XAI for Medicine September - December 2021

                                  

Introduction 

In recent years, there has been a significant increase in the use of machine learning models and artificial intelligence (AI) in the field of medicine. However, as these models become more complex and are integrated into critical decision-making processes, there is a growing need for explainable AI (XAI). In this blog post, we will explore the importance of XAI in medicine and how it can help ensure the safety, reliability, and ethical use of AI in healthcare.

1. Deep learning for epileptogenic zone delineation from the invasive EEG: challenges and lookouts

The scientific commentary discusses the challenges of using deep learning (DL) approaches for delineation of the epileptogenic zone (EZ) from invasive EEG* (iEEG). The article refers to a recent study that used DL to classify dichotomously whether a high-frequency oscillation (HFO) was located on a spike, as spike HFOs are considered more epileptogenic than non-spike HFOs. The authors discuss the challenges faced in using DL approaches for EZ delineation and the types of learning in ML. They also highlight that the generalizability of DL approaches for EZ delineation can be questioned as these studies often use small datasets. The article emphasizes that patient-wise cross-validation is the most adequate assessment of a DL model's generalizability for clinical use in different patients. The article also notes that the automatic detection of HFOs in the iEEG can address the tedious work of manually marking these biomarkers, but their specificity as an EZ biomarker may be limited as HFOs can be generated by healthy tissue. Thus, the authors suggest refining epileptic HFOs using DL to address this issue.

*EEG – Electroencephalography

2. The messiness of the menstruator: assessing personas and functionalities of menstrual tracking apps

The article discusses the gaps in menstrual tracking apps by analyzing trends in the intended users and functionalities advertised by these technologies. The study found misalignments between descriptions of menstrual trackers and themes developed from the literature, with narrow characterizations of menstruators and designs for limited needs. The study synthesized gaps in the design of menstrual tracking apps and discussed implications for designing around irregular menstrual cycles as the norm, the embodied, leaky experience of menstruation, and the varied biologies, identities, and goals of menstruators. The article also highlights the need for human-centered AI in FemTech to address technical limitations, design choices, and policy concerns related to menstrual trackers. Finally, the article emphasizes the potential for human-centered AI to transform menstrual tracking apps from body management tools into a nurturing and empowering experience.

3. Leveraging the potential of machine learning for assessing vascular ageing: state-of-the-art and future research 

The article discusses the use of machine learning (ML) in assessing vascular age, which is an important predictor of cardiovascular disease (CVD). It describes the relationship between arterial stiffness and vascular age and the use of biomarkers such as pulse wave velocity (PWV) and central blood pressure (CBP) in assessing vascular age. Two types of ML models are outlined for vascular age assessment: parameter estimation models and risk classification models. The authors provide examples of clinical applications of these models and describe the capabilities of different supervised ML techniques for developing these models. The main highlights are the opportunities presented by ML for enhancing vascular age assessment in research and clinical practice and calls for further research in this field. To develop accurate and clinically relevant ML models for vascular age assessment, suitable datasets are required, and the article provides guidance on methodology, data quality, and reporting. Additionally, benchmark datasets that reflect the target population and arterial stiffness markers should be established. The article presents case studies on the use of ML in assessing vascular age. There are three potential research directions for assessing vascular age: using electronic health record data, the pulse wave as a source of information, and incorporating measures of vascular age into consumer devices. These methods can help identify patients with known risk factors and inform clinical decision-making, but challenges remain in contextualizing measurements and ensuring data quality. 

4. Deep learning-based brain transcriptomic signatures associated with the neuropathological and clinical severity of Alzheimer’s disease

The article discusses the application of deep learning methodologies in identifying pseudo-temporal trajectories in transcriptomic space for Alzheimer's disease (AD) progression. Despite numerous scientific advances, no disease-modifying treatments are currently available for AD, and the complicated molecular aetiology driving the disease calls for a broad search for effective therapeutics beyond the conventional amyloid cascade hypothesis. Recent efforts have begun to model AD progression as a continuous trajectory using cross-sectional transcriptomic data, leveraging the methods developed in single-cell genomics and machine learning. The authors leverage multidimensional, well-characterized and high-quality genomic, neuropathological and clinical data from the Accelerating Medicines Project for Alzheimer’s Disease (AMP-AD) programme and apply deep learning framework to identify pseudo-temporal trajectories in transcriptomic space and the underlying gene signatures for AD progression. They observe that the deep learning model trained on the ROSMAP cohort when applied to two independent AMP-AD data sets, the MAYO RNA-seq study cohort and The Mount Sinai Brain Bank (MSBB) study cohort, results in similar trajectories and sample distribution following a generalized pattern, and the estimated severity index values remain strongly correlated with pathological biomarkers and clinical function scores.

5. Enhancing the impact of Artificial Intelligence in Medicine: A joint AIFM-INFN Italian initiative for a dedicated cloud-based computing infrastructure 

Artificial Intelligence (AI) techniques have been implemented in the field of Medical Imaging for more than forty years. Medical Physicists, Clinicians and Computer Scientists have been collaborating since the beginning to realize software solutions to enhance the informative content of medical images, including AI-based support systems for image interpretation. Despite the recent massive progress in this field due to the current emphasis on Radiomics, Machine Learning and Deep Learning, there are still some barriers to overcome before these tools are fully integrated into the clinical workflows to finally enable a precision medicine approach to patients’ care. Nowadays, as Medical Imaging has entered the Big Data era, innovative solutions to efficiently deal with huge amounts of data and to exploit large and distributed computing resources are urgently needed. In the framework of a collaboration agreement between the Italian Association of Medical Physicists (AIFM) and the National Institute for Nuclear Physics (INFN), we propose a model of an intensive computing infrastructure, especially suited for training AI models, equipped with secure storage systems, compliant with data protection regulation, which will accelerate the development and extensive validation of AI-based solutions in the Medical Imaging field of research. This solution can be developed and made operational by Physicists and Computer Scientists working on complementary fields of research in Physics, such as High Energy Physics and Medical Physics, who have all the necessary skills to tailor the AI-technology to the needs of the Medical Imaging community and to shorten the pathway towards the clinical applicability of AI-based decision support systems. 

6. An accurate fuzzy rule-based classification systems for heart disease diagnosis 

The article discusses the need for physicians and healthcare providers to better understand clinical decision-making processes and methods, with a focus on the diagnosis of heart disease, a leading cause of death. A novel data-driven method for heart disease diagnosis using fuzzy clustering and linguistic modifiers is proposed to design a fuzzy rule-based classification system.  Some highlights are the importance of balancing interpretability and precision in machine learning models and the use of Explainable Artificial Intelligence (XAI) techniques to explain machine learning models. The article concludes that the proposed model is superior in terms of balancing interpretability and precision, and that XAI techniques, such as rule-based models, can help in understanding the inference process of the system and building confidence in its decision-making. 

7. Applied machine learning in cancer research: A systematic review for patient diagnosis, classification and prognosis 

Artificial Intelligence (AI) has recently altered the landscape of cancer research and medical oncology using traditional Machine Learning (ML) algorithms and cutting-edge Deep Learning (DL) architectures. In this review article the focus is on the ML aspect of AI applications in cancer research and present the most indicative studies with respect to the ML algorithms and data used. The PubMed and dblp databases were considered to obtain the most relevant research works of the last five years. Based on a comparison of the proposed studies and their research clinical outcomes concerning the medical ML application in cancer research, three main clinical scenarios were identified. The authors gave an overview of the well-known DL and Reinforcement Learning (RL) methodologies, as well as their application in clinical practice, and they briefly discuss Systems Biology in cancer research. They also provide a thorough examination of the clinical scenarios with respect to disease diagnosis, patient classification and cancer prognosis and survival. The most relevant studies identified in the preceding year are presented along with their primary findings. Furthermore, the effective implementation and the main points that need to be addressed in the direction of robustness, explainability and transparency of predictive models are examined. Finally, there is a summary of the most recent advances in the field of AI/ML applications in cancer research and medical oncology, as well as some of the challenges and open issues that need to be addressed before data-driven models can be implemented in healthcare systems to assist physicians in their daily practice. 

8. XAI Feature Detector for Ultrasound Feature Matching 

Feature matching is a crucial component of computer vision that has various applications. With the emergence of Computer-Aided Diagnosis (CAD), the need for feature matching has also emerged in the medical imaging field. In this paper, a novel algorithm using the Explainable Artificial Intelligence (XAI) approach is proposed in order to achieve feature detection for ultrasound images based on the Deep Unfolding Super-resolution Network (USRNET). Based on the experimental results, the presented method shows higher interpretability and robustness than existing traditional feature extraction and matching algorithms. The proposed method provides a new insight for medical image processing, and may achieve better performance in the future with advancements of deep neural networks.

9. Detection of spontaneous seizures in EEGs in multiple experimental mouse models of epilepsy 

Electroencephalography (EEG) is a key tool for non-invasive recording of brain activity and the diagnosis of epilepsy. EEG monitoring is also widely employed in rodent models to track epilepsy development and evaluate experimental therapies and interventions. Whereas automated seizure detection algorithms have been developed for clinical EEG, preclinical versions face challenges of inter-model differences and lack of EEG standardization, leaving researchers relying on time-consuming visual annotation of signals. Approach. In this study, a machine learning-based seizure detection approach, ‘Epi-AI’, which can semi-automate EEG analysis in multiple mouse models of epilepsy was developed. Twenty-six mice with a total EEG recording duration of 6451 h were used to develop and test the Epi-AI approach. EEG recordings were obtained from two mouse models of kainic acid-induced epilepsy (Models I and III), a genetic model of Dravet syndrome (Model II) and a pilocarpine mouse model of epilepsy (Model IV). The Epi-AI algorithm was compared against two threshold-based approaches for seizure detection, a local Teager-Kaiser energy operator (TKEO) approach and a global Teager-Kaiser energy operator-discrete wavelet transform (TKEO-DWT) combination approach. Epi-AI demonstrated a superior sensitivity, 91.4%–98.8%, and specificity, 93.1%–98.8%, in Models I–III, to both of the threshold-based approaches which performed well on individual mouse models but did not generalise well across models. The performance of the TKEO approach in Models I–III ranged from 66.9%–91.3% sensitivity and 60.8%–97.5% specificity to detect spontaneous seizures when compared with expert annotations. The sensitivity and specificity of the TKEO-DWT approach were marginally better than the TKEO approach in Models I–III at 73.2%–80.1% and 75.8%–98.1%, respectively. When tested on EEG from Model IV which was not used in developing the Epi-AI approach, Epi-AI was able to identify seizures with 76.3% sensitivity and 98.1% specificity. Epi-AI has the potential to provide fast, objective and reproducible semi-automated analysis of multiple types of seizure in long-duration EEG recordings in rodents.  

              

10. Efficient and Explainable Deep Neural Networks for Airway Symptom Detection in Support of Wearable Health Technology

This study discusses the use of deep neural networks (DNNs) optimized through evolutionary algorithms to classify airway-related symptoms based on mechano-acoustic data signals acquired from laboratory-generated and publicly available datasets. The aim is to create an efficient and explainable DNN that can be used for remote, long-term monitoring of airway symptoms using mechano-acoustic sensing wearables. The article found that the optimized DNNs had a low footprint and predicted airway symptoms with 83.7% accuracy on unseen data. An evolutionary algorithm is utilized to find a small and efficient CNN architecture for multi-class classification of airway symptoms using non-speech audio signals. The fitness for each architecture was calculated based on validation accuracy and inference time. The conclusion is that their work represents the first step in developing effective and explainable AI algorithms for long-term remote monitoring of airway symptoms by mechano-acoustic wearables. 

[1] https://tinyurl.com/ycs3u7c8

[2] https://tinyurl.com/7n6cn79f

[3] https://tinyurl.com/2p99wj95

[4] https://tinyurl.com/2mhdk3h4

[5] https://reader.elsevier.com/reader/sd/pii/S1120179721003203?token=BEA95144A8836E1B990ACDC99C18371B3BDE1222D1413D183A2644007951931FEFD61B5E87775B5A4EAF8FF9FF36922F&originRegion=eu-west-1&originCreation=20230312101038

[6] https://reader.elsevier.com/reader/sd/pii/S2468227621003203?token=4E62B5CBE52EA263D73220259718F2ED42A3B9E198B18688DEB9E064C0EA48D91DBAB524E517887606B0E296D4CFC961&originRegion=eu-west-1&originCreation=20230312101044

[7] https://reader.elsevier.com/reader/sd/pii/S2001037021004281?token=5E294CD11401B9CD5FE28ED8A1A2881E2190BCDAB1F269B27241999B0C5D1195570A73D23FFD4A52A95043A04CD2651C&originRegion=eu-west-1&originCreation=20230312101053

[8] https://ieeexplore.ieee.org/document/9629944

[9] https://iopscience.iop.org/article/10.1088/1741-2552/ac2ca0/pdf

[10] https://www.proquest.com/docview/2615378768/430F25AD46584ABEPQ/1

Team members: Meda Ioana, Petrascu Andreea, Pop Paula. 

 

 Medicine XAI January - April 2021

Created by Team Cat Ladies, Andreea-Raluca TRANDAFIR & Eliza-Maria CIOVICA for Expert Systems 2023

Introduction

    The use of Artificial Intelligence (AI) and Machine Learning (ML) techniques is rapidly expanding in various medical fields. AI has the potential to provide accurate and effective preventive and curative interventions. However, concerns have also been raised about potential risks, harm, and trust issues due to the opacity of some AI algorithms, which makes them unexplainable. If the decision-making logic cannot be adequately explained, how can the decisions from these AI-based systems be trusted? Explainable Artificial Intelligence (XAI) aims to address these questions. In this article, we examined developments in XAI within the medical domain for a precise period of time, January to April 2021. Our objective was to see what has been uncovered in the previously stated period of time.

1. Expert Level Evaluations for Explainable AI (XAI) Methods in the Medical Domain - February 2021

    In this article, the aim is to show on an experimental level the way the expert-level evaluation of XAI methods in medical applications can be used and coincide with the actual explanations of clinicians. This is approached by collecting annotations provided by expert subjects equipped with an eye-tracker as they classify medical images and compute an approach for comparing results with those provided by XAI methods. The effectiveness of this technique is being demonstrated through a several experiments.

2. Feature-Guided CNN for Denoising Images From Portable Ultrasound Devices - February 2021

    This article refers to the ultrasound, the non-invasive medical imaging scanning device that has significantly improved medical diagnosis accuracy and efficiency. Portable ultrasound devices have become more popular due to their convenience and lower cost. Patients and physicians can easily access scanned images via a wireless network, but the image quality of portable devices is often inferior to that of standard hospital ultrasound equipment. This is because portable devices capture images with significant noise, which can hinder diagnosis accuracy. Addressing this issue, the article presents the Feature-guided Denoising Convolutional Neural Network (FDCNN) which has been proposed to remove noise while retaining important feature information. This model employs a hierarchical denoising framework that uses a feature masking layer for medical images. Additionally, an Explainable Artificial Intelligence (XAI) based feature extraction algorithm has been developed for medical images. The experimental results show that this feature extraction method outperforms previous methods, and when combined with the new denoising neural network architecture, portable ultrasound devices can achieve better diagnostic performance.

(First, the original image is extracted with a feature mask layer via a U-net network based on Guided Backpropagation. Then, add noise to the featureless areas using the mask layer. Subsequently, feed the image into the noise reduction network and perform residual learning. Finally, merge the feature information and the denoised images by a Laplacian fusion algorithm. )

3.  A Roadmap towards Breast Cancer Therapies Supported by Explainable Artificial Intelligence - April 2021

    Personalized medicine has become increasingly important in recent years, particularly in the development of oncological therapies. Patients' profiling strategies have shown promising rewards. In this study is introduced an explainable artificial intelligence (XAI) framework based on an adaptive dimensional reduction that outlines the most important clinical features for oncological patients' profiling and determines the profile, i.e., the cluster a patient belongs to based on these features. The data was collected on 267 breast cancer patients for this purpose. The dimensional reduction method employed identifies the relevant subspace, and distances among patients are used by a hierarchical clustering procedure to identify the corresponding optimal categories. The results revealed  that the molecular subtype is the most important feature for clustering. It was also assessed the robustness of current therapies and guidelines and found that the available patients' profiles determined in an unsupervised manner correspond to either molecular subtypes or therapies chosen based on guidelines, highlighting the interpretability of explainable approaches to machine learning techniques. This study suggests that data-driven therapies can be designed to emphasize the differences observed among patients.

4. The fourth scientific discovery paradigm for precision medicine and healthcare: Challenges ahead - April 2021

    The advancement of modern information techniques, such as next-generation sequencing (NGS), smart sensors based on the Internet of Everything (IoE), and artificial intelligence (AI) algorithms, has led to the emergence of data-intensive research and applications as the fourth paradigm for scientific discovery. Despite this, there are various challenges in the practical implementation of this paradigm. This article summarizes some challenges related to data-intensive discovery and applications in precision medicine and healthcare, and discusses future perspectives on next-generation medicine.

5. Unraveling the deep learning gearbox in optical coherence tomography image segmentation towards explainable artificial intelligence - February 2021

    The use of machine learning in analyzing medical data is common, but the internal workings of the algorithms are often opaque. To address this, researchers have proposed an enhanced convolutional neural network for optical coherence tomography image segmentation, which incorporates a Traceable Relevance Explainability (T-REX) technique. This involves generating ground truth data by multiple graders, calculating the differences between human graders and the algorithm, and visualizing the results. The T-REX setup resulted in a small average variability between the human graders and the algorithm, and the convolutional neural network allowed for modifiable predictions depending on the compartment. By making machine learning processes more transparent and understandable, the T-REX setup may lead to optimized applications.

6.    Enhancing Human-Machine Teaming for Medical Prognosis Through Neural Ordinary Differential Equations (NODEs) February 2021

    The article discusses the importance of producing explainable AI (XAI) to progress in improving the interpretability of some ML models, but also taking into account that all the efforts suffer from limitations: they work best at identifying why a system fails, but do poorly at explaining when and why a model’s prediction is correct. The authors supposed that the acceptability of ML predictions in expert domains is limited by: the machine’s horizon of prediction and the inability for machine predictions to incorporate human intuition into their model., so they proposed: Neural Ordinary Differential Equations (NODEs) to enhance human understanding and encourage acceptability. Their approach prioritized human cognitive intuition as the center of the algorithm. This approach aims to improve human-machine collaboration even in medical prognoses.

7.    Situated Case Studies for a Human-Centered Design of Explanation User Interfaces March 2021

   

    The article makes an introduction to how the idea of a human-centered perspective in the design of machine-learning based applications, especially in the context of Explainable Artificial Intelligence (XAI) is increasingly researched. The authors revealed that the problems they saw with a clear methodological guidance were due to the fact that each new situation seems to require new setup. They proposed a collection of case studies for human-centered XAI that can provide methodological guidance. They proposed three case studies in which they apply a HCD approach in the context of XAI: 

• Explaining Privacy-Preserving Machine Learning. This case study is in the medical domain and focuses on the value-oriented data donation of patients. The goal of this is to balance the trade-off between the protection of patients’ personal data and the need for unrestricted data in individualized medicine.

• Explaining Interactive Clustering Results, on how ML techniques are used to handle large-scale data in qualitative research settings.

• Explanations in Narrative-based Decision-Making, as a collaboration with a medical ethicist. As in complex situations there is a need for a holistic perspective of the patient, their situation, preferences, and moral concepts of what a good life represents for them. What is best for the patient must be determined in each specific case using a narrative-structured decision-making process rather than relying solely on external information.`

8. A Comparative Approach to Explainable Artificial Intelligence Methods in Application to High-Dimensional Electronic Health Records: Examining the Usability of XAI March 2021

  

  As explanatory AI (XAI) is on the rise, it aims to produce a factor of trust (which for humans is achieved by communicative means, rather than ML algorithms that cannot solely produce). The authors approached the medical field, and encountered challenges when dealing with the involvement of human-subjects, as trusting a machine to tend towards the livelihood of a human was hard, leaving trust as the basis of the human-expert in acceptance to the machines decision. This article aims to demonstrate the usability of explainable architecture as a layer to the medical domain supporting ML predictions and human-expert opinion. They executed different algorithms on a given data sets, and the conclusion was to proceed with XGBoost algorithm, as it was the best performing across all the given datasets:

9. Deep Learning Based Decision Support for Medicine -- A Case Study on Skin Cancer Diagnosis March 2021

  

  This article gets the light on the fact that early detection of skin cancers like melanoma is crucial when referring to  the survival of humans. As the majority of work in the medical AI community focusses on a diagnosis setting, that is more relevant for autonomous operations, the authors see the importance of practical decision support to not only have plain diagnosis, but provide explanations. This article focuses on an overview of the works towards explainable, DL-based decision support in medical applications with skin cancer diagnosis from clinical, dhermoscopic and histopathologic images as examples. As analysis reveals that the current work is dominated by visual relevance maps, the authors aim to focus on explanation of the images.

10. A multilayer multimodal detection and prediction model based on explainable artificial intelligence for Alzheimer’s disease

    

As Alzheimer’s disease (AD) is the most common type of dementia, its diagnosis and progression detection have been increasingly studied. The authors developed an accurate and interpretable AD diagnosis and progression detection model, which provide accurate decisions along with a set of explanations for each decision. This study took 1048 subjects from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) real-world dataset: 294 cognitively normal, 254 stable mild cognitive impairment (MCI), 232 progressive MCI, and 268 AD. It is actually a two-layer model with random forest (RF) as classifier algorithm. The performance of the model is optimized with key markers selected from a large set of biological and clinical measures.

Bibliography

[1] Muddamsetty, S.M., Jahromi, M.N.S., Moeslund, T.B. (2021). Expert Level Evaluations for Explainable AI (XAI) Methods in the Medical Domain. In: , et al. Pattern Recognition. ICPR International Workshops and Challenges. ICPR 2021. Lecture Notes in Computer Science(), vol 12663. Springer, Cham. https://doi.org/10.1007/978-3-030-68796-0_3

[2] G. Dong, Y. Ma and A. Basu, "Feature-Guided CNN for Denoising Images From Portable Ultrasound Devices," in IEEE Access, vol. 9, pp. 28272-28281, 2021, doi: 10.1109/ACCESS.2021.3059003.

[3] Amoroso, Nicola, Domenico Pomarico, Annarita Fanizzi, Vittorio Didonna, Francesco Giotta, Daniele La Forgia, Agnese Latorre, Alfonso Monaco, Ester Pantaleo, Nicole Petruzzellis, Pasquale Tamborra, Alfredo Zito, Vito Lorusso, Roberto Bellotti, and Raffaella Massafra. 2021. "A Roadmap towards Breast Cancer Therapies Supported by Explainable Artificial Intelligence" Applied Sciences 11, no. 11: 4881. https://doi.org/10.3390/app11114881

[4] Li Shen, Jinwei Bai, Jiao Wang, Bairong Shen, The fourth scientific discovery paradigm for precision medicine and healthcare: Challenges ahead, Precision Clinical Medicine, Volume 4, Issue 2, June 2021, Pages 80–84, https://doi.org/10.1093/pcmedi/pbab007

[5]  Maloca, P.M., Müller, P.L., Lee, A.Y. et al. Unraveling the deep learning gearbox in optical coherence tomography image segmentation towards explainable artificial intelligence. Commun Biol 4, 170 (2021). https://doi.org/10.1038/s42003-021-01697-y

 [6] Fompeyrine, D.A., Vorm, E.S., Ricka, N., Rose, F. and Pellegrin, G., 2021. Enhancing human-machine teaming for medical prognosis through neural ordinary differential equations (NODEs). Human-Intelligent Systems Integration, 3(4), pp.263-275.

[7] Müller-Birn, C., Glinka, K., Sörries, P., Tebbe, M. and Michl, S., 2021. Situated Case Studies for a Human-Centered Design of Explanation User Interfaces. arXiv preprint arXiv:2103.15462.

[8] Duell, J.A., 2021. A comparative approach to explainable artificial intelligence methods in application to high-dimensional electronic health records: Examining the usability of xai. arXiv preprint arXiv:2103.04951.

[9] Lucieri, A., Dengel, A. and Ahmed, S., 2021. Deep Learning Based Decision Support for Medicine--A Case Study on Skin Cancer Diagnosis. arXiv preprint arXiv:2103.05112.

[10] El-Sappagh, S., Alonso, J.M., Islam, S.M.R. et al. A multilayer multimodal detection and prediction model based on explainable artificial intelligence for Alzheimer’s disease. Sci Rep 11, 2660 (2021). https://doi.org/10.1038/s41598-021-82098-3

Explainable AI for Bioinformatics: Methods, Tools, and Applications - Team: SummonersRift

Team: SummonersRift

Lesll Armand-Alessandro

Gaspar Eduard

Alexandru Raul-Ionut

 Presented in Week 2

Period: September-December 2022

 XAI for bioinformatics

      Explainable AI for Bioinformatics: Methods, Tools, and Applications

Short Introduction

Artificial intelligence and machine learning in general have demonstrated remarkable performances in many tasks, from image processing to natural language processing, especially with the advent of deep learning. 

Artificial intelligence (AI) systems that are built on machine learning (ML) and deep neural networks (DNNs) are increasingly deployed in numerous application domains such as military, cybersecurity, healthcare, etc. Further, ML and DNN models are applied to solving complex and emerging biomedical research problems: from text mining, drug discovery, and singlecell RNA sequencing to early disease diagnosis and prognosis. Another common application of AI in precision medicine is predicting what treatment protocols are likely to succeed on a patient based on patient phenotype, demographic, and treatment contexts .

AI has already surpassed human medical experts in specific areas such as detecting tumors and analyzing disease progression. However, complex DNNs or ML models, which are often perceived as opaque and black-box, can make it difficult to understand the reasoning behind their decisions. This lack of transparency can be a challenge for both end-users and decision-makers, as well as AI developers. Additionally, in sensitive areas like healthcare, explainability and accountability are not only desirable but also legally required for AI systems that can have a significant impact on human lives. Fairness is another growing concern, as algorithmic decisions should not show bias or discrimination towards certain groups or individuals based on sensitive attributes. Explainable artificial intelligence (XAI) aims to overcome the opaqueness of black-box models and provide transparency in how AI systems make decisions. Interpretable ML models can explain how they make predictions and the factors that influence their outcomes. However, most state-of-the-art interpretable ML methods are domain-agnostic and evolved from fields like computer vision, automated reasoning, or statistics, making direct application to bioinformatics problems challenging without customization and domain-specific adaptation. 

Importance of XAI in Bioinformatics

Handling large-scale biomedical data involves significant challenges, including heterogeneity, high dimensionality, unstructured data, and high levels of noise and uncertainty. Despite its data-driven nature and technical complexity, the adoption of data-driven approaches in many bioinformatics scenarios is hindered by the lack of efficient ML models capable of tackling these challenges. In order for AI systems to provide trustworthy and reliable decisions, the need for interpretable ML models has become increasingly important to ensure transparency, fairness, and accountability in critical situations. Although not all predictions need to be explained, having a model that is interpretable can make it easier for users to understand and trust its decisions. Is is  showed via some experiments that under the right conditions, augmentations based on XAI can provide significant, diverse, and reliable benefits advantages over black-box models.

Helps avoid practical consequences

One of the critical applications of AI is aiding diagnosis and treatment of various cancerous conditions. Early detection and classification of patients into high or low-risk groups is crucial for effective management of the illness. An example of this is a doctor diagnosing a patient with breast cancer. Given that breast cancer is a leading cause of death in women, it is important for the diagnosis to be thoroughly investigated. By utilizing omics data, such as genetic mutations, copy number variations (CNVs), gene expression (GE), DNA methylation, and miRNA expression, accurate diagnosis and treatment can be identified. The diagnosis of a breast cancer patient requires a careful examination of multiple sources of data, including omics information, bioimaging, and clinical records. A multimodal DNN model trained on this data can classify samples with high accuracy.

Reduces complexity and improves accuracy

The aim of genomics data analysis is to extract biologically relevant information and gain insight into the role of biomarkers, such as genes, in cancer development. However, biological processes are complex systems controlled by the interactions of thousands of genes, not single gene-based mechanisms.  It is crucial to select biologically significant features with high correlation to the target classes and low correlation between genes. Accurately identifying cancer-specific biomarkers:              i) enhances classification accuracy,                                                                                                        ii) enables biologists to study the interactions of relevant genes, and                                                    iii) helps understand their functional behavior, leading to further gene discovery.                              After identifying these biomarkers based on feature attributions, they can be ranked based on their relative or absolute importance. These identified genes can serve as cancer-specific marker genes that distinguish specific or multiple tumor classes

Improves decision fairness

With the widespread use of AI, it is essential to address fairness concerns, as AI systems can make significant and impactful decisions in sensitive environments. Bias is a major hindrance to fair decision-making and has been a subject of discussion in philosophy and psychology for a long time. Statistically, bias means a false representation of the truth with respect to the population, and it can occur at any stage of the ML pipeline, from data collection, feature selection, model training, hyperparameter setting, to interpretation of results for affected individuals, such as patients. For instance, in healthcare and biomedicine, representative, selection, and discriminatory biases can easily be present in biophysical data, raising fairness concerns and potentially leading to unfair outcomes in various learning tasks. ML algorithms are a type of statistical discrimination that may become problematic when they give certain privileged groups a systematic advantage and certain disadvantaged groups a systematic disadvantage. If the data used to train an ML model is biased, the model will also be biased and produce biased decisions if used in an AI system.

Internal governance and legal compliance

As AI is more widely used, there is a growing need for transparency and explanation of AI decisions for ethical, legal, and safety reasons. This is particularly important in sensitive domains such as healthcare where AI can affect human lives. The EU General Data Protection Regulation (EU GDPR) recognizes the importance of ethics, accountability, and robustness in AI and requires automated decision-making processes to have appropriate safeguards, including the right to understand and challenge the decision, and the right not to be solely subject to an AI-based decision that significantly impacts one's life. The regulation enforces that processing based on automated decision-making tools should be subject to suitable safeguards, including the right to obtain an explanation of the decision reached after such an assessment and to challenge the decision. The regulation also states that individuals have the right not to be subject to a decision based solely on automated processing and whenever human subjects have their lives significantly impacted by an automatic decision-making machine

Techniques and Methods for Interpretable ML

A variety of interpretable machine learning methods, both model-specific and model-agnostic, have been proposed and can be broadly classified into three categories: probing, perturbing, and model surrogation. Depending on the level of abstraction, they can also be categorized as local or global interpretability methods. This article covers a lot of important subjects for example bioimaging, cancer genomics, text mining, and reasoning.

                                                            - Conclusion -

Explainable Artificial Intelligence (XAI) aims to overcome the opaqueness of black-box models and provide transparency in how AI systems make decisions. It is important in bioinformatics because handling large-scale biomedical data involves significant challenges, including heterogeneity, high dimensionality, unstructured data, and high levels of noise and uncertainty. The adoption of data-driven approaches in many bioinformatics scenarios is hindered by the lack of efficient ML models capable of tackling these challenges. For AI systems to provide trustworthy and reliable decisions, the need for interpretable ML models has become increasingly important to ensure transparency, fairness, and accountability in critical situations. XAI can reduce complexity and improve accuracy, improve decision fairness, help avoid practical consequences, and ensure internal governance and legal compliance.

A Deep Learning Approach Considering Image Background for Pneumonia Identification Using Explainable AI (XAI)

1. Introduction

Deep learning technologies are currently used to detect pneumonia. Unfortunately, standard deep learning approaches for pneumonia detection take less into consideration the effects of the lung X-ray picture backdrop on the model’s testing effect, limiting the model’s accuracy development. The chosen paper focuses on a deep learning approach that takes image background elements into account and studies the suggested method with explainable deep learning for explainability. The main idea is to eliminate the image background, increase the accuracy of pneumonia recognition and use the Grad-CAM approach to create an explainable deep learning model for pneumonia identification. The proposed approach enhances pneumonia detection accuracy, with the best accuracy of VGG16 reaching 95.6%.

2. Materials and methods

2.1. Overview

The deep learning method proposed in our chosen paper considers image background factors, and explainable deep learning is used to analyse the proposed method.

Step 1. 

A collection of X-ray images datasets of pneumonia patients and normal patients from publicly-available websites is collected, then raw images are organized and cleaned. 

The authors obtained a dataset of 5840 chest X-ray images of children aged 1 to 5 years with pneumonia or without pneumonia from the Guangzhou Women's and Children's Medical Center. Out of these, 4265 images were of patients with pneumonia and 1575 were of healthy subjects.

The authors of the proposed paper obtained a dataset of almost 6000 chest X-rays images of children aged 1 to 5 years with or without pneumonia from a chinese women’s and children’s medical center. Out of these, a little over 4000 images were of patients with pneumonia and almost 1500 were healthy subjects. To ensure accurate training and testing results, they selected an equal number of images from both groups and randomly split them into an 80% training set and a 20% testing set. They then removed the background of all the images and retained only the heart and lung features, resulting in a chest X-ray image dataset without a background. Finally, before training and testing, the images were scaled to a uniform size of 224 x 224.

Data augmentation is a technique used to prevent overfitting and expand small datasets. It involves applying transformations such as changes in size, orientation, and color to the original images to create new variations, resulting in a more diverse and valuable dataset. This method is widely used in deep learning for image classification tasks.

There are various methods of data augmentation. These include flipping the image (horizontally or vertically), rotating it at different angles, scaling the image (enlarging or reducing without changing content), cropping it (centrally or randomly), adjusting brightness, contrast, and hue (with random perturbations). These methods are illustrated in Figure 2.

Data augmentation can improve model generalization, reduce overfitting, and improve model robustness. In medical image analysis, where datasets are typically small, like in this paper, data augmentation is used to expand the available data.

Step 2. 

ResNet50 and the VGG16 pretrained model, suitable deep learning models, are selected in order to build deep learning models for pneumonia X-ray image classification using the idea of transfer learning. The training and testing are first conducted with the background condition included. 

Transfer learning:

Transfer learning is a method used to transfer knowledge gained from solving one problem to another similar or related problem, especially when the amount of available data is limited for training. This approach can reduce the amount of training data required for deep learning, reduce data dependency, and solve problems with a small amount of data. In this study, transfer learning is used to build the deep learning model using pre-trained models, which reduces dependence on the amount of data and improves accuracy due to the small dataset. This principle is illustrated in Figure 3.

Deep learning model construction:

This paper employs a transfer learning approach to build deep learning models for identifying pneumonia, utilizing pre-trained ResNet50 and VGG16 models from the ImageNet dataset. The final models constructed are the ResNet50 model and the VGG16 model, which are explained in detail below.

The ResNet50 model, a deep residual network, was proposed by Kaiming He et al. in 2015. ResNet introduced the residual block to address the problem of gradient disappearance resulting from increasing the depth of the neural network. This block also speeds up the training process and improves model accuracy. The ResNet50 model with a residual blockage is used in this paper. This reduces the number of input channels by a 1x1 convolution, thereby decreasing the overall model parameters and increasing computation speed.

VGG16 model: 

The VGG16 model is designed with a simple structure and utilizes small 3×3 convolutional kernels instead of larger ones to increase the network's depth and reduce its parameters, thereby improving the network's fitting ability. Compared to the AlexNet model architecture, the VGG16 network structure is deeper and wider, but the increase in computation is controlled.

Step 3. 

The background of the image is removed, and the training and testing are performed using ResNet50 and VGG16 pretrained deep learning models for X-ray images without the background condition to obtain a deep learning model for pneumonia X-ray image recognition considering the image background.

Background removal:

The VGG16 model uses small 3x3 convolutional kernels to increase network depth and improve fitting ability, with a simpler structure than AlexNet but deeper and wider. The increase in computation is controlled.

Deep Learning Model Construction:

The proposed study uses two deep learning models, ResNet50 and VGG16 pretrained, for pneumonia recognition without considering the background. The models are trained and tested using a pneumonia X-ray image dataset with the background removed to develop a model for recognizing pneumonia in X-ray images.

Step 4.

A visual explainability analysis of the deep learning model is performed using the GradCAM method to further obtain a trustworthy deep learning model for practical application of pneumonia recognition.

The paper uses the Grad-CAM method for visualization to explain the deep learning model for pneumonia recognition, which leads to a trustworthy model. Grad-CAM is preferred over CAM as it does not require a global average pooling operation, which can be difficult for CNN models with multiple fully connected layers. The CAM method is only applicable if the CNN model structure is modified by replacing the fully connected layer with global average pooling.

The Grad-CAM method calculates the weights of each channel of the feature map using the weighted gradient class activation map, which is based on the network back propagation gradient and the output vector. The weights are then used to weight and sum the corresponding feature map, and the resulting class activation map is obtained using the ReLU activation function. This method visualizes the attention of the neural network in the form of a heat map that shows the important regions of the image for target prediction and classification.

The Grad-CAM method can be applied to different CNN models without altering their structure, and it provides better visualization and more fine-grained details compared to the CAM method. It also avoids the tradeoff between model accuracy and explainability, and does not require retraining. A heatmap generated using the Grad-CAM method for pneumonia identification is presented in Figure 4, where red areas indicate regions of high concern. The heatmap highlights the lungs as the most significant area for pneumonia identification.

3. Results

The experimental data used had to first be filtered and cleaned, as well as ensure an equal number of pneumonia and normal images. This was done because only the binary pneumonia and normal X-ray image recognition was performed.

In total, the dataset contains 2500 images, from which 80% of them are randomly selected for the training set. Each image then had to be processed using manual methods to remove the background.

The deep learning models used are ResNet50 pretraining model and VGG16 pretraining model. The loss function used was torch.nn.CrossEntropyLoss() and torch.optim.SGD() as the optimizer.

Results without X-Ray Background Removal

In the figure below can be seen a comparison between 10 epochs for the ResNet50 model and VGG16 model. What can be observed here is that the highest accuracy for both models after data augmentation is higher than before data augmentation, which is due to the small number of the samples in the dataset. Therefore, data augmentation has to be used to improve the accuracy of the model. Furthermore, based on the comparison, the accuracy of the VGG16 model is higher than the ResNet50 model, regardless of data augmentation. This is because the problem of binary classification is simple, and the dataset is smaller.

After data augmentation, the data set is extended, thus leading to better model training and better accuracy for both models. Even so, in this particular situation, VGG16 model is better for pneumonia identifications on small datasets because it has a higher accuracy rate.

Results considering X-Ray Image Background

From the images, only the lung and heart components are kept. As in the case before, ResNet50 and VGG16 models are used.

1. Analysis results when removing the X-Ray Image Background

Comparing the two models before and after data augmentation with 10 epochs, as seen in the graphical representation above, the accuracy of ResNet50 after the background removal increases with each epoch after data augmentation and then it tends to stabilize. The highest accuracy rate for the ResNet50 after data augmentation is improved when compared to the case without data augmentation, but the effect is not really obvious. When comparing the accuracy increase for the VGG16 model when having or not data augmentation, the accuracy increases greatly when having data augmentation.

After data enhancement, the VGG16 model still outperforms the ResNet50 model in overall results and accuracy. Even so, the VGG16 model manifests some fluctuations, but overall, the accuracy is high and it has strong prospects for the development of pneumonia identification.

2. Before and after background removal

For this part the graphical data was not added, but the results show that the ResNet50 model’s accuracy for identification of pneumonia after the background removal is the same as before removing the background. This happens because the model has a deeper depth. thus being able to achieve higher accuracy for small datasets. On the other hand, the VGG16 model’s accuracy for pneumonia identification is greatly improved after removing the background regardless of data augmentation. The highest accuracy that was obtained for the VGG16 model reached 95.6% (mainly due to the simple structure of the model when compared to ResNet50).

The discussed paper compares the proposed model with previous architectures related works and out of all of them it scores the highest accuracy of 95.6% with the note that the datasets are not the same. Even so, this means that pneumonia identification problem can achieve better results.

Explainability Analysis of Deep Learning Model

Using the Grad-CAM method, five normal and pneumonia lung X-ray images were randomly selected before and after the removal of the background. The heatmaps of ResNet50 and VGG16 models were generated before and after background removal to identify pneumonia X-ray images, as shown in figures 9-12. Figure 9 displays the heatmap of normal lung X-ray images before the background removal, while figure 10 shows the heatmap of pneumonia X-ray images after the background removal. The heatmap of healthy lung X-ray images after background removal is illustrated in figure 11, while figure 12 shows the heatmap of pneumonia X-ray images after the background removal. The red color in the heatmap signifies high attention, blue signifies low attention, and yellow indicates the attention between the two.

After analyzing Fig. 9∼12, it is evident that both ResNet50 and VGG16 models can detect important lung areas. However, the ResNet50 model has a broader focus range, and some images are recognized inaccurately. On the other hand, the VGG16 model has a finer focus range, accurately identifying key areas of the lung texture, thus achieving more precise classification of pneumonia X-ray images. This could explain why the VGG16 model performs better than the ResNet50 model in terms of accuracy.

The ResNet50 and VGG16 models identified important lung areas before removing the background but also generated unnecessary attention in the abdomen and mandible, among other areas, which could affect the final model's accuracy. To address this issue, the researchers removed irrelevant background factors, such as the mandible, abdomen, and arm, and retained only the heart and lung parts relevant to pneumonia identification. As a result, attention to irrelevant locations was significantly reduced, and the models focused more accurately on key lung areas, resulting in improved accuracy in pneumonia recognition and enhanced model performance.

In addition, there are limitations in interpreting the two deep learning models, particularly regarding their internal structure. The small dataset used in this study also limits the accuracy of the models for large datasets. The manual processing method used to obtain accurate lung X-ray images after background removal is accurate but less efficient, and processing large sample datasets can be challenging. To address these limitations, more deep learning models will be analyzed, and the models will be modified and selected to create a more efficient and accurate model for pneumonia classification. Acquiring more pneumonia X-ray image data and classifying pneumonia types such as bacterial and viral pneumonia can also refine the diagnosis of pneumonia. The COVID-19 pandemic highlights the need for rapid diagnosis of COVID-19 X-ray images using deep learning methods to improve identification efficiency and accuracy. In the future, a more accurate image segmentation and extraction method will be proposed to obtain lung X-ray images with higher accuracy after background removal.

4. Discussions

Transfer learning and data augmentation are used to reduce the problem of deep learning models requiring massive training data and improve the accuracy of pneumonia X-ray image recognition. However, there are still limitations, such as the small dataset used, and better test results cannot be obtained for some models that apply to large sample datasets. In the future, more models will be compared and analyzed, and a deep learning model that is more suitable for the classification of lung X-ray images will be selected. The authors hope to achieve a rapid diagnosis of X-ray pictures of COVID-19 using deep learning methods in the future and propose a more accurate image segmentation and extraction method for efficient extraction of lung X-ray images with the background removed.

5. Conclusions

The method uses pre-existing deep learning models, removes the image background, and applies the Grad-CAM method to obtain an explainable deep learning model. The results show that considering the image background improves the accuracy of the deep learning model for pneumonia identification. Additionally, after removing the image background factors, the attention to key locations of the lungs increases, as shown by the Grad-CAM heatmap results. Future work is planned to establish a high-precision and trusted pneumonia recognition model.

Bibliography:

[1] https://ieeexplore.ieee.org/abstract/document/9827572

Bioinformatics XAI May - August 2021

1.

    The article describes the development of MISTy, a machine learning framework that extracts interactions from spatial omics data in a flexible, scalable, and explainable way. MISTy builds multiple views focusing on different spatial or functional contexts to dissect different effects and can be applied to different spatially resolved omics data without the need for cell-type annotation. The performance of MISTy is evaluated on an in silico dataset and three breast cancer datasets, demonstrating its applicability and potential for improving patient stratification. The flexibility of MISTy is also demonstrated by integrating different views to analyze intercellular signaling. Overall, MISTy provides a powerful tool for analyzing highly multiplexed spatial data in a way that can uncover novel insights into tumor progression and clinical outcomes.

      Fig. 1 - CNN Architecture.

2.

    The article discusses the need for improved methods to explain multimodal biomedical models. The authors propose using a gradient-based feature attribution approach called layer-wise relevance propagation (LRP) to explain the relative importance of different modalities in automated sleep stage classification. They trained a 1-D CNN with EEG, EOG, and EMG data and found that EEG was the most important modality for classification, followed by EOG and EMG. LRP provided consistent levels of importance for correctly classified samples and inconsistent levels for incorrectly classified samples, demonstrating its potential for explaining multimodal electrophysiology classifiers.

3.

    APRILE is an explainable AI framework that uses graph neural networks to identify molecular mechanisms underlying polypharmacy side effects. By identifying a set of proteins and associated Gene Ontology terms as mechanistic explanations of side effects caused by drug combinations, APRILE generates explanations for over 843,000 side effect-drug pair events across 20 disease categories. The framework provides new insights into molecular mechanisms of adverse drug reactions and facilitates the use of AI in healthcare.

4.

    The study proposes a pipeline for analyzing DNA methylation datasets for classification of controls and patients with different diseases. The pipeline includes data harmonization, machine learning classification models, dimensionality reduction, missing value imputation, and explainable artificial intelligence algorithms. Results show that harmonization improves classification accuracy, and dimensionality reduction and imputation methods achieve high accuracy. The approach is demonstrated using Parkinson’s disease and schizophrenia as examples, and explainable artificial intelligence approaches allow for explaining model predictions from both populational and individual perspectives. The proposed pipeline allows for solving data harmonization problems, imputing missing values, and building classification models of small dimensionality.

5.

    The study examined gender differences in risk factors for Cannabis Use Disorder (CUD) using a data-driven approach with machine learning and brain data. The study found that environmental factors (such as low education level and instrumental support) contributed more to CUD in women, while individual factors like personality, mental health, and neurocognitive factors played a more important role in men. The study suggests that these differences should be further investigated to develop more effective treatment approaches for CUD.

6.

    The study aimed to improve the explainability of deep learning models used for sleep staging with multimodal data. The researchers trained a convolutional neural network using EEG, EOG, and EMG data and proposed an ablation approach that replaces each modality with values that approximate line-related noise. The relative importance of each modality was consistent with sleep staging guidelines, with EEG being important for most stages and EOG for REM and nonREM stages. EMG had low importance across classes. The study suggests that a careful selection of an ablation approach may provide a clearer indicator of modality importance, and provides guidance for future research using explainability methods with multimodal electrophysiology data.

7.

    The study explores the role of variations in cis-regulatory elements (CREs) in driving the establishment of lineage-specific traits in plants. However, predicting expression behaviors from CRE patterns is challenging due to the complex biological processes involved. The researchers used cistrome datasets and explainable convolutional neural network (CNN) frameworks to predict genome-wide expression patterns in tomato fruits from the DNA sequences in gene regulatory regions. They developed a prediction model for a key expression pattern for the initiation of tomato fruit ripening by fixing the effects of trans-elements using single cell-type spatiotemporal transcriptome data for the response variables. The CNNs identified critical nucleotide residues for the expression pattern in each gene, which were validated experimentally in ripening tomato fruits. This framework will help understand regulatory networks derived from CREs and transcription factor interactions and provide a flexible way of designing alleles with optimized expression.

Fig. 2 - Visualizing the cis-regulatory elements or other nucleotide residues with the help of an explainable CNN.

8. 

    The article discusses the development of connectomics, which involves reconstructing neural connections at a nanometer scale. The article focuses on the use of deep learning methods in image processing to analyze connectomic data, specifically the development of an annotated Ultra-high Resolution Image Segmentation dataset for Cell membrane called U-RISC. The article highlights that the performance of current deep learning methods still lags behind human-level performance in analyzing the dataset, and attribution analysis is used to identify why this is the case. The article then provides a new benchmark for cell membrane segmentation on U-RISC and proposes suggestions for developing deep learning algorithms in the field of connectomics.

9. 

    The article presents the use of deep learning models as a biomarker for brain health through the estimation of brain age (BA) based on magnetic resonance imaging (MRI) data. The models accurately estimated age, revealing both small and large-scale changes associated with aging throughout the brain, as well as cardiovascular risk factors and accelerated aging in the frontal lobe. This study highlights the potential of deep learning models to detect brain-aging in healthy and at-risk individuals, providing valuable insight into brain health and serving as a useful tool for assessing brain health.

10.

    The article discusses the importance of using unsupervised artificial intelligence (AI) for analyzing large amounts of genetic data, such as genome sequences. The authors developed a batch-learning self-organizing map (BLSOM) for oligonucleotide compositions, which allowed for the separation of SARS-CoV-2 sequences into known clades and revealed subgrouping within those clades. The BLSOM was also used to analyze sequences from other microorganisms, eukaryotes, and the human genome, providing insights into their genetic characteristics. The authors explain the methodological strategies used in their analysis and highlight the advantages of using explainable AI like BLSOM for mining big data.

Bibliography

[1]    C. A. Ellis, R. L. Miller, and V. D. Calhoun, “A Novel Local Explainability Approach for Spectral Insight into Raw Eeg-Based Deep Learning Classifiers.” bioRxiv, p. 2021.06.10.447983, Jun. 11, 2021. doi: 10.1101/2021.06.10.447983.

[2]    C. A. Ellis, D. A. Carbajal, R. Zhang, R. L. Miller, V. D. Calhoun, and M. D. Wang, “An Explainable Deep Learning Approach for Multimodal Electrophysiology Classification.” bioRxiv, p. 2021.05.12.443594, Jun. 08, 2021. doi: 10.1101/2021.05.12.443594.

[3]    H. Xu et al., “APRILE: Exploring the Molecular Mechanisms of Drug Side Effects with Explainable Graph Neural Networks.” bioRxiv, p. 2021.07.02.450937, Sep. 20, 2021. doi: 10.1101/2021.07.02.450937.

[4]    A. Kalyakulina, I. Yusipov, M. G. Bacalini, C. Franceschi, M. Vedunova, and M. Ivanchenko, “Disease classification for whole blood DNA methylation: meta-analysis, missing values imputation, and XAI.” bioRxiv, p. 2022.05.10.491404, May 23, 2022. doi: 10.1101/2022.05.10.491404.

[5]    G. Niklason et al., “Explainable Machine Learning Analysis Reveals Gender Differences in the Phenotypic and Neurobiological Markers of Cannabis Use Disorder.” bioRxiv, p. 2021.08.30.458245, Sep. 01, 2021. doi: 10.1101/2021.08.30.458245.

[6]    C. A. Ellis, R. Zhang, D. A. Carbajal, R. L. Miller, V. D. Calhoun, and M. D. Wang, “Explainable Sleep Stage Classification with Multimodal Electrophysiology Time-series.” bioRxiv, p. 2021.05.04.442658, Jun. 08, 2021. doi: 10.1101/2021.05.04.442658.

[7]    T. Akagi et al., “Genome-wide cis-decoding for expression designing in tomato using cistrome data and explainable deep learning.” bioRxiv, p. 2021.06.01.446518, Jun. 01, 2021. doi: 10.1101/2021.06.01.446518.

[8]    R. Shi et al., “U-RISC: an ultra-high-resolution electron microscopy dataset challenging existing deep learning algorithms.” bioRxiv, p. 2021.05.30.446334, Jun. 15, 2021. doi: 10.1101/2021.05.30.446334.

[9]    S. M. Hofmann et al., “Towards the Interpretability of Deep Learning Models for Multi-modal Neuroimaging: Finding Structural Changes of the Ageing Brain.” bioRxiv, p. 2021.06.25.449906, Jun. 08, 2022. doi: 10.1101/2021.06.25.449906.

[10]  T. Ikemura, Y. Iwasaki, K. Wada, Y. Wada, and T. Abe, “Unsupervised explainable AI for the collective analysis of a massive number of genome sequences: various examples from the small genome of pandemic SARS-CoV-2 to the human genome.” bioRxiv, p. 2021.05.23.445371, May 24, 2021. doi: 10.1101/2021.05.23.445371.