Diagnosis Everywhere: Lightweight AI That Detects Disease from Scans and Records on Limited Hospital Hardware

Federated learning enables collaborative medical model training without centralizing raw institutional data, yet its practical use remains constrained by non IID distributions, communication overhead, limited computing resources, and the need to support multiple diagnostic tasks across different modalities. This paper presents the proposed model, an efficiency aware federated multi task learning framework for medical imaging and clinical text applications. For imaging tasks, the model uses a frozen SqueezeNet encoder with task specific prediction heads and applies entropy adaptive structured pruning, where each client’s pruning rate is adjusted according to local label distribution heterogeneity. For clinical text, BioClinicalBERT is adapted using LoRA modules inserted into the query and value projections, followed by structured top 3 layer pruning to reduce inference and communication cost. Experiments are conducted on fundus based eye disease classification, COVID 19 radiography classification, and five unified clinical text datasets. Under non IID conditions, the proposed model achieves 90.27 ± 0.44% accuracy and 0.903 F1 for eye disease classification, and 78.51 ± 0.53% accuracy and 0.784 F1 for COVID 19 radiography. Compared with FedAvg, FedProx, SCAFFOLD, FedBN, AutoFLIP, and fixed pruning, the adaptive variant provides the strongest imaging performance while reducing communication to 3.14 MB per round and active parameters to 0.87M. In clinical text modeling, FLUTE based LoRA fine tuning achieves 91.76 ± 0.36% accuracy and 0.913 F1. With top 3 layer pruning, the model retains 90.94 ± 0.40% accuracy while reducing parameters to 86M, FLOPs to 79.4G, latency to 79 ms, communication to 0.85 MB per round, and estimated energy demand by 22.6%. These results indicate that adaptive compression and parameter efficient adaptation can improve federated medical learning under heterogeneous and resource constrained conditions.