Anomaly detection in time series plays a vital role in diverse domains such as healthcare, finance, and industrial monitoring, where identifying deviations from normal behavior can signal critical events. While traditional methods often focus on univariate time series and assume fixed temporal dynamics, real-world systems are typically multivariate and characterized by complex interdependencies. Ignoring these relationships can lead to suboptimal detection of system-level anomalies. This paper proposes a novel graph-based framework for multivariate time series anomaly detection that explicitly captures temporal patterns and structural relationships among variables. Individual univariate time series are first transformed into Horizontal Visibility Graphs (HVGs), which are then combined into multiplex networks to preserve inter-layer interactions. Additionally, we construct feature-based similarity graphs derived from statistical properties of the series to model inter-series relations. Anomalies are identified by comparing the neighborhood structure of each series against a historical reference set, enabling the detection of subtle and coordinated deviations. Computational experiments on real-world healthcare data illustrate the behavior and practical relevance of the proposed approach in capturing complex anomalies, offering a robust and interpretable alternative to traditional techniques.

Heart attacks are among the leading causes of death globally, and the earliest possible identification of at-risk patients is critical to lowering deaths. Advanced machine learning and deep learning algorithms have been effectively used to predict the presence of heart attack based on clinical and laboratory markers. This study used five explainable artificial intelligence techniques (XAI) to ensure that predictions made by the model are understandable and interpretable to facilitate clinical decisions. Fourteen nature-inspired feature selection algorithms were applied to identify the most informative markers while optimizing the predictive models for greater accuracy and reliability. Mutual information achieved a maximum testing accuracy of 90% and highest precision of 94%. The Whale Optimization Algorithm, Jaya Algorithm, Grey Wolf Optimizer and Sine Cosine Algorithm were the next best performing algorithms. The XAI results showed that the most important markers were ST slope, Oldpeak, exercise-induced angina, chest pain type, and fasting blood sugar. These models can be implemented in healthcare institutions to predict heart attack risks early, allowing timely interventions to reduce the likelihood of severe cardiovascular diseases. By supporting healthcare professionals with computer-aided diagnostic tools, these systems can enhance patient-specific decision-making while alleviating strain on healthcare resources. • Develop an explainable machine learning model for heart attack prediction. • Optimize predictive accuracy using nature-inspired feature selection. • Identify key clinical markers influencing heart attack risk. • Achieve 90% accuracy with mutual information for feature selection. • Support healthcare professionals with interpretable predictive analytics.

Liver disease poses a significant global health challenge requiring accurate and timely diagnosis. This research develops a novel deep learning model, named AFLID-Liver, to improve the classification of liver diseases from medical data. The AFLID-Liver model integrates three key techniques: an Attention Mechanism to focus on the most relevant data features, Long Short-Term Memory (LSTM) networks to process potential sequential information, and Focal Loss to effectively handle imbalances between different disease classes in the dataset. This combination enhances the model's ability to learn complex patterns and make robust predictions. We evaluated AFLID-Liver using a dataset of various patient records, including biomarkers and demographics. Our proposed model achieved superior performance, with 99.9 % accuracy, 99.9 % precision, and a 99.9 % F-score, significantly outperforming a baseline Gated Recurrent Unit (GRU) model (99.7 % accuracy, 97.9 % F-score) and existing state-of-the-art approaches. These results demonstrate AFLID-Liver's potential for highly accurate liver disease detection. To validate the generalizability of the proposed model, we performed cross validation using an external dataset which also yielded a good performance depicting the potential of the proposed model. The novelty lies in the synergistic integration of these techniques, offering a robust approach for clinical decision support and improved patient outcomes. Future research will aim to enhance the computational efficiency, paving the way for its adoption in real-time clinical applications.