Advancing Healthcare Through AI: Innovations in Monitoring and Diagnostic Technologies at the Augmented Reality for Health Monitoring Laboratory (ARHeMLab) Giovanni Annuzzi3 , Andrea Apicella1 , Pasquale Arpaia1,* , Lutgarda Bozzetto3 , Umberto Bracale6 , Egidio De Benedetto1 , Paolo De Blasiis2 , Antonio Esposito1 , Francesco Isgrò1 , Giacomo Lus4 , Nicola Moccaldi1 , Roberto Peltrini7 , Roberto Prevete1 and Simona Raimo5 1 ARHeMLab, Dipartimento di Ingegneria Elettrica e delle Tecnologie dell’Informazione, Università di Napoli Federico II, via Claudio 21, Naples, 80125, Italy 2 Università della Basilicata, Via dell’Ateneo Lucano 10, 85100 Potenza, Italy 3 Dipartimento di Medicina Clinica e Chirurgia, Università di Napoli Federico II via Pansini 5, Napoli, 80131, Italy 4 Dipartimento di Scienze Mediche e Chirurgiche Avanzate, Università della Campania Luigi Vanvitelli, p.zza L. Miraglia, 2, Napoli, 80138, Italy 5 Dipartimento di Scienze Mediche e Chirurgiche, Università Magna Graecia di Catanzaro, viale Europa, Catanzaro, 88100, Italy 6 Dipartimento di Medicina, Chirurgia e Odontoiatria, Università di Salerno, Via Giovanni Paolo II, 132, 84084 Fisciano (SA) 7 Dipartimento di Sanità Pubblica, Università di Napoli Federico II via Pansini 5, Napoli, 80131, Italy Abstract The growing sophistication of Artificial Intelligence (AI) and machine learning technologies presents exciting possibilities for advancements in healthcare diagnostics and monitoring. This paper explores our research activities at the Augmented Reality for Health Monitoring Laboratory (ARHeMLab) at the Università di Napoli Federico II. The focus is on our integration of AI, machine learning, and augmented reality technologies to improve healthcare practices. Our research encompasses a broad spectrum of areas. We are developing advanced EEG-based systems for real-time monitoring of cognitive function. Additionally, we are investigating the application of machine learning algorithms to enhance the accuracy of blood perfusion assessment during laparoscopic surgeries. Furthermore, we are exploring the potential of AI to personalise non-invasive treatments like transcranial Electrical Stimulation (tES) for neurological conditions. This paper outlines our core research areas, the methodologies we employ, and the potential impact of our work on improving healthcare practices. By presenting our current projects and initiatives, the paper illustrates ARHeMLab’s commitment to advancing medical technology. Ultimately, our goal is to enhance patient outcomes and contribute to a more responsive healthcare system. Keywords AI in Healthcare, Diagnostic Technologies, Patient Monitoring Systems, Precision Medicine 1. Introduction portunities AI presents within the healthcare domain. We recognise its complexity and are committed to con- The application of Artificial Intelligence (AI) in health- ducting thorough, ethical research to uncover real-world care is a burgeoning field with the potential to revolu- solutions. tionise clinical practices and patient outcomes [1, 2]. Our This research group has been involved in developing laboratory, the Augmented Reality for Health Monitor- AI-powered systems for non-invasive cardiovascular risk ing Laboratory (ARHeMLab) at the Università di Napoli assessment with wearable technology [3] and automated Federico II, sits at the forefront of this exciting research fracture detection in maxillofacial trauma patients [4]. landscape. These projects contribute to the ongoing exploration of ArnhemLab explores the potential applications of arti- various artificial intelligence applications in improving ficial intelligence and augmented reality within a schol- patient care and diagnostic procedures. arly setting, focusing on advancements in healthcare As the use of artificial intelligence in healthcare is knowledge and development of novel tools. Our research still developing, ArnhemLab operates in an environment is guided by a deep awareness of the challenges and op- characterised by both unknowns and potential applica- tions. Our projects, encompassing cognitive monitoring Ital-IA 2024: 4th National Conference on Artificial Intelligence, orga- with EEG-based systems and AI deployment for complex nized by CINI, May 29-30, 2024, Naples, Italy disease diagnosis, represent steps towards understand- * Corresponding author. ing how technology can be effectively integrated into $ pasquale.arpaia@unina.it (P. Arpaia) healthcare. © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings The following sections detail our current research ac- contribute to more data-driven and responsive healthcare tivities. We focus on harnessing AI’s power to push practices. boundaries in healthcare, particularly through innova- tive monitoring and diagnostic technologies. This pa- 2.2. Enhancing Medical Interventions per is structured to first introduce ARHeMLab’s core re- search areas, highlighting our significant advancements with Machine Learning in applying AI and machine learning within healthcare. Here the focus is on applying machine learning tech- We will then explore specific examples of how we in- niques to refine and improve the effectiveness of medical tegrate these technologies into practical healthcare so- treatments and procedures. Our current research projects lutions. Through this exploration, we aim to provide explore innovative ways to leverage data-driven insights a clear overview of ARHeMLab’s contribution to AI- in both surgical and non-invasive therapeutic contexts. driven healthcare advancement, offering insights into One area of focus is optimising blood perfusion qual- our methodologies, achievements, and future research ity during laparoscopic colorectal surgeries. Machine directions. learning algorithms analyse intraoperative data to pre- dict tissue blood flow adequacy, assisting surgeons in making real-time decisions that can directly impact sur- 2. Research Focus Areas of gical outcomes and patient recovery. ARHeMLab We are also conducting research to understand the ef- fects of non-invasive treatments like transcranial Electri- This section outlines our core research initiatives, each cal Stimulation (tES) on brain activity. Machine learning addressing a distinct topic critical to the broader field of helps identify patterns and correlations between treat- AI in healthcare. These topics encompass the develop- ment parameters and neurophysiological responses. This ment of EEG-based systems for cognitive function moni- research aims to tailor treatments to better suit individual toring and the application of AI to improve gait analysis patient profiles and enhance therapeutic efficacy. and rehabilitation, among others. Each subsection pro- vides a brief introduction to our contributions in these areas, laying the groundwork for a more in-depth explo- 3. AI and Machine Learning for ration of their significance, methodologies, and potential Enhanced Diagnostics and impact on transforming healthcare practices and patient outcomes. These projects represent our efforts to harness Monitoring the power of machine learning for advancing medical The increasing application of Artificial Intelligence (AI) interventions. By integrating sophisticated analytical and Machine Learning (ML) in healthcare is driving sig- techniques, we aim, for instance, to achieve higher preci- nificant advancements in diagnostic accuracy and patient sion in surgeries and customise non-invasive treatments monitoring. This research area focuses on utilising these for more personalised and effective patient care. computational technologies to gain a deeper understand- ing of complex health conditions and optimise healthcare 2.1. AI and Machine Learning for interventions. Enhanced Diagnostics and The research initiatives under this theme exemplify Monitoring a broader shift in healthcare towards data-driven, per- sonalised medicine. These projects, by connecting com- The research within this topic explores the application putational science with clinical practice, are not simply of artificial intelligence, machine learning, and electroen- theoretical exercises; they are laying the foundation for cephalography (EEG) in healthcare. The goal is to im- potential transformations in healthcare delivery. As these prove diagnostic accuracy and enable real-time patient technologies develop and integrate more seamlessly into monitoring. healthcare systems, they could usher in a new era of We focus on analysing complex health data, such as diagnostics and patient monitoring. This era might be the link between nutrition and chronic conditions, and characterised by increased precision and a potential focus monitoring cognitive states in high-pressure medical pro- on tailoring care to individual patient needs and contexts. fessions. The potential benefits include personalised Looking ahead, there is a possibility that AI and machine medicine approaches and improved performance and learning technologies could become central to improve- safety for healthcare professionals. ments in healthcare delivery and patient outcomes. We are developing systems for predicting health states and assessing cognitive load. By identifying specific biomarkers and cognitive indicators, this research aims to Figure 1: The influence of nutrition on diabetic health using explainable Deep Neural Networks (DNNs). We analyse pre-meal glucose data and other factors for 15 patients with Type 1 Diabetes Mellitus to predict post-meal glucose levels. To understand the model’s predictions, we employ Shapley value analysis. 3.1. Technical Exploration of Nutritional capture real-time brain activity across various frequency Impact on Diabetic Health bands (see Figure 2). The high-dimensional nature of the data necessitates One area of research within this field focuses on the rela- sophisticated signal processing techniques. Algorithms tionship between dietary intake and blood glucose levels perform spectral analysis, transforming EEG signals into in individuals with type 1 diabetes [5, 6]. Traditional power densities across theta, alpha, and beta frequency models for predicting glycemic response have limitations, bands. These bands are known to be associated with dif- often using linear approaches that don’t account for the ferent cognitive states, ranging from deep concentration complex interplay between various dietary components to approaching fatigue. and individual metabolic variations. A key aspect of this approach involves applying ma- Current research explores the use of advanced ma- chine learning classifiers. These classifiers are trained on chine learning algorithms, such as Random Forest and labelled datasets to distinguish between these cognitive Support Vector Machines. These algorithms are trained states with high accuracy. The resulting dynamic moni- on large datasets encompassing diverse nutritional pro- toring tool aims to alert surgeons to the onset of cogni- files, glycemic indices, and patient-specific metabolic tive fatigue, potentially improving surgical precision and responses. The algorithms identify subtle correlations reducing the risk of errors associated with diminished between these factors that may not be captured by con- cognitive capacity. ventional analysis (see Figure 1. Furthermore, the research incorporates explainable Ar- tificial Intelligence (XAI) principles to ensure the model’s 3.3. EEG Feature Selection for Enhanced outputs are interpretable. This provides both patients Cognitive Monitoring and healthcare professionals with actionable insights intoTo further refine the detail of cognitive load assessment, how different foods and meal timing affect glycemic con- some projects are exploring the use of Sequential Feature trol. This personalised dietary planning tool could rep- Selector (SFS) algorithms. These algorithms identify the resent a significant advancement in managing T1DM by most informative EEG features that reflect the cognitive offering a tailored approach that potentially mitigates the demands specific to neurosurgical tasks [8]. Unlike sim- risk of glycemic spikes and improves long-term health pler methods, SFS takes a more nuanced approach. It outcomes. iteratively evaluates the predictive power of each feature and its interaction with others, ultimately constructing 3.2. Advancements in Wearable EEG a subset of features that maximises the model’s perfor- Systems for Cognitive State mance. Monitoring This meticulous selection process, combined with ma- chine learning classifiers such as Deep Neural Networks Alongside research on nutrition, another area of focus and Gradient Boosting Machines, facilitates the devel- explores the use of wearable EEG-based systems to moni- opment of robust models for real-time cognitive load tor cognitive load and fatigue in neurosurgeons [7]. This assessment. These targeted monitoring systems offer research utilises high-resolution EEG caps designed to potential benefits not only in improving surgical out- Figure 2: The architecture of the system for EFs fatigue assessment. The system records brain activity (EEG) during a cognitive task and uses machine learning to classify the type and intensity of the mental load. comes but also for applications in other high-pressure precision and a greater focus on tailoring care to indi- professions where cognitive performance is crucial [9]. vidual patient needs and contexts. The future suggests a possibility where AI and machine learning technolo- 3.4. Exploratory Use of XAI in Cognitive gies become central to improving healthcare delivery and patient outcomes. Function Analysis Research using Explainable Artificial Intelligence (XAI) to analyse EEG features associated with critical cognitive 4. Enhancing Medical functions, such as inhibition and working memory activa- Interventions with Machine tion, represents a new area of investigation in cognitive neuroscience. By employing models that provide insights Learning into how algorithms make decisions, researchers can link This research area explores the application of artificial specific EEG patterns to cognitive processes, leading to a intelligence and machine learning techniques in medical deeper understanding of brain function. interventions. This approach aims to improve the pre- This convergence of AI and neuroscience not only cision, efficiency, and personalisation of both surgical advances our understanding of cognitive health but also and non-invasive treatments. Current research projects opens doors for developing interventions to promote within this domain utilise ML algorithms to investigate cognitive resilience, potentially improving professional new possibilities in medical treatments while also setting performance across various fields. a focus on improving patient care and safety. The research initiatives under the theme "AI and Ma- chine Learning for Enhanced Diagnostics and Monitor- ing" exemplify a broader shift in healthcare towards data- 4.1. Technical Advancements in Surgical driven, personalised medicine. These projects, by con- Perfusion Assessment necting computational science with clinical practice, are A significant portion of this research is dedicated to not simply theoretical exercises; they are laying the foun- improving outcomes in laparoscopic colorectal surgery dation for transformative healthcare solutions. through the machine learning-assisted assessment of As these technologies develop and integrate more blood perfusion quality [10]. Perfusion, the process of seamlessly into healthcare systems, they hold promise blood delivery to tissue, is a critical determinant of tissue for ushering in a new era of diagnostics and patient mon- health and recovery post-surgery. Traditional methods itoring. This era would be characterised by increased for assessing perfusion rely on visual inspection, which Figure 3: Surgical perfusion assessment. Four frames highlighting respective Regions of Interest (ROIs): Frames (a) and (d) display ROIs with adequate perfusion (high green intensity) and are predicted as 1. Frames (b) and (c) show inadequately perfused ROIs (low green intensity and/or uneven indocyanine green (ICG) distribution) and are predicted as 0. can be subjective and variable. The integration of ML invasive treatments like transcranial Electrical Stimu- offers a paradigm shift towards a more objective, data- lation (tES). tES has shown promise for various neurolog- driven approach. ical conditions by modulating brain activity. However, Technical methodologies involve the utilisation of in- the variability in patient response poses a challenge to traoperative imaging technologies, such as fluorescence its widespread adoption. angiography, combined with advanced image process- This challenge is met with the development of ML mod- ing algorithms. Machine learning models, particularly els capable of analysing electroencephalography data convolutional neural networks (CNNs), are trained on to identify biomarkers predictive of treatment success. vast datasets comprising images labelled with perfusion By employing supervised learning techniques, models outcomes. These models learn to identify features and are trained on pre- and post-treatment EEG recordings, patterns correlated with optimal and sub-optimal per- alongside clinical outcome measures. Feature selection fusion, such as tissue colour, brightness, and contrast algorithms, such as principal component analysis (PCA) changes indicative of blood flow. and mutual information, reduce dimensionality and iso- We are currently engaged in a comprehensive research late the most predictive features of treatment response, effort aimed at exploring and identifying potential meth- such as specific frequency bands or connectivity patterns ods to accurately predict and estimate the risk factors between brain regions. associated with Anastomotic Leakage following colorec- Advanced classification algorithms, including support tal surgery. Anastomotic Leakage is a significant and vector machines and gradient boosting machines, are serious postoperative complication, where the connec- then utilised to classify patients based on their likelihood tion between two sections of the intestines (anastomosis) of benefiting from tES. This personalised approach not fails to heal properly, leading to the leakage of intesti- only enhances patient outcomes but also contributes to nal contents into the abdominal cavity. This can result the understanding of the underlying mechanisms of ac- in severe infection, sepsis, and in some cases, can be tion of tES, paving the way for optimised protocols and life-threatening. broader applicability [11]. 4.2. Machine Learning in Non-Invasive 5. Conclusions Treatment Optimization ARHeMLab’s research applies machine learning to im- Parallel to surgical innovations, research efforts are also prove clinical practices, focusing on EEG-based systems concentrated on enhancing the effectiveness of non- and ML algorithms for detecting cognitive decline. This ture Medicine 25 (2019) 30–36. doi:10.1038/ enhances diagnostic accuracy and monitoring, especially s41591-018-0307-0. for medical professionals in high-stress environments. [3] P. Arpaia, R. Cuocolo, F. Donnarumma, A. Esposito, Initial studies with healthy subjects performing cognitive N. Moccaldi, A. Natalizio, R. Prevete, Conceptual tasks show promise for real-time cognitive state assess- design of a machine learning-based wearable soft ment during complex activities like surgery. sensor for non-invasive cardiovascular risk assess- Our research identifies specific EEG features linked to ment, Measurement 169 (2021) 108551. cognitive activation levels, paving the way for preven- [4] M. Amodeo, V. Abbate, P. Arpaia, R. Cuocolo, tive measures and targeted cognitive rehabilitation pro- G. Dell’Aversana Orabona, M. Murero, M. Parvis, grams for at-risk populations. Additionally, ARHeMLab R. Prevete, L. Ugga, Transfer learning for an auto- explores ML to assess blood perfusion quality during la- mated detection system of fractures in patients with paroscopic surgeries, leading to a novel decision-support maxillofacial trauma, Applied Sciences 11 (2021). system to increase surgical safety and efficiency. [5] G. Annuzzi, A. Apicella, P. Arpaia, L. Bozzetto, Looking ahead, ARHeMLab’s research will involve re- S. Criscuolo, E. De Benedetto, M. Pesola, R. Prevete, cruiting a diverse participant pool and utilizing a broader Exploring nutritional influence on blood glucose spectrum of EEG features to refine detection capabilities forecasting for type 1 diabetes using explainable AI, and broaden system applicability. We will also investigate IEEE JBHI (2023). wearable EEG systems to assess cognitive load during [6] G. Annuzzi, A. Apicella, P. Arpaia, L. Bozzetto, motor tasks, aiming for a comprehensive understanding S. Criscuolo, E. De Benedetto, M. Pesola, R. Pre- of cognitive states in dynamic environments. vete, E. Vallefuoco, Impact of nutritional factors in Future research will explore neural correlates of treat- blood glucose prediction in type 1 diabetes through ments like transcranial Electrical Stimulation (tES) for machine learning, IEEE Access 11 (2023) 17104– conditions such as Multiple Sclerosis, aiming to corre- 17115. late EEG measurements with treatment outcomes and [7] A. Apicella, P. Arpaia, P. De Blasiis, A. D. Calce, develop adaptive, personalized tES protocols. Addition- A. Fullin, L. Gargiulo, L. Maffei, F. Mancino, N. Moc- ally, we aim to improve the ML-based decision-support caldi, A. Pollastro, E. Vallefuoco, EEG-based system system for blood perfusion assessment in surgery by in- for executive function fatigue detection, in: 2022 creasing resolution and automating ROI selection. MetroXRAINE, 2022, pp. 656–660. Finally, we plan to further investigate the application of [8] P. Arpaia, R. Ayadi, G. Carone, N. Castelli, machine learning (ML) in complex medical assessments. A. Della Calce, I. Del Chicca, M. Frosolone, This expanded research will focus on more intricate and L. Gargiulo, G. Mastrati, N. Moccaldi, M. Nalin, multifaceted evaluations, including analysing how un- A. Perin, M. Picciafuoco, Toward an eeg-based derlying medical conditions might influence the results system for monitoring cognitive load in neurosur- of standard procedures and assessments. geons, in: 2023 IEEE MetroXRAINE, 2023, pp. 456– 461. [9] P. Arpaia, M. Frosolone, L. Gargiulo, N. Moccaldi, Acknowledgements M. Nalin, A. Perin, C. Puttilli, Specific feature selec- tion in wearable EEG-based transducers for moni- This work was financially supported by the Italian toring high cognitive load in neurosurgeons, under Ministry of Health, through the project HubLife Sci- review (2024). ence – Digital Health (LSH-DH) PNC-E3-2022-23683267 [10] P. Arpaia, U. Bracale, F. Corcione, E. De Benedetto, - DHEAL-COM – CUP E63C22003790001, within the “Na- A. Di Bernardo, V. Di Capua, L. Duraccio, R. Pel- tional Plan for Complementary Investments - Innovative trini, R. Prevete, Assessment of blood perfusion Health Ecosystem” - Unique Investment Code: PNC-E.3. quality in laparoscopic colorectal surgery by means of machine learning, Scientific Reports 12 (2022) References 14682. [11] P. Arpaia, L. Ammendola, M. Cropano, M. De Luca, [1] F. Jiang, Y. Jiang, H. Zhi, Y. Dong, H. Li, S. Ma, A. Della Calce, L. Gargiulo, G. Lus, L. Maffei, Y. Wang, Q. Dong, H. Shen, Y. Wang, Artificial D. Malangone, N. Moccaldi, S. Raimo, E. Signoriello, intelligence in healthcare: past, present and future, P. De Blasiis, Machine learning-based identifica- Stroke and vascular neurology 2 (2017). tion of tES-treatment neurocorrelates, under review [2] J. He, S. L. Baxter, J. Xu, J. Xu, X. Zhou, (2024). K. Zhang, The practical implementation of arti- ficial intelligence technologies in medicine, Na-