Inductive biases in theory-based reinforcement learning

Understanding the inductive biases that allow humans to learn in complex environments has been an important goal of cognitive science. Yet, while we have discovered much about human biases in specific learning domains, much of this research has focused on simple tasks that lack the complexity of the real world. In contrast, video games involving agents and objects embedded in richly structured systems provide an experimentally tractable proxy for real-world complexity. Recent work has suggested that key aspects of human learning in domains like video games can be captured by model-based reinforcement learning (RL) with object-oriented relational models-what we term theory-based RL. Restricting the model class in this way provides an inductive bias that dramatically increases learning efficiency, but in this paper we show that humans employ a stronger set of biases in addition to syntactic constraints on the structure of theories. In particular, we catalog a set of semantic biases that constrain the content of theories. Building these semantic biases into a theory-based RL system produces more human-like learning in video game environments.

Development and training of a machine learning algorithm to identify patients at risk for recurrence following an arthroscopic Bankart repair (CLEARER): protocol for a retrospective, multicentre, cohort study

INTRODUCTION

Shoulder instability is a common injury, with a reported incidence of 23.9 per 100 000 person-years. There is still an ongoing debate on the most effective treatment strategy. Non-operative treatment has recurrence rates of up to 60%, whereas operative treatments such as the Bankart repair and bone block procedures show lower recurrence rates (16% and 2%, respectively) but higher complication rates (<2% and up to 30%, respectively). Methods to determine risk of recurrence have been developed; however, patient-specific decision-making tools are still lacking. Artificial intelligence and machine learning algorithms use self-learning complex models that can be used to make patient-specific decision-making tools. The aim of the current study is to develop and train a machine learning algorithm to create a prediction model to be used in clinical practice-as an online prediction tool-to estimate recurrence rates following a Bankart repair.

METHODS AND ANALYSIS

This is a multicentre retrospective cohort study. Patients with traumatic anterior shoulder dislocations that were treated with an arthroscopic Bankart repair without remplissage will be included. This study includes two parts. Part 1, collecting all potential factors influencing the recurrence rate following an arthroscopic Bankart repair in patients using multicentre data, aiming to include data from >1000 patients worldwide. Part 2, the multicentre data will be re-evaluated (and where applicable complemented) using machine learning algorithms to predict outcomes. Recurrence will be the primary outcome measure.

ETHICS AND DISSEMINATION

For safe multicentre data exchange and analysis, our Machine Learning Consortium adhered to the WHO regulation 'Policy on Use and Sharing of Data Collected by WHO in Member States Outside the Context of Public Health Emergencies'. The study results will be disseminated through publication in a peer-reviewed journal. No Institutional Review Board is required for this study.

Comparing gradient boosting machine and Bayesian threshold BLUP for genome-based prediction of categorical traits in wheat breeding

Genomic selection (GS) is a predictive methodology that is changing plant breeding. Genomic selection trains a statistical machine-learning model using available phenotypic and genotypic data with which predictions are performed for individuals that were only genotyped. For this reason, some statistical machine-learning methods are being implemented in GS, but in order to improve the selection of new genotypes early in the prediction process, the exploration of new statistical machine-learning algorithms must continue. In this paper, we performed a benchmarking study between the Bayesian threshold genomic best linear unbiased predictor model (TGBLUP; popular in GS) and the gradient boosting machine (GBM). This comparison was done using four real wheat (Triticum aestivum L.) data sets with categorical traits measured in terms of two metrics: the proportion of cases correctly classified (PCCC) and the Kappa coefficient in the testing set. Under 10 random partitions with four different sizes of testing proportions (20, 40, 60, and 80%), we compared the two algorithms and found that in three of the four data sets, the GBM outperformed the TGBLUP model in terms of both metrics (PCCC and Kappa coefficient). In the larger data sets (Data Sets 3 and 4), the gain in terms of prediction accuracy of the GBM was considerably significant. For this reason, we encourage more research using the GBM in GS to evaluate its virtues in terms of prediction performance in the context of GS.

Meta-Learning in Neural Networks: A Survey

The field of meta-learning, or learning-to-learn, has seen a dramatic rise in interest in recent years. Contrary to conventional approaches to AI where tasks are solved from scratch using a fixed learning algorithm, meta-learning aims to improve the learning algorithm itself, given the experience of multiple learning episodes. This paradigm provides an opportunity to tackle many conventional challenges of deep learning, including data and computation bottlenecks, as well as generalization. This survey describes the contemporary meta-learning landscape. We first discuss definitions of meta-learning and position it with respect to related fields, such as transfer learning and hyperparameter optimization. We then propose a new taxonomy that provides a more comprehensive breakdown of the space of meta-learning methods today. We survey promising applications and successes of meta-learning such as few-shot learning and reinforcement learning. Finally, we discuss outstanding challenges and promising areas for future research.

Identification of key predictors of hospital mortality in critically ill patients with embolic stroke using machine learning

Embolic stroke (ES) is characterized by high morbidity and mortality. Its mortality predictors remain unclear. The present study aimed to use machine learning (ML) to identify the key predictors of mortality for ES patients in the intensive care unit (ICU). Data were extracted from two large ICU databases: Medical Information Mart for Intensive Care (MIMIC)-IV for training and internal validation, and eICU Collaborative Research Database (eICU-CRD) for external validation. We developed predictive models of ES mortality based on 15 ML algorithms. We relied on the synthetic minority oversampling technique (SMOTE) to address class imbalance. Our main performance metric was area under the receiver operating characteristic (AUROC). We adopted recursive feature elimination (RFE) for feature selection. We assessed model performance using three disease-severity scoring systems as benchmarks. Of the 1566 and 207 ES patients enrolled in the two databases, there were 173 (15.70%), 73 (15.57%), and 36 (17.39%) hospital mortality in the training, internal validation, and external validation cohort, respectively. The random forest (RF) model had the largest AUROC (0.806) in the internal validation phase and was chosen as the best model. The AUROC of the RF compact (RF-COM) model containing the top six features identified by RFE was 0.795. In the external validation phase, the AUROC of the RF model was 0.838, and the RF-COM model was 0.830, outperforming other models. Our findings suggest that the RF model was the best model and the top six predictors of ES hospital mortality were Glasgow Coma Scale, white blood cell, blood urea nitrogen, bicarbonate, age, and mechanical ventilation.

Ranking community detection algorithms for complex social networks using multilayer network design approach

Purpose

Community detection is a significant research field in the study of social networks and analysis because of its tremendous applicability in multiple domains such as recommendation systems, link prediction and information diffusion. The majority of the present community detection methods considers either node information only or edge information only, but not both, which can result in loss of important information regarding network structures. In real-world social networks such as Facebook and Twitter, there are many heterogeneous aspects of the entities that connect them together such as different type of interactions occurring, which are difficult to study with the help of homogeneous network structures. The purpose of this study is to explore multilayer network design to capture these heterogeneous aspects by combining different modalities of interactions in single network.

Design/methodology/approach

In this work, multilayer network model is designed while taking into account node information as well as edge information. Existing community detection algorithms are applied on the designed multilayer network to find the densely connected nodes. Community scoring functions and partition comparison are used to further analyze the community structures. In addition to this, analytic hierarchical processing-technique for order preference by similarity to ideal solution (AHP-TOPSIS)-based framework is proposed for selection of an optimal community detection algorithm.

Findings

In the absence of reliable ground-truth communities, it becomes hard to perform evaluation of generated network communities. To overcome this problem, in this paper, various community scoring functions are computed and studied for different community detection methods.

Research limitations/implications

In this study, evaluation criteria are considered to be independent. The authors observed that the criteria used are having some interdependencies, which could not be captured by the AHP method. Therefore, in future, analytic network process may be explored to capture these interdependencies among the decision attributes.

Practical implications

Proposed ranking can be used to improve the search strategy of algorithms to decrease the search time of the best fitting one according to the case study. The suggested study ranks existing community detection algorithms to find the most appropriate one.

Social implications

Community detection is useful in many applications such as recommendation systems, health care, politics, economics, e-commerce, social media and communication network.

Originality/value

Ranking of the community detection algorithms is performed using community scoring functions as well as AHP-TOPSIS methods.

Explainable online ensemble of deep neural network pruning for time series forecasting

Both the complex and evolving nature of time series data make forecasting among one of the most challenging tasks in machine learning. Typical methods for forecasting are designed to model time-evolving dependencies between data observations. However, it is generally accepted that none of them are universally valid for every application. Therefore, methods for learning heterogeneous ensembles by combining a diverse set of forecasters together appears as a promising solution to tackle this task. While several approaches in the context of time series forecasting have focused on how to combine individual models in an ensemble, ranging from simple and enhanced averaging tactics to applying meta-learning methods, few works have tackled the task of ensemble pruning, i.e. individual model selection to take part in the ensemble. In addition, in classical ML literature, ensemble pruning techniques are mostly restricted to operate in a static manner. To deal with changes in the relative performance of models as well as changes in the data distribution, we employ gradient-based saliency maps for online ensemble pruning of deep neural networks. This method consists of generating individual models’ performance saliency maps that are subsequently used to prune the ensemble by taking into account both aspects of accuracy and diversity. In addition, the saliency maps can be exploited to provide suitable explanations for the reason behind selecting specific models to construct an ensemble that plays the role of a forecaster at a certain time interval or instant. An extensive empirical study on many real-world datasets demonstrates that our method achieves excellent or on par results in comparison to the state-of-the-art approaches as well as several baselines. Our code is available on Github (https://github.com/MatthiasJakobs/os-pgsm/tree/ecml_journal_2022).

TCMI: a non-parametric mutual-dependence estimator for multivariate continuous distributions

The identification of relevant features, i.e., the driving variables that determine a process or the properties of a system, is an essential part of the analysis of data sets with a large number of variables. A mathematical rigorous approach to quantifying the relevance of these features is mutual information. Mutual information determines the relevance of features in terms of their joint mutual dependence to the property of interest. However, mutual information requires as input probability distributions, which cannot be reliably estimated from continuous distributions such as physical quantities like lengths or energies. Here, we introduce total cumulative mutual information (TCMI), a measure of the relevance of mutual dependences that extends mutual information to random variables of continuous distribution based on cumulative probability distributions. TCMI is a non-parametric, robust, and deterministic measure that facilitates comparisons and rankings between feature sets with different cardinality. The ranking induced by TCMI allows for feature selection, i.e., the identification of variable sets that are nonlinear statistically related to a property of interest, taking into account the number of data samples as well as the cardinality of the set of variables. We evaluate the performance of our measure with simulated data, compare its performance with similar multivariate-dependence measures, and demonstrate the effectiveness of our feature-selection method on a set of standard data sets and a typical scenario in materials science.

Doppelgänger Spotting in Biomedical Gene Expression Data

Doppelgänger effects (DEs) occur when samples exhibit chance similarities such that, when split across training and validation sets, inflates the trained machine learning (ML) model performance. This inflationary effect causes misleading confidence on the deployability of the model. Thus, so far, there are no tools for doppelgänger identification or standard practices to manage their confounding implications. We present doppelgangerIdentifier, a software suite for doppelgänger identification and verification. Applying doppelgangerIdentifier across a multitude of diseases and data types, we show the pervasive nature of DEs in biomedical gene expression data. We also provide guidelines toward proper doppelgänger identification by exploring the ramifications of lingering batch effects from batch imbalances on the sensitivity of our doppelgänger identification algorithm. We suggest doppelgänger verification as a useful procedure to establish baselines for model evaluation that may inform on whether feature selection and ML on the data set may yield meaningful insights.

•

Doppelgänger effects inflate the machine learning performance

•

Doppelgänger effects exist in RNA-Seq and microarray gene expression data

•

Developed doppelgangerIdentifier, a software to identify and verify doppelgängers

•

Provide guidelines for proper doppelgänger identification

Machine learning methods to predict attrition in a population-based cohort of very preterm infants

The timely identification of cohort participants at higher risk for attrition is important to earlier interventions and efficient use of research resources. Machine learning may have advantages over the conventional approaches to improve discrimination by analysing complex interactions among predictors. We developed predictive models of attrition applying a conventional regression model and different machine learning methods. A total of 542 very preterm (< 32 gestational weeks) infants born in Portugal as part of the European Effective Perinatal Intensive Care in Europe (EPICE) cohort were included. We tested a model with a fixed number of predictors (Baseline) and a second with a dynamic number of variables added from each follow-up (Incremental). Eight classification methods were applied: AdaBoost, Artificial Neural Networks, Functional Trees, J48, J48Consolidated, K-Nearest Neighbours, Random Forest and Logistic Regression. Performance was compared using AUC- PR (Area Under the Curve—Precision Recall), Accuracy, Sensitivity and F-measure. Attrition at the four follow-ups were, respectively: 16%, 25%, 13% and 17%. Both models demonstrated good predictive performance, AUC-PR ranging between 69 and 94.1 in Baseline and from 72.5 to 97.1 in Incremental model. Of the whole set of methods, Random Forest presented the best performance at all follow-ups [AUC-PR₁: 94.1 (2.0); AUC-PR₂: 91.2 (1.2); AUC-PR₃: 97.1 (1.0); AUC-PR₄: 96.5 (1.7)]. Logistic Regression performed well below Random Forest. The top-ranked predictors were common for both models in all follow-ups: birthweight, gestational age, maternal age, and length of hospital stay. Random Forest presented the highest capacity for prediction and provided interpretable predictors. Researchers involved in cohorts can benefit from our robust models to prepare for and prevent loss to follow-up by directing efforts toward individuals at higher risk.

Applied Machine Learning for Spine Surgeons: Predicting Outcome for Patients Undergoing Treatment for Lumbar Disc Herniation Using PRO Data

STUDY DESIGN

Retrospective/prospective study.

OBJECTIVE

Models based on preoperative factors can predict patients' outcome at 1-year follow-up. This study measures the performance of several machine learning (ML) models and compares the results with conventional methods.

METHODS

Inclusion criteria were patients who had lumbar disc herniation (LDH) surgery, identified in the Danish national registry for spine surgery. Initial training of models included 16 independent variables, including demographics and presurgical patient-reported measures. Patients were grouped by reaching minimal clinically important difference or not for EuroQol, Oswestry Disability Index, Visual Analog Scale (VAS) Leg, and VAS Back and by their ability to return to work at 1 year follow-up. Data were randomly split into training, validation, and test sets by 50%/35%/15%. Deep learning, decision trees, random forest, boosted trees, and support vector machines model were trained, and for comparison, multivariate adaptive regression splines (MARS) and logistic regression models were used. Model fit was evaluated by inspecting area under the curve curves and performance during validation.

RESULTS

Seven models were arrived at. Classification errors were within ±1% to 4% SD across validation folds. ML did not yield superior performance compared with conventional models. MARS and deep learning performed consistently well. Discrepancy was greatest among VAS Leg models.

CONCLUSIONS

Five predictive ML and 2 conventional models were developed, predicting improvement for LDH patients at the 1-year follow-up. We demonstrate that it is possible to build an ensemble of models with little effort as a starting point for further model optimization and selection.

Rational arbitration between statistics and rules in human sequence processing

Detecting and learning temporal regularities is essential to accurately predict the future. A long-standing debate in cognitive science concerns the existence in humans of a dissociation between two systems, one for handling statistical regularities governing the probabilities of individual items and their transitions, and another for handling deterministic rules. Here, to address this issue, we used finger tracking to continuously monitor the online build-up of evidence, confidence, false alarms and changes-of-mind during sequence processing. All these aspects of behaviour conformed tightly to a hierarchical Bayesian inference model with distinct hypothesis spaces for statistics and rules, yet linked by a single probabilistic currency. Alternative models based either on a single statistical mechanism or on two non-commensurable systems were rejected. Our results indicate that a hierarchical Bayesian inference mechanism, capable of operating over distinct hypothesis spaces for statistics and rules, underlies the human capability for sequence processing.

Harnessing feature extraction capacities from a pre-trained convolutional neural network (VGG-16) for the unsupervised distinction of aortic outflow velocity profiles in patients with severe aortic stenosis

Aims

Hypothesizing that aortic outflow velocity profiles contain more valuable information about aortic valve obstruction and left ventricular contractility than can be captured by the human eye, features of the complex geometry of Doppler tracings from patients with severe aortic stenosis (AS) were extracted by a convolutional neural network (CNN).

Methods and results

After pre-training a CNN (VGG-16) on a large data set (ImageNet data set; 14 million images belonging to 1000 classes), the convolutional part was employed to transform Doppler tracings to 1D arrays. Among 366 eligible patients [age: 79.8 ± 6.77 years; 146 (39.9%) women] with pre-procedural echocardiography and right heart catheterization prior to transcatheter aortic valve replacement (TAVR), good quality Doppler tracings from 101 patients were analysed. The convolutional part of the pre-trained VGG-16 model in conjunction with principal component analysis and k-means clustering distinguished two shapes of aortic outflow velocity profiles. Kaplan-Meier analysis revealed that mortality in patients from Cluster 2 (n = 40, 39.6%) was significantly increased [hazard ratio (HR) for 2-year mortality: 3; 95% confidence interval (CI): 1-8.9]. Apart from reduced cardiac output and mean aortic valve gradient, patients from Cluster 2 were also characterized by signs of pulmonary hypertension, impaired right ventricular function, and right atrial enlargement. After training an extreme gradient boosting algorithm on these 101 patients, validation on the remaining 265 patients confirmed that patients assigned to Cluster 2 show increased mortality (HR for 2-year mortality: 2.6; 95% CI: 1.4-5.1, P-value: 0.004).

Conclusion

Transfer learning enables sophisticated pattern recognition even in clinical data sets of limited size. Importantly, it is the left ventricular compensation capacity in the face of increased afterload, and not so much the actual obstruction of the aortic valve, that determines fate after TAVR.

Predicting How Well Adolescents Get Along with Peers and Teachers: A Machine Learning Approach

How well adolescents get along with others such as peers and teachers is an important aspect of adolescent development. Current research on adolescent relationship with peers and teachers is limited by classical methods that lack explicit test of predictive performance and cannot efficiently discover complex associations with potential non-linearity and higher-order interactions among a large set of predictors. Here, a transparently reported machine learning approach is utilized to overcome these limitations in concurrently predicting how well adolescents perceive themselves to get along with peers and teachers. The predictors were 99 items from four instruments examining internalizing and externalizing psychopathology, sensation-seeking, peer pressure, and parent-child conflict. The sample consisted of 3232 adolescents (M = 14.0 years, SD = 1.0 year, 49% female). Nonlinear machine learning classifiers predicted with high performance adolescent relationship with peers and teachers unlike classical methods. Using model explainability analyses at the item level, results identified influential predictors related to somatic complaints and attention problems that interacted in nonlinear ways with internalizing behaviors. In many cases, these intrapersonal predictors outcompeted in predictive power many interpersonal predictors. Overall, the results suggest the need to cast a much wider net of variables for understanding and predicting adolescent relationships, and highlight the power of a data-driven machine learning approach with implications on a predictive science of adolescence research.

Simple Models in Complex Worlds: Occam’s Razor and Statistical Learning Theory

The idea that “simplicity is a sign of truth”, and the related “Occam’s razor” principle, stating that, all other things being equal, simpler models should be preferred to more complex ones, have been long discussed in philosophy and science. We explore these ideas in the context of supervised machine learning, namely the branch of artificial intelligence that studies algorithms which balance simplicity and accuracy in order to effectively learn about the features of the underlying domain. Focusing on statistical learning theory, we show that situations exist for which a preference for simpler models (as modeled through the addition of a regularization term in the learning problem) provably slows down, instead of favoring, the supervised learning process. Our results shed new light on the relations between simplicity and truth approximation, which are briefly discussed in the context of both machine learning and the philosophy of science.

Prediction of treatment outcome in clinical trials under a personalized medicine perspective

A central problem in most data-driven personalized medicine scenarios is the estimation of heterogeneous treatment effects to stratify individuals into subpopulations that differ in their susceptibility to a particular disease or response to a specific treatment. In this work, with an illustrative example on type 2 diabetes we showed how the increasing ability to access and analyzed open data from randomized clinical trials (RCTs) allows to build Machine Learning applications in a framework of personalized medicine. An ensemble machine learning predictive model is first developed and then applied to estimate the expected treatment response according to the medication that would be prescribed. Machine learning is quickly becoming indispensable to bridge science and clinical practice, but it is not sufficient on its own. A collaborative effort is requested to clinicians, statisticians, and computer scientists to strengthen tools built on machine learning to take advantage of this evidence flow.

Propension to customer churn in a financial institution: a machine learning approach

This paper examines churn prediction of customers in the banking sector using a unique customer-level dataset from a large Brazilian bank. Our main contribution is in exploring this rich dataset, which contains prior client behavior traits that enable us to document new insights into the main determinants predicting future client churn. We conduct a horserace of many supervised machine learning algorithms under the same cross-validation and evaluation setup, enabling a fair comparison across algorithms. We find that the random forests technique outperforms decision trees, k-nearest neighbors, elastic net, logistic regression, and support vector machines models in several metrics. Our investigation reveals that customers with a stronger relationship with the institution, who have more products and services, who borrow more from the bank, are less likely to close their checking accounts. Using a back-of-the-envelope estimation, we find that our model has the potential to forecast potential losses of up to 10% of the operating result reported by the largest Brazilian banks in 2019, suggesting the model has a significant economic impact. Our results corroborate the importance of investing in cross-selling and upselling strategies focused on their current customers. These strategies can have positive side effects on customer retention.

Automatic detection of passing and shooting in water polo using machine learning: a feasibility study

There is currently no efficient way to quantify overhead throwing volume in water polo. Therefore, this study aimed to test the feasibility of a method to detect passes and shots in water polo automatically using inertial measurement units (IMU) and machine-learning algorithms. Eight water polo players wore one IMU sensor on the wrist (dominant hand) and one on the sacrum during six practices each. Sessions were filmed with a video camera and manually tagged for individual shots or passes. Data were synchronised between video tagging and IMU sensors using a cross-correlation approach. Support vector machine (SVM) and artificial neural networks (ANN) were compared based on sensitivity and specificity for identifying shots and passes. A total of 7294 actions were identified during the training sessions, including 945 shots and 5361 passes. Using SVM, passes and shots together were identified with 94.4% (95%CI = 91.8-96.4) sensitivity and 93.6% (95%CI = 91.4-95.4) specificity. Using ANN yielded similar sensitivity (93.0% [95%CI = 90.1-95.1]) and specificity (93.4% [95%CI = 91.1 = 95.2]). The results suggest that this method of identifying overhead throwing motions with IMU has potential for future field applications. A set-up with one single sensor at the wrist can suffice to measure these activities in water polo.

Solving the Pulmonary Hypertension Paradox in Patients With Severe Tricuspid Regurgitation by Employing Artificial Intelligence

OBJECTIVES

This study aimed to improve echocardiographic assessment of pulmonary hypertension (PH) in patients presenting with severe tricuspid regurgitation (TR).

BACKGROUND

Echocardiographic assessment of PH in patients with severe TR carries several pitfalls for underestimation, hence concealing the true severity of PH in very sick patients in particular, and ultimately obscuring the impact of PH on survival after transcatheter tricuspid valve intervention (TTVI).

METHODS

All patients in this study underwent TTVI for severe TR between 2016 and 2020. To predict the mean pulmonary artery pressure (mPAP) solely based on echocardiographic parameters, we trained an extreme gradient boosting (XGB) algorithm. The derivation cohort was constituted by 116 out of 162 patients with both echocardiography and right heart catheterization data, preprocedurally obtained, from a bicentric registry. Moreover, 142 patients from an independent institution served for external validation.

RESULTS

Systolic pulmonary artery pressure was consistently underestimated by echocardiography in comparison to right heart catheterization (40.3 ± 15.9 mm Hg vs 44.1 ± 12.9 mm Hg; P = 0.0066), and the assessment was most discrepant among patients with severe defects of the tricuspid valve and impaired right ventricular systolic function. Using 9 echocardiographic parameters as input variables, an XGB algorithm could reliably predict mPAP levels (R = 0.96, P < 2.2 × 10^-16). Moreover, patients with elevations in predicted mPAP levels ≥29.9 mm Hg showed significantly reduced 2-year survival after TTVI (58.3% [95% CI: 41.7%-81.6%] vs 78.8% [95% CI: 68.7%-90.5%]; P = 0.026). Importantly, the poor prognosis associated with elevation in predicted mPAP levels was externally confirmed (HR for 2-year mortality: 2.9 [95% CI: 1.5-5.7]; P = 0.002).

CONCLUSIONS

PH in patients with severe TR can be reliably assessed based on echocardiographic parameters in conjunction with an XGB algorithm, and elevations in predicted mPAP levels translate into increased mortality after TTVI.

A Comparative Assessment of Machine Learning Models for Landslide Susceptibility Mapping in the Rugged Terrain of Northern Pakistan

This study investigated the performances of different techniques, including random forest (RF), support vector machine (SVM), maximum entropy (maxENT), gradient-boosting machine (GBM), and logistic regression (LR), for landslide susceptibility mapping (LSM) in the rugged terrain of northern Pakistan. Initially, a landslide inventory of 200 samples was produced along with an additional 200 samples indicating nonlandslide areas and divided into training (70%) and validation (30%) groups using a stratified loop-based random sampling approach. Then, a geospatial database of 12 possible landslide influencing factors (LIFs) was generated, including elevation, slope, aspect, topographic wetness index (TWI), topographic position index (TPI), distance to drainage, distance to fault, distance to road, normalized difference vegetation index (NDVI), rainfall, land cover/land use (LCLU), and a geological map of the study area. None of the LIFs were redundant for the modeling, as indicated by the multicollinearity test (tolerance > 0.1) and information gain ratio (IGR > 0). We extended the evaluation measures of each algorithm from area-under-the-curve (AUC) analysis to the calculation of performance overall (POA) with the help of precision, recall, F1 score, accuracy (ACC), and Matthew’s correlation coefficient (MCC). The results showed that the SVM was the most promising model (AUC = 0.969, POA = 2669) for the LSM, followed by RF (AUC = 0.967, POA = 2656), GBM (AUC = 0.967, POA = 2623), maxENT (AUC = 0.872, POA = 1761), and LR (AUC = 0.836, POA = 1299). It is important to note that the SVM, RF, and GBM were the top performers, with almost similar accuracy. Thus, each of these could be equally effective for LSM and can be used for risk reduction and mitigation measures in the rugged terrain of Pakistan and other regions with similar topography.

Ranks underlie outcome of combining classifiers: Quantitative roles for diversity and accuracy

Combining classifier systems potentially improves predictive accuracy, but outcomes have proven impossible to predict. Classification most commonly improves when the classifiers are "sufficiently good" (generalized as " accuracy ") and "sufficiently different" (generalized as " diversity "), but the individual and joint quantitative influence of these factors on the final outcome remains unknown. We resolve these issues. Beginning with simulated data, we develop the DIRAC framework (DIversity of Ranks and ACcuracy), which accurately predicts outcome of both score-based fusions originating from exponentially modified Gaussian distributions and rank-based fusions, which are inherently distribution independent. DIRAC was validated using biological dual-energy X-ray absorption and magnetic resonance imaging data. The DIRAC framework is domain independent and has expected utility in far-ranging areas such as clinical biomarker development/personalized medicine, clinical trial enrollment, insurance pricing, portfolio management, and sensor optimization.

Harnessing multimodal data integration to advance precision oncology

Advances in quantitative biomarker development have accelerated new forms of data-driven insights for patients with cancer. However, most approaches are limited to a single mode of data, leaving integrated approaches across modalities relatively underdeveloped. Multimodal integration of advanced molecular diagnostics, radiological and histological imaging, and codified clinical data presents opportunities to advance precision oncology beyond genomics and standard molecular techniques. However, most medical datasets are still too sparse to be useful for the training of modern machine learning techniques, and significant challenges remain before this is remedied. Combined efforts of data engineering, computational methods for analysis of heterogeneous data and instantiation of synergistic data models in biomedical research are required for success. In this Perspective, we offer our opinions on synthesizing complementary modalities of data with emerging multimodal artificial intelligence methods. Advancing along this direction will result in a reimagined class of multimodal biomarkers to propel the field of precision oncology in the coming decade.

Energy Efficiency of Inference Algorithms for Clinical Laboratory Data Sets: Green Artificial Intelligence Study

Background

The use of artificial intelligence (AI) in the medical domain has attracted considerable research interest. Inference applications in the medical domain require energy-efficient AI models. In contrast to other types of data in visual AI, data from medical laboratories usually comprise features with strong signals. Numerous energy optimization techniques have been developed to relieve the burden on the hardware required to deploy a complex learning model. However, the energy efficiency levels of different AI models used for medical applications have not been studied.

Objective

The aim of this study was to explore and compare the energy efficiency levels of commonly used machine learning algorithms—logistic regression (LR), k-nearest neighbor, support vector machine, random forest (RF), and extreme gradient boosting (XGB) algorithms, as well as four different variants of neural network (NN) algorithms—when applied to clinical laboratory datasets.

Methods

We applied the aforementioned algorithms to two distinct clinical laboratory data sets: a mass spectrometry data set regarding Staphylococcus aureus for predicting methicillin resistance (3338 cases; 268 features) and a urinalysis data set for predicting Trichomonas vaginalis infection (839,164 cases; 9 features). We compared the performance of the nine inference algorithms in terms of accuracy, area under the receiver operating characteristic curve (AUROC), time consumption, and power consumption. The time and power consumption levels were determined using performance counter data from Intel Power Gadget 3.5.

Results

The experimental results indicated that the RF and XGB algorithms achieved the two highest AUROC values for both data sets (84.7% and 83.9%, respectively, for the mass spectrometry data set; 91.1% and 91.4%, respectively, for the urinalysis data set). The XGB and LR algorithms exhibited the shortest inference time for both data sets (0.47 milliseconds for both in the mass spectrometry data set; 0.39 and 0.47 milliseconds, respectively, for the urinalysis data set). Compared with the RF algorithm, the XGB and LR algorithms exhibited a 45% and 53%-60% reduction in inference time for the mass spectrometry and urinalysis data sets, respectively. In terms of energy efficiency, the XGB algorithm exhibited the lowest power consumption for the mass spectrometry data set (9.42 Watts) and the LR algorithm exhibited the lowest power consumption for the urinalysis data set (9.98 Watts). Compared with a five-hidden-layer NN, the XGB and LR algorithms achieved 16%-24% and 9%-13% lower power consumption levels for the mass spectrometry and urinalysis data sets, respectively. In all experiments, the XGB algorithm exhibited the best performance in terms of accuracy, run time, and energy efficiency.

Conclusions

The XGB algorithm achieved balanced performance levels in terms of AUROC, run time, and energy efficiency for the two clinical laboratory data sets. Considering the energy constraints in real-world scenarios, the XGB algorithm is ideal for medical AI applications.

Machine learning accurately predicts the multivariate performance phenotype from morphology in lizards

Completing the genotype-to-phenotype map requires rigorous measurement of the entire multivariate organismal phenotype. However, phenotyping on a large scale is not feasible for many kinds of traits, resulting in missing data that can also cause problems for comparative analyses and the assessment of evolutionary trends across species. Measuring the multivariate performance phenotype is especially logistically challenging, and our ability to predict several performance traits from a given morphology is consequently poor. We developed a machine learning model to accurately estimate multivariate performance data from morphology alone by training it on a dataset containing performance and morphology data from 68 lizard species. Our final, stacked model predicts missing performance data accurately at the level of the individual from simple morphological measures. This model performed exceptionally well, even for performance traits that were missing values for >90% of the sampled individuals. Furthermore, incorporating phylogeny did not improve model fit, indicating that the phenotypic data alone preserved sufficient information to predict the performance based on morphological information. This approach can both significantly increase our understanding of performance evolution and act as a bridge to incorporate performance into future work on phenomics.

BERT-Kcr: prediction of lysine crotonylation sites by a transfer learning method with pre-trained BERT models

MOTIVATION

As one of the most important post-translational modifications (PTMs), protein lysine crotonylation (Kcr) has attracted wide attention, which involves in important physiological activities, such as cell differentiation and metabolism. However, experimental methods are expensive and time-consuming for Kcr identification. Instead, computational methods can predict Kcr sites in silico with high efficiency and low cost.

RESULTS

In this study, we proposed a novel predictor, BERT-Kcr, for protein Kcr sites prediction, which was developed by using a transfer learning method with pre-trained bidirectional encoder representations from transformers (BERT) models. These models were originally used for natural language processing (NLP) tasks, such as sentence classification. Here, we transferred each amino acid into a word as the input information to the pre-trained BERT model. The features encoded by BERT were extracted and then fed to a BiLSTM network to build our final model. Compared with the models built by other machine learning and deep learning classifiers, BERT-Kcr achieved the best performance with AUROC of 0.983 for 10-fold cross validation. Further evaluation on the independent test set indicates that BERT-Kcr outperforms the state-of-the-art model Deep-Kcr with an improvement of about 5% for AUROC. The results of our experiment indicate that the direct use of sequence information and advanced pre-trained models of NLP could be an effective way for identifying PTM sites of proteins.

AVAILABILITY AND IMPLEMENTATION

The BERT-Kcr model is publicly available on http://zhulab.org.cn/BERT-Kcr_models/.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Training and evaluating machine learning algorithms for ocean microplastics classification through vibrational spectroscopy

Microplastics are contaminants of emerging concern - not only environmental, but also to human health. Characterizing them is of fundamental importance to evaluate their potential impacts and target specific actions aiming to reduce potential harming effects. This study extends the exploration of machine learning classification algorithms applied to FTIR spectra of microplastics collected at sea. A comparison of successful classification models was made in order to evaluate prediction performance for 13 classes of polymers. A rigorous methodology was applied using a pipeline scheme to avoid bias in the training and selection phases. The application of an oversampling technique also contributed by compensating unbalanceness in the dataset. The log-loss was used as the minimization function target and to assess performance. In our analysis, Support Vector Machine Classifier provides a good relationship between simplicity and performance, for a fast and useful automatic characterization of microplastics.

Dis-Cover AI Minds to Preserve Human Knowledge

Modern AI technologies make use of statistical learners that lead to self-empiricist logic, which, unlike human minds, use learned non-symbolic representations. Nevertheless, it seems that it is not the right way to progress in AI. The structure of symbols—the operations by which the intellectual solution is realized—and the search for strategic reference points evoke important issues in the analysis of AI. Studying how knowledge can be represented through methods of theoretical generalization and empirical observation is only the latest step in a long process of evolution. For many years, humans, seeing language as innate, have carried out symbolic theories. Everything seems to have skipped ahead with the advent of Machine Learning. In this paper, after a long analysis of history, the rule-based and the learning-based vision, we would investigate the syntax as possible meeting point between the different learning theories. Finally, we propose a new vision of knowledge in AI models based on a combination of rules, learning, and human knowledge.

A New Deep Learning Calibration Method Enhances Genome-Based Prediction of Continuous Crop Traits

Genomic selection (GS) has the potential to revolutionize predictive plant breeding. A reference population is phenotyped and genotyped to train a statistical model that is used to perform genome-enabled predictions of new individuals that were only genotyped. In this vein, deep neural networks, are a type of machine learning model and have been widely adopted for use in GS studies, as they are not parametric methods, making them more adept at capturing nonlinear patterns. However, the training process for deep neural networks is very challenging due to the numerous hyper-parameters that need to be tuned, especially when imperfect tuning can result in biased predictions. In this paper we propose a simple method for calibrating (adjusting) the prediction of continuous response variables resulting from deep learning applications. We evaluated the proposed deep learning calibration method (DL_M2) using four crop breeding data sets and its performance was compared with the standard deep learning method (DL_M1), as well as the standard genomic Best Linear Unbiased Predictor (GBLUP). While the GBLUP was the most accurate model overall, the proposed deep learning calibration method (DL_M2) helped increase the genome-enabled prediction performance in all data sets when compared with the traditional DL method (DL_M1). Taken together, we provide evidence for extending the use of the proposed calibration method to evaluate its potential and consistency for predicting performance in the context of GS applied to plant breeding.

Foundations of Machine Learning-Based Clinical Prediction Modeling: Part I-Introduction and General Principles

We provide explanations on the general principles of machine learning, as well as analytical steps required for successful machine learning-based predictive modeling, which is the focus of this series. In particular, we define the terms machine learning, artificial intelligence, as well as supervised and unsupervised learning, continuing by introducing optimization, thus, the minimization of an objective error function as the central dogma of machine learning. In addition, we discuss why it is important to separate predictive and explanatory modeling, and most importantly state that a prediction model should not be used to make inferences. Lastly, we broadly describe a classical workflow for training a machine learning model, starting with data pre-processing and feature engineering and selection, continuing on with a training structure consisting of a resampling method, hyperparameter tuning, and model selection, and ending with evaluation of model discrimination and calibration as well as robust internal or external validation of the fully developed model. Methodological rigor and clarity as well as understanding of the underlying reasoning of the internal workings of a machine learning approach are required, otherwise predictive applications despite being strong analytical tools are not well accepted into the clinical routine.

Sensor-based particle mass prediction of lightweight packaging waste using machine learning algorithms

Sensor-based material flow characterization (SBMC) promises to improve the performance of future-generation sorting plants by enabling new applications like automatic quality monitoring or process control. Prerequisite for this is the derivation of mass-based material flow characteristics from pixel-based sensor data, which requires known individual particle masses. Since particle masses cannot be measured inline, the prediction of particle masses of lightweight packaging (LWP) waste using machine learning (ML) algorithms is investigated. Five LWP material classes were sampled, preprocessed, and scanned on a custom-made test rig, resulting in a dataset containing 3D laser triangulation (3DLT) images, RGB images, and corresponding masses of n = 3,830 particles. Based on 66 extracted shape measurements, six ML models were trained for particle mass prediction (PMP). Their performance was compared with two state-of-the-art reference models using (i) material-specific mean particle masses and (ii) grammages. Obtained particle masses showed a high variation and significant differences between material classes and particle size classes. After feature selection, both reference models achieving R²-scores of (i) 0.422 ± 0.121 and (ii) 0.533 ± 0.224 were outperformed by all investigated ML models. A random forest regressor with an R²-score of 0.763 ± 0.091 and a normalized mean absolute error of 0.243 ± 0.050 achieved the most accurate PMP. In contrast to studies on primary raw materials, PMP of LWP waste is challenging due to influences of packaging design and post-consumer disposal behavior. ML algorithms are a promising approach for PMP that outperform state-of-the-art methods by 43% higher R²-scores.

The Impact of Meta-Induction: From Skepticism to Optimality

In the first section, five major attempts to solve the problem of induction and their failures are discussed. In the second section, an account of meta-induction is introduced. It offers a novel solution to the problem of induction, based on mathematical theorems about the predictive optimality of attractivity-weighted meta-induction. In the third section, how the a priori justification of meta-induction provides a non-circular a posteriori justification of object-induction, based on its superior track record, is explained. In the fourth section, four important extensions and refinements of the method of meta-induction are presented. The final section, summarizes the major impacts of the program of meta-induction for epistemology, the philosophy of science and cognitive science.

Facial Recognition Intensity in Disease Diagnosis Using Automatic Facial Recognition

Artificial intelligence (AI) technology is widely applied in different medical fields, including the diagnosis of various diseases on the basis of facial phenotypes, but there is no evaluation or quantitative synthesis regarding the performance of artificial intelligence. Here, for the first time, we summarized and quantitatively analyzed studies on the diagnosis of heterogeneous diseases on the basis on facial features. In pooled data from 20 systematically identified studies involving 7 single diseases and 12,557 subjects, quantitative random-effects models revealed a pooled sensitivity of 89% (95% CI 82% to 93%) and a pooled specificity of 92% (95% CI 87% to 95%). A new index, the facial recognition intensity (FRI), was established to describe the complexity of the association of diseases with facial phenotypes. Meta-regression revealed the important contribution of FRI to heterogeneous diagnostic accuracy (p = 0.021), and a similar result was found in subgroup analyses (p = 0.003). An appropriate increase in the training size and the use of deep learning models helped to improve the diagnostic accuracy for diseases with low FRI, although no statistically significant association was found between accuracy and photographic resolution, training size, AI architecture, and number of diseases. In addition, a novel hypothesis is proposed for universal rules in AI performance, providing a new idea that could be explored in other AI applications.

MarketBrain: An EEG based intelligent consumer preference prediction system

The traditional marketing research tools (Personal Depth Interview, Surveys, FGD, etc.) are cost-prohibitive and often criticized for not extracting true consumer preferences. Neuromarketing tools promise to overcome such limitations. In this study, we proposed a framework, MarketBrain, to predict consumer preferences. In our experiment, we administered marketing stimuli (five products with endorsements), collected EEG signals by EMOTIV EPOC+, and used signal processing and classification algorithms to develop the prediction system. Wavelet Packet Transform was used to extract frequency bands (δ, θ, α, β₁, β₂, γ) and then statistical features were extracted for classification. Among the classifiers, Support Vector Machine (SVM) achieved the best accuracy (96.01±0.71) using 5-fold cross-validation. Results also suggested that specific target consumers and endorser appearance affect the prediction of the preference. So, it is evident that EEG-based neuromarketing tools can help brands and businesses effectively predict future consumer preferences. Hence, it will lead to the development of an intelligent market driving system for neuromarketing applications.

The application of pangenomics and machine learning in genomic selection in plants

Genomic selection approaches have increased the speed of plant breeding, leading to growing crop yields over the last decade. However, climate change is impacting current and future yields, resulting in the need to further accelerate breeding efforts to cope with these changing conditions. Here we present approaches to accelerate plant breeding and incorporate nonadditive effects in genomic selection by applying state-of-the-art machine learning approaches. These approaches are made more powerful by the inclusion of pangenomes, which represent the entire genome content of a species. Understanding the strengths and limitations of machine learning methods, compared with more traditional genomic selection efforts, is paramount to the successful application of these methods in crop breeding. We describe examples of genomic selection and pangenome-based approaches in crop breeding, discuss machine learning-specific challenges, and highlight the potential for the application of machine learning in genomic selection. We believe that careful implementation of machine learning approaches will support crop improvement to help counter the adverse outcomes of climate change on crop production.