Testing EvoForest Beyond Structural Breaks: An Independent Evaluation on Atrial-Fibrillation Detection

EvoForest uses large-language-model-guided search to evolve computational-graph representations evaluated by a cross-validated Ridge readout. This study independently reimplements the architecture described in the public EvoForest paper and evaluates its transfer to atrial-fibrillation discrimination in short single-lead ECG recordings from PhysioNet/CinC 2017. On a sealed internal test partition, EvoForest improved ROC-AUC from 0.7131 for its unevolved seed to 0.7698 and improved average precision from 0.1867 to 0.2588. A fixed 2,048-feature random-convolution representation achieved higher discrimination, with 0.8784 ROC-AUC and 0.4966 average precision. The results show transferable signal from adaptive graph evolution while identifying substantial redundancy, proposal failures, computational burden, and limits to the present evidence. Architecture reimplementation: https://github.com/Gabriel-Kahen/evoforest-reimplementation