Anomaly detection is an important task in the field of medical diagnostics, especially when it comes to ECG signals. Anomalies in ECG signals can be indicative of various cardiac conditions, such as arrhythmias, atrial fibrillation, ventricular tachycardia, and other, possibly life threatening, cardiac diseases. In medicine AI-assisted anomaly detection is favorable compared to diagnoses prediction because it can identify all out-of-the-ordinary patterns and not only those that are well represented in the available data bases. Also, AI-assisted anomaly detection can help reduce the risk of medical errors due to its ability to detect subtle abnormalities that may not be easily detected by humans. However, traditional anomaly detection methods are often limited in their ability to explain why an anomaly was detected. We propose an approach to an explainable anomaly detection and denoising of 12-lead ECG signals using a denoising autoencoder. The proposed approach is based on the idea of reconstructing the original signal from a noisy version of it, and then using the reconstruction error to detect anomalies, and also pinpoint them on the ECG in a visual way. We evaluate the proposed approach on a publicly available data sets and show that it is able to detect anomalies with high accuracy and explain why they were detected. The developed framework was also implemented as a cloud-based service that enables user-friendly ECG anomaly detection with minimal software and hardware capabilities.

Traffic forecasting is crucial for the development of an intelligent traffic prediction system that can help manage the traffic and thus reducing travel time with better utilization of existing road capacities. It is especially important that spatiotemporal dependencies can be accurately modeled together with ability to incorporate other external data that can have effects on traffic patterns. In this paper, we first introduce and make publicly available a new traffic data set obtained over the period of multiple years on different road segments in the city of Ljubljana. With available data, we build a diverse set of models that are able to effectively learn from historical traffic patterns and incorporate external data to produce accurate predictions while still respecting the manual data collection process. Constructed models are further evaluated on their robustness to different hyper-parameters. We also investigate if the models can capture different events such as holidays and how the models' performance degrade over time.