Artificial Intelligence in Medicine: Knowledge Representation and Transparent and Explainable Systems

The formalization of clinical guideline knowledge is a prerequisite for the development of guideline-based decision support tools that can be used in clinical practice. Several guideline representation languages have been developed to formalize clinical guidelines and execute them over individual patient data. However, no standard has emerged from these efforts, and the core guideline elements to be represented have not been agreed upon in practice. One result is that there is little support when it comes to re-engineer a guideline modelled in a specific language into another language with different features. In this paper we describe a practical exercise consisting in modelling a guideline fragment in a target representation language starting from the same fragment modelled in a source language, having the source and target languages very different features. Concretely, we used PROforma as the source language and GDL as the target one. We also describe a methodological approach to facilitate this task. The lessons learnt from this work can be of interest not only to modellers tackling a similar task but also to developers of guideline transformation methods.

Goal-based reasoning may be used to support clinical decision making for multimorbidity patients; medical knowledge originating in different computer-interpretable guidelines (CIGs) and from medical ontologies may be matched at the goal level. This matching may be based on CIG metaproperty specifications referring to standard medical ontologies (e.g., as the U.S. Department of Veterans Affairs National Drug File - Reference Terminology) and adhering to standard patient information models (e.g., HL7’s Fast Healthcare Interoperability Resources). To support such knowledge and data integration, we developed a method for specifying metaproperty annotations within PROforma CIGs. We positioned this specification step within an existing method for CIG knowledge elicitation/specification, known as the Consensus method. Because clinicians time is costly, the research question that we evaluated in this study was whether knowledge engineers could successfully use this method to specify clinical practice guideline consensus documents in goal-annotated PROforma terms. The preliminary evaluation with nine information systems students taking an advanced knowledge representation course indicates is encouraging. We discuss the technical and conceptual modeling errors and how they could guide instruction of the goal-oriented CIG modeling.

Differential diagnosis of bacterial and viral meningitis remains an important clinical problem, particularly in the initial hours of hospitalization, before obtaining results of lumbar punction. We conducted a retrospective analysis of the medical records of 193 children hospitalized in St. Joseph Children’s Hospital in Poznan. In this study, we applied the original methodology of dominance-based rough set approach (DRSA) to induce diagnostic patterns from meningitis data and to represent them by decision rules useful in discriminating between bacterial and viral meningitis. The rule induction algorithm applied to this end is VC-DomLEM from jRS library. In the studied group of 193 patients, there were 124 boys and 69 girls, and the mean age was 94 months. The patients were characterized by 10 attributes, of which only 5 were used in 5 rules able to discriminate between bacterial and viral meningitis with an average precision of 98%, where C-reactive protein attribute (CRP) appeared to be the most valuable. Factors associated with bacterial meningitis were: CRP level ≥ 85 mg/l, or age < 2 months. Factors associated with viral meningitis were CRP level ≤ 60 mg/l and procalcytonin level < 0.5 ng/ml, or CRP level ≤ 84 mg/l and the presence of vomiting. We established a minimum set of attributes significant for classification of patients with bacterial or viral meningitis. These attributes are analyzed in just 5 rules able to distinguish almost perfectly between bacterial and viral meningitis without the need of lumbar punction.

The acute respiratory distress syndrome (ARDS) is a frequent type of respiratory failure observed in intensive care units. The Berlin classification identifies three severity levels of ARDS (mild, moderate, and severe), but this classification is under controversy in the medical community because it reflects neither the care requirements nor the expected clinical outcome of the patients. Here, the database MIMIC III (MetaVision) was used to investigate the similarity of patients within each one of the Berlin severity groups. We also ranked the relevance of common ARDS descriptive features and proposed four alternative classifiers to improve Berlin’s classification in the prediction of the duration of mechanical ventilation and mortality. One of these classifiers proved to be significantly better than current proposals and, therefore, it can be considered as a robust model to potentially improve health care processes and quality in the management of ARDS patients in Intensive Care Units (ICUs).

Overtreatment or mistreatment of patients is a phenomenon commonly encountered in health care and especially in the Intensive Care Unit (ICU) resulting in increased morbidity and mortality. We explore the MIMIC-III intensive care unit database and conduct experiments on an interpretable feature space based on the fusion of severity subscores, commonly used to predict mortality in an ICU setting. Clustering of medication and procedure context vectors based on a semantic representation has been performed to find common and individual treatment patterns. Two-day patient health state trajectories of a cohort of congestive heart failure patients are clustered and correlated with the treatment and evaluated based on an increase or reduction of probability of mortality on the second day of stay. Experimental results show differences in treatments and outcomes and the potential for using patient health state trajectories as a starting point for further evaluation of medical treatments and interventions.

Understanding medication effects is an important activity in pharmacovigilance in which patients are the most important contributor. Social media, where users share their personal experiences of medication effects, have been recommended as an alternative data source of gathering signal information of suspected medication effects. To discover potential medication-effect relations from Twitter data, we devised a method employing analogical reasoning with neural embedding of Twitter text. The process involves learning the neural embedding from unlabeled tweets and performing vector arithmetic, making it obscure to understand how an inferred relation is derived. To make the process understandable and interpretable and to facilitate the decision making on accepting or rejecting any inferred medication-effect relations, we added explanation(s) to each step of the process. An example of inferred relation is provided to demonstrate the effectiveness of our approach in explaining how the result of each step is derived.

An accurate and timely prediction of whether an infection is going to be resistant to a particular antibiotic could improve the clinical outcome of the patient as well as reduce the risk of spreading resistant microorganisms.

From a data analysis perspective, four key factors are present in antimicrobial resistance prediction: the high dimensionality of the data available, the imbalance present in the datasets, the concept drift along time and the need for their acceptance and implantation by clinical staff.

To date, no study has looked specifically at combining different strategies to deal with each of these four key factors. We believe interpretable prediction models are required. This study was undertaken to evaluate the impact of baseline interpretable predicting approaches using a dataset of real hospital data. In particular, we study the capacity of logistic regression, conditional trees and C5.0 rule-based models to improve the prediction when they are combined with oversampling, filtering and sliding windows.

Machine learning based predictive models have been used in different areas of everyday life for decades. However, with the recent availability of big data, new ways emerge on how to interpret the decisions of machine learning models. In addition to global interpretation focusing on the general prediction model decisions, this paper emphasizes the importance of local interpretation of predictions. Local interpretation focuses on specifics of each individual and provides explanations that can lead to a better understanding of the feature contribution in smaller groups of individuals that are often overlooked by the global interpretation techniques. In this paper, three machine learning based prediction models were compared: Gradient Boosting Machine (GBM), Random Forest (RF) and Generalized linear model with regularization (GLM). No significant differences in prediction performance, measured by mean average error, were detected: GLM: 0.573 (0.569 − 0.577); GBM: 0.579 (0.575 − 0.583); RF: 0.579 (0.575 − 0.583). Similar to other studies that used prediction models for screening in type 2 diabetes mellitus, we found a strong contribution of features like age, gender and BMI on the global interpretation level. On the other hand, local interpretation technique discovered some features like depression, smoking status or physical activity that can be influential in specific groups of patients. This study outlines the prospects of using local interpretation techniques to improve the interpretability of prediction models in the era of personalized healthcare. At the same time, we try to warn the users and developers of prediction models that prediction performance should not be the only criteria for model selection.

In this paper, a unified computational framework towards medical image explanation is proposed to promote the ability of computers on understanding and interpreting medical images. Four complementary modules are included, such as the construction of Medical Image-Text Joint Embedding (MITE) based on large-scale medical images and related texts; a Medical Image Semantic Association (MISA) mechanism based on the MITE multimodal knowledge representation; a Hierarchical Medical Image Caption (HMIC) module that is visually understandable to radiologists; and a language-independent medical imaging report generation prototype system by integrating the HMIC and transfer learning method. As an initial study of automatic medical image explanation, preliminary experiments were carried out to verify the feasibility of the proposed framework, including the extraction of large scale medical image-text pairs, semantic concept detection from medical images, and automatic medical imaging reports generation. However, there is still a great challenge to produce medical image interpretations clinically usable, and further research is needed to empower machines explaining medical images like a human being.

Sensor-based methods for human gait analysis often utilize electromyography capturing rich time-series data. Then, for transparent and explainable analysis interpretable methods are of prime importance. This paper presents analytical approaches in a framework for interpretable anomaly detection and classification of multivariate time series for human gait analysis. We exemplify the application utilizing a real-world medical dataset in the biomechanical orthopedics domain.

Bipolar disorder (BD) is a serious mental disorder characterized by manic episodes of elevated mood and overactivity, interspersed with periods of depression. Typically, the psychiatric assessment of affective state is carried out by a psychiatrist during routine check-up visits. However, diagnostics of a phase change can be facilitated by monitoring data collected by the patient’s smartphone. Previous studies concentrated primarily on the phase detection formulated as a classification task. In this study, we introduce a new approach to predict the phase change of BD patients using acoustic features and a combination of the Kohonen’s self-organizing maps and random forests. The primary goal is to predict the forthcoming change of patient’s state. We report on preliminary results that confirm the existence of a relation between the outcome of unsupervised learning (clustering) and the psychiatric assessment. Next, we evaluate the out-of-sample accuracy to predict the patient’s state with random forests. Finally, we discuss the potential of unsupervised learning for monitoring BD patients.

In recent years we see an increasing interest in applications of complex machine learning methods to medical problems. Black box models based on deep neural networks or ensembles are more and more popular in diagnostic, personalized medicine (Hamet and Tremblay 2017) or screening studies (Scheeder et al. 2018). Partially because they are accurate and easy to train. Nevertheless such models may be hard to understand and interpret. In high stake decisions, especially in medicine, the understanding of factors that drive model decisions is crucial. Lack of model understanding creates a serious risk in applications.

In our study we propose and validate new approaches to exploration and explanation of predictive models for early postoperative mortality in lung cancer patients. Models are created on the Domestic Lung Cancer Database run by the National Institute of Tuberculosis and Lung Diseases. We show how explainable machine learning techniques can be used to combine data driven signals with domain knowledge. Additionally we explore whether the insight provided by model explainers give valuable information for physicians.