Knowledge Discovery and Data Mining in Biomedical Informatics: The Future Is in Integrative, Interactive Machine Learning Solutions

Identification of Explainable Structures in Data with a Human-in-the-Loop

Explainable AIs (XAIs) often do not provide relevant or understandable explanations for a domain-specific human-in-the-loop (HIL). In addition, internally used metrics have biases that might not match existing structures in the data. The habilitation thesis presents an alternative solution approach by deriving explanations from high dimensional structures in the data rather than from predetermined classifications. Typically, the detection of such density- or distance-based structures in data has so far entailed the challenges of choosing appropriate algorithms and their parameters, which adds a considerable amount of complex decision-making options for the HIL. Central steps of the solution approach are a parameter-free methodology for the estimation and visualization of probability density functions (PDFs); followed by a hypothesis for selecting an appropriate distance metric independent of the data context in combination with projection-based clustering (PBC). PBC allows for subsequent interactive identification of separable structures in the data. Hence, the HIL does not need deep knowledge of the underlying algorithms to identify structures in data. The complete data-driven XAI approach involving the HIL is based on a decision tree guided by distance-based structures in data (DSD). This data-driven XAI shows initial success in the application to multivariate time series and non-sequential high-dimensional data. It generates meaningful and relevant explanations that are evaluated by Grice’s maxims.

Human-in-the-loop machine learning: a state of the art

Researchers are defining new types of interactions between humans and machine learning algorithms generically called human-in-the-loop machine learning. Depending on who is in control of the learning process, we can identify: active learning, in which the system remains in control; interactive machine learning, in which there is a closer interaction between users and learning systems; and machine teaching, where human domain experts have control over the learning process. Aside from control, humans can also be involved in the learning process in other ways. In curriculum learning human domain experts try to impose some structure on the examples presented to improve the learning; in explainable AI the focus is on the ability of the model to explain to humans why a given solution was chosen. This collaboration between AI models and humans should not be limited only to the learning process; if we go further, we can see other terms that arise such as Usable and Useful AI. In this paper we review the state of the art of the techniques involved in the new forms of relationship between humans and ML algorithms. Our contribution is not merely listing the different approaches, but to provide definitions clarifying confusing, varied and sometimes contradictory terms; to elucidate and determine the boundaries between the different methods; and to correlate all the techniques searching for the connections and influences between them.

Novel Machine-Learning Based Framework Using Electroretinography Data for the Detection of Early-Stage Glaucoma

Purpose

Early-stage glaucoma diagnosis has been a challenging problem in ophthalmology. The current state-of-the-art glaucoma diagnosis techniques do not completely leverage the functional measures' such as electroretinogram's immense potential; instead, focus is on structural measures like optical coherence tomography. The current study aims to take a foundational step toward the development of a novel and reliable predictive framework for early detection of glaucoma using machine-learning-based algorithm capable of leveraging medically relevant information that ERG signals contain.

Methods

ERG signals from 60 eyes of DBA/2 mice were grouped for binary classification based on age. The signals were also grouped based on intraocular pressure (IOP) for multiclass classification. Statistical and wavelet-based features were engineered and extracted. Important predictors (ERG tests and features) were determined, and the performance of five machine learning-based methods were evaluated.

Results

Random forest (bagged trees) ensemble classifier provided the best performance in both binary and multiclass classification of ERG signals. An accuracy of 91.7 and 80% was achieved for binary and multiclass classification, respectively, suggesting that machine-learning-based models can detect subtle changes in ERG signals if trained using advanced features such as those based on wavelet analyses.

Conclusions

The present study describes a novel, machine-learning-based method to analyze ERG signals providing additional information that may be used to detect early-stage glaucoma. Based on promising performance metrics obtained using the proposed machine-learning-based framework leveraging an established ERG data set, we conclude that the novel framework allows for detection of functional deficits of early/various stages of glaucoma in mice.

Machine Reading Comprehension with Rich Knowledge

Machine reading comprehension (MRC) is a crucial and challenging task in natural language processing (NLP). With the development of deep learning, language models have achieved excellent results. However, these models still cannot answer complex questions. Currently, researchers often utilize structured knowledge, such as knowledge bases (KBs), as external knowledge by directly extracting triples to enhance the results of machine reading. Although they can support certain background knowledge, the triples are limited to the interrelationships among entities or words. Unlike structured knowledge, unstructured knowledge is rich and extensive. However, these methods ignore unstructured knowledge resources, such as Wikipedia. In addition, the effect of combining the two types of knowledge is still not known. In this study, we first attempt to explore the usefulness of combining them. We introduce a fusion mechanism into a rich knowledge fusion layer (RKF) to obtain more useful and relevant knowledge from different external knowledge resources. Further to promote interaction among different types of knowledge, a bi-matching layer is added. We propose the RKF-NET framework based on BERT, and our experimental results demonstrate the effectiveness of two classic datasets: SQuAD1.1 and the Easy-Challenge (ARC).

Using Machine Learning to Identify Metabolomic Signatures of Pediatric Chronic Kidney Disease Etiology

BACKGROUND

Untargeted plasma metabolomic profiling combined with machine learning (ML) may lead to discovery of metabolic profiles that inform our understanding of pediatric CKD causes. We sought to identify metabolomic signatures in pediatric CKD based on diagnosis: FSGS, obstructive uropathy (OU), aplasia/dysplasia/hypoplasia (A/D/H), and reflux nephropathy (RN).

METHODS

Untargeted metabolomic quantification (GC-MS/LC-MS, Metabolon) was performed on plasma from 702 Chronic Kidney Disease in Children study participants (n: FSGS=63, OU=122, A/D/H=109, and RN=86). Lasso regression was used for feature selection, adjusting for clinical covariates. Four methods were then applied to stratify significance: logistic regression, support vector machine, random forest, and extreme gradient boosting. ML training was performed on 80% total cohort subsets and validated on 20% holdout subsets. Important features were selected based on being significant in at least two of the four modeling approaches. We additionally performed pathway enrichment analysis to identify metabolic subpathways associated with CKD cause.

RESULTS

ML models were evaluated on holdout subsets with receiver-operator and precision-recall area-under-the-curve, F1 score, and Matthews correlation coefficient. ML models outperformed no-skill prediction. Metabolomic profiles were identified based on cause. FSGS was associated with the sphingomyelin-ceramide axis. FSGS was also associated with individual plasmalogen metabolites and the subpathway. OU was associated with gut microbiome-derived histidine metabolites.

CONCLUSION

ML models identified metabolomic signatures based on CKD cause. Using ML techniques in conjunction with traditional biostatistics, we demonstrated that sphingomyelin-ceramide and plasmalogen dysmetabolism are associated with FSGS and that gut microbiome-derived histidine metabolites are associated with OU.

Advanced fuzzy cognitive maps: state-space and rule-based methodology for coronary artery disease detection

According to the World Health Organization, 50% of deaths in European Union are caused by Cardiovascular Diseases (CVD), while 80% of premature heart diseases and strokes can be prevented. In this study, a Computer-Aided Diagnostic model for a precise diagnosis of Coronary Artery Disease (CAD) is proposed. The methodology is based on State Space Advanced Fuzzy Cognitive Maps (AFCMs), an evolution of the traditional Fuzzy Cognitive Maps. Also, a rule-based mechanism is incorporated, to further increase the knowledge of the proposed system and the interpretability of the decision mechanism. The proposed method is evaluated utilizing a CAD dataset from the Department of Nuclear Medicine of the University Hospital of Patras, in Greece. Several experiments are conducted to define the optimal parameters of the proposed AFCM. Furthermore, the proposed AFCM is compared with the traditional FCM approach and the literature. The experiments highlight the effectiveness of the AFCM approach, obtaining 85.47% accuracy in CAD diagnosis, showing an improvement of +7% over the traditional approach. It is demonstrated that the AFCM approach in developing Fuzzy Cognitive Maps outperforms the conventional approach, while it constitutes a reliable method for the diagnosis of Coronary Artery Disease.

Explainable AI and Multi-Modal Causability in Medicine

Progress in statistical machine learning made AI in medicine successful, in certain classification tasks even beyond human level performance. Nevertheless, correlation is not causation and successful models are often complex “black-boxes”, which make it hard to understand why a result has been achieved. The explainable AI (xAI) community develops methods, e. g. to highlight which input parameters are relevant for a result; however, in the medical domain there is a need for causability: In the same way that usability encompasses measurements for the quality of use, causability encompasses measurements for the quality of explanations produced by xAI. The key for future human-AI interfaces is to map explainability with causability and to allow a domain expert to ask questions to understand why an AI came up with a result, and also to ask “what-if” questions (counterfactuals) to gain insight into the underlying independent explanatory factors of a result. A multi-modal causability is important in the medical domain because often different modalities contribute to a result.

An Integrated Platform for Skin Cancer Heterogenous and Multilayered Data Management

Electronic health record (EHR) systems improve health care services by allowing the combination of health data with clinical decision support features and clinical image analyses. This study presents a modular and distributed platform that is able to integrate and accommodate heterogeneous, multidimensional (omics, histological images and clinical) data for the multi-angled portrayal and management of skin cancer patients. The proposed design offers a layered analytical framework as an expansion of current EHR systems, which can integrate high-volume molecular -omics data, imaging data, as well as relevant clinical observations. We present a case study in the field of dermatology, where we attempt to combine the multilayered information for the early detection and characterization of melanoma. The specific architecture aspires to lower the barrier for the introduction of personalized therapeutic approaches, towards precision medicine. The paper describes the technical issues of implementation, along with an initial evaluation of the system and discussion.

Machine learning and clinical epigenetics: a review of challenges for diagnosis and classification

BACKGROUND

Machine learning is a sub-field of artificial intelligence, which utilises large data sets to make predictions for future events. Although most algorithms used in machine learning were developed as far back as the 1950s, the advent of big data in combination with dramatically increased computing power has spurred renewed interest in this technology over the last two decades.

MAIN BODY

Within the medical field, machine learning is promising in the development of assistive clinical tools for detection of e.g. cancers and prediction of disease. Recent advances in deep learning technologies, a sub-discipline of machine learning that requires less user input but more data and processing power, has provided even greater promise in assisting physicians to achieve accurate diagnoses. Within the fields of genetics and its sub-field epigenetics, both prime examples of complex data, machine learning methods are on the rise, as the field of personalised medicine is aiming for treatment of the individual based on their genetic and epigenetic profiles.

CONCLUSION

We now have an ever-growing number of reported epigenetic alterations in disease, and this offers a chance to increase sensitivity and specificity of future diagnostics and therapies. Currently, there are limited studies using machine learning applied to epigenetics. They pertain to a wide variety of disease states and have used mostly supervised machine learning methods.

The entropy of RR intervals is associated to gestational age in full-term newborns with adequate weight for gestational age

Introduction: The variability of successive RR intervals has been pointed out as an indicator of systemic homeostasis. The entropy of successive RR intervals is associated with a greater adaptive capacity, which is essential after childbirth, characterized by a change from an intrauterine environment that constantly adapts to the fetal demands, to an extrauterine environment that requires constant biological adaptations by the neonate.Objectives: To analyze the association between gestational age (GA) and the entropy of RR intervals in term infants with adequate birth weight in the first hours of extrauterine life.Methods: In a cross-sectional study design maternal, labor and neonatal characteristics were collected from the obstetric records. Successive RR intervals were recorded from neonates up to 72 hours (i.e. 3 days) of birth.Subjects: Fifty term infants, healthy and with adequate birth weight. Outcome measures: the variability of RR intervals was analyzed obtaining the entropy of 1000 successive RR intervals. Pearson's correlation was used to evaluate the association between GA and the entropy of successive RR intervals, while linear regression was used to obtain the coefficient of determination (r²) as well as a prediction model. The adequacy of the prediction model was evaluated using the Komolgorov-Smirnov test to evaluate the residuals distribution.Results: There was a positive and significant association between the studied variables (r = 0.437; p = .002). The coefficient of determination allowed us to infer that approximately 19.3% of the RR interval entropy from the studied sample can be explained by the GA (r² = 0.193; p = .002). The analysis of the residuals distribution confirmed that the regression model obtained here was adequate.Conclusion: Our results indicate that, even within a normal range of GA (≥37 a < 42 weeks) and birth weight, a longer intrauterine life allows a greater entropy of successive RR intervals, indicating a greater maturation of biological systems and adaptive capacity.

Why imaging data alone is not enough: AI-based integration of imaging, omics, and clinical data

Artificial intelligence (AI) is currently regaining enormous interest due to the success of machine learning (ML), and in particular deep learning (DL). Image analysis, and thus radiomics, strongly benefits from this research. However, effectively and efficiently integrating diverse clinical, imaging, and molecular profile data is necessary to understand complex diseases, and to achieve accurate diagnosis in order to provide the best possible treatment. In addition to the need for sufficient computing resources, suitable algorithms, models, and data infrastructure, three important aspects are often neglected: (1) the need for multiple independent, sufficiently large and, above all, high-quality data sets; (2) the need for domain knowledge and ontologies; and (3) the requirement for multiple networks that provide relevant relationships among biological entities. While one will always get results out of high-dimensional data, all three aspects are essential to provide robust training and validation of ML models, to provide explainable hypotheses and results, and to achieve the necessary trust in AI and confidence for clinical applications.

Impact of big data on oral health outcomes

Biomedical big data amasses from different sources such as electronic health records, health research, wearable devices and social media. Recent advances in data capturing, storage and analysis techniques have facilitated conversion of a wealth of knowledge in biomedical big data into evidence-based actionable plans to enhance population health and well-being. The delay in reaping the benefits of biomedical big data in dentistry is mainly due to the slow adoption of electronic health record systems, unstructured clinical records, tattered communication between data silos and perceiving oral health as a separate entity from general health. Recent recognition of the complex interplay between oral and general health has acknowledged the power of oral health big data to glean new insights on disease prevention and management. This review paper summarizes recent advances, limitations and challenges in biomedical big data in health care with emphasis on oral health and discusses the potential future applications of oral health big data to improve the quality and efficiency of personalized health care.

The effects of periodontal treatment on diabetic patients: The DIAPERIO randomized controlled trial

AIM

To assess whether periodontal treatment can lead to clinical, glycaemic control and quality of life improvements in metabolically unbalanced diabetic patients (type 1 or type 2) diagnosed with periodontitis.

METHODS

In this open-labelled randomized controlled trial, diabetic subjects (n = 91) were given "immediate" or "delayed" periodontal treatment (full-mouth non-surgical scaling and root planing, systemic antibiotics, and oral health instructions). The main outcome was the effect on glycated haemoglobin (HbA_1C ) and fructosamine levels. The General Oral Health Assessment Index and the SF-36 index were used to assess quality of life (QoL).

RESULTS

Periodontal health significantly improved after periodontal treatment (p < 0.001). Periodontal treatment seemed to be safe but had no significant effects on glycaemic control based on HbA_1C (adjusted mean difference with a 95% confidence interval (aMD) of 0.04 [-0.16;0.24]) and fructosamine levels (aMD 5.0 [-10.2;20.2]). There was no obvious evidence of improvement in general QoL after periodontal treatment. However, there was significant improvement in oral health-related QoL (aMD 7.0 [2.4;11.6], p = 0.003).

CONCLUSION

Although periodontal treatment showed no clinical effect on glycaemic control in this trial, important data were provided to support periodontal care among diabetic patients. Periodontal treatment is safe and improves oral health-related QoL in patients living with diabetes. ISRCTN15334496.

An Innovative Framework for Bioimage Annotation and Studies

The collection and analysis of clinical data are needed to investigate diseases and to define medical protocols and treatments. Bioimages, medical annotations and patient history are clinical data acquired and studied to perform a correct diagnosis and to propose an appropriate therapy. Currently, hospital departments manage these data using legacy systems which do not often allow data integration among different departments or health structures. Thus, in many cases clinical information sharing and exchange are difficult to implement. This is also the case for biomedical images for which data integration or data overlapping is usually not available. Image annotations and comparison can be crucial for physicians in many case studies. In this paper, a general purpose framework for bioimage management and annotations is proposed. Moreover, a simple-to-use information system has been developed to integrate clinical and diagnosis codes. The framework allows physicians (1) to integrate DICOM images from different platforms and (2) to report notes and highlights directly on images, thus offering, among the others, to query and compare similar clinical cases. This contribution is the result of a framework aimed to support oncologists in managing DICOM images and clinical data from different departments. Data integration is performed using a here-proposed XML-based module also utilized to trace temporal changes in image annotations.

A Review of User Interface Design for Interactive Machine Learning

Interactive Machine Learning (IML) seeks to complement human perception and intelligence by tightly integrating these strengths with the computational power and speed of computers. The interactive process is designed to involve input from the user but does not require the background knowledge or experience that might be necessary to work with more traditional machine learning techniques. Under the IML process, non-experts can apply their domain knowledge and insight over otherwise unwieldy datasets to find patterns of interest or develop complex data-driven applications. This process is co-adaptive in nature and relies on careful management of the interaction between human and machine. User interface design is fundamental to the success of this approach, yet there is a lack of consolidated principles on how such an interface should be implemented. This article presents a detailed review and characterisation of Interactive Machine Learning from an interactive systems perspective. We propose and describe a structural and behavioural model of a generalised IML system and identify solution principles for building effective interfaces for IML. Where possible, these emergent solution principles are contextualised by reference to the broader human-computer interaction literature. Finally, we identify strands of user interface research key to unlocking more efficient and productive non-expert interactive machine learning applications.

Empirical advances with text mining of electronic health records

Background

Korian is a private group specializing in medical accommodations for elderly and dependent people. A professional data warehouse (DWH) established in 2010 hosts all of the residents’ data. Inside this information system (IS), clinical narratives (CNs) were used only by medical staff as a residents’ care linking tool.

The objective of this study was to show that, through qualitative and quantitative textual analysis of a relatively small physiotherapy and well-defined CN sample, it was possible to build a physiotherapy corpus and, through this process, generate a new body of knowledge by adding relevant information to describe the residents’ care and lives.

Methods

Meaningful words were extracted through Standard Query Language (SQL) with the LIKE function and wildcards to perform pattern matching, followed by text mining and a word cloud using R® packages. Another step involved principal components and multiple correspondence analyses, plus clustering on the same residents’ sample as well as on other health data using a health model measuring the residents’ care level needs.

Results

By combining these techniques, physiotherapy treatments could be characterized by a list of constructed keywords, and the residents’ health characteristics were built. Feeding defects or health outlier groups could be detected, physiotherapy residents’ data and their health data were matched, and differences in health situations showed qualitative and quantitative differences in physiotherapy narratives.

Conclusions

This textual experiment using a textual process in two stages showed that text mining and data mining techniques provide convenient tools to improve residents’ health and quality of care by adding new, simple, useable data to the electronic health record (EHR). When used with a normalized physiotherapy problem list, text mining through information extraction (IE), named entity recognition (NER) and data mining (DM) can provide a real advantage to describe health care, adding new medical material and helping to integrate the EHR system into the health staff work environment.

Electronic supplementary material

The online version of this article (doi:10.1186/s12911-017-0519-0) contains supplementary material, which is available to authorized users.

Infer Cause of Death for Population Health Using Convolutional Neural Network

In biomedical data analysis, inferring the cause of death is a challenging and important task, which is useful for both public health reporting purposes, as well as improving patients' quality of care by identifying severer conditions. Causal inference, however, is notoriously difficult. Traditional causal inference mainly relies on analyzing data collected from experiment of specific design, which is expensive, and limited to a certain disease cohort, making the approach less generalizable. In our paper, we adopt a novel data-driven perspective to analyze and improve the death reporting process, to assist physicians identify the single underlying cause of death. To achieve this, we build state-of-the-art deep learning models, convolution neural network (CNN), and achieve around 75% accuracy in predicting the single underlying cause of death from a list of relevant medical conditions. We also provide interpretations for the black-box neural network models, so that death reporting physicians can apply the model with better understanding of the model.

Introduction to MAchine Learning & Knowledge Extraction (MAKE)

The grand goal of Machine Learning is to develop software which can learn from previous experience—similar to how we humans do. Ultimately, to reach a level of usable intelligence, we need (1) to learn from prior data, (2) to extract knowledge, (3) to generalize—i.e., guessing where probability function mass/density concentrates, (4) to fight the curse of dimensionality, and (5) to disentangle underlying explanatory factors of the data—i.e., to make sense of the data in the context of an application domain. To address these challenges and to ensure successful machine learning applications in various domains an integrated machine learning approach is important. This requires a concerted international effort without boundaries, supporting collaborative, cross-domain, interdisciplinary and transdisciplinary work of experts from seven sections, ranging from data pre-processing to data visualization, i.e., to map results found in arbitrarily high dimensional spaces into the lower dimensions to make it accessible, usable and useful to the end user. An integrated machine learning approach needs also to consider issues of privacy, data protection, safety, security, user acceptance and social implications. This paper is the inaugural introduction to the new journal of MAchine Learning & Knowledge Extraction (MAKE). The goal is to provide an incomplete, personally biased, but consistent introduction into the concepts of MAKE and a brief overview of some selected topics to stimulate future research in the international research community.

Interactive knowledge discovery with the doctor-in-the-loop: a practical example of cerebral aneurysms research

Established process models for knowledge discovery find the domain-expert in a customer-like and supervising role. In the field of biomedical research, it is necessary to move the domain-experts into the center of this process with far-reaching consequences for both their research output and the process itself. In this paper, we revise the established process models for knowledge discovery and propose a new process model for domain-expert-driven interactive knowledge discovery. Furthermore, we present a research infrastructure which is adapted to this new process model and demonstrate how the domain-expert can be deeply integrated even into the highly complex data-mining process and data-exploration tasks. We evaluated this approach in the medical domain for the case of cerebral aneurysms research.