Urban Vegetation Mapping from Aerial Imagery Using Explainable AI (XAI)

Urban vegetation mapping is critical in many applications, i.e., preserving biodiversity, maintaining ecological balance, and minimizing the urban heat island effect. It is still challenging to extract accurate vegetation covers from aerial imagery using traditional classification approaches, because urban vegetation categories have complex spatial structures and similar spectral properties. Deep neural networks (DNNs) have shown a significant improvement in remote sensing image classification outcomes during the last few years. These methods are promising in this domain, yet unreliable for various reasons, such as the use of irrelevant descriptor features in the building of the models and lack of quality in the labeled image. Explainable AI (XAI) can help us gain insight into these limits and, as a result, adjust the training dataset and model as needed. Thus, in this work, we explain how an explanation model called Shapley additive explanations (SHAP) can be utilized for interpreting the output of the DNN model that is designed for classifying vegetation covers. We want to not only produce high-quality vegetation maps, but also rank the input parameters and select appropriate features for classification. Therefore, we test our method on vegetation mapping from aerial imagery based on spectral and textural features. Texture features can help overcome the limitations of poor spectral resolution in aerial imagery for vegetation mapping. The model was capable of obtaining an overall accuracy (OA) of 94.44% for vegetation cover mapping. The conclusions derived from SHAP plots demonstrate the high contribution of features, such as Hue, Brightness, GLCM_Dissimilarity, GLCM_Homogeneity, and GLCM_Mean to the output of the proposed model for vegetation mapping. Therefore, the study indicates that existing vegetation mapping strategies based only on spectral characteristics are insufficient to appropriately classify vegetation covers.

Machine learning for prediction of diabetes risk in middle-aged Swedish people

Aims

To study if machine learning methodology can be used to detect persons with increased type 2 diabetes or prediabetes risk among people without known abnormal glucose regulation.

Methods

Machine learning and interpretable machine learning models were applied on research data from Stockholm Diabetes Preventive Program, including more than 8000 people initially with normal glucose tolerance or prediabetes to determine high and low risk features for further impairment in glucose tolerance at follow-up 10 and 20 years later.

Results

The features with the highest importance on the outcome were body mass index, waist-hip ratio, age, systolic and diastolic blood pressure, and diabetes heredity. High values of these features as well as diabetes heredity conferred increased risk of type 2 diabetes. . The machine learning model was used to generate individual, comprehensible risk profiles, where the diabetes risk was obtained for each person in the data set. Features with the largest increasing or decreasing effects on the risk were determined.

Conclusions

The primary application of this machine learning model is to predict individual type 2 diabetes risk in people without diagnosed diabetes, and to which features the risk relates. However, since most features affecting diabetes risk also play a role for metabolic control in diabetes, e.g. body mass index, diet composition, tobacco use, and stress, the tool can possibly also be used in diabetes care to develop more individualized, easily accessible health care plans to be utilized when encountering the patients.

An Explainable AI-Based Fault Diagnosis Model for Bearings

In this paper, an explainable AI-based fault diagnosis model for bearings is proposed with five stages, i.e., (1) a data preprocessing method based on the Stockwell Transformation Coefficient (STC) is proposed to analyze the vibration signals for variable speed and load conditions, (2) a statistical feature extraction method is introduced to capture the significance from the invariant pattern of the analyzed data by STC, (3) an explainable feature selection process is proposed by introducing a wrapper-based feature selector-Boruta, (4) a feature filtration method is considered on the top of the feature selector to avoid the multicollinearity problem, and finally, (5) an additive Shapley explanation followed by k-NN is proposed to diagnose and to explain the individual decision of the k-NN classifier for debugging the performance of the diagnosis model. Thus, the idea of explainability is introduced for the first time in the field of bearing fault diagnosis in two steps: (a) incorporating explainability to the feature selection process, and (b) interpretation of the classifier performance with respect to the selected features. The effectiveness of the proposed model is demonstrated on two different datasets obtained from separate bearing testbeds. Lastly, an assessment of several state-of-the-art fault diagnosis algorithms in rotating machinery is included.

Improving Risk Identification of Adverse Outcomes in Chronic Heart Failure Using SMOTE+ENN and Machine Learning

PURPOSE

This study sought to develop models with good identification for adverse outcomes in patients with heart failure (HF) and find strong factors that affect prognosis.

PATIENTS AND METHODS

A total of 5004 qualifying cases were selected, among which 498 cases had adverse outcomes and 4506 cases were discharged after improvement. The study subjects were hospitalized patients diagnosed with HF from a regional cardiovascular hospital and the cardiology department of a medical university hospital in Shanxi Province of China between January 2014 and June 2019. Synthesizing minority oversampling technology combined with edited nearest neighbors (SMOTE+ENN) was used to pre-process unbalanced data. Traditional logistic regression (LR), k-nearest neighbor (KNN), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) were used to build risk identification models, and each model was repeated 100 times. Model discrimination and calibration were estimated using F1-score, the area under the receiver-operating characteristic curve (AUROC), and Brier score. The best performing of the five models was used to identify the risk of adverse outcomes and evaluate the influencing factors.

RESULTS

The SME-XGBoost was the best performing model with means of F1-score (0.3673, 95% confidence interval [CI]: 0.3633-0.3712), AUC (0.8010, CI: 0.7974-0.8046), and Brier score (0.1769, CI: 0.1748-0.1789). Age, N-terminal pronatriuretic peptide, pulmonary disease, etc. were the most significant factors of adverse outcomes in patients with HF.

CONCLUSION

The combination of SMOTE+ENN and advanced machine learning methods effectively improved the discrimination efficacy of adverse outcomes in HF patients, accurately stratified patients at risk of adverse outcomes, and found the top factors of adverse outcomes. These models and factors emphasize the importance of health status data in determining adverse outcomes in patients with HF.

A new method for exploring gene-gene and gene-environment interactions in GWAS with tree ensemble methods and SHAP values

Background

The identification of gene–gene and gene–environment interactions in genome-wide association studies is challenging due to the unknown nature of the interactions and the overwhelmingly large number of possible combinations. Parametric regression models are suitable to look for prespecified interactions. Nonparametric models such as tree ensemble models, with the ability to detect any unspecified interaction, have previously been difficult to interpret. However, with the development of methods for model explainability, it is now possible to interpret tree ensemble models efficiently and with a strong theoretical basis.

Results

We propose a tree ensemble- and SHAP-based method for identifying as well as interpreting potential gene–gene and gene–environment interactions on large-scale biobank data. A set of independent cross-validation runs are used to implicitly investigate the whole genome. We apply and evaluate the method using data from the UK Biobank with obesity as the phenotype. The results are in line with previous research on obesity as we identify top SNPs previously associated with obesity. We further demonstrate how to interpret and visualize interaction candidates.

Conclusions

The new method identifies interaction candidates otherwise not detected with parametric regression models. However, further research is needed to evaluate the uncertainties of these candidates. The method can be applied to large-scale biobanks with high-dimensional data.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12859-021-04041-7.

To trust or not to trust an explanation: using LEAF to evaluate local linear XAI methods

The main objective of eXplainable Artificial Intelligence (XAI) is to provide effective explanations for black-box classifiers. The existing literature lists many desirable properties for explanations to be useful, but there is a scarce consensus on how to quantitatively evaluate explanations in practice. Moreover, explanations are typically used only to inspect black-box models, and the proactive use of explanations as a decision support is generally overlooked. Among the many approaches to XAI, a widely adopted paradigm is Local Linear Explanations-with LIME and SHAP emerging as state-of-the-art methods. We show that these methods are plagued by many defects including unstable explanations, divergence of actual implementations from the promised theoretical properties, and explanations for the wrong label. This highlights the need to have standard and unbiased evaluation procedures for Local Linear Explanations in the XAI field. In this paper we address the problem of identifying a clear and unambiguous set of metrics for the evaluation of Local Linear Explanations. This set includes both existing and novel metrics defined specifically for this class of explanations. All metrics have been included in an open Python framework, named LEAF. The purpose of LEAF is to provide a reference for end users to evaluate explanations in a standardised and unbiased way, and to guide researchers towards developing improved explainable techniques.

Combat COVID-19 infodemic using explainable natural language processing models

Misinformation of COVID-19 is prevalent on social media as the pandemic unfolds, and the associated risks are extremely high. Thus, it is critical to detect and combat such misinformation. Recently, deep learning models using natural language processing techniques, such as BERT (Bidirectional Encoder Representations from Transformers), have achieved great successes in detecting misinformation. In this paper, we proposed an explainable natural language processing model based on DistilBERT and SHAP (Shapley Additive exPlanations) to combat misinformation about COVID-19 due to their efficiency and effectiveness. First, we collected a dataset of 984 claims about COVID-19 with fact-checking. By augmenting the data using back-translation, we doubled the sample size of the dataset and the DistilBERT model was able to obtain good performance (accuracy: 0.972; areas under the curve: 0.993) in detecting misinformation about COVID-19. Our model was also tested on a larger dataset for AAAI2021 — COVID-19 Fake News Detection Shared Task and obtained good performance (accuracy: 0.938; areas under the curve: 0.985). The performance on both datasets was better than traditional machine learning models. Second, in order to boost public trust in model prediction, we employed SHAP to improve model explainability, which was further evaluated using a between-subjects experiment with three conditions, i.e., text (T), text+SHAP explanation (TSE), and text+SHAP explanation+source and evidence (TSESE). The participants were significantly more likely to trust and share information related to COVID-19 in the TSE and TSESE conditions than in the T condition. Our results provided good implications for detecting misinformation about COVID-19 and improving public trust.

Prognostic Assessment of COVID-19 in the Intensive Care Unit by Machine Learning Methods: Model Development and Validation

BACKGROUND

Patients with COVID-19 in the intensive care unit (ICU) have a high mortality rate, and methods to assess patients' prognosis early and administer precise treatment are of great significance.

OBJECTIVE

The aim of this study was to use machine learning to construct a model for the analysis of risk factors and prediction of mortality among ICU patients with COVID-19.

METHODS

In this study, 123 patients with COVID-19 in the ICU of Vulcan Hill Hospital were retrospectively selected from the database, and the data were randomly divided into a training data set (n=98) and test data set (n=25) with a 4:1 ratio. Significance tests, correlation analysis, and factor analysis were used to screen 100 potential risk factors individually. Conventional logistic regression methods and four machine learning algorithms were used to construct the risk prediction model for the prognosis of patients with COVID-19 in the ICU. The performance of these machine learning models was measured by the area under the receiver operating characteristic curve (AUC). Interpretation and evaluation of the risk prediction model were performed using calibration curves, SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), etc, to ensure its stability and reliability. The outcome was based on the ICU deaths recorded from the database.

RESULTS

Layer-by-layer screening of 100 potential risk factors finally revealed 8 important risk factors that were included in the risk prediction model: lymphocyte percentage, prothrombin time, lactate dehydrogenase, total bilirubin, eosinophil percentage, creatinine, neutrophil percentage, and albumin level. Finally, an eXtreme Gradient Boosting (XGBoost) model established with the 8 important risk factors showed the best recognition ability in the training set of 5-fold cross validation (AUC=0.86) and the verification queue (AUC=0.92). The calibration curve showed that the risk predicted by the model was in good agreement with the actual risk. In addition, using the SHAP and LIME algorithms, feature interpretation and sample prediction interpretation algorithms of the XGBoost black box model were implemented. Additionally, the model was translated into a web-based risk calculator that is freely available for public usage.

CONCLUSIONS

The 8-factor XGBoost model predicts risk of death in ICU patients with COVID-19 well; it initially demonstrates stability and can be used effectively to predict COVID-19 prognosis in ICU patients.

An Interpretable Prediction Model for Identifying N7-Methylguanosine Sites Based on XGBoost and SHAP

Recent studies have increasingly shown that the chemical modification of mRNA plays an important role in the regulation of gene expression. N⁷-methylguanosine (m7G) is a type of positively-charged mRNA modification that plays an essential role for efficient gene expression and cell viability. However, the research on m7G has received little attention to date. Bioinformatics tools can be applied as auxiliary methods to identify m7G sites in transcriptomes. In this study, we develop a novel interpretable machine learning-based approach termed XG-m7G for the differentiation of m7G sites using the XGBoost algorithm and six different types of sequence-encoding schemes. Both 10-fold and jackknife cross-validation tests indicate that XG-m7G outperforms iRNA-m7G. Moreover, using the powerful SHAP algorithm, this new framework also provides desirable interpretations of the model performance and highlights the most important features for identifying m7G sites. XG-m7G is anticipated to serve as a useful tool and guide for researchers in their future studies of mRNA modification sites.

Toward safer highways, application of XGBoost and SHAP for real-time accident detection and feature analysis

Detecting traffic accidents as rapidly as possible is essential for traffic safety. In this study, we use eXtreme Gradient Boosting (XGBoost)-a Machine Learning (ML) technique-to detect the occurrence of accidents using a set of real time data comprised of traffic, network, demographic, land use, and weather features. The data used from the Chicago metropolitan expressways was collected between December 2016 and December 2017, and it includes 244 traffic accidents and 6073 non-accident cases. In addition, SHAP (SHapley Additive exPlanation) is employed to interpret the results and analyze the importance of individual features. The results show that XGBoost can detect accidents robustly with an accuracy, detection rate, and a false alarm rate of 99 %, 79 %, and 0.16 %, respectively. Several traffic related features, especially difference of speed between 5 min before and 5 min after an accident, are found to have relatively more impact on the occurrence of accidents. Furthermore, a feature dependency analysis is conducted for three pairs of features. First, average daily traffic and speed after accidents/non-accidents time at the upstream location are interpreted jointly. Then, distance to Central Business District and residential density are analyzed. Finally, speed after accidents/non-accidents time at upstream location and speed after accidents/non-accidents time at downstream location are evaluated with respect to the model's output.

Development and Interpretation of Multiple Machine Learning Models for Predicting Postoperative Delayed Remission of Acromegaly Patients During Long-Term Follow-Up

Background: Some patients with acromegaly do not reach the remission standard in the short term after surgery but achieve remission without additional postoperative treatment during long-term follow-up; this phenomenon is defined as postoperative delayed remission (DR). DR may complicate the interpretation of surgical outcomes in patients with acromegaly and interfere with decision-making regarding postoperative adjuvant therapy. Objective: We aimed to develop and validate machine learning (ML) models for predicting DR in acromegaly patients who have not achieved remission within 6 months of surgery. Methods: We enrolled 306 acromegaly patients and randomly divided them into training and test datasets. We used the recursive feature elimination (RFE) algorithm to select features and applied six ML algorithms to construct DR prediction models. The performance of these ML models was validated using receiver operating characteristics analysis. We used permutation importance, SHapley Additive exPlanations (SHAP), and local interpretable model-agnostic explanation (LIME) algorithms to determine the importance of the selected features and interpret the ML models. Results: Fifty-five (17.97%) acromegaly patients met the criteria for DR, and five features (post-1w rGH, post-1w nGH, post-6m rGH, post-6m IGF-1, and post-6m nGH) were significantly associated with DR in both the training and the test datasets. After the RFE feature selection, the XGboost model, which comprised the 15 important features, had the greatest discriminatory ability (area under the curve = 0.8349, sensitivity = 0.8889, Youden's index = 0.6842). The XGboost model showed good discrimination ability and provided significantly better estimates of DR of patients with acromegaly compared with using only the Knosp grade. The results obtained from permutation importance, SHAP, and LIME algorithms showed that post-6m IGF-1 is the most important feature in XGboost algorithm prediction and showed the reliability and the clinical practicability of the XGboost model in DR prediction. Conclusions: ML-based models can serve as an effective non-invasive approach to predicting DR and could aid in determining individual treatment and follow-up strategies for acromegaly patients who have not achieved remission within 6 months of surgery.

Machine-learning to stratify diabetic patients using novel cardiac biomarkers and integrative genomics

BACKGROUND

Diabetes mellitus is a chronic disease that impacts an increasing percentage of people each year. Among its comorbidities, diabetics are two to four times more likely to develop cardiovascular diseases. While HbA1c remains the primary diagnostic for diabetics, its ability to predict long-term, health outcomes across diverse demographics, ethnic groups, and at a personalized level are limited. The purpose of this study was to provide a model for precision medicine through the implementation of machine-learning algorithms using multiple cardiac biomarkers as a means for predicting diabetes mellitus development.

METHODS

Right atrial appendages from 50 patients, 30 non-diabetic and 20 type 2 diabetic, were procured from the WVU Ruby Memorial Hospital. Machine-learning was applied to physiological, biochemical, and sequencing data for each patient. Supervised learning implementing SHapley Additive exPlanations (SHAP) allowed binary (no diabetes or type 2 diabetes) and multiple classification (no diabetes, prediabetes, and type 2 diabetes) of the patient cohort with and without the inclusion of HbA1c levels. Findings were validated through Logistic Regression (LR), Linear Discriminant Analysis (LDA), Gaussian Naïve Bayes (NB), Support Vector Machine (SVM), and Classification and Regression Tree (CART) models with tenfold cross validation.

RESULTS

Total nuclear methylation and hydroxymethylation were highly correlated to diabetic status, with nuclear methylation and mitochondrial electron transport chain (ETC) activities achieving superior testing accuracies in the predictive model (~ 84% testing, binary). Mitochondrial DNA SNPs found in the D-Loop region (SNP-73G, -16126C, and -16362C) were highly associated with diabetes mellitus. The CpG island of transcription factor A, mitochondrial (TFAM) revealed CpG24 (chr10:58385262, P = 0.003) and CpG29 (chr10:58385324, P = 0.001) as markers correlating with diabetic progression. When combining the most predictive factors from each set, total nuclear methylation and CpG24 methylation were the best diagnostic measures in both binary and multiple classification sets.

CONCLUSIONS

Using machine-learning, we were able to identify novel as well as the most relevant biomarkers associated with type 2 diabetes mellitus by integrating physiological, biochemical, and sequencing datasets. Ultimately, this approach may be used as a guideline for future investigations into disease pathogenesis and novel biomarker discovery.

Explainable extreme gradient boosting tree-based prediction of toluene, ethylbenzene and xylene wet deposition

Current research suggests that, apart from photochemical reactions, toluene, ethylbenzene and xylene (TEX) removal from ambient air might be affected by atmospheric precipitation, depending on the concentrations and water solubility of the compounds, Henry's law, physico-chemical properties of the water, as well as the frequency and intensity of precipitation events. Nevertheless, existing knowledge of the role that wet deposition plays in biogeochemical cycles of volatile species remains insufficient, and this topic requires more scientific effort to be explored and understood. In this study, we employed the eXtreme Gradient Boosting tree ensemble for revealing TEX transfer from ambient air to rainwater, and applied a novel SHapley Additive exPlanations feature attribution framework to examine the relevance of the monitored parameters and identify key factors that govern wet deposition of TEX. According to the results, main impacts, including ambient air TEX concentrations, and rainwater and air temperatures, and occasional, but less important impacts, including wind speed, air pressure, turbidity, and total organic carbon, NO₃^-, Cl^- and K⁺ rainwater concentration, shaped TEX partition between gaseous and aqueous phases during rain events.

The Southampton Hand Assessment Procedure revisited: A transparent linear scoring system, applied to data of experienced prosthetic users

STUDY DESIGN

Cross-sectional.

INTRODUCTION

Southampton Hand Assessment Procedure (SHAP) provides function scores for hand grips (prehensile patterns) and an overall score, the index of function (IOF). The underlying equations of SHAP are not publicly available, which induces opacity. Furthermore, SHAP has been scarcely tested in prosthetic users.

METHODS

Issues with SHAP-IOF are discussed; an alternative scoring system, that is, linear index of function (LIF) and a weighted version (W-LIF) are presented. In LIF, task times are transformed linearly, relative to SHAP norms, and are computed into LIF-prehensile patterns (LIF_PP). LIF and IOF were compared using data of 27 experienced prosthetic users.

RESULTS

High correlation and agreement between LIF and IOF was found: LIF_PP vs IOF_PP ranged from r = 0.880 to r = 0.988, and W-LIF vs IOF had a correlation coefficient of r = 0.984.

DISCUSSION

SHAP data of prosthetic users are valuable benchmarks for health care professionals. LIF calculations are a good and cost free alternative for IOF scores.

CONCLUSION(S)

Measurements with LIF and IOF may be considered similar, but LIF is transparent and easier to use than IOF.

LEVEL OF EVIDENCE

Clinical measurement and cross-sectional.

Translating Research on Myoelectric Control into Clinics-Are the Performance Assessment Methods Adequate?

Missing an upper limb dramatically impairs daily-life activities. Efforts in overcoming the issues arising from this disability have been made in both academia and industry, although their clinical outcome is still limited. Translation of prosthetic research into clinics has been challenging because of the difficulties in meeting the necessary requirements of the market. In this perspective article, we suggest that one relevant factor determining the relatively small clinical impact of myocontrol algorithms for upper limb prostheses is the limit of commonly used laboratory performance metrics. The laboratory conditions, in which the majority of the solutions are being evaluated, fail to sufficiently replicate real-life challenges. We qualitatively support this argument with representative data from seven transradial amputees. Their ability to control a myoelectric prosthesis was tested by measuring the accuracy of offline EMG signal classification, as a typical laboratory performance metrics, as well as by clinical scores when performing standard tests of daily living. Despite all subjects reaching relatively high classification accuracy offline, their clinical scores varied greatly and were not strongly predicted by classification accuracy. We therefore support the suggestion to test myocontrol systems using clinical tests on amputees, fully fitted with sockets and prostheses highly resembling the systems they would use in daily living, as evaluation benchmark. Agreement on this level of testing for systems developed in research laboratories would facilitate clinically relevant progresses in this field.