RUN
Repositório Institucional da Universidade NOVA de Lisboa
Entradas recentes
Development and Validation of a Food Frequency Questionnaire to Assess Polyphenol Intake and Its Association with Inflammation in the Portuguese Population: Study Plan
Publication . Hilman, Lizaveta; Santos, Cláudia Nunes; Mendonça, Nuno; NOVA Medical School|Faculdade de Ciências Médicas (NMS|FCM); Comprehensive Health Research Centre (CHRC) - pólo NMS; iNOVA4Health - pólo NMS
Multi-LLM Ensemble Framework for Quality Assurance in AI-Generated Content
Publication . Bernardino, Carlos; Aparicio, Manuela; Gonçalves, Paula; Information Management Research Center (MagIC) - NOVA Information Management School; NOVA Information Management School (NOVA IMS)
The rapid adoption of Large Language Models (LLMs) in digitized systems and services creates urgent quality assurance challenges across high-stakes sectors such as healthcare, journalism, education, and enterprise operations. A fundamental question is whether LLMs can reliably evaluate their own outputs, or whether self-bias undermines verification in digital transformation pipelines. This paper presents the Atomic Evaluation Framework (AEF), developed using Design Science Research Methodology (DSRM), for near-real-time quality assurance of AI-generated content with Human-in-the-Loop oversight and auditability. AEF performs multi-model cross-validation by querying independent LLMs in parallel for each generated item, computing agreement-based uncertainty signals to flag risky outputs, and persisting evaluation metadata for compliance-oriented audit trails. Our experiments reveal measurable self-evaluation bias, with generators systematically overrating their own content relative to external evaluators. The multi-model ensemble demonstrates strong alignment with human annotator judgments, achieving near-human accuracy at a fraction of the cost and latency of manual review. A per-profile analysis further shows that self-bias varies significantly with content type, with controversial and opinion-laden material exhibiting substantially higher overrating. This work contributes a DSRM-grounded architecture and evaluation protocol for multi-LLM cross-validation, quantifying self-bias and operationalizing Human-in-the-Loop QA at nearreal- time latency.
Application of Ultrasonography in Stratifying Malignancy Risk for Indeterminate Thyroid Nodules as per TBSRTC 2023
Publication . Guerreiro, Sofia; Mourão, Mariana; Loureiro, Isabel; Eusébio, Rosário; Canberk, Sule; Pinto Marques, Hugo; Escola Nacional de Saúde Pública (ENSP)
Introduction: Thyroid nodules are extremely common and require complex management to prevent unnecessary surgical intervention and ensure that no malignant disease is overlooked. Several diagnostic tools and scoring systems are available to evaluate the risk of malignancy (ROM). The goal is to assess variables that can aid and support the clinical recommendations suggested by the updated Bethesda System for Reporting Thyroid Cytopathology (TBSRTC-2023), such as the ultrasonographic features of thyroid nodules, particularly for the indeterminate categories III (atypia of undetermined significance) and IV (follicular neoplasm). Methods: We retrospectively analysed the correlation of the demographic and ultrasonographic characteristics of thyroid nodules with the cytopathological and histopathological diagnoses of TBSRTC categories III (atypia of undetermined significance), IV (follicular neoplasm), V (suspicious for malignancy), and VI (malignant) in patients who underwent surgery in a single Portuguese centre over a 10-year period. Results: In total, 360 nodules were evaluated in 341 patients, and 57% were histopathologically malignant or borderline. The majority were included in the TBSRTC indeterminate categories III and IV, with ROMs of 44% and 43%, respectively. The ultrasonographic characteristics associated with a higher TBSRTC category and a greater ROM value were hypoechogenicity, the presence of microcalcifications, irregular margins, and the presence of cervical adenopathy. When correlating with a malignant histology, only adenopathy and the presence of microcalcifications were observed to be statistically significant. Discussion: The indeterminate categories of the TBSRTC have been the most challenging ones to manage. The new TBSRTC (2023) guidelines, as well as the ultrasonographic characteristics of a patient’s nodule, can be helpful in assessing the ROM and deciding on an appropriate course of treatment. Other resources, such as molecular tests, are also playing a more important role in the clinical decision process and may become crucial in the future. Conclusions: The worrisome ultrasound features that this study found to statistically correlate with a malignant histology were the presence of microcalcifications and adenopathy. The clinical management of thyroid nodules requires a careful analysis of clinical history and an evaluation of demographic details, personal and family history, ultrasonographic features, and the results of cytopathology, thyroid function, and molecular/genetic tests.
Psoriatic arthritis digital phenotyping and inflammation drivers (PDPID) study
Publication . Hojeij, Batoul; Tchetverikov, Ilja; Vasileiou, Eleni; Apostolidis, George; Dimitroulas, Theodoros; Mytilinaiou, Maria; Gonçalves, Cátia; Rodrigues, Ana M; Coates, Laura; Konstantinidis, Dimitrios; Dimitropoulos, Kosmas; Melanitis, Nikos; Foolen, Jasper; Wagenaar, Wendy; Charisis, Vasileios; Hadjileontiadis, Leontios J; Luime, Jolanda J; NOVA Medical School|Faculdade de Ciências Médicas (NMS|FCM); Comprehensive Health Research Centre (CHRC) - pólo NMS; BMJ Publishing Group
INTRODUCTION: Changes in psoriatic arthritis (PsA) disease activity, particularly flares, are unpredictable and can impact patients' quality of life. Digital biomarkers derived from data collected via smart devices provide an opportunity for unobtrusive continuous monitoring of symptoms reflecting disease activity. Besides, our understanding of the mechanisms and triggers behind flares is limited. The primary objectives of the Psoriatic Arthritis Digital Phenotyping and Inflammation Drivers (PDPID) study are: (1) to develop digital biomarker capable of detecting PsA flare using data collected from patient's smartphone and smartwatch, and (2) to develop machine learning models for PsA flare prediction using clinical, biological, environmental and digital data. METHODS AND ANALYSIS: The PDPID study is a 12-month multi-centre prospective cohort study conducted across four countries. Study visits are scheduled at baseline (T0), and 3, 6, 9 and 12 months follow-up visits, with additional patient-initiated visits due to flare. At inclusion, patients have a study app installed on their smartphone and receive a smartwatch. Flare information is collected using physician- and patient-reported flare questionnaires administered at each visit, and via the flare button in the app that patients can activate. The study collects clinical data (physical measurements, questionnaires), biological markers (salivary DNA, gut microbiome, hair cortisol, C-reactive protein) and environmental exposure data (weather and air pollution). The app includes in-app questionnaires, active photo and video tests and passively captures digital data from the smartphone and smartwatch sensors. ETHICS AND DISSEMINATION: Ethical approval has been granted from each of the participating countries (Erasmus MC, the Netherlands: MEC-2023-0470; HRA and Health and Care Research Wales, UK: 332916 NOVA Medical School|Faculdade de Ciências Médicas (NMS|FCM) Ethical Committee (CEFCM), Portugal (124/2023/CEFCM); MREC Hipokrateion Hospital Thessaloniki, Greece: 5549/31.01.24). Findings of this study will be disseminated through reports to the funding body, the project website, newsletters, social media, national and international conferences and symposiums, and scientific publications. TRIAL REGISTRATION NUMBER: NCT06347237.
How Frontline Employees Navigate Complaints
Publication . Gottschalk, Sabrina; Rohden , Simoni F.; Wiertz, Caroline; Information Management Research Center (MagIC) - NOVA Information Management School; NOVA Information Management School (NOVA IMS)
While the literature offers many recommendations for optimal decision-making in service recovery, less is known about the actual decision strategies FLEs employ. We draw on a multi-method qualitative dataset including 40 semi-structured in-depth interviews, a diary study with 20 participants, and a survey, and identify five distinct decision heuristics used by FLEs in recovery situations, depending on their specific recovery context. We find that recovery contexts shape distinct cost considerations for FLEs, which, in turn, determine the recovery heuristics they resort to. Our study contributes to the literature on the role of FLEs in recovery by shedding light on FLE decision-making processes and by examining the role of costs in recovery contexts resulting from organizational parameters. Our findings show that the heuristics can play out differently for staff, managers, and customers, with each group experiencing different benefits and challenges.