Curriculum Vitae

Teaching assistance

Total number of hours: 143 (143 for SSD IBIO-01/A)

Research activity topics

Deep learning algorithms for blood glucose forecasting in type 1 diabetes (2024-Today)

Type 1 Diabetes (T1D) is a chronic metabolic condition characterized by an insufficient production of insulin, a key hormone for blood glucose (BG) homeostasis. As a result, individuals with T1D require frequent corrective actions to keep glucose within the target physiological range (70–180 mg/dL) and to mitigate the risk of both micro- and macrovascular complications. In this context, forecasting algorithms can provide valuable support by enabling proactive decision-making and improving overall glycemic control.

Existing reviews on BG forecasting do not fully reflect recent methodological advances, particularly the rapid adoption of deep learning (DL) models, which now represent the state of the art in the field. To address this gap, we conducted an updated systematic review focused specifically on DL algorithms for BG forecasting [J1].

A major limitation identified is the lack of a standardized evaluation framework, which hinders fair comparison of algorithms across studies. To address this issue, we benchmarked forecasting models on the T1DEXI dataset, one of the largest available for T1D. Our results showed DL models outperform traditional autoregressive with exogenous input (ARX) models [A1], and that incorporating insulin, meal, and exercise information leads to improved performance over CGM-only approaches [C1].

In addition, our review highlights the limited adoption of explainable AI (XAI), despite its critical role for safe and trustworthy deployment of DL models in clinical practice. To bridge this gap, we employed explainability techniques to evaluate the physiological fidelity of DL approaches for BG forecasting, and introduced PhyNet, a novel physiology-constrained monotonic neural network designed to mitigate safety and reliability issues [A2, SJ2]. Although predictive performance was comparable across models, our analyses showed that standard DL approaches often fail to capture fundamental physiological relationships—namely, that carbohydrates increase BG and insulin lowers it. In contrast, by carefully designing its architecture and weight constraints, PhyNet achieves both predictive accuracy comparable to DL baselines and physiologically coherent behavior—and does so at reduced computational cost—offering a safer, more reliable, and more efficient foundation for T1D applications. Overall, this study reveals and addresses critical safety risks in DL-based BG forecasting, providing insights to guide future research and accelerate the clinical translation of DL algorithms.

This research began during my Master’s thesis and continues as part of my doctoral studies.

A deep learning framework for real-time detection of insulin delivery failures in type 1 diabetes (2025-Today)

Insulin pumps are increasingly used for insulin delivery in individuals with type 1 diabetes (T1D). However, these devices can experience mechanical failures that interrupt insulin delivery. Early identification of insulin pump faults (IPF) is therefore essential to enable timely intervention and reduce the risk of hyperglycemia and diabetic ketoacidosis, a potentially life-threatening complication.

To address this, we propose a population-level, dual-stage deep learning framework for online IPF detection [SJ2]. While details of the methodology cannot yet be disclosed, results obtained using the FDA-approved UVA/Padova T1D simulator are highly promising. Specifically, our approach achieves the best performance reported in the literature in terms of recall, detection delay, and false positives per day, demonstrating it can enable more accurate and timely interventions while keeping a low false-alarm burden. While these results highlight the effectiveness of supervised DL for IPF detection and underscore its potential to significantly improve the safety and reliability of insulin delivery devices, future work should validate the framework on real-world clinical datasets to assess its practical applicability.

An open-science-oriented simulator for bacterial communities’ evolution exploiting agent-based modeling (2022-2023)

Understanding the dynamics of bacterial communities is essential across several domains, from human health and microbiome research to environmental ecology and biotechnology. These communities emerge from complex interactions between bacterial species and their environment, including nutrient competition, metabolite exchange, and spatial organization. Experimental investigation of these processes is often limited by the difficulty, time, and cost associated with culturing bacterial species, making computational modeling a valuable approach for simulating community behavior and enabling rapid hypothesis testing.

To address this challenge, we developed BactLife [SJ2, C3], an open-science-oriented simulator for studying bacterial community evolution through agent-based modeling. Bacteria are modelled as autonomous agents interacting with nutrients and metabolites in discretized environments, where each species is characterized by its own metabolism, growth dynamics, and mobility. The framework supports both synthetic simulations and biologically grounded configurations of bacterial metabolisms and environmental conditions through integration with publicly available biological databases [SJ2].

BactLife was designed as a modular and extensible open-source platform that can be readily adapted to diverse biological scenarios. To improve usability and accessibility, we also developed an interactive graphical user interface [C2], enabling users without programming expertise to configure simulations and visualize results. Using BactLife, we reproduced representative microbial ecological scenarios such as competitive exclusion, unfavorable environmental fitness, and commensalistic cross-feeding interactions, without explicitly defining interaction networks [AN1, CN1].

The resulting platform provides a flexible tool for hypothesis testing, education, and microbial ecology research, with potential future applications in microbiome engineering, synthetic community design, and microbial network inference. This work was initiated during my Bachelor’s internship and thesis and was subsequently completed during my Master’s studies.

Reviewer activity

Total number: 7 for journals, 3 for conferences

Participation to orientation activities for prospective students