Research in Medicine
At TOELT AI Lab, our medical research is focused on creating AI-driven solutions that elevate diagnostic accuracy and patient care. Leveraging machine learning, we develop advanced predictive models for adverse health events, from neonatal imaging for developmental diagnostics, to challenges such as data loss in eye-tracking for neurodevelopmental research. Our models combine machine learning with expert clinical insights, providing robust tools for healthcare providers and supporting more personalized, data-informed patient outcomes.
Selected Papers
Symbrain: A large-scale dataset of MRI images for neonatal brain symmetry analysis (2024)
Gucciardi, A., Ghazouali, S. E., Venturini, F., Groznik, V., & Michelucci, U. (2024). Symbrain: A large-scale dataset of MRI images for neonatal brain symmetry analysis. arXiv preprint arXiv:2401.11814.
This paper introduces a comprehensive MRI dataset, SYMBRAIN, specifically designed to analyze brain symmetry in neonatal subjects. Brain symmetry, a key factor in diagnosing developmental anomalies, can be difficult to measure due to neonatal brain size and rapid changes. SYMBRAIN provides high-resolution, annotated images from the Developing Human Connectome Project, containing T1 and T2 MRI scans annotated with midline markers, to assist in detecting symmetry variations.
The dataset supports research in automated anomaly detection, as the midline annotations enable machine learning models to identify and assess asymmetry in brain structure. SYMBRAIN’s structure allows for deep learning applications, where models trained on this data can autonomously identify abnormal patterns associated with developmental issues in neonates.
This annotated dataset fills a gap in neonatal imaging by offering resources that support accurate, automated diagnostic tools, potentially enhancing clinical efficiency and diagnostic precision. The open access to SYMBRAIN via HuggingFace further promotes innovation, allowing researchers to apply and test new algorithms for symmetry analysis across various medical and developmental studies.
A new median filter application to deal with large windows of missing data in eye-gaze measurements (2023)
Gucciardi, A., Crotti, M., Ben Itzhak, N., Mailleux, L., Ortibus, E., Michelucci, U., and Sadikov, A. (2023, May). A new median filter application to deal with large windows of missing data in eye-gaze measurements. In CEUR Workshop Proceedings (CEUR-WS. org) (Vol. 3363, pp. 1-17). CEUR-WS. org.
This paper proposes an innovative approach to handling missing data in eye-gaze tracking, especially within datasets where tracking can be frequently interrupted. Targeting the challenges of analyzing eye-hand coordination in children with cerebral palsy, the study utilizes the Kinarm Exoskeleton and eye-tracking technology to gather detailed motion data. However, these datasets are often incomplete, with missing data "gaps" that hinder accurate interpretation of visual-motor activities.
To address this, the authors developed a custom median filter tailored to large missing data windows, which can effectively handle continuous gaps without distorting the dataset's temporal integrity. Traditional filtering techniques are less effective with large gaps, as they can introduce bias or eliminate essential details within the dataset. The newly designed filter algorithm interpolates data by excluding extreme outliers while preserving the natural movement trends within the data. This filter enables researchers to reconstruct a clearer picture of eye-gaze and hand-movement dynamics despite data interruptions.
Through empirical testing on eye-tracking data from children with unilateral cerebral palsy, the study demonstrates the filter’s effectiveness in providing a smoother, more analyzable dataset. The resulting tool, called KiPy, is open-source and supports real-time processing for applications in neurodevelopmental diagnostics. This advancement not only improves the robustness of visual-motor data analysis but also offers researchers a valuable resource for assessing coordination challenges in clinical settings.
Neural correlates of visuomotor functions in preterm children: a literature review focused on unilateral Cerebral Palsy (2023)
Crotti, Monica, Nofar Ben Itzhak, Lisa Mailleux, Umberto Michelucci, and Els Ortibus. "Neural correlates of visuomotor functions in preterm children: a literature review focused on unilateral Cerebral Palsy." In CEUR Workshop Proceedings, vol. 3363, pp. 65-87. CEUR Workshop Proceedings, 2023.
This paper reviews current research on the neural basis of visuomotor functions in preterm children diagnosed with unilateral cerebral palsy (uCP). Given the prevalence of uCP in children born preterm, the study addresses how impairments in visual and motor functions, stemming from early brain injuries, affect movement and perception. The close anatomical proximity of motor and visual pathways in these children often leads to overlapping dysfunctions, complicating movement tasks that rely on visual feedback.
The paper highlights that while motor impairments in children with uCP have been extensively studied, the specific impacts on visual perception and coordination remain less explored. Structural and diffusion MRI (dMRI) findings provide insights into brain areas commonly affected in uCP, including the optic radiations and the primary visual cortex. Damage to these regions correlates with visuomotor deficits, such as poor depth perception and compromised eye-hand coordination, underscoring the importance of visual processing for executing motor tasks.
This review emphasizes the need for further investigation into the relationship between brain damage and visual impairments in uCP, especially using advanced imaging techniques. The authors propose that more detailed mapping of neuroanatomical correlates could improve diagnostics and lead to better-targeted interventions. Future research that bridges the gap between neuroimaging and functional outcomes is essential for enhancing quality of life and independence in children with uCP.
Machine learning-based prediction of adverse events following an acute coronary syndrome (PRAISE): a modelling study of pooled datasets (2021)
D'Ascenzo, F., De Filippo, O., Gallone, G., Mittone, G., Deriu, M.A., Iannaccone, M., Ariza-Solé, A., Liebetrau, C., Manzano-Fernández, S., Quadri, G. and Kinnaird, T., …, Michelucci, U., … 2021. Machine learning-based prediction of adverse events following an acute coronary syndrome (PRAISE): a modelling study of pooled datasets. The Lancet, 397(10270), pp.199-207.
This paper introduces the PRAISE score, a machine learning-based tool designed to predict major adverse events, including mortality, recurrent myocardial infarction, and major bleeding, within a year following an acute coronary syndrome (ACS). Developed using data from two large cohorts, BleeMACS and RENAMI, the model evaluates 25 clinical variables to generate individualized risk scores, outperforming traditional linear models in prediction accuracy.
Through adaptive boosting, the PRAISE score achieved strong performance metrics, with AUC values reaching 0.92 for mortality prediction in external validation. This tool stratifies patients into low, intermediate, and high-risk categories, offering practical guidance for tailored post-ACS care. The risk stratification highlights potential candidates for more intensive follow-up or adjusted antithrombotic therapies, addressing an unmet need for precision in secondary prevention strategies.
By demonstrating superior accuracy over existing risk assessment scores, PRAISE holds promise for improving patient outcomes through personalized treatment planning. Its ability to handle complex clinical data and capture non-linear relationships makes it a valuable asset in managing ACS patient care, suggesting a shift toward more data-driven, individualized decision-making in cardiology.