When head and neck cancer returns after surgery, radiation and chemotherapy, oncologists still have to choose between two PD-1 blockers, nivolumab and pembrolizumab. A team at University Hospital Leipzig tested whether a Bayesian network could make that call in a form clinicians can actually read.
The model matched 95.1% of real tumor board decisions across 82 patients treated between 2018 and 2025, up from 84% for an earlier prototype. The study appears in the International Journal of Computer Assisted Radiology and Surgery.
The typo that mattered
On the first pass the network predicted pembrolizumab with only 74.6% accuracy. Instead of blaming the data, the researchers read their own probability tables line by line and found an inverted entry, a 0.2 where the clinical literature supports 0.8, which pushed recommendations away from pembrolizumab. A smaller earlier cohort never activated that branch, so nothing looked wrong until a larger, more varied group exposed it.
After the correction, the network caught 18 of 19 nivolumab cases and 60 of 63 pembrolizumab cases, with Cohen’s kappa of 0.859. Performance status, PD-L1 scores and platinum resistance were the variables that moved decisions.
That auditability is the point. A deep learning model buries a bad assumption inside millions of weights. A readable graph lets a human find one and fix it. The authors flag real limits: a single center, a modest sample, hand-specified probabilities and agreement with past decisions rather than patient outcomes.
