Assisting Clinicians with AI: Hierarchical Multi-Label Approaches to International Classification of Diseases Coding
Thesis Proposal by Riccardo Gibello and Enrico Gianluca Caiani (Advisor)
This thesis explores hierarchical multi-label approaches for automatic ICD coding using real clinical datasets.
This thesis addresses the challenge of automatic ICD coding, aiming to support clinicians by predicting diagnostic codes from clinical text through hierarchical multi-label classification. You will survey recent methods, adapt and benchmark state-of-the-art models, and evaluate them on real datasets using hierarchy-aware metrics. This thesis provides a hands-on opportunity to use Python and cutting-edge NLP tools (HuggingFace, PyTorch) on real biomedical data. You will build expertise in machine learning, model analysis, and reproducible research, while contributing to impactful AI solutions that enhance clinical practice.
Development and testing of a Deep Learning approach for estimating VO2max from 1-lead ECG
Thesis Proposal by Sarah Solbiati and Enrico Gianluca Caiani (Advisor)
This thesis will continue a previous work (https://www.nature.com/articles/s41526-025-00542-4) in which 12-leads 24h-Holter ECG was utilized to estimate longitudinal VO2max changes , both for space and terrestrial medicine.
Exploiting a retrospective database on Holter ECGs acquired during multiple head-down bed rest campaigns, and VO2max pre-post results, this thesis addresses the challenge of estimating VO2max only from short periods of 1-lead ECG, so to move this method towards higher applicability through the analysis of ECG data acquired using simple wearables. You will explore and compare performance of different Deep Learning models, as well as optimize the duration of the needed ECG signal for a realiable VO2max extimate. This thesis provides a hands-on opportunity to use Python on real biomedical data. You will build expertise in machine learning, model analysis, and reproducible research, while contributing to impactful AI solutions that enhance clinical practice.
Development and testing of a AI approach for detecting heart rate from head-ballistocardiography signals
Thesis Proposal by Sarah Solbiati, Angela Cortese, Federica Mozzini and Enrico Gianluca Caiani (Advisor)
This thesis will exploit on the results of a previous work (https://doi.org/10.1016/j.compbiomed.2025.111297) in which thoracic a DL method was developed to detect heart beats from the seismocardiographic signal (SCG) in an ECG-free approach. Trasfer learning methods will be applied to evaluate such model when applied to head-based signals, instead than thoracic.
This thesis addresses the challenge of estimating heart rate and related variability parameters from inertial signals acquired at the level of the head, through an ECG-free approach. Such signals have been acquired by different devices (VR headset, smart eyewear prototypes) in different settings, including real-life. Trasfer learning methods will be studied and applied to improve automated beat detection and extraction of useful biomarkers. This thesis provides a hands-on opportunity to use Python on real biomedical data. You will build expertise in machine learning, model analysis, and reproducible research, while contributing to impactful AI solutions that enhance clinical practice.
