ABSTRACT. Determining whether to proceed with a clinical intervention can be a challenging endeavor due to the potentially numerous variables at play. One of the most crucial piece of information for making this decision is a precise assessment of the intervention's effectiveness, but it tends to be a complex calculation for healthcare professionals. In hemodialysis patients, the presence of a functional arteriovenous fistula (AVF) is essential to achieve a sufficient dialysis dosage and prevent various complications. Percutaneous transluminal angioplasty (PTA) is a commonly employed procedure to restore the patency of AVFs. However, it carries the disadvantage of causing long-term vessel damage, thereby reducing the lifespan of the AVF. In this preliminary study we explore Dynamic Bayesian Network (DBN) to estimate the effectiveness of the next PTA from the elaboration of routinely collected clinical data. We build a DBN to predict the risk of problems of AVF and simulate how the next PTA could impact this prediction. The outcomes of this research could potentially contribute to the development of a decision support system for vascular surgeons, aiding in the optimization of the decision-making process regarding whether to proceed with a PTA and/or consider alternative solutions.

ABSTRACT. The ability of new AI models to assist humans in performing tasks is creating new business models and transforming existing ones at breakneck speed. One of the application areas benefiting from this technology is healthcare. The work presented in this article falls within this domain. In this sense, our work focuses on how AI can be used to facilitate the work of therapists responsible for the physical rehabilitation of stroke patients. In particular, we present a decision support system integrated in a global remote rehabilitation system composed of two interconnected applications: the one used by the therapist to define routines and monitor patients and the one used by the patient to perform rehabilitation exercises autonomously. The decision support system is based on the use of fuzzy logic, which significantly increases its scalability and interpretability. The proposed system is capable of automatically suggesting personalised modifications to the rehabilitation routine assigned to a patient by the therapist, based on the patient's performance. In addition, this system integrates aspects of XAI by being able to justify why it suggests such modifications, so that the therapist has more information when validating or not validating the modifications proposed by the artificial system. The paper discusses a case study describing how a stroke patient's routine is automatically adjusted by the system.

ABSTRACT. Over the last decades, many prognostic models based on artificial intelligence techniques have been used to provide detailed prognostic predictions in healthcare. Unfortunately, the real-world observational data used to train and validate these models are almost always affected by biases that can strongly impact the outcomes validity: two examples are values missing not-at-random and selection bias. Addressing them is a key element in achieving transportability and in studying the causal relationships that are critical in clinical decision making, going beyond simpler statistical approaches based on probabilistic association. In this context, we propose a novel approach that combines selection diagrams, missingness graphs, causal discovery and prior knowledge into a single graphical model to estimate the cardiovascular risk of adolescent and young females who survived breast cancer. We learn this model from data comprising two different cohorts of patients. The resulting causal network model is validated by expert clinicians in terms of risk assessment, accuracy and explainability, and provides a prognostic model that outperforms competing machine learning methods

ABSTRACT. Survival analysis is a crucial tool in healthcare, allowing us to understand and predict time-to-event occurrences using statistical and machine-learning techniques. As deep learning gains traction in this domain, a specific challenge emerges: neural network-based survival models often produce discrete-time outputs, with the number of discretization points being much fewer than the unique time points in the dataset, leading to potentially inaccurate survival functions. To this end, our study explores post-processing techniques for survival functions. Specifically, interpolation and smoothing can act as effective regularization, enhancing performance metrics integrated over time, such as the Integrated Brier Score and the Cumulative Area-Under-the-Curve. We employed various regularization techniques on diverse real-world healthcare datasets to validate this claim. Empirical results suggest a significant performance improvement when using these post-processing techniques, underscoring their potential as a robust enhancement for neural network-based survival models. These findings suggest that integrating the strengths of neural networks with the non-discrete nature of survival tasks can yield more accurate and reliable survival predictions in clinical scenarios.

ABSTRACT. Causal inference for testing clinical hypotheses from observational data presents many difficulties because the underlying data-generating model and the associated causal graph are not usually available. Furthermore, observational data may contain missing values, which impact the recovery of the causal graph by causal discovery algorithms: a crucial issue often ignored in clinical studies. In this work, we use data from a multi-centric study on endometrial cancer to analyze the impact of different missingness mechanisms on the recovered causal graph. This is achieved by extending state-of-the-art causal discovery algorithms to exploit expert knowledge without sacrificing theoretical soundness. We validate the recovered graph with expert physicians, showing that our approach finds clinically-relevant solutions. Finally, we discuss the goodness of fit of our graph and its consistency from a clinical decision-making perspective using graphical separation to validate causal pathways.

ABSTRACT. Dementia is a set of mental diseases affecting millions of people worldwide. Similarly to all the other mental health issues, it is often difficult to forecast the trend of the disease for patients suffering from it. In this context, data of patients suffering from mental health are usually collected through questionnaires, psychological and cognitive tests, over several timepoints. This way, longitudinal data can help identify disease trajectories and allow medical doctors to forecast specific treatments. In this study, we analyze an open, unrestricted dataset of electronic health records (EHRs) of patients suffering from dementia, called OASIS-2, through several unsupervised machine learning methods (k-means, Hierarchical Clustering, Gaussian Mixture Model, and Spectral Clustering). This dataset contains demographic data and psychological test data collected over five independent visits, and having 142 patients at the first visit and ten features. Our goal is to identify patients’ clusters that stay stable over the visits, and then to characterize these clusters by studying their variables. We also measure the performances of the clustering methods through conventional metrics for internal and external validation. Our preliminary results show that unsupervised techniques can identify significant clusters of patients with mental health issues in this dataset and that Hierarchical Clustering outperforms the other algorithms to this end.

ABSTRACT. AI in healthcare: A view from the industry (Roche Diagnostics)

ABSTRACT. The scheduling of periodic treatments consists of planning a care path over a period of several weeks, in which patients have to perform different treatments respecting a certain periodicity. Treatments must be assigned to a day taking into account patients' preferences, operators' availability, the possibility of needing instruments/machinery, and of having different facilities in which operators can work and move around during the working time. This problem has been formulated to have a certain degree of flexibility and to be suitable for different contexts where periodicity is required, e.g. rehabilitation or planning a cycle of drug therapies. In this paper, we present a solution to this problem based on Answer Set Programming (ASP). We consider two different possible scenarios: scheduling one patient at a time, and scheduling blocks of patients. We have also conducted an experimental analysis showing that ASP is a suitable solution to this problem.

ABSTRACT. Predictive process monitoring is a process mining task aimed at forecasting information about a running process trace, such as the most correct next activity to be executed. In medical domains, predictive process monitoring can provide valuable decision support in atypical and nontrivial situations. In this paper, we propose a predictive process monitoring approach relying on the use of a Transformer, a deep learning architecture based on the attention mechanism, that we are testing in the domain of stroke management. A major contribution of our work lies in the incorporation of ontological domain-specific knowledge, carried out through a graph positional encoding technique. The paper also presents and discusses the first encouraging experimental result we are collecting.

ABSTRACT. The problem of finding a Master Surgical Schedule (MSS) consists of scheduling different specialties to the operating rooms of a hospital clinic. To produce a proper MSS, each specialty must be assigned to some operating rooms. The number of assignments is different for each specialty and can vary during the considered planning horizon. Realizing a satisfying schedule is of upmost importance for a hospital clinic. Recently, a compact solution based on Answer Set Programming (ASP) to the MSS problem has been introduced and tested with satisfying results, but only on synthetic data. In this paper, we also adapt the encoding and test our overall solution on real data from ASL1 Liguria in Italy. The experiments show that our ASP solution provides satisfying results also when tested on real data.

ABSTRACT. In this paper we leverage our past work to outline and describe, also using examples, a comprehensive architecture that goes from PMAs, i.e., Patient Monitoring Agents, to a novel notion of Hybrid Society (HS) instantiated in this paper to the healthcare domain. The HS is meant to encompass PMAs, and also offices of medical specialists, healthcare companies such as medical centers, and institutions such as hospitals. Our approach is based on computational logic due to its nice formal properties.

ABSTRACT. The human brain processes a vast amount of visual information daily, with complex neural mechanisms underlying the perception and interpretation of these stimuli. Recent advances in functional magnetic resonance imaging (fMRI) have allowed researchers to decode visual information from brain activity patterns in humans. We introduce a pioneering method for decoding brain activity into meaningful images and captions, with a specific emphasis on brain captioning because of increased flexibility rather than images. Our approach leverages the latest advancements in image captioning models, along with a novel image reconstruction pipeline based on latent diffusion models and depth estimation. By combining these techniques, we demonstrate significant progress in brain decoding, showcasing the enormous potential of integrating vision and language to better understand human cognition.

ABSTRACT. In recent years, thanks to improved computational power and the collection of data, AI has become a fundamental tool in basic research and industry. Despite this very rapid development, AI and deep neural networks remain black boxes that are difficult to explain. While a multitude of explainability (xAI) methods have been developed, their effectiveness and usefulness in realistic use cases is understudied. This is a major limitation in the application of these algorithms in sensitive fields such as clinical diagnosis, where the robustness, transparency and reliability of the algorithm are indispensable for its use. In addition, the majority of works have focused on feature attribution (e.g., saliency maps) techniques, neglecting other interesting families of xAI methods such as data influence models. The aim of this work is to implement, extend and test, for the first time, data influence functions in a challenging clinical problem, namely, the segmentation of tumour brains in MRI. We present a new methodology to calculate an influence score that is generalizable for all semantic segmentation tasks where the different labels are mutually exclusive.

ABSTRACT. Image registration is the process of bringing different images into a common coordinate system — a technique widely used in various applications of computer vision, such as remote sensing, image retrieval, and, most commonly, medical imaging. Deep learning based techniques have been applied successfully to tackle various complex medical image processing problems, including medical image registration. Over the years, several image registration techniques have been proposed using deep learning. Deformable image registration techniques such as Voxelmorph have been successful in capturing finer changes and providing smoother deformations. However, Voxelmorph, as well as ICNet and FIRE, do not explicitly encode global dependencies (i.e. the overall anatomical view of the supplied image) and, therefore, cannot track large deformations. In order to tackle the aforementioned problems, this paper extends the Voxelmorph approach in three different ways. To improve the performance in case of small as well as large deformations, supervision of the model at different resolutions has been integrated using a multi-scale UNet. To support the network to learn and encode the minute structural co-relations of the given image-pairs, a self-constructing graph network (SCGNet) has been used as the latent of the multi-scale UNet — which can improve the learning process of the model and help the model to generalise better. And finally, to make the deformations inverse-consistent, cycle consistency loss has been employed. On the task of registration of brain MRIs, the proposed method achieved significant improvements over ANTs and VoxelMorph, obtaining a Dice score of 0.8013 ± 0.0243 for intramodal and 0.6211 ± 0.0309 for intermodal, while VoxelMorph achieved 0.7747 ± 0.0260 and 0.6071 ± 0.0510, respectively.

ABSTRACT. Neuroblastoma is a childhood cancer that affects thousands of kids worldwide every years. Electronic health records (EHRs) of neuroblastoma patients contain valuable data for physicians and researchers because they collect both well-known clinical factors and factors whose clinical value has never been investigated. In this study, we analyzed data from EHRs of 3,034 patients with neuroblastoma from the Registro Italiano dei Tumori Neuroblastici Periferici. To perform the analysis we applied a supervised machine learning approach based on Random Forests for predicting patients’ outcome and relapse/progression, and a recursive feature elimination (RFE) approach for feature ranking. Feature ranking indicated the time to maximum response to first-line treatment in addition to the maximum response to first-line treatment as one of the most predictive factors of patient outcome, thus providing to physicians new potential treatment indications for patients affected by neuroblastoma.

Published article The complete, open-access, peer-review article on this study can be found at: Davide Chicco, et al., European Journal of Cancer, 193(113291), pages 1-11, 2023. https://doi.org/10.1016/j.ejca.2023.113291

ABSTRACT. Technologies like passive brain-computer interfaces (BCI) can enhance human-machine interaction. Anyhow, there are still shortcomings in terms of easiness of use, reliability, and generalizability that prevent passive-BCI from entering real-life situations. The current work aimed to technologically and methodologically design a new gel-free passive-BCI system for out-of-the-lab employment. The choice of the water-based electrodes and the design of a new lightweight headset met the need for easy-to-wear, comfortable, and highly acceptable technology. The proposed system showed high reliability in both laboratory and realistic settings, performing not significantly different from the gold standard based on gel electrodes. In both cases, the proposed system allowed effective discrimination (AUC > 0.9) between low and high levels of workload, vigilance, and stress even for high temporal resolution (<10 s). Finally, the generalizability of the proposed system has been tested through a cross-task calibration. The system calibrated with the data recorded during the laboratory tasks was able to discriminate the targeted human factors during the realistic task reaching AUC values higher than 0.8 at 40 s of temporal resolution in case of vigilance and workload, and 20 s of temporal resolution for the stress monitoring. These results pave the way for ecologic use of the system, where calibration data of the realistic task are difficult to obtain.

ABSTRACT. Metabolomics has emerged as a promising discipline in pharmaceuticals and preventive healthcare, holding great potential for disease detection and drug testing. However, analysing large metabolic datasets remains challenging, with available methods generally relying on limited and incompletely annotated biological pathways. In this study, we introduce a novel approach that leverages machine learning classifiers trained on molecular fingerprints of metabolites to predict their responses under specific experimental conditions. We evaluate this approach using a cellular model for the genetic disease Ataxia Telangiectasia. Our analysis reveals several challenges associated with training machine learning models on metabolite fingerprints, including data imbalance, high dimensionality, and duplicate fingerprints. After addressing these issues, our models achieve satisfactory performance, providing evidence that the structural properties of metabolites hold predictive power over their response to experimental conditions. Additionally, we perform feature importance analysis to identify chemical configurations contributing to the classification process, shedding light on impacted biological pathways. Remarkably, this analysis not only identifies groups of metabolites known to participate in affected pathways but also discovers metabolites not previously associated with the disease, opening up novel opportunities for further exploration. These findings offer new avenues for research and underscore the potential of machine learning in elucidating complex metabolic responses in genetic disorders, ultimately contributing to improved disease understanding and therapeutic development.

ABSTRACT. Computed tomography (CT) and magnetic resonance imaging (MRI) are two widely used clinical imaging modalities for non-invasive diagnosis. However, both of these modalities come with certain problems. CT uses harmful ionising radiation, and MRI suffers from slow acquisition speed. Both problems can be tackled by undersampling, such as sparse sampling. However, such undersampled data leads to lower resolution and introduces artefacts. Several techniques, including deep learning based methods, have been proposed to reconstruct such data. However, the undersampled reconstruction problem for these two modalities was always considered as two different problems and tackled separately by different research works. This paper proposes a unified solution for both sparse CT and undersampled radial MRI reconstruction, achieved by applying Fourier transform-based pre-processing on the radial MRI and then finally reconstructing both modalities using sinogram upsampling combined with filtered back-projection. The Primal–Dual network is a deep learning based method for reconstructing sparsely-sampled CT data. This paper introduces Primal–Dual UNet, which improves the Primal–Dual network in terms of accuracy and reconstruction speed. The proposed method resulted in an average SSIM of 0.932±0.021 while performing sparse CT reconstruction for fan-beam geometry with a sparsity level of 16, achieving a statistically significant improvement over the previous model, which resulted in 0.919±0.016. Furthermore, the proposed model resulted in 0.903±0.019 and 0.957±0.023 average SSIM while reconstructing undersampled brain and abdominal MRI data with an acceleration factor of 16, respectively - statistically significant improvements over the original model, which resulted in 0.867±0.025 and 0.949±0.025. Finally, this paper shows that the proposed network not only improves the overall image quality, but also improves the image quality for the regions-of-interest: liver, kidneys, and spleen; as well as generalises better than the baselines in presence the of a needle.

ABSTRACT. This study explores the use of Federated Learning (FL) for stenosis detection in coronary angiography images (CA). Two heterogeneous datasets from two institutions were considered: Dataset 1 includes 1219 images from 200 patients, which we acquired at the Ospedale Riuniti of Ancona (Italy); Dataset 2 includes 7492 sequential images from 90 patients from a previous study available in the literature. Stenosis detection was performed by using a Faster R-CNN model. In our FL framework, only the weights of the model backbone were shared among the two client institutions, using Federated Averaging (FedAvg) for weight aggregation. We assessed the performance of stenosis detection using Precision (Prec), Recall (Rec), and F1 score (F1). Our results showed that the FL framework does not substantially affects clients 2 performance, which already achieved good performance with local training; for client 1, insTead, FL framework increases the performance with respect to local model of +3.76%, +17.21% and +10.80%, respectively, reaching Prec = 73.56, Rec = 67.01 and F1 = 70.13. With such results, we showed that FL may enable multicentric studies relevant to automatic stenosis detection in CA by addressing data heterogeneity from various institutions, while preserving patient privacy.

ABSTRACT. At the end of 2019, the World Health Organization (WHO) referred that the Public Health Commission of Hubei Province, China, re- ported cases of severe and unknown pneumonia, characterized by fever, malaise, dry cough, dyspnoea and respiratory failure, which occurred in the urban area of Wuhan. A new coronavirus, SARS- CoV-2, was identified as responsible for the lung infection, now called COVID-19 (coronavirus disease 2019). Since then there has been an exponential growth of infections and at the beginning of March 2020 the WHO declared the epidemic a global emergency. An early diagnosis of those carrying the virus becomes crucial to contain the spread, morbidity and mortality of the pandemic. The definitive diagnosis is made through specific tests, among which imaging tests play an important role in the care path of the pa- tient with suspected or confirmed COVID-19. Patients with serious COVID-19 typically experience viral pneumonia. In this paper we launch the idea to use the Multiple Instance Learning paradigm to classify pneumonia X-ray images, consider- ing three different classes: radiographies of healthy people, radio- graphies of people with bacterial pneumonia and of people with viral pneumonia. The proposed algorithms, which are very fast in practice, appear promising especially if we take into account that no preprocessing technique has been used.

ABSTRACT. Lateral radiography is one of the most important records for patients’ evaluation in orthodontics and cephalometric analysis is fundamental to conduct correct diagnosis and treatment plan. This analysis includes both linear and angular measurements that quantitatively describe cranial and intermaxillary relationships. In order to obtain such measurements, anatomical landmarks are used. These reference points can be found on the soft tissue profile and on hard tissues such as teeth and skeletal contour. It is important to be extremely precise in the identification of these landmarks to compute correct measurements: even the slightest discrepancy could result in wrong values leading to different and possibly erroneous treatment plan. The automatic computerized identification of such anatomical landmarks on lateral cephalograms would greatly simplify this important step in the diagnostic process. Our aim is to apply artificial intelligence techniques for the automatic detection of these landmarks, with the final objective of developing a software, THERE (auTomatic HElpeR for cEphalometry), which exploits a predictive model that analyses teleradiographs, returns the coordinates of the anatomical landmarks, and automatically calculates the measurements necessary for diagnosis. This short paper describes the system interface and the first results obtained towards the training of the model(s) for landmarks prediction.

ABSTRACT. Artificial Intelligence (AI) and Machine Learning (ML) applied to health-care related problems hold the promise to revolutionize medicine. However, exploiting AI/ML predictions to drive clinical decisions requires the implementation of safeguard measures to reduce the risk of patient harm. Lack of AI/ML generalization ability across different populations has been widely reported and it is often caused by dataset shift between training and the populations predicted by the AI/ML during test and deployment, leading to unreliable predictions in underrepresented subpopulations and a potential lack of trust in AI/ML. ML reliability can provide a measure of the degree of trust that the prediction for a new instance is correct, thus allowing the human user to accept or reject the prediction when the reliability is high or low. Here, we propose a method to assess ML classification reliability using generative models, in particular Autoencoders. We show that this approach can identify instances for which the classification is wrong at a higher rate (unreliable set) both on a simulated scenario and on a genomics variant interpretation problem, where each variant in a patient genome is classified as pathogenic or not in the context of Rare Disease diagnosis. This approach can support clinicians to spot potential ML failures during deployment. We also make available a Python package, named RelAI, to the scientific community interested in embedding reliability measures into their ML pipelines.

ABSTRACT. Although the availability of a large amount of data is usually given for granted, there are relevant scenarios where this is not the case; for instance, in the biomedi-cal/healthcare domain, some applications require to build huge datasets of proper images, but the acquisition of such images is often hard for different reasons (e.g., accessibility, costs, pathology-related variability), thus causing limited and usually imbalanced datasets. Hence, the need for synthesizing photo-realistic images via advanced Data Augmentation techniques is crucial. In this paper we propose a hybrid inductive-deductive approach to the problem; in particular, starting from a limited set of real labeled images, the proposed framework makes use of logic programs for declaratively specifying the structure of new images, that is guaranteed to comply with both a set of constraints coming from the domain knowledge and some specific desiderata. The resulting labeled images undergo a dedicated process based on Deep Learning in charge of creating photo-realistic images that comply with the generated label.

ABSTRACT. In recent years, Artificial Intelligence has witnessed a deep transformation, primarily driven by advancements in deep learning architectures. Among these, the Transformer architecture has emerged as a pivotal milestone, revolutionizing natural language processing and several other tasks and domains. The Transformer’s ability to capture contextual dependencies across sequences, paired with its parallelizable design, made it exceptionally versatile. This plays a fundamental role in the healthcare field, where the ability to integrate and process data from various modalities, such as medical images, clinical notes and patient records, is of paramount importance in order to enable AI models to provide more informed answers. This complexity raise the demand for models that can integrate information from multiple modalities, such as text, images and audio such as multimodal transformers, which are sophisticated architectures able to process and fuse information across different modalities. Furthermore, an important goal to be achieved in the healthcare domain is to focus on pre-trained models, given the scarcity of large datasets in this field, and the need to minimise the computational resources, since healthcare organizations are not equipped with high-performance computation devices. This paper presents a methodology for harnessing pre-trained large language models based on the transformer architecture, in order to facilitate the integration of different data sources, with a specific focus on the fusion of radiological images and textual reports. The ensuing approach involves the fine-tuning of pre-existing textual models, enabling their seamless extension into diverse domains.

ABSTRACT. The importance of explanations in decision-making, particularly in the medical domain, has been widely recognized. However, the evaluation of the quality of these explanations remains a challenging task. In this work, we propose a novel approach for assessing and evaluating the reasons provided in explanations about clinical cases. Our approach leverages an external knowledge base and a defined prevalence function to score each reason based on its pertinence in the domain. By applying a deterministic prevalence function, we ensure total transparency of the reasons' assessment, facilitating a precise explanation of the rationale behind the scoring hierarchy of each reason. We demonstrate the effectiveness of our approach in clinical cases, where medical experts explain the rationale behind a specific diagnosis and why other potential diagnoses are dismissed. Our methodology provides a nuanced and detailed evaluation of the explanation, contributing to a more comprehensive understanding of the decision-making process.

ABSTRACT. Recent advances in neural network (NN) architectures have demonstrated that complex topologies possess the potential to surpass the performance of conventional feedforward networks. Nonetheless, previous studies investigating the relationship between network topology and model performance have yielded inconsistent results, complicating their applicability in contexts beyond those scrutinized. In this study, we examine the utility of directed acyclic graphs (DAGs) for modeling intricate relationships among neurons within NNs. We introduce a novel algorithm for the efficient training of DAG-based networks and assess their performance relative to multilayer perceptrons (MLPs). Through experimentation on synthetic datasets featuring varying levels of difficulty and noise, we observe that complex networks founded on pertinent graphs outperform MLPs in terms of accuracy, particularly within high-difficulty scenarios. Additionally, we explore the theoretical underpinnings of these observations and explore the potential trade-offs associated with employing complex networks. Our research offers valuable insights into the capabilities and constraints of complex NN architectures, thus contributing to the ongoing pursuit of designing more potent and efficient deep learning models.