The truth will come to light: directions and challenges in extracting the knowledge embedded within trained artificial neural networks

A Machine Learning Approach to Assess Patients with Deep Neck Infection Progression to Descending Mediastinitis: Preliminary Results

BACKGROUND

Deep neck infection (DNI) is a serious infectious disease, and descending mediastinitis is a fatal infection of the mediastinum. However, no study has applied artificial intelligence to assess progression to descending mediastinitis in DNI patients. Thus, we developed a model to assess the possible progression of DNI to descending mediastinitis.

METHODS

Between August 2017 and December 2022, 380 patients with DNI were enrolled; 75% of patients (n = 285) were assigned to the training group for validation, whereas the remaining 25% (n = 95) were assigned to the test group to determine the accuracy. The patients' clinical and computed tomography (CT) parameters were analyzed via the k-nearest neighbor method. The predicted and actual progression of DNI patients to descending mediastinitis were compared.

RESULTS

In the training and test groups, there was no statistical significance (all p > 0.05) noted at clinical variables (age, gender, chief complaint period, white blood cells, C-reactive protein, diabetes mellitus, and blood sugar), deep neck space (parapharyngeal, submandibular, retropharyngeal, and multiple spaces involved, ≥3), tracheostomy performance, imaging parameters (maximum diameter of abscess and nearest distance from abscess to level of sternum notch), or progression to mediastinitis. The model had a predictive accuracy of 82.11% (78/95 patients), with sensitivity and specificity of 41.67% and 87.95%, respectively.

CONCLUSIONS

Our model can assess the progression of DNI to descending mediastinitis depending on clinical and imaging parameters. It can be used to identify DNI patients who will benefit from prompt treatment.

Use of machine learning to predict bladder cancer survival outcomes: a systematic literature review

INTRODUCTION

The objective of this systematic review is to summarize the use of machine learning (ML) in predicting overall survival (OS) in patients with bladder cancer.

METHODS

Search terms for bladder cancer, ML algorithms, and mortality were used to identify studies in PubMed and Web of Science as of February 2022. Notable inclusion/exclusion criteria contained the inclusion of studies that utilized patient-level datasets and exclusion of primary gene expression-related dataset studies. Study quality and bias were assessed using the International Journal of Medical Informatics (IJMEDI) checklist.

RESULTS

Of the 14 included studies, the most common algorithms were artificial neural networks (n = 8) and logistic regression (n = 4). Nine articles described missing data handling, with five articles removing patients with missing data entirely. With respect to feature selection, the most common sociodemographic variables were age (n = 9), gender (n = 9), and smoking status (n = 3), with clinical variables most commonly including tumor stage (n = 8), grade (n = 7), and lymph node involvement (n = 6). Most studies (n = 10) were of medium IJMEDI quality, with common areas of improvement being the descriptions of data preparation and deployment.

CONCLUSIONS

ML holds promise for optimizing bladder cancer care through accurate OS predictions, but challenges related to data processing, feature selection, and data source quality must be resolved to develop robust models. While this review is limited by its inability to compare models across studies, this systematic review will inform decision-making by various stakeholders to improve understanding of ML-based OS prediction in bladder cancer and foster interpretability of future models.

A primer on artificial intelligence in plant digital phenomics: embarking on the data to insights journey

Artificial intelligence (AI) has emerged as a fundamental component of global agricultural research that is poised to impact on many aspects of plant science. In digital phenomics, AI is capable of learning intricate structure and patterns in large datasets. We provide a perspective and primer on AI applications to phenome research. We propose a novel human-centric explainable AI (X-AI) system architecture consisting of data architecture, technology infrastructure, and AI architecture design. We clarify the difference between post hoc models and 'interpretable by design' models. We include guidance for effectively using an interpretable by design model in phenomic analysis. We also provide directions to sources of tools and resources for making data analytics increasingly accessible. This primer is accompanied by an interactive online tutorial.

Contestable AI by Design: Towards a Framework

As the use of AI systems continues to increase, so do concerns over their lack of fairness, legitimacy and accountability. Such harmful automated decision-making can be guarded against by ensuring AI systems are contestable by design: responsive to human intervention throughout the system lifecycle. Contestable AI by design is a small but growing field of research. However, most available knowledge requires a significant amount of translation to be applicable in practice. A proven way of conveying intermediate-level, generative design knowledge is in the form of frameworks. In this article we use qualitative-interpretative methods and visual mapping techniques to extract from the literature sociotechnical features and practices that contribute to contestable AI, and synthesize these into a design framework.

Deep Learning Artificial Intelligence to Predict the Need for Tracheostomy in Patients of Deep Neck Infection Based on Clinical and Computed Tomography Findings—Preliminary Data and a Pilot Study

Background: Deep neck infection (DNI) can lead to airway obstruction. Rather than intubation, some patients need tracheostomy to secure the airway. However, no study has used deep learning (DL) artificial intelligence (AI) to predict the need for tracheostomy in DNI patients. Thus, the purpose of this study was to develop a DL framework to predict the need for tracheostomy in DNI patients. Methods: 392 patients with DNI were enrolled in this study between August 2016 and April 2022; 80% of the patients (n = 317) were randomly assigned to a training group for model validation, and the remaining 20% (n = 75) were assigned to the test group to determine model accuracy. The k-nearest neighbor method was applied to analyze the clinical and computed tomography (CT) data of the patients. The predictions of the model with regard to the need for tracheostomy were compared with actual decisions made by clinical experts. Results: No significant differences were observed in clinical or CT parameters between the training group and test groups. The DL model yielded a prediction accuracy of 78.66% (59/75 cases). The sensitivity and specificity values were 62.50% and 80.60%, respectively. Conclusions: We demonstrated a DL framework to predict the need for tracheostomy in DNI patients based on clinical and CT data. The model has potential for clinical application; in particular, it may assist less experienced clinicians to determine whether tracheostomy is necessary in cases of DNI.

Use of an artificial intelligence-based rule extraction approach to predict an emergency cesarean section

Objective

One of the major problems with artificial intelligence (AI) is that it is generally known as a “black box”. Therefore, the present study aimed to construct an emergency cesarean section (CS) prediction system using an AI‐based rule extraction approach as a “white box” to detect the cause for the emergency CS.

Methods

Data were collected from all perinatal records of all delivery outcomes at Osaka Medical College between December 2014 and July 2019. We identified the delivery method for all deliveries after 36 gestational weeks as either (1) vaginal delivery or scheduled CS, or (2) emergency CS. From among these, we selected 52 risk factors to feed into an AI‐based rule extraction algorithm to extract rules to predict an emergency CS.

Results

We identified 1513 singleton deliveries (1285 [84.9%] vaginal deliveries, 228 emergency CS [15.1%]) and extracted 15 rules. We achieved an average accuracy of 81.90% using five‐fold cross‐validation and an area under the receiving operating characteristic curve of 71.46%.

Conclusion

To our knowledge, this is the first study to use interpretable AI‐based rule extraction technology to predict an emergency CS. This system appears to be useful for identifying hidden factors for emergency CS.

This is the first study to construct a prediction system for an emergency cesarean section using an artificial intelligence‐based “white box” rule extraction approach.

The ethics of algorithms: key problems and solutions

Research on the ethics of algorithms has grown substantially over the past decade. Alongside the exponential development and application of machine learning algorithms, new ethical problems and solutions relating to their ubiquitous use in society have been proposed. This article builds on a review of the ethics of algorithms published in 2016 (Mittelstadt et al. Big Data Soc 3(2), 2016). The goals are to contribute to the debate on the identification and analysis of the ethical implications of algorithms, to provide an updated analysis of epistemic and normative concerns, and to offer actionable guidance for the governance of the design, development and deployment of algorithms.

Detection of Lower Albuminuria Levels and Early Development of Diabetic Kidney Disease Using an Artificial Intelligence-Based Rule Extraction Approach

The aim of the present study was to determine the lowest cut-off value for albuminuria levels, which can be used to detect diabetic kidney disease (DKD) using the urinary albumin-to-creatinine ratio (UACR). National Health and Nutrition Examination Survey (NHANES) data for US adults were used, and participants were classified as having diabetes or prediabetes based on a self-report and physiological measures. The study dataset comprised 942 diabetes and 524 prediabetes samples. This study clarified the significance of the lower albuminuria (UACR) levels, which can detect DKD, using an artificial intelligence-based rule extraction approach. The diagnostic rules (15 concrete rules) for both samples were extracted using a recursive-rule eXtraction (Re-RX) algorithm with continuous attributes (continuous Re-RX) to discriminate between prediabetes and diabetes datasets. Continuous Re-RX showed high test accuracy (77.56%) and a large area under the receiver operating characteristics curve (75%), which derived the two cut-off values (6.1 mg/g Cr and 71.00 mg/g Cr) for the lower albuminuria level in the UACR to detect early development of DKD. The early cut-off values for normoalbuminuria (NA) and microalbuminuria (MA) will be determined to help detect CKD and DKD, and to detect diabetes before MA develop and to prevent diabetic complications.

The Right Direction Needed to Develop White-Box Deep Learning in Radiology, Pathology, and Ophthalmology: A Short Review

The popularity of deep learning (DL) in the machine learning community has been dramatically increasing since 2012. The theoretical foundations of DL are well-rooted in the classical neural network (NN). Rule extraction is not a new concept, but was originally devised for a shallow NN. For about the past 30 years, extensive efforts have been made by many researchers to resolve the “black box” problem of trained shallow NNs using rule extraction technology. A rule extraction technology that is well-balanced between accuracy and interpretability has recently been proposed for shallow NNs as a promising means to address this black box problem. Recently, we have been confronting a “new black box” problem caused by highly complex deep NNs (DNNs) generated by DL. In this paper, we first review four rule extraction approaches to resolve the black box problem of DNNs trained by DL in computer vision. Next, we discuss the fundamental limitations and criticisms of current DL approaches in radiology, pathology, and ophthalmology from the black box point of view. We also review the conversion methods from DNNs to decision trees and point out their limitations. Furthermore, we describe a transparent approach for resolving the black box problem of DNNs trained by a deep belief network. Finally, we provide a brief description to realize the transparency of DNNs generated by a convolutional NN and discuss a practical way to realize the transparency of DL in radiology, pathology, and ophthalmology.

Applications of artificial neural networks in health care organizational decision-making: A scoping review

Health care organizations are leveraging machine-learning techniques, such as artificial neural networks (ANN), to improve delivery of care at a reduced cost. Applications of ANN to diagnosis are well-known; however, ANN are increasingly used to inform health care management decisions. We provide a seminal review of the applications of ANN to health care organizational decision-making. We screened 3,397 articles from six databases with coverage of Health Administration, Computer Science and Business Administration. We extracted study characteristics, aim, methodology and context (including level of analysis) from 80 articles meeting inclusion criteria. Articles were published from 1997–2018 and originated from 24 countries, with a plurality of papers (26 articles) published by authors from the United States. Types of ANN used included ANN (36 articles), feed-forward networks (25 articles), or hybrid models (23 articles); reported accuracy varied from 50% to 100%. The majority of ANN informed decision-making at the micro level (61 articles), between patients and health care providers. Fewer ANN were deployed for intra-organizational (meso- level, 29 articles) and system, policy or inter-organizational (macro- level, 10 articles) decision-making. Our review identifies key characteristics and drivers for market uptake of ANN for health care organizational decision-making to guide further adoption of this technique.

A contemporary review of machine learning in otolaryngology-head and neck surgery

One of the key challenges with big data is leveraging the complex network of information to yield useful clinical insights. The confluence of massive amounts of health data and a desire to make inferences and insights on these data has produced a substantial amount of interest in machine-learning analytic methods. There has been a drastic increase in the otolaryngology literature volume describing novel applications of machine learning within the past 5 years. In this timely contemporary review, we provide an overview of popular machine-learning techniques, and review recent machine-learning applications in otolaryngology-head and neck surgery including neurotology, head and neck oncology, laryngology, and rhinology. Investigators have realized significant success in validated models with model sensitivities and specificities approaching 100%. Challenges remain in the implementation of machine-learning algorithms. This may be in part the unfamiliarity of these techniques to clinician leaders on the front lines of patient care. Spreading awareness and confidence in machine learning will follow with further validation and proof-of-value analyses that demonstrate model performance superiority over established methods. We are poised to see a greater influx of machine-learning applications to clinical problems in otolaryngology-head and neck surgery, and it is prudent for providers to understand the potential benefits and limitations of these technologies. Laryngoscope, 130:45-51, 2020.

Artificial Neural Networks

This chapter examines the history of artificial neural networks research through the present day. The components of artificial neural network architectures and both unsupervised and supervised learning methods are discussed. Although a step-by-step tutorial of how to develop artificial neural networks is not included, additional reading suggestions covering artificial neural network development are provided. The advantages and disadvantages of artificial neural networks for research and real-world applications are presented as well as potential solutions to many of the disadvantages. Future research directions for the field of artificial neural networks are presented.

Neural system prediction and identification challenge

Can we infer the function of a biological neural network (BNN) if we know the connectivity and activity of all its constituent neurons?This question is at the core of neuroscience and, accordingly, various methods have been developed to record the activity and connectivity of as many neurons as possible. Surprisingly, there is no theoretical or computational demonstration that neuronal activity and connectivity are indeed sufficient to infer the function of a BNN. Therefore, we pose the Neural Systems Identification and Prediction Challenge (nuSPIC). We provide the connectivity and activity of all neurons and invite participants (1) to infer the functions implemented (hard-wired) in spiking neural networks (SNNs) by stimulating and recording the activity of neurons and, (2) to implement predefined mathematical/biological functions using SNNs. The nuSPICs can be accessed via a web-interface to the NEST simulator and the user is not required to know any specific programming language. Furthermore, the nuSPICs can be used as a teaching tool. Finally, nuSPICs use the crowd-sourcing model to address scientific issues. With this computational approach we aim to identify which functions can be inferred by systematic recordings of neuronal activity and connectivity. In addition, nuSPICs will help the design and application of new experimental paradigms based on the structure of the SNN and the presumed function which is to be discovered.

NEURAL NETWORK RULE EXTRACTION BY A NEW ENSEMBLE CONCEPT AND ITS THEORETICAL AND HISTORICAL BACKGROUND: A REVIEW

This paper presents theoretical and historical backgrounds related to neural network rule extraction. It also investigates approaches for neural network rule extraction by ensemble concepts. Bologna pointed out that although many authors had generated comprehensive models from individual networks, much less work had been done to explain ensembles of neural networks. This paper carefully surveyed the previous work on rule extraction from neural network ensembles since 1988. We are aware of three major research groups i.e., Bologna' group, Zhou' group and Hayashi' group. The reason of these situations is obvious. Since the structures of previous neural network ensembles were quite complicated, the research on the efficient rule extraction algorithm from neural network ensembles was few although their learning capability was extremely high. Thus, these issues make rule extraction algorithm for neural network ensemble difficult task. However, there is a practical need for new ideas for neural network ensembles in order to realize the extremely high-performance needs of various rule extraction problems in real life. This paper successively explain nature of artificial neural networks, origin of neural network rule extraction, incorporating fuzziness in neural network rule extraction, theoretical foundation of neural network rule extraction, computational complexity of neural network rule extraction, neuro-fuzzy hybridization, previous rule extraction from neural network ensembles and difficulties of previous neural network ensembles. Next, this paper address three principles of proposed neural network rule extraction: to increase recognition rates, to extract rules from neural network ensembles, and to minimize the use of computing resources. We also propose an ensemble-recursive-rule extraction (E-Re-RX) by two or three standard backpropagation to train multi-layer perceptrons (MLPs), which enabled extremely high recognition accuracy and the extraction of comprehensible rules. Furthermore, this enabled rule extraction that resulted in fewer rules than those in previously proposed methods. This paper summarizes experimental results of rule extraction using E-Re-RX by multiple standard backpropagation MLPs and provides deep discussions. The results make it possible for the output from a neural network ensemble to be in the form of rules, thus open the "black box" of trained neural networks ensembles. Finally, we provide valuable conclusions and as future work, three open questions on the E-Re-RX algorithm.

Investigating biocomplexity through the agent-based paradigm

Capturing the dynamism that pervades biological systems requires a computational approach that can accommodate both the continuous features of the system environment as well as the flexible and heterogeneous nature of component interactions. This presents a serious challenge for the more traditional mathematical approaches that assume component homogeneity to relate system observables using mathematical equations. While the homogeneity condition does not lead to loss of accuracy while simulating various continua, it fails to offer detailed solutions when applied to systems with dynamically interacting heterogeneous components. As the functionality and architecture of most biological systems is a product of multi-faceted individual interactions at the sub-system level, continuum models rarely offer much beyond qualitative similarity. Agent-based modelling is a class of algorithmic computational approaches that rely on interactions between Turing-complete finite-state machines--or agents--to simulate, from the bottom-up, macroscopic properties of a system. In recognizing the heterogeneity condition, they offer suitable ontologies to the system components being modelled, thereby succeeding where their continuum counterparts tend to struggle. Furthermore, being inherently hierarchical, they are quite amenable to coupling with other computational paradigms. The integration of any agent-based framework with continuum models is arguably the most elegant and precise way of representing biological systems. Although in its nascence, agent-based modelling has been utilized to model biological complexity across a broad range of biological scales (from cells to societies). In this article, we explore the reasons that make agent-based modelling the most precise approach to model biological systems that tend to be non-linear and complex.

Symbolic representation of recurrent neural network dynamics

Simple recurrent error backpropagation networks have been widely used to learn temporal sequence data, including regular and context-free languages. However, the production of relatively large and opaque weight matrices during learning has inspired substantial research on how to extract symbolic human-readable interpretations from trained networks. Unlike feedforward networks, where research has focused mainly on rule extraction, most past work with recurrent networks has viewed them as dynamical systems that can be approximated symbolically by finite-state machine (FSMs). With this approach, the network's hidden layer activation space is typically divided into a finite number of regions. Past research has mainly focused on better techniques for dividing up this activation space. In contrast, very little work has tried to influence the network training process to produce a better representation in hidden layer activation space, and that which has been done has had only limited success. Here we propose a powerful general technique to bias the error backpropagation training process so that it learns an activation space representation from which it is easier to extract FSMs. Using four publicly available data sets that are based on regular and context-free languages, we show via computational experiments that the modified learning method helps to extract FSMs with substantially fewer states and less variance than unmodified backpropagation learning, without decreasing the neural networks' accuracy. We conclude that modifying error backpropagation so that it more effectively separates learned pattern encodings in the hidden layer is an effective way to improve contemporary FSM extraction methods.

Combined rule extraction and feature elimination in supervised classification

There are a vast number of biology related research problems involving a combination of multiple sources of data to achieve a better understanding of the underlying problems. It is important to select and interpret the most important information from these sources. Thus it will be beneficial to have a good algorithm to simultaneously extract rules and select features for better interpretation of the predictive model. We propose an efficient algorithm, Combined Rule Extraction and Feature Elimination (CRF), based on 1-norm regularized random forests. CRF simultaneously extracts a small number of rules generated by random forests and selects important features. We applied CRF to several drug activity prediction and microarray data sets. CRF is capable of producing performance comparable with state-of-the-art prediction algorithms using a small number of decision rules. Some of the decision rules are biologically significant.

An attribute weight assignment and particle swarm optimization algorithm for medical database classifications

In this research, a hybrid model is developed by integrating a case-based reasoning approach and a particle swarm optimization model for medical data classification. Two data sets from UCI Machine Learning Repository, i.e., Liver Disorders Data Set and Breast Cancer Wisconsin (Diagnosis), are employed for benchmark test. Initially a case-based reasoning method is applied to preprocess the data set thus a weight vector for each feature is derived. A particle swarm optimization model is then applied to construct a decision-making system for diseases identified. The PSO algorithm starts by partitioning the data set into a relatively large number of clusters to reduce the effects of initial conditions and then reducing the number of clusters into two. The average forecasting accuracy for breast cancer of CBRPSO model is 97.4% and for liver disorders is 76.8%. The proposed case-based particle swarm optimization model is able to produce more accurate and comprehensible results for medical experts in medical diagnosis.

THE EXTRACTION AND COMPARISON OF KNOWLEDGE FROM LOCAL FUNCTION NETWORKS

Extracting rules from RBFs is not a trivial task because of nonlinear functions or high input dimensionality. In such cases, some of the hidden units of the RBF network have a tendency to be "shared" across several output classes or even may not contribute to any output class. To address this we have developed an algorithm called LREX (for Local Rule EXtraction) which tackles these issues by extracting rules at two levels: hREX extracts rules by examining the hidden unit to class assignments while mREX extracts rules based on the input space to output space mappings. The rules extracted by our algorithm are compared and contrasted against a competing local rule extraction system. The central claim of this paper is that local function networks such as radial basis function (RBF) networks have a suitable architecture based on Gaussian functions that is amenable to rule extraction.

F. DE CARVALHO ANDRÉCP, MONARD MARIAC. AN APPROACH TO EXPLAIN NEURAL NETWORKS USING SYMBOLIC ALGORITHMS

Although Artificial Neural Networks have been satisfactorily employed in several problems, such as clustering, pattern recognition, dynamic systems control and prediction, they still suffer from significant limitations. One of them is that the induced concept representation is not usually comprehensible to humans. Several techniques have been suggested to extract meaningful knowledge from trained networks. This paper proposes the use of symbolic learning algorithms, commonly used by the Machine Learning community, such as C4.5, C4.5rules and CN2, to extract symbolic representations from trained networks. The approach proposed is similar to that used by the Trepan algorithm, which extracts symbolic representations, expressed as decision trees, from trained networks. Experimental results are presented and discussed in order to compare the knowledge extracted from Artificial Neural Networks using the proposed approach and the Trepan approach. Results are compared regarding two aspects: fidelity and comprehensibility.

A new formulation for feedforward neural networks

Feedforward neural network is one of the most commonly used function approximation techniques and has been applied to a wide variety of problems arising from various disciplines. However, neural networks are black-box models having multiple challenges/difficulties associated with training and generalization. This paper initially looks into the internal behavior of neural networks and develops a detailed interpretation of the neural network functional geometry. Based on this geometrical interpretation, a new set of variables describing neural networks is proposed as a more effective and geometrically interpretable alternative to the traditional set of network weights and biases. Then, this paper develops a new formulation for neural networks with respect to the newly defined variables; this reformulated neural network (ReNN) is equivalent to the common feedforward neural network but has a less complex error response surface. To demonstrate the learning ability of ReNN, in this paper, two training methods involving a derivative-based (a variation of backpropagation) and a derivative-free optimization algorithms are employed. Moreover, a new measure of regularization on the basis of the developed geometrical interpretation is proposed to evaluate and improve the generalization ability of neural networks. The value of the proposed geometrical interpretation, the ReNN approach, and the new regularization measure are demonstrated across multiple test problems. Results show that ReNN can be trained more effectively and efficiently compared to the common neural networks and the proposed regularization measure is an effective indicator of how a network would perform in terms of generalization.

A new data mining scheme using artificial neural networks

Classification is one of the data mining problems receiving enormous attention in the database community. Although artificial neural networks (ANNs) have been successfully applied in a wide range of machine learning applications, they are however often regarded as black boxes, i.e., their predictions cannot be explained. To enhance the explanation of ANNs, a novel algorithm to extract symbolic rules from ANNs has been proposed in this paper. ANN methods have not been effectively utilized for data mining tasks because how the classifications were made is not explicitly stated as symbolic rules that are suitable for verification or interpretation by human experts. With the proposed approach, concise symbolic rules with high accuracy, that are easily explainable, can be extracted from the trained ANNs. Extracted rules are comparable with other methods in terms of number of rules, average number of conditions for a rule, and the accuracy. The effectiveness of the proposed approach is clearly demonstrated by the experimental results on a set of benchmark data mining classification problems.

Extracting rules from neural networks as decision diagrams

Rule extraction from neural networks (NNs) solves two fundamental problems: it gives insight into the logic behind the network and in many cases, it improves the network's ability to generalize the acquired knowledge. This paper presents a novel eclectic approach to rule extraction from NNs, named LOcal Rule Extraction (LORE), suited for multilayer perceptron networks with discrete (logical or categorical) inputs. The extracted rules mimic network behavior on the training set and relax this condition on the remaining input space. First, a multilayer perceptron network is trained under standard regime. It is then transformed into an equivalent form, returning the same numerical result as the original network, yet being able to produce rules generalizing the network output for cases similar to a given input. The partial rules extracted for every training set sample are then merged to form a decision diagram (DD) from which logic rules can be extracted. A rule format explicitly separating subsets of inputs for which an answer is known from those with an undetermined answer is presented. A special data structure, the decision diagram, allowing efficient partial rule merging is introduced. With regard to rules' complexity and generalization abilities, LORE gives results comparable to those reported previously. An algorithm transforming DDs into interpretable boolean expressions is described. Experimental running times of rule extraction are proportional to the network's training time.

On the importance of comprehensible classification models for protein function prediction

The literature on protein function prediction is currently dominated by works aimed at maximizing predictive accuracy, ignoring the important issues of validation and interpretation of discovered knowledge, which can lead to new insights and hypotheses that are biologically meaningful and advance the understanding of protein functions by biologists. The overall goal of this paper is to critically evaluate this approach, offering a refreshing new perspective on this issue, focusing not only on predictive accuracy but also on the comprehensibility of the induced protein function prediction models. More specifically, this paper aims to offer two main contributions to the area of protein function prediction. First, it presents the case for discovering comprehensible protein function prediction models from data, discussing in detail the advantages of such models, namely, increasing the confidence of the biologist in the system's predictions, leading to new insights about the data and the formulation of new biological hypotheses, and detecting errors in the data. Second, it presents a critical review of the pros and cons of several different knowledge representations that can be used in order to support the discovery of comprehensible protein function prediction models.