Development of multimodal systems for monitoring paediatric brain disorders

In the last years, artificial intelligence (AI) methods are extensively applied in several fields, including healthcare, with several applications to support diagnostic approaches or treatments. The research activities carried on during my PhD work have been devoted to the development of AI methods to support neonatologists and paediatric neurologists in the detection, characterization, and monitoring of brain disorders in paediatric subjects. Specifically, the PhD work was focused on the development of multimodal systems for: neonatal and absence seizure detection; quantitative characterization of the speech phenotype for some genetic syndromes; prediction of the neurodevelopmental scales in newborns with sepsis. In the first part of this PhD work, absence seizure detectors have been developed both for online and offline applications based on Electroencephalographic (EEG) signals and sonification algorithms. Following the encouraging results obtained for absence seizures, first attempts were made to validate EEG-based Neonatal Seizure Detectors (NSDs), a still tricky and time-consuming issue in the clinical practice. Moreover, Heart rate variability (HRV) analysis was proposed as an alternative approach for the detection of neonatal seizures. Experimental results confirmed the involvement of the Autonomic Nervous System during or close to neonatal seizures. The comparison between EEG-based NSDs and HRV ones confirmed that the best approach to detect neonatal seizures is still the EEG. However, when EEG techniques are not available, the use of HRV-based NSDs could be a promising alternative. In the second part of this PhD work, quantitative acoustical analysis has been applied to the definition of the speech phenotype for four genetic syndromes: Down, Noonan, Costello and Smith-Magenis. Preliminary results confirm that acoustical measures could add helpful information for several syndromes with well-known language/voice impairments. Being completely non-invasive, acoustical analysis and AI methods might significantly contribute to the clinical assessment of such pathologies, also after surgical, pharmacological or logopaedic treatments and for long-term monitoring of the acoustical characteristics of the voice of these subjects. The last part of this PhD thesis exploits the possibility of forecasting neurodevelopmental scores in preterm newborns with and without sepsis. Using AI regression models, reliable results at different time steps of the follow-up were obtained, both with EEG and HRV features. The BAYLEY-III test was used to compute the scores in three different domains: cognitive, language and motor. Results suggest that both EEG and HRV quantitative analysis could be helpful for the clinical staff, identifying the newborns at risk of neurodevelopmental delays. Summing up, this PhD thesis shows how AI methods could be a valid support to clinicians in neurological paediatrics. Several experimental results are presented, showing possible applications and factual integration between AI techniques and clinical knowledge and needs, providing novel solutions and tools to support the clinical staff in the detection and characterization of brain diseases in infants and children.