Transparent Machine Learning Research Articles

Beyond Predictive Learning Analytics Modelling and onto Explainable Artificial Intelligence with Prescriptive Analytics and ChatGPT

A significant body of recent research in the field of Learning Analytics has focused on leveraging machine learning approaches for predicting at-risk students in order to initiate timely interventions and thereby elevate retention and completion rates. The overarching feature of the majority of these research studies has been on the science of prediction only. The component of predictive analytics concerned with interpreting the internals of the models and explaining their predictions for individual cases to stakeholders has largely been neglected. Additionally, works that attempt to employ data-driven prescriptive analytics to automatically generate evidence-based remedial advice for at-risk learners are in their infancy. eXplainable AI is a field that has recently emerged providing cutting-edge tools which support transparent predictive analytics and techniques for generating tailored advice for at-risk students. This study proposes a novel framework that unifies both transparent machine learning as well as techniques for enabling prescriptive analytics, while integrating the latest advances in large language models for communicating the insights to learners. This work demonstrates a predictive modelling framework for identifying learners at risk of qualification non-completion based on a real-world dataset comprising ∼\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\sim $$\\end{document}7000 learners with their outcomes, covering 2018 - 2022. The study further demonstrates how predictive modelling can be augmented with prescriptive analytics on two case studies to generate human-readable prescriptive feedback for those who are at risk using ChatGPT.

Transparent machine learning suggests a key driver in the decision to startinsulin therapy in individuals with type 2 diabetes.

The objective of this study is to establish a predictive model using transparent machine learning (ML) to identify any drivers that characterize therapeutic inertia. Data in the form of both descriptive and dynamic variables collected from electronic records of 1.5 million patients seen at clinics within the Italian Association of Medical Diabetologists between 2005-2019 were analyzed using logic learning machine (LLM), a "clear box" ML technique. Data were subjected to a first stage of modeling to allow ML to automatically select the most relevant factors related to inertia, and then four further modeling steps individuated key variables that discriminated the presence or absence of inertia. The LLM model revealed a key role for average glycated hemoglobin (HbA1c) threshold values correlated with the presence or absence of insulin therapeutic inertia with an accuracy of 0.79. The model indicated that a patient's dynamic rather than static glycemic profile has a greater effect on therapeutic inertia. Specifically, the difference in HbA1c between two consecutive visits, what we call the HbA1c gap, plays a crucial role. Namely, insulin therapeutic inertia is correlated with an HbA1c gap of <6.6 mmol/mol (0.6%), but not with an HbA1c gap of >11 mmol/mol (1.0%). The results reveal, for the first time, the interrelationship between a patient's glycemic trend defined by sequential HbA1c measurements and timely or delayed initiation of insulin therapy. The results further demonstrate that LLM can provide insight in support of evidence-based medicine using real world data.

Transparent machine learning models for predicting decisions to undertake energy retrofits in residential buildings

Transparent machine learning models for predicting decisions to undertake energy retrofits in residential buildings

A transparent machine learning algorithm to manage diabetes: TDMSML

Diabetes is nowadays a very common medical problem among the people worldwide. The disease is becoming more prevalent with the modern and hectic lifestyle followed by people. As a result, designing an adequate medical expert system to assist physicians in treating the disease on time is critical. Expert systems are required to identify the major cause(s) of the disease, so that precautionary measures can be taken ahead of time. Several medical expert systems have already been proposed, but each has its own set of shortcomings, such as the use of trial and error methods, trivial decision-making procedures, and so on. As a result, this paper proposes a Transparent Diabetes Management System Using Machine Learning (TDMSML) expert system that uses decision tree rules to identify the major factor(s) of diabetes. The TDMSML model comprises of three phases: rule generation, transparent rule selection, and major factor identification. The rule generation phase generates rules using decision tree. Transparent rule selection stage selects the transparent rules followed by pruning the redundant rules to get the minimized rule-set. The major factor identification stage extracts the major factor(s) with range(s) from the minimized rule-set. These factor(s) with certain range(s) are characterized as major cause(s) of diabetes disease. The model is validated with the Pima Indian diabetes data set collected from Kaggle.

Large scale energy labelling with models: The EU TABULA model versus machine learning with open data

In the European Union (EU), the built environment is responsible for 40% of energy consumption. To reduce this energy consumption, policymakers require insight into current efficiency of the building stock and potential for improvement. The EU Energy Performance of Buildings Directive provides guidelines for surveying building performance and assigning energy labels; however, implementation has been slow. To address the backlog, assigning labels based on modelling rather than individual surveying might be an attractive, quicker alternative. The EU TABULA building typologies is such a model; however, in our validation against 2.5 million houses surveyed individually already, the instrument correctly classifies only 26% of buildings. This study proposes that machine learning can help leverage the abundance of (spatial) open data to suggest energy labels more closely matching the quality of individual surveying. Various transparent machine learning algorithms were tested and optimized to select the most suitable technique to model energy labels. This study shows that a random forest classifier trained on open data can reach an accuracy of 71%, thereby demonstrating the potential of open data and machine learning to quickly generate better energy labels for an entire building stock.

Advancing Reproducibility and Accountability of Unsupervised Machine Learning in Text Mining: Importance of Transparency in Reporting Preprocessing and Algorithm Selection

Machine learning (ML) enables the analysis of large datasets for pattern discovery. ML methods and the standards for their use have recently attracted increasing attention in organizational research; recent accounts have raised awareness of the importance of transparent ML reporting practices, especially considering the influence of preprocessing and algorithm choice on analytical results. However, efforts made thus far to advance the quality of ML research have failed to consider the special methodological requirements of unsupervised machine learning (UML) separate from the more common supervised machine learning (SML). We confronted these issues by studying a common organizational research dataset of unstructured text and discovered interpretability and representativeness trade-offs between combinations of preprocessing and UML algorithm choices that jeopardize research reproducibility, accountability, and transparency. We highlight the need for contextual justifications to address such issues and offer principles for assessing the contextual suitability of UML choices in research settings.

Computational Analysis of Pathological Image Enables Interpretable Prediction for Microsatellite Instability.

BackgroundMicrosatellite instability (MSI) is associated with several tumor types and has become increasingly vital in guiding patient treatment decisions; however, reasonably distinguishing MSI from its counterpart is challenging in clinical practice.MethodsIn this study, interpretable pathological image analysis strategies are established to help medical experts to identify MSI. The strategies only require ubiquitous hematoxylin and eosin–stained whole-slide images and perform well in the three cohorts collected from The Cancer Genome Atlas. Equipped with machine learning and image processing technique, intelligent models are established to diagnose MSI based on pathological images, providing the rationale of the decision in both image level and pathological feature level.FindingsThe strategies achieve two levels of interpretability. First, the image-level interpretability is achieved by generating localization heat maps of important regions based on deep learning. Second, the feature-level interpretability is attained through feature importance and pathological feature interaction analysis. Interestingly, from both the image-level and feature-level interpretability, color and texture characteristics, as well as their interaction, are shown to be mostly contributed to the MSI prediction.InterpretationThe developed transparent machine learning pipeline is able to detect MSI efficiently and provide comprehensive clinical insights to pathologists. The comprehensible heat maps and features in the intelligent pipeline reflect extra- and intra-cellular acid–base balance shift in MSI tumor.

Combining explainable machine learning, demographic and multi-omic data to inform precision medicine strategies for inflammatory bowel disease.

Inflammatory bowel diseases (IBDs), including ulcerative colitis and Crohn’s disease, affect several million individuals worldwide. These diseases are heterogeneous at the clinical, immunological and genetic levels and result from complex host and environmental interactions. Investigating drug efficacy for IBD can improve our understanding of why treatment response can vary between patients. We propose an explainable machine learning (ML) approach that combines bioinformatics and domain insight, to integrate multi-modal data and predict inter-patient variation in drug response. Using explanation of our models, we interpret the ML models’ predictions to infer unique combinations of important features associated with pharmacological responses obtained during preclinical testing of drug candidates in ex vivo patient-derived fresh tissues. Our inferred multi-modal features that are predictive of drug efficacy include multi-omic data (genomic and transcriptomic), demographic, medicinal and pharmacological data. Our aim is to understand variation in patient responses before a drug candidate moves forward to clinical trials. As a pharmacological measure of drug efficacy, we measured the reduction in the release of the inflammatory cytokine TNFα from the fresh IBD tissues in the presence/absence of test drugs. We initially explored the effects of a mitogen-activated protein kinase (MAPK) inhibitor; however, we later showed our approach can be applied to other targets, test drugs or mechanisms of interest. Our best model predicted TNFα levels from demographic, medicinal and genomic features with an error of only 4.98% on unseen patients. We incorporated transcriptomic data to validate insights from genomic features. Our results showed variations in drug effectiveness (measured by ex vivo assays) between patients that differed in gender, age or condition and linked new genetic polymorphisms to patient response variation to the anti-inflammatory treatment BIRB796 (Doramapimod). Our approach models IBD drug response while also identifying its most predictive features as part of a transparent ML precision medicine strategy.

Probabilistic machine learning for breast cancer classification.

A probabilistic neural network has been implemented to predict the malignancy of breast cancer cells, based on a data set, the features of which are used for the formulation and training of a model for a binary classification problem. The focus is placed on considerations when building the model, in order to achieve not only accuracy but also a safe quantification of the expected uncertainty of the calculated network parameters and the medical prognosis. The source code is included to make the results reproducible, also in accordance with the latest trending in machine learning research, named Papers with Code. The various steps taken for the code development are introduced in detail but also the results are visually displayed and critically analyzed also in the sense of explainable artificial intelligence. In statistical-classification problems, the decision boundary is the region of the problem space in which the classification label of the classifier is ambiguous. Problem aspects and model parameters which influence the decision boundary are a special aspect of practical investigation considered in this work. Classification results issued by technically transparent machine learning software can inspire more confidence, as regards their trustworthiness which is very important, especially in the case of medical prognosis. Furthermore, transparency allows the user to adapt models and learning processes to the specific needs of a problem and has a boosting influence on the development of new methods in relevant machine learning fields (transfer learning).

MedAI: Transparency in Medical Image Segmentation

MedAI: Transparency in Medical Image Segmentation is a challenge held for the first time at the Nordic AI Meet that focuses on medical image segmentation and transparency in machine learning (ML)-based systems. We propose three tasks to meet specific gastrointestinal image segmentation challenges collected from experts within the field, including two separate segmentation scenarios and one scenario on transparent ML systems. The latter emphasizes the need for explainable and interpretable ML algorithms. We provide a development dataset for the participants to train their ML models, tested on a concealed test dataset.

Mining Primary Care Electronic Health Records for Automatic Disease Phenotyping: A Transparent Machine Learning Framework.

(1) Background: We aimed to develop a transparent machine-learning (ML) framework to automatically identify patients with a condition from electronic health records (EHRs) via a parsimonious set of features. (2) Methods: We linked multiple sources of EHRs, including 917,496,869 primary care records and 40,656,805 secondary care records and 694,954 records from specialist surgeries between 2002 and 2012, to generate a unique dataset. Then, we treated patient identification as a problem of text classification and proposed a transparent disease-phenotyping framework. This framework comprises a generation of patient representation, feature selection, and optimal phenotyping algorithm development to tackle the imbalanced nature of the data. This framework was extensively evaluated by identifying rheumatoid arthritis (RA) and ankylosing spondylitis (AS). (3) Results: Being applied to the linked dataset of 9657 patients with 1484 cases of rheumatoid arthritis (RA) and 204 cases of ankylosing spondylitis (AS), this framework achieved accuracy and positive predictive values of 86.19% and 88.46%, respectively, for RA and 99.23% and 97.75% for AS, comparable with expert knowledge-driven methods. (4) Conclusions: This framework could potentially be used as an efficient tool for identifying patients with a condition of interest from EHRs, helping clinicians in clinical decision-support process.

Exploring syntactic and semantic features for authorship attribution

Authorship attribution is to extract features to identify authors of anonymous documents. Many previous works on authorship attribution focus on statistical style features (e.g., sentence/word length), content features (e.g., frequent words, n-grams). Modeling these features by regression or some transparent machine learning methods gives a portrait of the authors’ writing style. But these methods do not capture the syntactic (e.g., dependency relationship) or semantic (e.g., topics) information. In recent years, some researchers model syntactic trees or latent semantic information by neural networks. However, few works take them together. In this paper, we propose a novel Multi-Channel Self-Attention Network (MCSAN) incorporating both the inter-channel and inter-positional interaction to extract n-grams of the characters, words, parts of speech (POS), phrase structures, dependency relationships, and topics from multiple dimensions (style, content, syntactic and semantic features) to distinguish different authors. And then we incorporate these extracted features with logistic regression (LR) to do experiments, and the experimental results manifest that our extracted features are effective. Our methods improve 2.1% and 3.0% on CCAT10 and CCAT50, respectively, comparing with state-of-the-art models.

Transparent Machine Learning Models to Diagnose Suspicious Thoracic Lesions Leveraging CT Guided Biopsy Data

To train and validate machine learning models capable of classifying suspicious thoracic lesions as benign or malignant and to further classify malignant lesions by pathologic subtype while quantifying feature importance for each classification. 796 patients who had undergone CT guided thoracic biopsy for a concerning thoracic lesion (79.3% lung, 11.4% mediastinum, 6.5% pleura, 2.7% chest wall) were retrospectively enrolled. Lesions were classified as malignant or benign based on ground-truth pathology result, and malignant lesions were classified as primary or secondary cancer. Clinical variables were extracted from EMR and radiology reports. Supervised binary and multiclass classification models were trained to classify lesions based on the input features and evaluated on a held-out test set. Model specific feature analyses were performed to identify variables most predictive of each class, as well as to assess the independent importance of clinical, and imaging features. Binary classification models achieved a top accuracy of 80.6%, with predictive features included smoking history, age, lesion size, and lesion location. Multiclass classification models achieved a top weighted average f1-score of 0.73. Features predictive of primary cancer included smoking history, race, and age, while features predictive of secondary cancer included lesion location, and a history of cancer. Machine learning models enable classification of suspicious thoracic lesions based on clinical and imaging variables, achieving clinically useful performance while identifying importance of individual input features on a pathology-proven dataset. We believe models such as these are more likely to be trusted and adopted by clinicians.

Revealing the Impact of Global Heating on North Atlantic Circulation Using Transparent Machine Learning

AbstractThe North Atlantic ocean is key to climate through its role in heat transport and storage. Climate models suggest that the circulation is weakening but the physical drivers of this change are poorly constrained. Here, the root mechanisms are revealed with the explicitly transparent machine learning (ML) method Tracking global Heating with Ocean Regimes (THOR). Addressing the fundamental question of the existence of dynamical coherent regions, THOR identifies these and their link to distinct currents and mechanisms such as the formation regions of deep water masses, and the location of the Gulf Stream and North Atlantic Current. Beyond a black box approach, THOR is engineered to elucidate its source of predictive skill rooted in physical understanding. A labeled data set is engineered using an explicitly interpretable equation transform and k‐means application to model data, allowing theoretical inference. A multilayer perceptron is then trained, explaining its skill using a combination of layerwise relevance propagation and theory. With abrupt CO2 quadrupling, the circulation weakens due to a shift in deep water formation regions, a northward shift of the Gulf Stream and an eastward shift in the North Atlantic Current. If CO2 is increased 1% yearly, similar but weaker patterns emerge influenced by natural variability. THOR is scalable and applicable to a range of models using only the ocean depth, dynamic sea level and wind stress, and could accelerate the analysis and dissemination of climate model data. THOR constitutes a step toward trustworthy ML called for within oceanography and beyond, as its predictions are physically tractable.

Combining genetic algorithm with machine learning strategies for designing potent antimicrobial peptides

BackgroundCurrent methods in machine learning provide approaches for solving challenging, multiple constraint design problems. While deep learning and related neural networking methods have state-of-the-art performance, their vulnerability in decision making processes leading to irrational outcomes is a major concern for their implementation. With the rising antibiotic resistance, antimicrobial peptides (AMPs) have increasingly gained attention as novel therapeutic agents. This challenging design problem requires peptides which meet the multiple constraints of limiting drug-resistance in bacteria, preventing secondary infections from imbalanced microbial flora, and avoiding immune system suppression. AMPs offer a promising, bioinspired design space to targeting antimicrobial activity, but their versatility also requires the curated selection from a combinatorial sequence space. This space is too large for brute-force methods or currently known rational design approaches outside of machine learning. While there has been progress in using the design space to more effectively target AMP activity, a widely applicable approach has been elusive. The lack of transparency in machine learning has limited the advancement of scientific knowledge of how AMPs are related among each other, and the lack of general applicability for fully rational approaches has limited a broader understanding of the design space.MethodsHere we combined an evolutionary method with rough set theory, a transparent machine learning approach, for designing antimicrobial peptides (AMPs). Our method achieves the customization of AMPs using supervised learning boundaries. Our system employs in vitro bacterial assays to measure fitness, codon-representation of peptides to gain flexibility of sequence selection in DNA-space with a genetic algorithm and machine learning to further accelerate the process.ResultsWe use supervised machine learning and a genetic algorithm to find a peptide active against S. epidermidis, a common bacterial strain for implant infections, with an improved aggregation propensity average for an improved ease of synthesis.ConclusionsOur results demonstrate that AMP design can be customized to maintain activity and simplify production. To our knowledge, this is the first time when codon-based genetic algorithms combined with rough set theory methods is used for computational search on peptide sequences.

Transparent Machine Learning: Building Trust in Data Analytics

Transparent Machine Learning: Building Trust in Data Analytics

Data‐driven approach reveals heterogeneity and region‐specific association of white matter hyperintensities with the APOE genotype

AbstractBackgroundWe used a transparent machine learning algorithm to identify subtypes of white matter damage in aging and to characterize their clinical representations. White matter hyperintensities (WMH), a common finding on brain magnetic resonance imaging (MRI), are known to be associated with cognitive impairment and dementia. Previous studies have shown spatial heterogeneity of WMH appearance, but the contribution of Alzheimer disease (AD) dementia and related genetic risk factors to this heterogeneity remains unclear. We sought to investigate the heterogeneity of WMH, in the general population, leveraging state of the art pattern analysis methods and exploring whether the APOE genotype is contributing to this heterogeneity.MethodA sample of N=1836 people was obtained from the Study of Health in Pomerania (SHIP), Germany, covering a wide age range (22–84 years, median age 52.2±13.16). WMH were automatically segmented on multi‐modal MRI scans and decomposed into four spatially distinctive components using Non‐Negative Matrix Factorization (NNMF), a machine‐learning based structural covariance method. Associations of these components with the APOE genotype (available for n=303) was investigated in subjects older than 60 years of age, using a multivariable linear regression model for each component, with WMH component loadings as the outcome variable and APOE genotype (encoded as shown in Table 1) as the independent variable. Age, sex and ship sub‐cohort were included as covariates. All reported P‐values were Bonferroni corrected for multiple comparisons.ResultData‐driven approach identified four WMH components, consistent with clinical categorization in deep and periventricular lesions, while further dividing periventricular WMH into posterior, frontal and dorsal components (Figure 1). APOE genotype was significantly associated with the dorsal periventricular WMH component (P=0.035, Bonferroni corrected), but not with the other three WMH components (Table 1).ConclusionWe obtained a high consistency of data‐driven WMH components and established clinical categorizations. Our data demonstrated that the appearance of WMH follows heterogeneous regional distribution patterns. Additionally, the APOE genotype, the strongest genetic risk factor for sporadic AD, is specifically associated with one of the components indicating spatial specificity and a potential link to early MRI changes in the predementia phase.

Deep learning-level melanoma detection by interpretable machine learning and imaging biomarker cues.

.Significance: Melanoma is a deadly cancer that physicians struggle to diagnose early because they lack the knowledge to differentiate benign from malignant lesions. Deep machine learning approaches to image analysis offer promise but lack the transparency to be widely adopted as stand-alone diagnostics.Aim: We aimed to create a transparent machine learning technology (i.e., not deep learning) to discriminate melanomas from nevi in dermoscopy images and an interface for sensory cue integration.Approach: Imaging biomarker cues (IBCs) fed ensemble machine learning classifier (Eclass) training while raw images fed deep learning classifier training. We compared the areas under the diagnostic receiver operator curves.Results: Our interpretable machine learning algorithm outperformed the leading deep-learning approach 75% of the time. The user interface displayed only the diagnostic imaging biomarkers as IBCs.Conclusions: From a translational perspective, Eclass is better than convolutional machine learning diagnosis in that physicians can embrace it faster than black box outputs. Imaging biomarkers cues may be used during sensory cue integration in clinical screening. Our method may be applied to other image-based diagnostic analyses, including pathology and radiology.

Using human in vitro transcriptome analysis to build trustworthy machine learning models for prediction of animal drug toxicity

During the development of new drugs or compounds there is a requirement for preclinical trials, commonly involving animal tests, to ascertain the safety of the compound prior to human trials. Machine learning techniques could provide an in-silico alternative to animal models for assessing drug toxicity, thus reducing expensive and invasive animal testing during clinical trials, for drugs that are most likely to fail safety tests. Here we present a machine learning model to predict kidney dysfunction, as a proxy for drug induced renal toxicity, in rats. To achieve this, we use inexpensive transcriptomic profiles derived from human cell lines after chemical compound treatment to train our models combined with compound chemical structure information. Genomics data due to its sparse, high-dimensional and noisy nature presents significant challenges in building trustworthy and transparent machine learning models. Here we address these issues by judiciously building feature sets from heterogenous sources and coupling them with measures of model uncertainty achieved through Gaussian Process based Bayesian models. We combine the use of insight into the feature-wise contributions to our predictions with the use of predictive uncertainties recovered from the Gaussian Process to improve the transparency and trustworthiness of the model.

Operational framework and training standard requirements for AI-empowered robotic surgery.

For autonomous robot-delivered surgeries to ever become a feasible option, we recommend the combination of human-centered artificial intelligence (AI) and transparent machine learning (ML), with integrated Gross anatomy models. This can be supplemented with medical imaging data of cadavers for performance evaluation. We reviewed technological advances and state-of-the-art documented developments. We undertook a literature search on surgical robotics and skills, tracing agent studies, relevant frameworks, and standards for AI. This embraced transparency aspects of AI. We recommend "a procedure/skill template" for teaching AI that can be used by a surgeon. Similar existing methodologies show that when such a metric-based approach is used for training surgeons, cardiologists, and anesthetists, it results in a >40% error reduction in objectively assessed intraoperative procedures. The integration of Explainable AI and ML, and novel tissue characterization sensorics to tele-operated robotic-assisted procedures with medical imaged cadavers, provides robotic guidance and refines tissue classifications at a molecular level.