User activity detection (UAD) is a key challenge in grant-free random access (GFRA) for massive MIMO systems, especially when the number of active users exceed the pilot sequence length. Classical Fisher information matrix (FIM)–based detectors degrade rapidly in such overloaded scenarios, and increasing the pilot sequence length is often impractical for massive Machine-type Communications (mMTC) and Internet of Things (IoT). This letter proposes a scalable cell-free distributed access point framework that combines local FIM-based covariance metrics with learning-based decision fusion. Metrics from local observations are aggregated and processed by gradient-boosting models (CatBoost, XGBoost) to capture nonlinear dependencies, enabling robust activity detection under dense access conditions with minimal inference overhead. Results indicate that in overcrowded regimes, where the number of active users exceeds the sequence length, increasing the sequence length yields only marginal improvements in the FIM performance. In contrast, scaling the number of access points (APs) substantially enhances detection accuracy, with error reductions exceeding 97% observed in certain scenarios. Furthermore, under a fixed antenna budget, increasing the number of distributed APs is more effective for UAD than allocating additional antennas per AP.
