Contribution Of Feature Research Articles

Brain Imaging Features in Patients with Gunshot Wounds to the Head.

To introduce the UChicago PBI Imaging score, a novel characterization of imaging features using head computed tomography (HCT) in patients with gunshot wounds to the head (GSWH) resulting in penetrating brain injury (PBI) and to quantify the association with mortality. We retrospectively collected and analyzed data from 230 patients with GSWH admitted to our Level 1 trauma center between May 1, 2018, and October 31, 2023. HCT images obtained on hospital arrival were evaluated for predefined imaging features by two blinded readers and arbitrated, when needed, by a third. The average contribution of each radiological feature to mortality at hospital discharge was assessed using a SuperLearner ensemble model trained on ∼77% of the cohort. Each feature's contribution was scaled to ensure the additive final score per patient ranged between 0 and 100. The HCT features predicting in-hospital mortality, ranked from highest to lowest importance, were transhemispheric projectile below the level of the third ventricle (18 [16.8, 19.9]), presence of blood in the lateral ventricles (ventricles casted) (18[16.8, 19.6]), brainstem involvement (14 [12.7, 15.1]), transhemispheric projectile above the level of the third ventricle (11 [9.7, 11.6]), presence of any amount of blood in the ambient cistern (9[8.2, 10]), presence of any amount of blood in the lateral ventricles (9 [7.9, 9.8]), cerebellar involvement (9 [7.9, 9.5]), any evidence of ventricular effacement (4 [3.4, 4.6]), midline shift (MLS) >0 mm (4 [3.4, 4.4]), perforating injury (3 [2.4, 3.2]), and presence of an intracerebral hematoma (ICH) >20 mm in the largest diameter (2 [1.4, 1.9]). The UChicago PBI Imaging score showed a strong performance, achieving an area under the curve (AUC) of 0.86 (95% CI: [0.77, 0.96]) on a test set of 56 patients who were not included in model training. This indicates better prediction accuracy compared to both the Rotterdam score (AUC 0.8, 95% CI: [0.68, 0.96]) and the Marshall score (AUC 0.66, 95% CI: [0.52, 0.81]). Our model performed particularly well for patients with a Glasgow Coma Scale (GCS) score between 5 and 9. In this range, our model's performance (AUC 0.86) remained stable, while the Rotterdam and Marshall Scores showed notably lower predictive accuracy, with AUCs of 0.61 and 0.52, respectively. A dedicated evaluation of GSWH HCT reveals an association between disease burden, as quantified by unique features not native to blunt TBI imaging models, and mortality. Specifically, transhemispheric injury below the level of the third ventricle along with blood-casting bilateral ventricles and brainstem involvement was highly associated with mortality. The model is optimized for intermediate GCS scores where greater prognostic uncertainty exists. This study parallels efforts to refine TBI classification, underscoring the necessity for precise imaging-based classification in PBI to identify imaging biomarkers and ultimately enhance prognostication and targeted treatment.

Explainable predictions of a machine learning model to forecast the postoperative length of stay for severe patients: machine learning model development and evaluation.

Predicting the length of stay in advance will not only benefit the hospitals both clinically and financially but enable healthcare providers to better decision-making for improved quality of care. More importantly, understanding the length of stay of severe patients who require general anesthesia is key to enhancing health outcomes. Here, we aim to discover how machine learning can support resource allocation management and decision-making resulting from the length of stay prediction. A retrospective cohort study was conducted from January 2018 to October 2020. A total cohort of 240,000 patients' medical records was collected. The data were collected exclusively for preoperative variables to accurately analyze the predictive factors impacting the postoperative length of stay. The main outcome of this study is an analysis of the length of stay (in days) after surgery until discharge. The prediction was performed with ridge regression, random forest, XGBoost, and multi-layer perceptron neural network models. The XGBoost resulted in the best performance with an average error within 3days. Moreover, we explain each feature's contribution over the XGBoost model and further display distinct predictors affecting the overall prediction outcome at the patient level. The risk factors that most importantly contributed to the stay after surgery were as follows: a direct bilirubin laboratory test, department change, calcium chloride medication, gender, and diagnosis with the removal of other organs. Our results suggest that healthcare providers take into account the risk factors such as the laboratory blood test, distributing patients, and the medication prescribed prior to the surgery. We successfully predicted the length of stay after surgery and provide explainable models with supporting analyses. In summary, we demonstrate the interpretation with the XGBoost model presenting insights on preoperative features and defining higher risk predictors to the length of stay outcome. Our development in explainable models supports the current in-depth knowledge for the future length of stay prediction on electronic medical records that aids the decision-making and facilitation of the operation department.

Carbon emission prediction of 275 cities in China considering artificial intelligence effects and feature interaction: A heterogeneous deep learning modeling framework

High technology and artificial intelligence (AI) are crucial for achieving urban Dual Carbon Goals. This study proposes a heterogeneous deep learning framework with analysis and prediction phases to explore AI technology's impact on urban carbon emissions. In the analysis phase, fixed effect models address differences in AI development and time heterogeneity among cities. In the prediction phase, an Attention Deep & Cross Network (ADCN) model leveraging feature interactions is proposed to enhance prediction precision and robustness. The Shapley Additive Explanations (SHAP) method quantifies each feature's contribution to ADCN's predictions, elucidating factors' impacts on carbon emissions. This study investigates AI development levels and other variables across 275 Chinese cities to test model performance and uncover the AI-carbon emissions relationship. Results show that fixed effects models significantly improve prediction accuracy, with ADCN outperforming statistical and machine learning models (RMSE: 646.262, MAE: 474.818, R²: 0.993). SHAP analysis reveals that AI technology level (11.85 %), smart city (12.35 %), energy consumption (11.60 %), population (9.38 %), urbanization rate (8.89 %), and GDP (8.40 %) significantly influence carbon emissions. Especially, the interaction between AI technology and smart city or intelligent manufacturing proportion increases their carbon reduction by 1.059 × 1021 or 4.992 × 1019 tons. AI technology moderates the impact of increasing energy consumption and urbanization, reducing their potential emissions by 20 % and 1 %. The framework offers high accuracy and scalability, providing valuable insights for strategy development.

Prediction of bike-sharing station demand using explainable artificial intelligence

Bike-sharing systems have grown in popularity in metropolitan areas, providing a handy and environmentally friendly transportation choice for commuters and visitors alike. As demand for bike-sharing programs grows, efficient capacity planning becomes critical to ensuring good user experience and system sustainability in terms of demand. The random forest model was used in this study to predict bike-sharing station demand and is considered a strong ensemble learning approach that can successfully capture complicated nonlinear correlations and interactions between input variables. This study employed data from the Smart Location Database (SLD) to test the model accuracy in estimating station demand and used a form of explainable artificial intelligence (XAI) function to further understand machine learning (ML) prediction outcomes owing to the blackbox tendencies of ML models. Vehicle Miles of Travel (VMT) and Greenhouse Gas (GHG) emissions were the most important features in predicting docking station demand individually but not holistically based on the datasets. The percentage of zero-car households, gross residential density, road network density, aggregate frequency of transit service, and gross activity density were found to have a moderate influence on the prediction model. Further, there may be a better prediction model generating sensible results for every type of explanatory variable, but their contributions are minimum to the prediction outcome. By measuring each feature's contribution to demand prediction in feature engineering, bike-sharing operators can acquire a better understanding of the bike-sharing station capacity and forecast future demands during planning. At the same time, ML models will need further assessment before a holistic conclusion.

Decoding pulsatile patterns of cerebrospinal fluid dynamics through enhancing interpretability in machine learning

Analyses of complex behaviors of Cerebrospinal Fluid (CSF) have become increasingly important in diseases diagnosis. The changes of the phase-contrast magnetic resonance imaging (PC-MRI) signal formed by the velocity of flowing CSF are represented as a set of velocity-encoded images or maps, which can be thought of as signal data in the context of medical imaging, enabling the evaluation of pulsatile patterns throughout a cardiac cycle. However, automatic segmentation of the CSF region in a PC-MRI image is challenging, and implementing an explained ML method using pulsatile data as a feature remains unexplored. This paper presents lightweight machine learning (ML) algorithms to perform CSF lumen segmentation in spinal, utilizing sets of velocity-encoded images or maps as a feature. The Dataset contains 57 PC-MRI slabs by 3T MRI scanner from control and idiopathic scoliosis participants are involved to collect data. The ML models are trained with 2176 time series images. Different cardiac periods image (frame) numbers of PC-MRIs are interpolated in the preprocessing step to align to features of equal size. The fivefold cross-validation procedure is used to estimate the success of the ML models. Additionally, the study focusses on enhancing the interpretability of the highest-accuracy eXtreme gradient boosting (XGB) model by applying the shapley additive explanations (SHAP) technique. The XGB algorithm presented its highest accuracy, with an average fivefold accuracy of 0.99% precision, 0.95% recall, and 0.97% F1 score. We evaluated the significance of each pulsatile feature's contribution to predictions, offering a more profound understanding of the model's behavior in distinguishing CSF lumen pixels with SHAP. Introducing a novel approach in the field, develop ML models offer comprehension into feature extraction and selection from PC-MRI pulsatile data. Moreover, the explained ML model offers novel and valuable insights to domain experts, contributing to an enhanced scholarly understanding of CSF dynamics.

Lane change decision prediction: an efficient BO-XGB modelling approach with SHAP analysis

The lane-change decision (LCD) is a critical aspect of driving behaviour. This study proposes an LCD model based on a Bayesian optimization (BO) framework and extreme gradient boosting (XGBoost) to predict whether a vehicle should change lanes. First, an LCD point extraction method is proposed to refine the exact LCD points with a highD dataset to increase model learning accuracy. Subsequently, an efficient XGBoost with BO (BO-XGB) was used to learn the LCD principles. The prediction accuracy on the highD dataset was 99.14% with a computation time of 66.837s. The accuracy on the CQSkyEyeX dataset was 99.45%. Model explanation using the shapley additive explanation (SHAP) method was developed to analyse the mechanism of the BO-XGB’s LCD prediction results, including global and sample explanations. The former indicates the particular contribution of each feature to the model prediction throughout the entire dataset. The latter denotes each feature's contribution to a single sample.

Exploring Nutritional Influence on Blood Glucose Forecasting for Type 1 Diabetes Using Explainable AI.

Type 1 diabetes mellitus (T1DM) is characterized by insulin deficiency and blood sugar control issues. The state-of-the-art solution is the artificial pancreas (AP), which integrates basal insulin delivery and glucose monitoring. However, APs are unable to manage postprandial glucose response (PGR) due to limited knowledge of its determinants, requiring additional information for accurate bolus delivery, such as estimated carbohydrate intake. This study aims to quantify the influence of various meal-related factors on predicting postprandial blood glucose levels (BGLs) at different time intervals (15 min, 60 min, and 120 min) after meals by using deep neural network (DNN) models. The prediction models incorporate preprandial blood glucose values, insulin dosage, and various meal-related nutritional factors such as intake of energy, carbohydrates, proteins, lipids, fatty acids, fibers, glycemic index, and glycemic load as input variables. The impact of input features was assessed by exploiting eXplainable Artificial Intelligence (XAI) methodologies, specifically SHapley Additive exPlanations (SHAP), which provide insights into each feature's contribution to the model predictions. By leveraging XAI methodologies, this study aims to enhance the interpretability and transparency of BGL prediction models and validate clinical literature hypotheses. The findings can aid in the development of decision-support tools for individuals with T1DM, facilitating PGR management and reducing the risks of adverse events. The improved understanding of PGR determinants may lead to advancements in AP technology and improve the overall quality of life for T1DM patients.

Machine Learning-Based Hyperglycemia Prediction: Enhancing Risk Assessment in a Cohort of Undiagnosed Individuals.

Noncommunicable diseases continue to pose a substantial health challenge globally, with hyperglycemia serving as a prominent indicator of diabetes. This study employed machine learning algorithms to predict hyperglycemia in a cohort of individuals who were asymptomatic and unraveled crucial predictors contributing to early risk identification. This dataset included an extensive array of clinical and demographic data obtained from 195 adults who were asymptomatic and residing in a suburban community in Nigeria. The study conducted a thorough comparison of multiple machine learning algorithms to ascertain the most effective model for predicting hyperglycemia. Moreover, we explored feature importance to pinpoint correlates of high blood glucose levels within the cohort. Elevated blood pressure and prehypertension were recorded in 8 (4.1%) and 18 (9.2%) of the 195 participants, respectively. A total of 41 (21%) participants presented with hypertension, of which 34 (83%) were female. However, sex adjustment showed that 34 of 118 (28.8%) female participants and 7 of 77 (9%) male participants had hypertension. Age-based analysis revealed an inverse relationship between normotension and age (r=-0.88; P=.02). Conversely, hypertension increased with age (r=0.53; P=.27), peaking between 50-59 years. Of the 195 participants, isolated systolic hypertension and isolated diastolic hypertension were recorded in 16 (8.2%) and 15 (7.7%) participants, respectively, with female participants recording a higher prevalence of isolated systolic hypertension (11/16, 69%) and male participants reporting a higher prevalence of isolated diastolic hypertension (11/15, 73%). Following class rebalancing, the random forest classifier gave the best performance (accuracy score 0.89; receiver operating characteristic-area under the curve score 0.89; F1-score 0.89) of the 26 model classifiers. The feature selection model identified uric acid and age as important variables associated with hyperglycemia. The random forest classifier identified significant clinical correlates associated with hyperglycemia, offering valuable insights for the early detection of diabetes and informing the design and deployment of therapeutic interventions. However, to achieve a more comprehensive understanding of each feature's contribution to blood glucose levels, modeling additional relevant clinical features in larger datasets could be beneficial.

Interpretable Machine Learning for Predicting Symptomatic Pneumonitis in Locally Advanced Non-Small Cell Lung Cancer Patients Treated with Concurrent Chemoradiotherapy and Immune Checkpoint Inhibitor Consolidation.

The rate of grade 2 and higher pneumonitis has increased with the use of immune checkpoint inhibitors (ICI) following chemoradiotherapy (CRT) for lung cancer, which may alter previously established dose-volume constraints (DVC). In this study, we used an interpretable machine learning model with clinical and dosimetric features to predict grade 2+ pneumonitis and determine DVC associated with pneumonitis for locally advanced non-small cell lung cancer (LA-NSCLC) radiotherapy (RT). Between October 2017 and December 2021, 223 consecutively treated patients with LANSCLC treated with CRT and ICI were retrospectively reviewed. The dataset was split into training and test sets (n = 144/79). Clinical features included age, sex, smoking status, pack-years, BMI, ECOG PS, COPD, tumor location, delivered dose, RT technique, chemotherapy agent and volume of GTVp/GTVn. A total of 228 dosimetric features from the heart, contralateral/ipsilateral lung and lungs-IGTV were extracted, including the minimum/mean dose to the hottest x% volume (Dx%[Gy]/MOHx%[Gy]; x was 5-95 in 5% increments) and minimum/mean/maximum dose and percent volume receiving at least xGy (VxGy [%]; x was 5-60 in 5Gy increments), as well as the overlapping volume of each structure with PTV and the distance from each structure to GTVp/GTVn. Feature selection was performed using Boruta, followed by collinearity removal based on the variance inflation factor. The explainable boosting machine (EBM) was trained on the selected features. The performance of EBM on the test set was evaluated using the area under the receiver operating characteristic curve (AUC) and compared with that of blackbox (BB) models, including extreme gradient boosting (XGB), random forest (RF), and supporting vector machine (SVM). The global explanation of each feature's contribution to the predictions provided by the EBM was used to determine DVC. Shapley additive explanations (SHAP) were used to explain BB predictions. Selected features, ranked in order of EBM's overall feature importance, were V25Gy [%] and MOH65%[Gy] in the ipsilateral lung, the maximum dose in the heart, MOH30%[Gy] in the contralateral lung, and BMI. No dosimetric features in the lungs-IGTV were selected. The SHAP values of three BB models showed similar trends to the feature importance of the EBM. The global explanations of the EBM suggested that to mitigate the risk of pneumonitis, the ipsilateral lung should have V25Gy [%] < 36.8% and MOH65%[Gy] < 39.5Gy, and the heart should have D0.03cc [Gy] < 66.0Gy. Furthermore, an increased risk of pneumonitis was indicated with an increase in BMI, and, surprisingly, a decrease in MOH30%[Gy] in the contralateral lung. The EBM showed the best performance for predicting grade 2+ pneumonitis (AUC = 0.739), followed by RF, SVM, and XGB (AUC = 0.735, 0.733, and 0.717). EBM has the potential to predict grade 2+ pneumonitis in LA-NSCLC patients treated with CRT and ICI, while providing guidance on DVC.

Modeling uncertainty of specimens employing spines and force‐limiting connections tested at E‐defense shake table

AbstractIn light of the significant damage observed after earthquakes in Japan and New Zealand, enhanced performing seismic force‐resisting systems and energy dissipation devices are increasingly being utilized in buildings. Numerical models are needed to estimate the seismic response of these systems for seismic design or assessment. While there have been studies on modeling uncertainty, selecting the model features most important to response can remain ambiguous, especially if the structure employs less well‐established lateral force‐resisting systems and components. Herein, a global sensitivity analysis was used to address modeling uncertainty in specimens with elastic spines and force‐limiting connections (FLCs) physically tested at full‐scale at the E‐Defense shake table in Japan. Modeling uncertainty was addressed for both model class and model parameter uncertainty by varying primary models to develop several secondary models according to pre‐established uncertainty groups. Numerical estimates of peak story drift ratio and floor acceleration were compared to the results from the experimental testing program using confidence intervals and root‐mean‐square error. Metrics such as the coefficient of variation, variance, linear Pearson correlation coefficient, and Sobol index were used to gain intuition about each model feature's contribution to the dispersion in estimates of the engineering demands. Peak floor acceleration was found to be more sensitive to modeling uncertainty compared to story drift ratio. Assumptions for the spine‐to‐frame connection significantly impacted estimates of peak floor accelerations, which could influence future design methods for spines and FLC in enhanced lateral‐force resisting systems.

Exploring the effect of various factors for ash content estimation via ensemble learning: Color-texture features, particle size, and magnification

Online estimation of ash content in coal based on machine vision has been paid more attention to by academia and industry. Existing research has mainly focused on feature extraction and model design for estimating ash content, but the exploration of the feature's contribution to the model is rarely reported. Therefore, this paper presented a method to analyze features' contribution to microscopic images. Expressly, ensemble learning could explicitly point out the assistance of different features to ash content estimation according to the results of weak regressors, which possessed good interpretability. Firstly, we explored the relationship among particle size, magnification, and precision to determine Adaboost as the optimal regressor. Secondly, we discussed the contribution of different features to the regressor quantitatively and qualitatively on the 0.5–0.25 (mm) particle size of datasets under magnification of 0.3, which we called the optimal condition. Ultimately, the experimental conclusions were validated by feature importance analysis and correlation analysis. The experimental results show that the magnification significantly influenced the precision of ash content estimation. Still, coal microscopic images' 0.5–0.25 (mm) particle size estimation accuracy was optimal under all magnifications. When magnification was 0.3 and particle size was 0.5–0.25 (mm), there was the best performance to ash content estimation (In the testing set, 98% of the samples had an error less than 0.5% between ash content estimated value and ash content value). Color features played a dominant role in ash content estimation in the optimal regressors. Only the building color feature regressor could achieve that 96.4% of samples had an estimated error of less than 0.5%. In contrast, only 68% of the samples using the texture feature regressors had an estimated error of less than 0.5%. Moreover, the building mixture of color and texture regressor possessed mutually reinforcing effects under specific conditions. The summarized features' contribution was essential for model building and design.

Prediction of delayed cerebral ischemia after cerebral aneurysm rupture using explainable machine learning approach.

Aneurysmal subarachnoid hemorrhage results in significant mortality and disability, which is worsened by the development of delayed cerebral ischemia. Tests to identify patients with delayed cerebral ischemia prospectively are of high interest. We created a machine learning system based on clinical variables to predict delayed cerebral ischemia in aneurysmal subarachnoid hemorrhage patients. We also determined which variables have the most impact on delayed cerebral ischemia prediction using SHapley Additive exPlanations method. 500 aneurysmal subarachnoid hemorrhage patients were identified and 369 met inclusion criteria: 70 patients developed delayed cerebral ischemia (delayed cerebral ischemia+) and 299 did not (delayed cerebral ischemia-). The algorithm was trained based upon age, sex, hypertension (HTN), diabetes, hyperlipidemia, congestive heart failure, coronary artery disease, smoking history, family history of aneurysm, Fisher Grade, Hunt and Hess score, and external ventricular drain placement. Random Forest was selected for this project, and prediction outcome of the algorithm was delayed cerebral ischemia+. SHapley Additive exPlanations was used to visualize each feature's contribution to the model prediction. The Random Forest machine learning algorithm predicted delayed cerebral ischemia: accuracy 80.65% (95% CI: 72.62-88.68), area under the curve 0.780 (95% CI: 0.696-0.864), sensitivity 12.5% (95% CI: -3.7 to 28.7), specificity 94.81% (95% CI: 89.85-99.77), PPV 33.3% (95% CI: -4.39 to 71.05), and NPV 84.1% (95% CI: 76.38-91.82). SHapley Additive exPlanations value demonstrated Age, external ventricular drain placement, Fisher Grade, and Hunt and Hess score, and HTN had the highest predictive values for delayed cerebral ischemia. Lower age, absence of hypertension, higher Hunt and Hess score, higher Fisher Grade, and external ventricular drain placement increased risk of delayed cerebral ischemia. Machine learning models based upon clinical variables predict delayed cerebral ischemia with high specificity and good accuracy.

Contribution of meteorological factors based on explainable artificial intelligence in predicting wind farm power production using machine learning algorithms

Wind fields are intermittent and nonlinear to meteorological factors and external environmental conditions. Statistical models have been proposed based on various approaches to precisely predict wind speed and energy production. However, determining the most suitable approach is challenging, regardless of the conditions. Currently, only wind speed, wind direction, temperature, atmospheric pressure, and humidity have been used as input features of models in most wind-power forecasting studies. However, few studies have described each feature's contribution to prediction performance when using meteorological factors, such as atmospheric stability and turbulence components, as input features. This study predicted the 10 min average power and daily energy production of a wind farm using four machine learning (ML) algorithms and 13 meteorological factors. The ultimate goal was to present the individual prediction contribution of meteorological factors using the Shapley additive explanations algorithm, which is an explainable artificial intelligence technique, based on the prediction results. Wind speed showed a dominant influence in the determination of energy production, followed by turbulent kinetic energy, turbulence intensity, and turbulence dissipation rate. Thus, insights into the detailed contribution of turbulence components to predict the performance facilitate the advancement of ML-based approaches, which can yield significant benefits in increasing the predictability of actual wind energy, thereby ensuring efficiency and stability in wind farm operations.

Explaining the Outputs of Convolutional Neural Network - Recurrent Neural Network (CNN-RNN) based Apparent Personality Detection Models using the Class Activation Maps

This study aims to use the Class Activation Map (CAM) visualisation technique to understand the outputs of apparent personality detection models based on a combination of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). The ChaLearn Looking at People First Impression (CVPR'17) dataset is used for experimentation in this study. The dataset consists of short video clips labelled with the Big Five personality traits. Two deep learning models were designed to predict apparent personality with VGG19 and ResNet152 as base models. Then the models were trained using the raw frames extracted from the videos. The highest accurate models from each architecture were chosen for feature visualisation. The test dataset of the CVPR'17 dataset is used for feature visualisation. To identify the feature's contribution to the network's output, the CAM XAI technique was applied to the test dataset and calculated the heatmap. Next, the bitwise intersection between the heatmap and background removed frames was measured to identify how much features from the human body (including facial and non-facial data) affected the network output. The findings revealed that nearly 35%-40% of human data contributed to the output of both models. Additionally, after analysing the heatmap with high-intensity pixels, the ResNet152 model was found to identify more human-related data than the VGG19 model, achieving scores of 46%-51%. The two models have different behaviour in identifying the key features which influence the output of the models based on the input.

The utility of wearable devices in assessing ambulatory impairments of people with multiple sclerosis in free-living conditions

Background and objectivesMultiple sclerosis (MS) is a progressive inflammatory and neurodegenerative disease of the central nervous system affecting over 2.5 million people globally. In-clinic six-minute walk test (6MWT) is a widely used objective measure to evaluate the progression of MS. Yet, it has limitations such as the need for a clinical visit and a proper walkway. The widespread use of wearable devices capable of depicting patients’ activity profiles has the potential to assess the level of MS-induced disability in free-living conditions. MethodsIn this work, we extracted 96 features in different temporal granularities (from minute-level to day-level) from wearable data and explored their utility in estimating 6MWT scores in a European (Italy, Spain, and Denmark) MS cohort of 337 participants over an average of 10 months’ duration. We combined these features with participants’ demographics using three regression models including elastic net, gradient boosted trees and random forest. In addition, we quantified the individual feature's contribution using feature importance in these regression models, linear mixed-effects models, generalized estimating equations, and correlation-based feature selection (CFS). ResultsThe results showed promising estimation performance with R2 of 0.30, which was derived using random forest after CFS. This model was able to distinguish the participants with low disability from those with high disability. Furthermore, we observed that the minute-level (≤ 8 minutes) step count, particularly those capturing the upper end of the step count distribution, had a stronger association with 6MWT. The use of a walking aid was indicative of ambulatory function measured through 6MWT. ConclusionsThis study demonstrates the utility of wearables devices in assessing ambulatory impairments in people with MS in free-living conditions and provides a basis for future investigation into the clinical relevance.

Explainable Boosting Machine for Choosing Radiation Dose-Volume Constraints on Cardio-Pulmonary Substructures Associated with Overall Survival in the NRG Oncology RTOG 0617 Clinical Trial

<h3>Purpose/Objective(s)</h3> Black box (BB) machine learning models are frequently used for decision making in healthcare. Although posthoc model-agnostic explanations can improve the BB model's interpretability, they are often unreliable. Explainable boosting machine (EBM) is a glass box model that allows for model-specific interpretation of each feature's contribution to outcome prediction. We used EBM for choosing radiation dose-volume constraints (DVCs) on cardio-pulmonary substructures (CPSs) associated with overall survival (OS) in locally advanced non-small cell lung cancer (NSCLC) radiotherapy. <h3>Materials/Methods</h3> A total of 379 NSCLC patients enrolled in RTOG 0617 were randomly split into training/validation subsets (n=230/149). A deep learning autosegmentation model was used for segmenting 12 CPSs: heart, atria, ventricles, aorta, left ventricle/atrium (LV/LA), right ventricle/atrium (RV/RA), inferior/superior vena cava, pulmonary artery, and pericardium. Contours of peri (=pericardium-heart), rest [=heart-(LV|RV|LA|RA)], clinical target volume and lungs were also obtained. From each structure, dose-volume histogram (DVH) features including min/mean dose to the hottest x% volume (Dx%[Gy]/MOHx%[Gy]; x was 5-100 in 5% increments) and min/mean/max dose as well as percent volume receiving at least xGy (VxGy[%]; x was 5-75 in 5Gy increments) were extracted. Unsupervised multivariate multiblock regression was used to select DVH features in each structure that were canonically correlated with 38 clinical/dosimetric/outcome variables taken from the NTCN/NCORP data archive for RTOG 0617. Additional feature selection for OS analysis was performed using stability selection supervised principal component analysis and Boruta. Algorithm for survival stacking was implemented for training EBM by casting survival analysis as a classification problem. Discrimination performance of EBM for OS in the validation set was evaluated by Harrell's c-index and compared to that of Cox proportional hazards (CPH) and 3 BB models: random survival forest (RSF), deep neural network (DeepSurv) and extreme gradient boosting (XGB). Shapley additive explanations (SHAP) were used to explain their predictions. <h3>Results</h3> Selected features included RA_V55Gy[%], aorta_V15Gy[%], LV_D5%[Gy] and pericardium_D20%[Gy]. EBM and all 3 BB models ranked LV_D5%[Gy] as the most important feature, while CPH model ranked second. EBM achieved the best performance for predicting OS, followed by DeepSurv, XGB, RSF and CPH (c-index=0.657, 0.646, 0.643, 0.632 and 0.630). Global explanations of EBM suggested that LV_D5%[Gy]<17.2Gy, pericardium_D20%[Gy]<52.6Gy, RA_V55Gy[%]<12.9% and aorta_V15Gy[%]<49.4% for improved OS. Although SHAP values of CPH and all 3 BB models showed similar trends, they were unable to assist in the precise selection of DVCs. <h3>Conclusion</h3> EBM can be used for modeling OS while also providing guidance on CPS DVC selection for optimal cardiac toxicity management in lung cancer radiotherapy.

A Machine Learning Approach for Predicting Early Phase Postoperative Hypertension in Patients Undergoing Carotid Endarterectomy

This study aimed to establish and validate a machine learning-based model for the prediction of early phase postoperative hypertension (EPOH) requiring the administration of intravenous vasodilators after carotid endarterectomy (CEA). Perioperative data from consecutive CEA procedures performed from January 2013 to August 2019 were retrospectively collected. EPOH was defined in post-CEA patients as hypertension involving a systolic blood pressure above 160mm Hg and requiring the administration of any intravenous vasodilator medications in the first 24hr after a return to the vascular ward. Gradient boosted regression trees were used to construct the predictive model, and the featured importance scores were generated by using each feature's contribution to each tree in the model. To evaluate the model performance, the area under the receiver operating characteristic curve was used as the main metric. Four-fold stratified cross-validation was performed on the data set, and the average performance of the 4 folds was reported as the final model performance. A total of 406 CEA operations were performed under general anesthesia. Fifty-three patients (13.1%) met the definition of EPOH. There was no significant difference in the percentage of postoperative stroke/death between patients with and without EPOH during the hospital stay. Patients with EPOH exhibited a higher incidence of postoperative cerebral hyperperfusion syndrome (7.5% vs. 0, P<0.001), as well as a higher incidence of cerebral hemorrhage (3.8% vs. 0, P<0.001). The gradient boosted regression trees prediction model achieved an average AUC of 0.77 (95% CI 0.62 to 0.92). When the sensitivity was fixed near 0.90, the model achieved an average specificity of 0.52 (95% CI 0.28 to 0.75). We have built the first-ever machine learning-based prediction model for EPOH after CEA. The validation result from our single-center database was very promising. This novel prediction model has the potential to help vascular surgeons identify high-risk patients and reduce related complications more efficiently.

Feature Selection Using Approximated High-Order Interaction Components of the Shapley Value for Boosted Tree Classifier

The Shapely value originates from the cooperative game theory and has been widely used in solving machine learning problems due to its high interpretability and consistency. The typical Shapley value evaluates the importance score of a feature as its average marginal contribution to a fully parameterized model under all possible feature combinations; as a result, its value includes the influences from both selected features and unselected ones. To better separate the corresponding sources, it is suggested to decompose the Shapley value into high-order interaction effect components such that the influences from each individual feature can be effectively untangled. A feature's contribution is decomposed into distinct interaction components, and each component corresponds to a joint contribution resulting from a particular feature combination. The feature ranking is therefore evaluated with respect to the selected feature subset and calculated as the total incremental contribution by summing up its corresponding decomposed interaction values. In this study, a computationally efficient model-dependent greedy search algorithm on the high-order interaction components is proposed to solve an optimal subset selection problem with hundreds of time-lagged interrelated input features. Our algorithm extends a recently developed low-order polynomial time method for calculating the interaction component values. The empirical analysis demonstrates that the proposed method always outperforms other methods that are based on the typical Shapley value, the gain or the split count criteria, in terms of in-sample representativeness and out-of-sample forecasting performance for handling a problem with hundreds of time-lagged input features.

Feature evaluation and selection for condition monitoring using a self-organizing map and spatial statistics

AbstractThis paper presents a novel approach to sensor-based feature evaluation and selection using a self-organizing map and spatial statistics as a combined technique applied to tool condition monitoring of the turning process. This approach takes advantage of the unique features of unsupervised neural networks combined with spatial statistics to perform analyses into the contributions of the different sensor-based features, carrying large quantities of noise, to achieve a classification of tool wear and a quantitative measure of each feature's suitability. This method does not assume a prior direct correlation between features avoiding misconstructions inherent to common approaches that assume that only obviously correlated features should be considered for condition monitoring. Instead, and taking advantage of neural networks ability to perform non-linear modeling, it has allowed a prior modeling of the process and then analyzed each feature's contribution toward classification. It was found that some of the commonly used features have proven to have a significant contribution to the classification of cutting tool wear, whereas others adversely affect classification performance. Further, it is demonstrated that the proposed combined technique can be used extensively to quantitatively evaluate the contribution of different features toward system monitoring in the presence of noisy data.

Robust face recognition against expressions and partial occlusions

Facial features under variant-expressions and partial occlusions could have degrading effect on overall face recognition performance. As a solution, we suggest that the contribution of these features on final classification should be determined. In order to represent facial features' contribution according to their variations, we propose a feature selection process that describes facial features as local independent component analysis (ICA) features. These local features are acquired using locally lateral subspace (LLS) strategy. Then, through linear discriminant analysis (LDA) we investigate the intraclass and interclass representation of each local ICA feature and express each feature's contribution via a weighting process. Using these weights, we define the contribution of each feature at local classifier level. In order to recognize faces under single sample constraint, we implement LLS strategy on locally linear embedding (LLE) along with the proposed feature selection. Additionally, we highlight the efficiency of the implementation of LLS strategy. The overall accuracy achieved by our approach on datasets with different facial expressions and partial occlusions such as AR, JAFFE, FERET and CK+ is 90.70%. We present together in this paper survey results on face recognition performance and physiological feature selection performed by human subjects.