Member-only story No Finding Is an Island C²A: co-occurrence aware class attention for multi-label chest X-ray classification, and why global average pooling is the wrong last step TL;DR: Standard chest X-ray classifiers end with global average pooling, which flattens the entire image into one shared summary vector and then scores every disease from it. That step destroys two things a radiologist relies on: where a finding sits on the film, and which other findings accompany it. C²A, presented at IEEE MLSP 2026, replaces that step with a classification head that pools separately for each disease under its own learned attention map, then lets the five resulting descriptors exchange information over a graph whose edges are warm-started from the empirical co-occurrence in the training labels. One bounded message-passing step is enough, and the resulting score decomposes algebraically into the evidence a disease found in its own region plus one printable term per related finding. On CheXpert the head reaches 0.895 macro-mean AUROC against 0.8874 for the strongest baseline, adds 0.01 GFLOPs and 0.23 milliseconds per image, and recovers 1.48 points on atelectasis, the one disease where the leading baseline actually regressed below CBAM. Ablations show both halves are load-bearing: removing the graph gives 0.889, and initializing it randomly instead of from…