AI-driven technologies in Digital Health & Well Being: early detection and intervention strategies Ilaria Amaro, Alessia Auriemma Citarella, Fabiola De Marco*,† , Attilio Della Greca, Luigi Di Biasi, Rita Francese, Domenico Rossi, Genoveffa Tortora and Cesare Tucci 1 University of Salerno, CAIS Lab, Department of Computer Science, Italy Abstract Artificial Intelligence (AI) is increasingly central to scientific research, especially in digital health and well-being. Technological advancements enable early detection of health issues and personalized treatments, facilitating real-time monitoring, promoting healthy lifestyles, and providing rehabilitation aids. The adoption of AI technologies requires reliability and promotes the development of symbiotic artificial intelligence systems, wherein humans and AI collaborate synergistically. This paper provides valuable insights into the study areas explored by our team. Our primary focus has centered on employing AI and explainable AI (XAI)-based methodologies to enable early detection and decision-making processes regarding treatments and rehabilitation for conditions such as skin melanoma, heart disease, and neurological disorders. It is crucial to recognize the importance of symbiotic systems and diagnostic support tools that rely on reliable technologies such as AI and XAI. The integration of these technologies not only improves the effectiveness of treatments and rehabilitation but also promotes greater transparency and understanding in the decision-making processes, underscoring their crucial role in the future of healthcare. Keywords Artificial Intelligence, eXplainable Artificial Intelligence, Symbiotic Artificial Intelligence, Mental Health, Alzheimer disease 1. Introduction has drawn attention to the critical role of Explainable AI (XAI) in ensuring transparency and trustworthiness in Artificial Intelligence (AI) is becoming increasingly the decision-making processes of these systems [1]. The prominent in scientific study, particularly in the field practical applications of AI in digital health and wellness of digital health and well-being. Indeed, the European are heterogeneous. Among these emerge: 1) Early de- community, recognizing the importance of the conver- tection and intervention: analyzing extensive data aids gence of technology and health, has adopted substantial in identifying precursor signs, enabling timely interven- investments to promote the development of cutting-edge tions; 2) Personalized treatment: AI tailors treatment technologies that leverage AI to improve the quality of plans to individual characteristics, enhancing effective- life and overall health of the population. Recent deploy- ness and patient satisfaction; 3) Real-time monitoring: ments under the coordinated AI framework make possi- continuous vital sign monitoring provides instant data ble the adoption of a new generation of technologies in for informed decision-making. 4) Promotion of healthy digital wellness, with an emphasis on health. Thanks to lifestyles: AI offers personalized advice through apps rapid advances in AI, computer systems have acquired and devices, fostering better habits. 5) Rehabilitation: AI analytical and predictive capabilities that can facilitate assists in designing tailored rehabilitation programs, and early detection of potential health problems by helping improving outcomes. to identify risky conditions while avoiding pathological Despite the significant inherent benefits, accepting degeneration. In parallel, the use of these technologies and adopting AI-driven technologies in the digital health and wellness fields may need more awareness and trust Ital-IA 2024: 4th National Conference on Artificial Intelligence, orga- nized by CINI, May 29-30, 2024, Naples, Italy among users. Thus, such technologies must be designed * Corresponding author. with deep consideration of the specific needs of the end † These authors contributed equally. users and be able to adapt dynamically to their changing $ iamaro@unisa.it (I. Amaro); aauriemmacitarella@unisa.it status and needs. The integration of AI systems with hu- (A. Auriemma Citarella); fdemarco@unisa.it (F. De Marco); man users creates a mutually beneficial relationship pro- adellagreca@unisa.it (A. Della Greca); ldibiasi@unisa.it moting the design of Symbiotic AI (SAI). In SAI, humans (L. Di Biasi); francese@unisa.it (R. Francese); dorossi@unisa.it (D. Rossi); tortora@unisa.it (G. Tortora); ctucci@unisa.it (C. Tucci) and AI work together synergistically, each leveraging the  0000-0002-6525-0217 (A. Auriemma Citarella); strengths of the other to achieve common goals. This col- 0000-0003-4285-9502 (F. De Marco); 0000-0002-4900-8666 laborative approach emphasizes trust, transparency, and (A. Della Greca); 0000-0002-9583-6681 (L. Di Biasi); effective communication between humans and AI sys- 0000-0002-6929-0056 (R. Francese); 0009-0005-6139-6920 (D. Rossi); tems. Symbiotic AI has the potential to enhance decision- 0000-0003-4765-8371 (G. Tortora) © 2024 Copyright for this paper by its authors. Use permitted under Creative Commons License making, problem-solving, and productivity across var- Attribution 4.0 International (CC BY 4.0). CEUR ceur-ws.org Workshop ISSN 1613-0073 Proceedings ious domains while maintaining control and oversight AI methods, highlighting the importance of reliable and over the decision-making process [2]. resilient models. There is a need for trustworthiness in In this context, our projects represent significant establishing confidence in AI systems, specifically in the progress in the development of innovative strategies for context of Digital Health. A transparent AI system helps early disease diagnosis and intervention methods. to build confidence among users by highlighting the sig- nificance of understanding AI decisions from the perspec- tive of SAI systems. To address this issue, we employed 2. Research Fields the GRADient-weighted Class Activation Mapping (Grad- CAM) algorithm to show how five Convolutional Neural Our primary research focus is medical image analysis, Networks (CNNs), AlexNet, GoogleNet, ShuffleNet, Mo- emphasizing the development of AI-driven and XAI al- bileNetV2, and SqueezeNet, perform in identifying the gorithms to understand their decisions. We aim to im- most relevant features in images for the final decision of prove diagnoses and outcomes across cardiovascular, skin the model [5]. Table 1 highlights that, overall, the best cancer (melanoma), Alzheimer’s (AD), and Parkinson’s performance is obtained by AlexNet, which reached an diseases (PD). Our research focuses on utilizing XAI tech- 82%, 87.4%, 75.8%, 81.5%, and 84.3% of Accuracy (ACC), niques to establish a trustability index for AI models ana- Sensitivity (SN), Specificity (SP), Precision (PRE) and F1 lyzing melanoma images and PVC signals. Through SAI SCORE (F1), respectively. In Figure 1, the behavior of systems, we aim to create reliable Computer-Aided Di- networks on the same images demonstrates how distinct agnosis (CAD) systems. Additionally, we employ neuro- networks learn vastly different features when the Grad- genetic Recurrent Neural Networks (RNNs) to identify CAM algorithm is applied. The results highlighted the PVCs and Deep Learning (DL) to classify COVID-19 in- necessity for consistency and robustness in AI models dividuals using Electrocardiogram (ECG) signals. Our and allowed us to define a trustability index, defined as research group designs and implements innovative AI (Λ), in the range (Λ ∈ [0, 1]), where a value near 1 methods, focusing on mental health. Our proposed re- indicates max trustability. We calculate the trustability search aims to enhance the diagnosis of schizophrenic index across all pairs and globally for all of the networks. syndrome [3] and rehabilitation of neurodegenerative dis- Results, validated by Λ calculation, indicate that relying orders like Alzheimer’s [4]. Our goal is to build decision on only one AI model for melanoma detection lacks reli- support tools for mental health professionals’ clinical di- ability without common shared patterns. It underscores agnoses to reduce health care expenditures and promote the importance of collaborative approaches, considering early diagnosis and intervention. Additionally, we are network interactions and synergies. creating AI-based cognitive and emotional rehabilitation solutions for AD patients to improve their quality of life and reduce cognitive impairment through individualized Table 1 systems that take into account their unique traits. Performance results of the five CNNs These activities were developed at the University of CNN ACC (%) SN (%) SP (%) PRE(%) F1 (%) Salerno, particularly at the CAIS Lab1 . Alexnet 82 87.4 75.8 81.5 84.3 GoogleNet 75 91.9 54.9 71.3 80.3 ShuffleNet 76 89.2 59.3 72.8 80.2 3. Skin Cancer MobileNetV2 SqueezeNet 76 85 82 92.8 68.1 67 75.8 75.8 78.8 78.8 Melanoma is one of the most common types of skin can- cer worldwide. In 2023, the American Cancer Society recorded 97,610 new cases, affecting 58,120 men and 39,490 women 2 . Despite representing just 1% of skin cancers, melanoma sustains a high mortality rate, under- scoring the essential need for early diagnosis enabled by CAD systems. 3.1. XAI on melanoma images One of the crucial aspects that makes building reliable Figure 1: Two different networks learn very different features CAD systems difficult is the fact that differences in neu- from Benign and malignant skin lesions (1 lesion at row). ral network structures might undermine confidence in 1 https://caislab.di.unisa.it 2 https://www.cancer.org/cancer/melanoma-skin- cancer/about/key-statistics.html Table 2 for avoiding misdiagnosis. We started from two main Final Results hypotheses: first, different types of PVCs correlate with Models ACC RE SP PRE F1 gmean AUC distinct patterns in the electrical signal; second if various ResNet-18 98,9% 95,3% 100% 100% 97,6% 97,6% 99,4% PVC types exist, it is feasible to divide an ECG dataset ResNet-50 SqueezeNet 98,5% 98,8% 96,9% 95,8% 99% 99,7% 96,7% 99,1% 96,8% 97,4% 97,9% 97,7% 99,4% 99,4% into clusters, each associated with a potential outcome. MobileNetV2 87,7% 93,5% 86,7% 55,5% 69,7% 90,1% 98,3% In this work [6], we investigated the potential of the Ex- AlexNet 86% 94% 84,6% 51,9% 66,9% 89,2% 98,1% OurCNN 75,2% 91,3% 72,4% 37% 52,6% 81,3% 97,5% tended Genetic Algorithms (EGA) approach to design RNNs. In particular, Long Short-Term Memory (LSTM) and bidirectional LSTM (BLSTM) models are employed 4. Cardiovascular Diseases to examine QRS complexes in the MIT-BIH Arrhythmia Database, which includes both non-PVC and PVC data, The prevalence of cardiovascular disease (CVD) is rapidly to understand the typical features of standard QRS sig- increasing over the world. According to the World Health nals and learn to reproduce them. Genetic algorithms Organization (WHO), CVD accounted for 17.9 million (GAs) are optimization techniques inspired by biological deaths in 2019. This data underscores the global effects of principles of the evolutionary theory. The neuroevolu- CVD and highlights the significance of preventive actions tion approach employs these evolutionary algorithms to reduce its impact on individuals and society. to produce artificial neural networks (ANN), parame- ters, and rules. They enable the discovery of optimal 4.1. SARS-CoV-2 detection through DL solutions to optimization problems by emulating natural selection, iterating through generations to improve can- The uncontrolled transmission of the COVID-19 pan- didate solutions. Consequently, these trained RNNs can demic required the adoption of preventive measures such accurately assess new ECGs with normal QRS complexes as surgical masks, hand sanitization, and social distanc- by minimizing Root mean square error (RMSE) between ing. Although molecular swabs and X-rays are recog- predictions and actual values. These architectures allow nized as the most accurate diagnostic methods, they have us to achieve a Root Mean Square Error (RMSE) of 15. Ini- disadvantages in terms of time, expense, and invasiveness. tial results suggest that PVC patterns likely reside in the Therefore, the significance of researching alternate meth- central segments of ECG signals (Figure 2). RSME values ods, such as ECG signals, increased. The objective of this in these segments support this observation, emphasizing study was to establish a correlation between ECGs and their importance for precise PVC detection. COVID-19 infection in a dataset of 1937 images divided into: patients with COVID-19 (COVID), with Myocar- dial Infarction (MI), with Previous History of Myocar- dial Infarction (HMI), Abnormal Heartbeat (ABN) and healthy (N). The study applied different Deep Learning algorithms, such as MobileNetV2, ResNet-18, ResNet-50, AlexNet, SqueezeNet, and a proprietary neural network. Training sessions progressed through comparisons of all five classes, then four (COVID-19, N, ABN, HMI), followed by three (COVID-19, N, ABN), and finally in a binary comparison (COVID-19 vs. N). The results (see Ta- Figure 2: Estimated ECG signal zones related to PVC with ble 2) demonstrated a high level of accuracy in classifica- higher RMSE variability compared to normal QRS complexes. tion, with ResNet-18 having an approximate accuracy of 98.94%, both for multiclass and binary classification. The study suggests a correlation between ECG signals and COVID-19 infection, potentially enabling the integration 4.3. CardioView: A XAI framework for of proposed neural networks into healthcare monitoring detecting PVCs systems for real-time detection. In this study, we proposed a visual framework, called CardioView, that integrates a human-in-the-loop (HITL) 4.2. Design of RNN with neuroevolution approach, ensuring continuous engagement and over- approach sight of cardiologists during the PVC detection process. In this initial phase of the work, our focus was on PVC PVCs disrupt normal heart rhythms, particularly the QRS classification, the implementation of XAI algorithms, and complex but their identification in ECG signals is crucial their clustering. Figure 3 outlines the workflow of the proposed framework, beginning with pre-processing on 2 https://www.who.int/health-topics/cardiovascular-diseases 5. Neurological Diseases Alzheimer’s disease (AD) is a condition characterized by the gradual accumulation of abnormal proteins, particu- larly amyloid- (A) and hyperphosphorylated tau, within the brain. This accumulation causes progressive impair- ment of synaptic function, neuronal health, and axonal integrity [7]. Regarding cognitive features, AD mainly af- Figure 3: CardioView workflow fects episodic memory (EM), semantic memory (SM), and spatial abilities (SA). However, the clinical presentation may vary in moderate or early cases. AD patients not the MIT-BIH Arrhythmia Database to enhance data qual- only experience memory and visuospatial disorders but ity. It involves two phases: training a CNN model on a also a range of symptoms that affect their emotional well- black-and-white (BW) dataset using classic and k-fold being. These symptoms include difficulties in problem- methods, followed by feature extraction. The RGB (red, solving, feelings of sadness, and lack of motivation, which green, and blue) dataset is then trained using the classic substantially impact the patient’s daily functioning. method, with the resulting model applied to the Grad- CAM algorithm for dataset modification and retraining. 5.1. Technology-driven rehabilitation Trained models extracted vital features, employed in k- means and Density-Based Spatial Clustering of Appli- strategies cations with Noise (DBSCAN) clustering algorithms to In recent years, due to advances in AI, research has turned identify patterns of PVC and non-PVC for early diag- increasing attention to the analysis and design of reha- nosis. Grad-CAM facilitated the visualization of ECG bilitative interventions targeting AD patients based on waveform segments, effectively distinguishing between the use of advanced technologies such as computerized PVC and non-PVC, highlighting potential multiple PVC cognitive training (CCT) and interventions that exploit classes, indicating diverse outcomes. The proposed CNN social robotics. achieved 96.21% of ACC, 98.09% of recall (RE), 94.74% of These approaches are aimed at providing adequate ther- PRE, and 99.28% of Area under the ROC Curve (AUC) apeutic support to counter the decline of cognitive func- on the GRADCAM dataset, highlighting promising re- tion in AD patients, presenting themselves as comple- sults in PVC detection. CardioView integrates cardiolo- mentary or alternative options to traditional therapies. gist surveys, crucial for the human-in-the-loop process The connection between sophisticated technological integral to reinforcement learning (RL). This approach tools and targeted rehabilitation methodologies is a grow- emphasizes the active participation of cardiologists, en- ing area of research in digital health and, in particular, in hancing the reliability and accuracy of PVC detection. caring for the elderly with neurodegenerative diseases. CardioView is structured to collect data via surveys dis- tributed to cardiologists, assisting in algorithm refine- 5.2. Mosaic of Memory: a serious game ment. The dynamic survey involving experienced car- diologists to express confidence levels in the AI model for Alzheimer’s patients outcomes (see Figure 4). Cardiologists evaluate 20 im- ages, providing feedback on XAI efficacy and suggesting improvements if needed, promoting iterative refinement in PVC detection. Figure 5: MOM experimental flow Figure 4: Step of the survey in CardioView Mosaic of Memory (MoM) is a serious game designed for AD patients inspired by the traditional "memory" card game. The game’s objective is to make correct matches of cards presented on a virtual grid, depicting the faces of the patient’s loved ones. MoM falls within the scope of CCTs and aims to slow down the deterioration of (i) spatial memory, as it is necessary to memorize the spa- tial positions of cards to proceed with the game, and (ii) autobiographical memory, due to the personalized con- tent of the game. MoM leverages a multimodal approach that combines visual and auditory stimuli to achieve its therapeutic goals and facilitate memory reinforcement. While the cards convey visual stimuli, the auditory stim- uli consist of audio recordings of the people depicted on the cards. This feature gives the game a high level of customization, reinforced by MoM’s ability to dynami- cally adapt to the user’s cognitive abilities through dif- ferent game difficulty levels (easy, medium, expert) or by selecting the automatic mode, which automatically adapts the game to patient capacity. This ensures that the game is always challenging but not frustrating for the player. When players experience high frustration levels, this negatively influences the gaming experience, leading to anger and stress. For this reason, MoM is equipped with AI, particularly an AU R-CNN model, designed to detect facial microexpressions through a multi-label clas- sification approach [8]. This improves MoM’s ability to detect and respond to player frustration during game- play. An additional distinguishing feature of the appli- Figure 6: RetroMind Framework cation is the dedicated patient profile for gaming and an administrative interface explicitly aimed at a loved one or the AD patient’s therapist. This interface allows their cognitive functioning and emotional well-being. you to customize the user experience by changing the 2) LLM customisation: Healthcare professionals per- background color and cards and modifying the game sonalise ChatGPT 3.5 Turbo based on the information content (images, names, and audio recordings of loved collected in phase one. enabling ChatGPT 3.5 Turbo to ones). The system also records detailed data from each personalise patient interaction. session, including games played and completed, average 3) Interaction phase: The robot Pepper, adapted to the frustration rate, aids used, average play times, and num- needs of the patient, supports the therapist in administer- ber of games played by difficulty. These features allow ing the AMI test. The robot captures the patient’s speech, healthcare providers to monitor the patient’s progress converts it into text and sends it to ChatGPT 3.5 Turbo. and tailor the gaming experience to her needs. Simultaneously, the content of the patient’s speech is transformed into visual representations by DALL-E3, pro- 5.3. RetroMind: a support tool for viding an image as output. reminiscence therapy. 4) Interaction and image phase: Pepper presents the patient with the images generated by DALL-E3 in the RetroMind framework is preliminary study which com- previous phase. With the support of the image, the pa- bines Large Language Models (LLMs) and social robots tient’s narration is stimulated, reinforcing the memory to enhance reminiscence therapy for Alzheimer’s disease of the events by arousing positive emotions. patients. Aims to support mental health professionals by 5) CSDD Administration: In the final phase, the men- providing visual representations of patients’ life memo- tal health professional administers the CSDD to monitor ries, facilitating personalized and empathic interactions changes in the patient’s emotional state compared to tailored for each of them. baseline levels, ensuring an ongoing assessment. The RetroMind framework procedure consists of five RetroMind is an innovative system that aims to improve steps as shown in figure 6: the effectiveness of reminiscence therapy. It acts as a 1) Traditional Therapy: Mental health profession- narrative support tool using generated images, while als administer the Autobiographical Memory Interview continuously monitoring the patient’s cognitive perfor- (AMI)[9] and the Cornell Scale for Depression in Demen- mance and emotional state. What distinguishes Retro- tia (CSDD)[10] tests to Alzheimer’s disease patients, col- Mind from other existing solutions in the literature is its lecting and transcribing patient responses to understand unique ability to recreate representations of memories References for individuals who may not have access to actual images of their past. [1] C. Trocin, P. Mikalef, Z. Papamitsiou, K. Conboy, Responsible ai for digital health: a synthesis and a research agenda, Information Systems Frontiers 25 5.4. Dynamic Visualization of Gene (2023) 2139–2157. Ontology Terms in AD and PD [2] B. Mahmud, G. Hong, B. Fong, A study of human–ai In the domain of systems biology, where intricate net- symbiosis for creative work: Recent developments works of biological molecules interact to regulate the and future directions in deep learning, ACM Trans- processes of an organism, the use of visual and in- actions on Multimedia Computing, Communica- teractive data representations is a critical aspect to tions and Applications 20 (2023) 1–21. aid in intuitively communicating complex knowledge. [3] I. Amaro, R. Francese, G. Tortora, C. Tucci, This is particularly true when navigating through vast L. D’Errico, M. Staffa, Supporting schizophrenia pa- multilevel data sets that encompass various omics sci- tients’ care with robotics and artificial intelligence, ences such as genomics, transcriptomics, proteomics, in: International Conference on Human-Computer and metabolomics. This study introduced a human- Interaction, Springer, 2023, pp. 482–495. interaction system for visualizing similarity data based [4] C. Tucci, I. Amaro, A. Della Greca, G. Tortora, Retro- on Gene Ontology (GO) functions (Cellular Component mind and the image of memories: a preliminary -CC, Molecular Function -MF, and Biological Process- BP) study of a support tool for reminiscence therapy, related to AD and PD proteins/genes [11]. Similarity in: Lecture Notes in Computer Science, Springer, data was generated using Lin and Wang distance mea- 2024. sures across all three areas of GO. The data was then [5] A. Auriemma Citarella, F. De Marco, L. Di Biasi, , clustered using the K-means algorithm, and a dynamic, G. Tortora, Trustability is all you need! symbiotic interactive view was developed using SigmaJS to allow ai strategies for melanoma diagnosis, in: ECCV, users to customize the analysis workflow interactively. 2024. To deepen our understanding of the functional relation- [6] F. De Marco, L. Di Biasi, A. Auriemma Citarella, ships between GO terms, we introduced a spatial distance G. Tortora, Improving pvc detection in ecg signals: metric, denoted as 𝑆𝐷, specifically utilized for visual- A recurrent neural network approach, in: Italian ization purposes in the rendering routines (see Figure 7). Workshop on Artificial Life and Evolutionary Com- This approach provides a more immediate visualization, putation, Springer, 2023, pp. 256–267. enabling users to capture the most relevant information [7] D. S. Knopman, H. Amieva, R. C. Petersen, G. Chéte- within the three vocabularies of GO. It facilitates an omic lat, D. M. Holtzman, B. T. Hyman, R. A. Nixon, D. T. view and enables multilevel analysis with finer details, Jones, Alzheimer disease, Nature reviews Disease compared to the traditional cluster view, enhancing un- primers 7 (2021) 33. derstanding of end-user knowledge. [8] C. Ma, L. Chen, J. Yong, Au r-cnn: Encoding expert prior knowledge into r-cnn for action unit detection, neurocomputing 355 (2019) 35–47. [9] M. D. Kopelman, B. Wilson, A. D. Baddeley, The autobiographical memory interview: a new assess- ment of autobiographical and personal semantic memory in amnesic patients, Journal of clinical and experimental neuropsychology 11 (1989) 724– 744. Figure 7: The contextual menu and an use case [10] G. S. Alexopoulos, R. C. Abrams, R. C. Young, C. A. Shamoian, Cornell scale for depression in dementia, Biological psychiatry 23 (1988) 271–284. [11] A. A. Citarella, F. De Marco, L. Di Biasi, M. Risi, G. Tortora, Gene ontology terms visualization with Acknowledgments dynamic distance-graph and similarity measures These studies were carried out within the FAIR - Fu- (s)., in: DMSVIVA, 2021, pp. 85–91. ture Artificial Intelligence Research and received funding from the European Union Next-GenerationEU (PIANO NAZIONALE DI RIPRESA E RESILIENZA (PNRR) – MIS- SIONE 4 COMPONENTE 2, INVESTIMENTO 1.3 – D.D. 1555 11/10/2022, PE00000013).