Thin patient records learn to borrow from a map of medicine
A model decides, visit by visit, whether a patient record needs outside knowledge or can stand alone.
Picture made for ABUREAU.
Hospital records are patchy: a visit here, a gap there. A new paper asks a simple question of each visit. Does this record need help, and how much?
The method, ReTA, picks one of three moves per visit: add light features from an outside medical knowledge graph, attach a small piece of the graph, or do nothing when the model is already confident. The authors, accepted to the EMNLP 2026 main conference, tested it on two public intensive-care data sets.
In the full text, the largest gain was on 30-day readmission in the newer data set, where the area under the curve reached 73.57 against 69.01 for the best rival. The authors say the lead grows as training labels get scarcer, but the text gives no numbers for that, only charts.
Both data sets are used under a credentialed health-data licence, and the authors call the work a research prototype, not for clinical use without more validation.
For people who own health data: sparse records are not worthless. Their value depends on how well they can be linked to trusted outside knowledge, so clear coding and documented links are what make a thin record useful.
