learned from data by causal discovery algorithms.
- Berkan built Lung Cancer Assistant, a decision support prototype for lung cancer teams that combines guideline rules and a Bayesian network in one platform.
- In a study published in BMJ Digital Health & AI in 2026, Berkan and co-authors compared a baseline language model, vector RAG over the NICE chronic kidney disease guideline, and GraphRAG on 70 simulated patients.
- The study scored answers for clinical correctness, patient specificity and clarity, using clinician assessment alongside an LLM judge.
Technology and approach
- Bayesian networks and the CAMML causal discovery algorithm
- Guideline rules combined with probabilistic decision support
- Vector retrieval-augmented generation (RAG) and GraphRAG over the NICE chronic kidney disease guideline
- An LLM judge compared against clinician assessment
Outcome
- A network structure learned by the CAMML algorithm reached an area under the ROC curve of 0.81, against 0.75 for the structure elicited from clinicians.
- Lung Cancer Assistant's guideline-based recommendations matched recorded treatments with exact and partial concordance rates of 0.57 and 0.79; its probabilistic side added survival estimates.
- In the 2026 study, guideline retrieval improved the correctness of recommendations; GraphRAG tailored answers more closely to the patient, but its longer answers were less clear.
- Agreement between the LLM judge and the clinician varied with the question, with more difficulty on complex management decisions.
- The lung cancer work received the British Thoracic Oncology Group's Best Research Prize in 2013, according to Berkan's site.
Decisions this experience speaks to
- We're being told the model performs well. How do we know that number is real and not an artefact of how it was tested?
- We're moving from research into a regulated or high-stakes environment. How do we validate it properly?
Related services and use cases
Sources
- Berkan Sesen’s profile on CTO Grand Rounds
- Bayesian networks for clinical decision support in lung cancer care (PLoS One, 2013) - Europe PMC (opens in a new tab)
- Lung Cancer Assistant (J R Soc Interface, 2014) - Europe PMC (opens in a new tab)
- RAG and GraphRAG for clinical decision support - sesen.ai (opens in a new tab)
- Berkan's publications and awards - sesen.ai (opens in a new tab)
This case describes Berkan Sesen’s own work before Grand Rounds, as Berkan Sesen’s profile and the public sources above describe it. It is not a Grand Rounds engagement and not a promise of the same result for another company.