Rebecca Salles

The scarcity of primary data and challenges in analyzing secondary data hinder comprehensive mental health monitoring and the development of evidence-based policies. In Brazil, although multiple health information systems capture critical mental health data, these systems lack integration, robust analysis, and systematic presentation of mental health indicators. This study introduces the Digital Public Mental Health Dataset, a comprehensive, transparent, and reproducible data resource that consolidates data from existing Brazilian health information systems. This dataset enables detailed tracking of mental health indicators, supports future research, and informs public health interventions to address mental health disparities.

Event detection in time series is essential for numerous real-world applications, from monitoring industrial systems to identifying health anomalies. Public annotated datasets are crucial for benchmarking, training, and validating detection models. Despite recent advances in the field, there is a lack of a standardized and unified repository for evaluating different event types, which limits progress in reproducibility, comparability, and model development. This paper presents the UniTED, a Unified Event Detection Dataset for time series. UniTED consolidates annotated series from diverse domains and offers a common format and protocol for evaluation. The repository supports three event types: anomalies, change points, and motifs. UniTED fosters reusability and reproducibility, contributing to improved performance assessment and model generalization across data analysis tasks. However, existing datasets have limitations, including poor standardization, a lack of annotation guidelines, limited support for different event types, and difficulties in automating performance evaluation. UniTED presents a harmonized ETL process, label and annotation conventions, and an open-source implementation. Three use cases are presented to demonstrate the applicability of the dataset.

Time series prediction is critical in domains such as economics, industry, and agriculture. Real-world scenarios often involve challenges like low data frequency, high variability, and non-repetitive patterns. Traditional statistical models and machine learning approaches, including Long Short-Term Memory (LSTM) networks, underperform in low-data contexts due to overfitting risks, intensive training requirements, and the lack of benchmarks tailored to such scenarios. Large language models (LLMs) have emerged as tools for time series forecasting, leveraging their ability to generalize and capture temporal dependencies and patterns across datasets without requiring extensive task-specific feature engineering. This study investigates the potential of time series foundation models (TSFMs), specifically Lag-Llama and Chronos, on low-frequency datasets by comparing zero-shot prediction across one-step-ahead and multi-step-ahead approaches. Our findings evaluate predictive error, robustness, efficiency, and applicability, showing how TSFMs address these limitations and enhance forecasting in data-scarce scenarios.