Aiming the problems that the clinical data of different patients is difficult for reasonable representation and the time interval between medical events is different, which lead to the difficulty of clinical prediction, a clinical prediction model based on the long short-term memory (LSTM) network optimized by fruit fly optimization algorithm in health time series data is proposed. First, FastText method is used to represent the interpretable vector of medical events, which can extract the concept relationship rich in medical information more effectively. Then, considering the strong dependence of clinical data on time stamp, LSTM network is used to model clinical events for better extraction of long-term and shortterm information, so as to improve the prediction performance of the model. Finally, the fruit fly optimization algorithm is used to find the optimal super parameters of LSTM network, which can improve the training efficiency and prediction precision of the network. Experimental results on MIMIC datasets show that the prediction precision, Recall@k and MAP@k of the proposed model are better than those of other models. The validity of the model is proved.
Class-incremental learning for 3D point cloud segmentation requires dense annotations for novel classes, which is prohibitively expensive and impractical for real-world applications with scarce labels. To bridge this gap, we propose a Semi-Supervised Class-Incremental Learning (SSCIL) framework that significantly reduces annotation costs. Our framework introduces three key innovations: (1) a relaxed geometry-aware feature-relation distillation module that transfers knowledge with relaxed constraints; (2) an unsupervised learning module, comprising color reconstruction and a view consistency module (CRVC), which leverages unlabeled data to learn robust and generalizable features; and (3) a multi-scale dynamic graph attention (MDG) backbone that captures both local geometric details and global contextual relationships. Extensive experiments on S3DIS and ScanNet benchmarks demonstrate that our method achieves competitive performance compared to the fully-supervised 3DPC-CISS, attaining approximately 97% of its mIoU at Cnovel = 5, S0 split while using only 10% of the labeled novel-class data. Our work establishes a practical and efficient paradigm for sustainable 3D scene understanding, with a 10× reduction in annotation effort, making incremental learning viable in label-scarce scenarios.