Expanded NPCBench checks whether a patient’s care answers hold together
The expanded NPCBench paper assesses artificial intelligence across linked steps in nasopharyngeal cancer care. Its CarePath cohort now covers 100 patients, combining pathology, imaging stage and care decisions. A strict patient-level score requires all selected answers to be right for the same person. The authors report a gap between individual decision accuracy and that joint measure.
Artificial Intelligence··Midday
A larger linked evaluation
NPCBench’s expanded paper appeared on arXiv on 5 November, assessing AI models along care pathways for nasopharyngeal cancer. The new version defines 4,777 component and consultation items. Its CarePath cohort covers 100 patients, with 100 pathology assessments, 100 imaging-stage assessments and 217 care decisions. Those linked tasks total 417. The expansion follows an earlier project description with fewer items and complete care episodes, making the enlarged evaluation the new development.[1]
Correct answers for one patient
Thirty models undertake the text tasks, and 21 also complete image-based and CarePath tests. The strict score counts a patient as successful only when pathology, imaging stage and every selected care answer are correct together. The authors report a highest strict score of 23 percent, alongside higher accuracy on individual decisions. Specialists in oncology, pathology and radiology review the clinical questions, evidence available at each step and reference answers. Patient material is de-identified.[1]
Limits of the patient records
Most records come from one center, and pathology is assigned in 75 care episodes. Class balance and the number of steps in an episode affect the assessment. The expanded audit of automated consultation scoring covers six cases. The authors call for cohorts from multiple centers with fully patient-matched materials and for prospective studies. NPCBench evaluates model answers against specialist pathways; it supplies no clinical authorization and does not test whether deploying these models benefits patients.[1]
