TrustAlert

Project Description

TrustAlert establishes a comprehensive, integrated AI platform for the near real-time analysis of healthcare data streams, providing early warning, monitoring and forecasting tools to public health response agencies and local healthcare services to anticipate medical needs before they become emergencies.

The project responds to a critical gap exposed by the COVID-19 pandemic: the absence of a timely, integrated system capable of combining structured healthcare administrative data with unstructured signals from news, social media, and open data platforms to generate actionable epidemiological intelligence. Fragmented data sources, infodemics, and the inability to adapt data collection to the pace of viral spread all hampered effective policy responses — TrustAlert proposes a coordinated AI-driven solution to address these shortcomings.

The platform pursues three interconnected goals: mapping existing patterns of morbidity and mortality in communities; developing NLP and deep learning tools to extract reliable clinical information from structured and unstructured data sources; and performing micro-simulations for what-if scenario planning to support better allocation of healthcare resources in emergencies. The solution is validated against retrospective data from COVID-19 and Chronic Airway Diseases, and is designed to support clinicians, hospital managers, public health policymakers, and citizens. TrustAlert fully adheres to the Open Science paradigm and the FAIR principles (Findable, Accessible, Interoperable, Reusable), making all data and outputs available as a public good.

LINKS Foundation Activities

LINKS Foundation participates as a subcontractor supporting the research activities of Fondazione Bruno Kessler (FBK), contributing expertise in AI, machine learning, and information extraction from news data and the Social Web applied to epidemiological early warning and event detection. The core activity entails the development of an AI-based event detection tool that monitors and analyses near real-time streams of information from news reports, online media, social media, and open platforms such as the GDELT 2.0 database.

Event Detection — Disease News Classification

Articles are collected from platforms, like GDELT, and analysed by Large Language Models to firstly distinguish between health-related news and non-health-related news.

The LLM is then used to classify each health-article based on the disease it covers, using standardised codes according to the International Classification of Diseases (ICD-9-CM).

Given the skewed distribution of these codes, impliying limited or non labeled data for some diseases due to their rarity, a Zero-Shot framework was chosen to conduct the experiments.

Event Detection - Disease Time Series Anomaly Detection

The classified news are aggregated by ICD-9 code and geolocalization, to generate time series representing the evolution of the number of news per disease across time.

Deep Learning models are used to perform Anomaly Detection for epidemic surveillance: anomalies in the news time series are taken as warning signals, potentially indicating emerging public health threaths, such as disease outbreaks or unusual health events.

Event Detection - Grafana Dashboard

The results of the previous two activies are showed through a web dashboard which allows in near real-time to monitor the evolution of the public attention towards an airway disease and the rise of warning signals indicating an anomalous pattern of some sort.

Each disease reports the corresponding news time series, and warning signals are highlighted as anomalies on the series.

Funding & Dates

SDGs — Sustainable Development Goals

🏥 Goal 3 — Good Health and Well-Being

🏗️ Goal 9 — Industry, Innovation and Infrastructure

Key Technologies

🤖 Large Language Models & Deep Learning for Epidemic Detection

LLM and Deep Learning-based analysis of near real-time news streams, social media posts, and the GDELT 2.0 open database to detect, geolocalize, and classify early epidemic signals, filtering reliable health information from infodemic noise to generate actionable alerts for public health authorities.

🧠 BERT & Deep Learning for Healthcare Trajectory Analysis

Application of BERT transformer architecture and deep learning to structured Healthcare Administrative Databases, enabling the learning of individual health trajectories (diagnoses, prescriptions, outpatient visits) and prediction of future healthcare needs and adverse outcomes at population scale.

📊 Explainable AI (XAI)

Explainable Artificial Intelligence techniques applied to predictive and clustering models to visualise model complexity, support clinical trust, and translate AI outputs into interpretable decision-support tools accessible to physicians, administrators, and policymakers.

🔮 Micro-Simulation & What-If Scenario Modelling

Monte Carlo simulations, mean field analysis, and stochastic process modelling for multi-granularity epidemic scenario forecasting, enabling healthcare managers and local authorities to evaluate containment strategies and optimise operational resource allocation in real time.

☁️ Secure HPC Cloud Platform (Portable Secure Tenants)

An open-source, privacy-compliant cloud PaaS deployed on the HPC4AI infrastructure at the University of Turin, providing secure multi-tenant environments for biomedical data analysis with authentication, encrypted storage, Kubernetes orchestration, and support for AI frameworks including TensorFlow and PyTorch.

🌐 GDELT & Social Web Data Integration

Integration of the GDELT 2.0 open event database (updated every 15 minutes) with social media, internet search, and WHO data streams into a unified near real-time analytical workflow, enabling continuous monitoring of global and local health events and cross-validation with structured clinical data.

Project Partners