Stefano Goria

Stefano Goria

Auto-generated profile, based on publicly available information

Stefano Goria

Thymia London UK

Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
12 December 2024 in JMIR Formative Research
Verified
Journal Article
Open access
AGAlexandra Livia Georgescu
Alexandra Livia Georgescu
NCNicholas Cummins
Nicholas Cummins
EMEmilia Molimpakis
Emilia Molimpakis
EGEduardo Giacomazzi
Eduardo Giacomazzi
JMJoana Rodrigues Marczyk
Joana Rodrigues Marczyk
SGStefano Goria
Stefano Goria
Background: Anxiety and depression represent prevalent yet frequently undetected mental health concerns within the older population. The challenge of identifying these conditions presents an opportunity for artificial intelligence (AI)–driven, remotely available, tools capable of screening and monitoring mental health. A critical criterion for such tools is their cultural adaptability to ensure effectiveness across diverse populations. Objective: This study aims to illustrate the preliminary transferability of two established AI models designed to detect high depression and anxiety symptom scores. The models were initially trained on data from a nonverbal working memory game (1- and 2-back tasks) in a dataset by thymia, a company that develops AI solutions for mental health and well-being assessments, encompassing over 6000 participants from the United Kingdom, United States, Mexico, Spain, and Indonesia. We seek to validate the models’ performance by applying it to a new dataset comprising older Brazilian adults, thereby exploring its transferability and generalizability across different demographics and cultures. Methods: A total of 69 Brazilian participants aged 51-92 years old were recruited with the help of Laços Saúde, a company specializing in nurse-led, holistic home care. Participants received a link to the thymia dashboard every Monday and Thursday for 6 months. The dashboard had a set of activities assigned to them that would take 10-15 minutes to complete, which included a 5-minute game with two levels of the n-back tasks. Two Random Forest models trained on thymia data to classify depression and anxiety based on thresholds defined by scores of the Patient Health Questionnaire (8 items) (PHQ-8) ≥10 and those of the Generalized Anxiety Disorder Assessment (7 items) (GAD-7) ≥10, respectively, were subsequently tested on the Laços Saúde patient cohort. Results: The depression classification model exhibited robust performance, achieving an area under the receiver operating characteristic curve (AUC) of 0.78, a specificity of 0.69, and a sensitivity of 0.72. The anxiety classification model showed an initial AUC of 0.63, with a specificity of 0.58 and a sensitivity of 0.64. This performance surpassed a benchmark model using only age and gender, which had AUCs of 0.47 for PHQ-8 and 0.53 for GAD-7. After recomputing the AUC scores on a cross-sectional subset of the data (the first n-back game session), we found AUCs of 0.79 for PHQ-8 and 0.76 for GAD-7. Conclusions: This study successfully demonstrates the preliminary transferability of two AI models trained on a nonverbal working memory task, one for depression and the other for anxiety classification, to a novel sample of older Brazilian adults. Future research could seek to replicate these findings in larger samples and other cultural contexts.

17 June 2024 in medRxiv
Verified
Preprint
Open access
PDF file
ANAgnes Norbury
Agnes Norbury
AGAlexandra Livia Georgescu
Alexandra Livia Georgescu
EMEmilia Molimpakis
Emilia Molimpakis
SGStefano Goria
Stefano Goria
NCNicholas Cummins
Nicholas Cummins
Background: Shift work, or working outside of normal circadian cycles, is associated with both experience of fatigue and poorer long-term physical and mental health. However, current methods for assessing fatigue present several barriers to building understanding of how these risks develop over time within individuals. Objective: Here, we explore the potential of momentary speech activity-based fatigue measurement in a large, multi-lingual cohort of shift workers, using an intensive longitudinal study design (twice-daily measurement over two weeks in N =1,197 individuals from six different countries). Methods: Paralinguistic speech features were used to predict different aspects of acute and chronic fatigue at each study time-point, with performance assessed in unseen (held-out) data. Results are reported both across the dataset as a whole, and for user-specific prediction models. Findings: In the cross-sectional analysis, good (close to or exceeding current state-of-the-art) performance was achieved for both current sleep deprivation and self-reported sleepiness levels. The within-user analysis revealed robust increases in performance, yielding the ability to detect more subjective aspects of fatigue such as pervasive physical and mental exhaustion. Conclusions: These findings offer preliminary support for the utility of using brief momentary speech samples as a low-burden, acceptable, and reliable measure of different aspects of fatigue in high-risk populations such as shift workers. Clinical Implications: Developing brief, accessible measures of different dimensions of fatigue is an important step towards building understanding of how risks for poorer health outcomes develop over time within individuals exposed to significant circadian disruption. Summary: What is already known on this topic: Previous studies have indicated that speech data may be a promising source of information about fatigue: however to date these have primarily been carried out in small, unrepresentative samples and in non-naturalistic settings, making results hard to generalize. What this study adds: Here, we present evidence from a large international study where participants both provided momentary speech activity data and reported levels of different aspects of fatigue, as they went about their usual working lives. Importantly, we use a modelling framework that explicitly takes into account potential influences of factors such as age, sex, and language on speech features and target fatigue measures, as well as assessing potential biases and reliability of model output. How this study might affect research, practice or policy: If individuals who are at heightened risk of fatigue-related health problems are able to monitor their fatigue levels regularly and in real-time, this may yield the opportunity to intervene prior to transitioning into less tractable states of poor physical and mental health.

6 February 2026 in Scientific Reports
Verified
Journal Article
Open access
PDF file
ANAgnes Norbury
Agnes Norbury
GFGeorge Fairs
George Fairs
AGAlexandra L. Georgescu
Alexandra L. Georgescu
MNMatthew M. Nour
Matthew M. Nour
EMEmilia Molimpakis
Emilia Molimpakis
SGStefano Goria
Stefano Goria
During psychiatric assessment, clinicians observe not only what patients report, but important nonverbal signs such as tone, speech rate, fluency, responsiveness, and body language. Weighing and integrating these different information sources is a challenging task and a good candidate for support by intelligence-driven tools—however this is yet to be realized in the clinic. Here, we argue that several important barriers to adoption can be addressed using Bayesian network modelling. To demonstrate this, we evaluate a model for depression and anxiety symptom prediction from voice and speech features in large-scale datasets (30,135 unique speakers). Alongside performance for conditions and symptoms (for depression, anxiety ROC-AUC = 0.842, 0.831 ECE = 0.018, 0.015; core individual symptom ROC-AUC > 0.74), we assess demographic fairness and investigate integration across and redundancy between different input modality types. Clinical usefulness metrics and acceptability to mental health service users are explored. When provided with sufficiently rich and large-scale multimodal data streams and specified to represent common mental conditions at the symptom rather than disorder level, such models are a principled approach for building robust assessment support tools: providing clinically-relevant outputs in a transparent and explainable format that is directly amenable to expert clinical supervision.

9 November 2026 in Global Change Biology
Verified
Journal Article
Open access
PDF file
DSDaniele Silvestro
Daniele Silvestro
SGStefano Goria
Stefano Goria
ERE‐Ping Rau
E‐Ping Rau
RLRenato Augusto Ferreira de Lima
Renato Augusto Ferreira de Lima
BGBen Groom
Ben Groom
AAAlexandre Antonelli
Alexandre Antonelli
The restoration of degraded ecosystems is critical for mitigating climate change and reversing biodiversity loss. Depending on the primary objective—such as maximizing carbon sequestration or protecting threatened species—and within the boundaries of budget limitations, different spatial priorities have been identified at global and regional scales. Funding mechanisms to support such work comprise public sources, philanthropy, and the private sector, including the sale of carbon and biodiversity credits. However, effectively exploring trade‐offs among restoration objectives and estimating the price of biodiversity and carbon credits to design financially viable projects remains scientifically challenging. Here we harness the power of artificial intelligence implemented in the software CAPTAIN to identify spatial priorities for ecological restoration that maximize multiple objectives at once and to allow a robust evaluation of biodiversity and climate outcomes. We find through a series of simulations that even low‐to‐moderate consideration of biodiversity in the design of restoration projects leads to the selection of restored areas that disproportionately improve the conservation of threatened species, while the carbon captured is only slightly reduced. We propose the bundling of biodiversity and carbon credits to support restoration efforts even in areas previously excluded for economic reasons. Applying our methodology to plant diversity in the Atlantic Forest of eastern South America, one of the most biodiverse and threatened ecosystems globally, we demonstrate its potential to guide restoration and to quantify trade‐offs between multiple optimization scenarios. Our study provides a robust, scalable method to maximize the joint benefits of ecological restoration for climate and nature.

17 March 2026 in medRxiv
Verified
Preprint
Open access
PDF file
EBElisa Brann
Elisa Brann
RPRoseline Polle
Roseline Polle
GČGiedrė Čepukaitytė
Giedrė Čepukaitytė
AGAlexandra L. Georgescu
Alexandra L. Georgescu
OPOwen Parsons
Owen Parsons
SGStefano Goria
Stefano Goria
Accessible screening for type 2 diabetes (T2D) is critical, with millions of cases remaining undiagnosed globally. Here, we present the largest known real-world validation study for a speech-based T2D prediction model, trained on speech data from over 21,000 individuals, that works on features extracted from 20-second speech recordings. The model was evaluated in two stages: 1) Against self-reported diagnoses in 7,319 English-speaking participants using AUC, and 2) Against HbA1c blood tests in a subset of 801 participants drawn from the full cohort. Performance was also compared against QDiabetes and in the presence of key confounding variables. The model demonstrated clinically useful predictive capacity on self-reported data (AUC = 0.80 ± 0.03), approaching QDiabetes (AUC = 0.86 ± 0.03). It was robust to most demographic confounds (e.g., age and sex) and medication use, with reduced performance in the presence of comorbidities (e.g., cardiovascular disease and hypertension). At diabetes threshold of HbA1c ≥48 mmol/mol, the model achieved an AUC of 0.75 (±0.07). This biomarker-validated speech-based tool demonstrates potential to complement existing methods through accessible, scalable screening requiring only a 20-second speech sample.

2 June 2025 in arXiv
Verified
Preprint
Open access
PDF file
RPRoseline Polle
Roseline Polle
ANAgnes Norbury
Agnes Norbury
AGAlexandra Livia Georgescu
Alexandra Livia Georgescu
NCNicholas Cummins
Nicholas Cummins
SGStefano Goria
Stefano Goria
Speaker-dependent modelling can substantially improve performance in speech-based health monitoring applications. While mixed-effect models are commonly used for such speaker adaptation, they require computationally expensive retraining for each new observation, making them impractical in a production environment. We reformulate this task as a meta-learning problem and explore three approaches of increasing complexity: ensemble-based distance models, prototypical networks, and transformer-based sequence models. Using pre-trained speech embeddings, we evaluate these methods on a large longitudinal dataset of shift workers (N=1,185, 10,286 recordings), predicting time since sleep from speech as a function of fatigue, a symptom commonly associated with ill-health. Our results demonstrate that all meta-learning approaches tested outperformed both cross-sectional and conventional mixed-effects models, with a transformer-based method achieving the strongest performance.