SHapley Additive exPlanations Values Research Articles

Evaluation of a Model to Target High-risk Psychiatric Inpatients for an Intensive Postdischarge Suicide Prevention Intervention

The months after psychiatric hospital discharge are a time of high risk for suicide. Intensive postdischarge case management, although potentially effective in suicide prevention, is likely to be cost-effective only if targeted at high-risk patients. A previously developed machine learning (ML) model showed that postdischarge suicides can be predicted from electronic health records and geospatial data, but it is unknown if prediction could be improved by adding additional information. To determine whether model prediction could be improved by adding information extracted from clinical notes and public records. Models were trained to predict suicides in the 12 months after Veterans Health Administration (VHA) short-term (less than 365 days) psychiatric hospitalizations between the beginning of 2010 and September 1, 2012 (299 050 hospitalizations, with 916 hospitalizations followed within 12 months by suicides) and tested in the hospitalizations from September 2, 2012, to December 31, 2013 (149 738 hospitalizations, with 393 hospitalizations followed within 12 months by suicides). Validation focused on net benefit across a range of plausible decision thresholds. Predictor importance was assessed with Shapley additive explanations (SHAP) values. Data were analyzed from January to August 2022. Suicides were defined by the National Death Index. Base model predictors included VHA electronic health records and patient residential data. The expanded predictors came from natural language processing (NLP) of clinical notes and a social determinants of health (SDOH) public records database. The model included 448 788 unique hospitalizations. Net benefit over risk horizons between 3 and 12 months was generally highest for the model that included both NLP and SDOH predictors (area under the receiver operating characteristic curve range, 0.747-0.780; area under the precision recall curve relative to the suicide rate range, 3.87-5.75). NLP and SDOH predictors also had the highest predictor class-level SHAP values (proportional SHAP = 64.0% and 49.3%, respectively), although the single highest positive variable-level SHAP value was for a count of medications classified by the US Food and Drug Administration as increasing suicide risk prescribed the year before hospitalization (proportional SHAP = 15.0%). In this study, clinical notes and public records were found to improve ML model prediction of suicide after psychiatric hospitalization. The model had positive net benefit over 3-month to 12-month risk horizons for plausible decision thresholds. Although caution is needed in inferring causality based on predictor importance, several key predictors have potential intervention implications that should be investigated in future studies.

Gray matter volume drives the brain age gap in schizophrenia: a SHAP study

Neuroimaging-based brain age is a biomarker that is generated by machine learning (ML) predictions. The brain age gap (BAG) is typically defined as the difference between the predicted brain age and chronological age. Studies have consistently reported a positive BAG in individuals with schizophrenia (SCZ). However, there is little understanding of which specific factors drive the ML-based brain age predictions, leading to limited biological interpretations of the BAG. We gathered data from three publicly available databases - COBRE, MCIC, and UCLA - and an additional dataset (TOPSY) of early-stage schizophrenia (82.5% untreated first-episode sample) and calculated brain age with pre-trained gradient-boosted trees. Then, we applied SHapley Additive Explanations (SHAP) to identify which brain features influence brain age predictions. We investigated the interaction between the SHAP score for each feature and group as a function of the BAG. These analyses identified total gray matter volume (group × SHAP interaction term β = 1.71 [0.53; 3.23]; pcorr < 0.03) as the feature that influences the BAG observed in SCZ among the brain features that are most predictive of brain age. Other brain features also presented differences in SHAP values between SCZ and HC, but they were not significantly associated with the BAG. We compared the findings with a non-psychotic depression dataset (CAN-BIND), where the interaction was not significant. This study has important implications for the understanding of brain age prediction models and the BAG in SCZ and, potentially, in other psychiatric disorders.

Explainable Convolutional Neural Networks Driven Knowledge Mining for Seismic Facies Classification

Seismic facies analysis is a crucial foundation for basin-fill studies and oil and gas exploration. With its rapid development, CNN-assisted interpretation is becoming increasingly popular. However, CNN models are often considered "black boxes" that lack transparency. To understand how CNN models classify seismic facies and visualize the contribution of each seismic attribute to the final predictive scoring, we have investigated class activation map (CAM) techniques and an explainable tool called Shapley additive explanations (SHAP) value. Based on real seismic data collected in the Sichuan basin, we compared the visualization performances of CAM and SHAP methods and found that the SHAP tool has better visualization capabilities than CAM methods, which only produce heat maps with positive values. Using SHAP values, we identified the importance of each seismic attribute and refined redundant attributes. This approach establishes a connection between seismic attributes and sedimentary environments and is a prime example of the capability of deep learning to discover knowledge beyond human experience. We applied the selected seismic attributes to generate a refined CNN model and compared it to the original CNN model, demonstrating the superiority of our proposed strategy. When we compared the predicted seismic facies using the refined CNN model based on SHAP features, the conventional K-means, SVM and Gaussian Naive Bayes methods, it is observed that our predicted map aligns well with geological knowledge with less prediction errors, demonstrating the effectiveness and feasibility of our developed strategy.

Determining Variables Associated with Annual Oil Palm Yield: An Explainable Gradient Boosting Approach

With the growing demands for Precision Agriculture (PA) in Indonesia, researchers have evaluated the utilization of Machine Learning for predicting oil palm yields and determining variables affecting them. Previous studies showed that meteorological variables, including rainfalls and wind speed, are associated with oil palm yields. In this research, the Extreme Gradient Boosting (XGBoost) model and the Shapley Additive exPlanations (SHAP) were deployed for analyzing the importance of 15 agrometeorological variables in predicting oil palm yield. The best model attained 1.911 RMSE and 0.855 R2. By analyzing the weights and gains of the XGBoost model along with the SHAP values, it was found that the yield in the previous year, the age and number of plants, the area of peat lands, and meteorological parameters such as rainfall rates and the number of rainy days in the previous three years were considered important. The previous year's yield in particular possesses the greatest weight and gain values according to the model, and the highest SHAP value among all input variables. However, the meteorological data used in this research are only limited to rainfall rates and the number of rainy days. In the future, more diverse variables can be analyzed.

Data-Driven Identification of Long-Term Glycemia Clusters and Their Individualized Predictors in Finnish Patients with Type 2 Diabetes.

To gain an understanding of the heterogeneous group of type 2 diabetes (T2D) patients, we aimed to identify patients with the homogenous long-term HbA1c trajectories and to predict the trajectory membership for each patient using explainable machine learning methods and different clinical-, treatment-, and socio-economic-related predictors. Electronic health records data covering primary and specialized healthcare on 9631 patients having T2D diagnosis were extracted from the North Karelia region, Finland. Six-year HbA1c trajectories were examined with growth mixture models. Linear discriminant analysis and neural networks were applied to predict the trajectory membership individually. Three HbA1c trajectories were distinguished over six years: "stable, adequate" (86.5%), "improving, but inadequate" (7.3%), and "fluctuating, inadequate" (6.2%) glycemic control. Prior glucose levels, duration of T2D, use of insulin only, use of insulin together with some oral antidiabetic medications, and use of only metformin were the most important predictors for the long-term treatment balance. The prediction model had a balanced accuracy of 85% and a receiving operating characteristic area under the curve of 91%, indicating high performance. Moreover, the results based on SHAP (Shapley additive explanations) values show that it is possible to explain the outcomes of machine learning methods at the population and individual levels. Heterogeneity in long-term glycemic control can be predicted with confidence by utilizing information from previous HbA1c levels, fasting plasma glucose, duration of T2D, and use of antidiabetic medications. In future, the expected development of HbA1c could be predicted based on the patient's unique risk factors offering a practical tool for clinicians to support treatment planning.

Quantification of the antagonistic and synergistic effects of Pb2+, Cu2+, and Zn2+ bioaccumulation by living Bacillus subtilis biomass using XGBoost and SHAP

Bioaccumulation and adsorption are efficient methods for removing heavy metal ions (HMIs) from aqueous environments. However, methods to quantifiably characterize the removal selectivity for co-existing HMIs are limited. In this study, we applied Shapley additive explanations (SHAP) following extreme gradient boosting (XGBoost) modeling, to generate SHAP values. We used these values to create an affinity interference index (AII) that quantitatively represented the interference between metal ions in a multi-metal bioaccumulation system. The selectivity for simultaneous bioaccumulation of Pb2+, Cu2+, and Zn2+ by living Bacillus subtilis biomass was then characterized as a proof of concept. The AII indicated that the bioaccumulation of Zn2+ was more strongly inhibited by Pb2+/Cu2+ (AII = 1) than that of Pb2+/Cu2+ by Zn2+. Moreover, the presence of Zn2+ promoted the bioaccumulation of Pb2+ (AII = 0.39), which was confirmed in further experiments where the bioaccumulation of Pb2+ (300 μM) was increased by 38% with Zn2+ (300 μM). This study demonstrated that the combination of XGBoost and SHAP is effective in the quantifiable characterization of the antagonistic and synergistic effects in a multi-metal simultaneous bioaccumulation system. This method could also be generalized to similar tasks for analyzing the selectivity effects in a multi-component system.

Prognostic Value of Combined Radiomic Features from Follow-Up DWI and T2-FLAIR in Acute Ischemic Stroke.

The biological pathways involved in lesion formation after an acute ischemic stroke (AIS) are poorly understood. Despite successful reperfusion treatment, up to two thirds of patients with large vessel occlusion remain functionally dependent. Imaging characteristics extracted from DWI and T2-FLAIR follow-up MR sequences could aid in providing a better understanding of the lesion constituents. We built a fully automated pipeline based on a tree ensemble machine learning model to predict poor long-term functional outcome in patients from the MR CLEAN-NO IV trial. Several feature sets were compared, considering only imaging, only clinical, or both types of features. Nested cross-validation with grid search and a feature selection procedure based on SHapley Additive exPlanations (SHAP) was used to train and validate the models. Considering features from both imaging modalities in combination with clinical characteristics led to the best prognostic model (AUC = 0.85, 95%CI [0.81, 0.89]). Moreover, SHAP values showed that imaging features from both sequences have a relevant impact on the final classification, with texture heterogeneity being the most predictive imaging biomarker. This study suggests the prognostic value of both DWI and T2-FLAIR follow-up sequences for AIS patients. If combined with clinical characteristics, they could lead to better understanding of lesion pathophysiology and improved long-term functional outcome prediction.

A machine learning approach to identifying patients with pulmonary hypertension using real-world electronic health records

BackgroundThis study aimed to develop a machine learning (ML) model to identify patients who are likely to have pulmonary hypertension (PH), using a large patient-level US-based electronic health record (EHR) database. MethodsA gradient boosting model, XGBoost, was developed using data from Optum's US-based de-identified EHR dataset (2007–2019). PH and disease control adult patients were identified using diagnostic, treatment and procedure codes and were randomly split into the training (90%) or test set (10%). Model features included patient demographics, physician visits, diagnoses, procedures, prescriptions, and laboratory test results. SHapley Additive exPlanations values were used to determine feature importance. ResultsWe identified 11,279,478 control and 115,822 PH patients (mean age, respectively: 62 and 68 years, both 53% female). The final model used 165 features, with the most important predictive features including diagnosis of heart failure, shortness of breath and atrial fibrillation. The model predicted PH with an area under the receiver operating characteristic curve (AUROC) of 0.92. AUROC remained above 0.80 for the prediction of PH up to and beyond 18 months before diagnosis. Among the PH patients, we also identified 955 pulmonary arterial hypertension (PAH) and 1432 chronic thromboembolic pulmonary hypertension (CTEPH) patients, and the range of AUROCs obtained for these cohorts was 0.79–0.90 and 0.87–0.96, respectively. ConclusionsThis model to detect PH based on patients’ EHR records is viable and performs well in subgroups of PAH and CTEPH patients. This approach has the potential to improve patient outcomes by reducing diagnostic delay in PH.

A Cluster‐Based Machine Learning Approach for Evaluating Factors Associated with Tau Protein in Plasma

AbstractBackgroundPlasma total tau (t‐tau) has shown acceptable diagnostic accuracy for Alzheimer’s disease (AD), but other factors than levels of t‐tau in the brain may influence peripheral levels. While little is known about peripheral factors that may influence plasma levels of t‐tau, such factors may be similar to those associated with AD, such as cardiovascular risk factors, which are often correlated. We have developed a machine learning pipeline to identify individual and clusters of correlated variables that might explain a proportion of the variance of plasma t‐tau levels in a well‐phenotyped population‐based cohort.MethodWe studied 992 dementia‐free older adults (mean age 76 years, 44% men) in the Age, Gene/Environment Susceptibility‐Reykjavik Study. We used agglomerative hierarchical clustering along with principal component (PC) analysis to group factors with moderate‐to‐high similarity from 290 potential predictors including socio‐demographic, anthropometric, laboratory, cardiometabolic and lifestyle factors, diseases and medications. We then calculated the PCs of the clusters and examined their relationship to plasma t‐tau. We trained two nonlinear models using t‐tau as the dependent outcome variable ‐ the extreme gradient boosting model (GBM) and support vector machine (SVM) ̶ on 80% of data using a ten‐fold cross‐validation approach; the models were evaluated on the remaining 20% of data. Model performance was assessed using the root mean square error (RMSE) and correlation. SHapley Additive exPlanations (SHAP) was used to rank and interpret the contribution of the clusters to the best‐performed model.ResultWe identified 25 distinct clusters. The best performing model was a SVM model including age, sex and the first PC from all 25 clusters(Figure 1). Based on the SHAP values in this model, we found the top 10 ranked clusters associated with p‐tau(Figure 2): platelets(5 variables), body size(33), physical activity‐walk(6), and tobacco(9) were negatively associated with t‐tau level, while body fat(8), sex, cognitive function(27), inflammation(7), blood pressure(6), and red blood cells(11) were positively associated with t‐tau.ConclusionOur novel approach suggests additional factors explain some of the variance in peripheral tau levels in older adults. Additional research is needed to confirm these findings and investigate whether these factors modulate or mediate tau‐related pathology.

Chest X-ray-based opportunistic screening of sarcopenia using deep learning.

Early detection and management of sarcopenia is of clinical importance. We aimed to develop a chest X-ray-based deep learning model to predict presence of sarcopenia. Data of participants who visited osteoporosis clinic at Severance Hospital, Seoul, South Korea, between January 2020 and June 2021 were used as derivation cohort as split to train, validation and test set (65:15:20). A community-based older adults cohort (KURE) was used as external test set. Sarcopenia was defined based on Asian Working Group 2019 guideline. A deep learning model was trained to predict appendicular lean mass (ALM), handgrip strength (HGS) and chair rise test performance from chest X-ray images; then the machine learning model (SARC-CXR score) was built using the age, sex, body mass index and chest X-ray predicted muscle parameters along with estimation uncertainty values. Mean age of the derivation cohort (n=926; women n=700, 76%; sarcopenia n=141, 15%) and the external test (n=149; women n=95, 64%; sarcopenia n=18, 12%) cohort was 61.4 and 71.6years, respectively. In the internal test set (a hold-out set, n=189, from the derivation cohort) and the external test set (n=149), the concordance correlation coefficient for ALM prediction was 0.80 and 0.76, with an average difference of 0.18±2.71 and 0.21±2.28, respectively. Gradient-weight class activation mapping for deep neural network models to predict ALM and HGS commonly showed highly weight pixel values at bilateral lung fields and part of the cardiac contour. SARC-CXR score showed good discriminatory performance for sarcopenia in both internal test set [area under the receiver-operating characteristics curve (AUROC) 0.813, area under the precision-recall curve (AUPRC) 0.380, sensitivity 0.844, specificity 0.739, F1-score 0.540] and external test set (AUROC 0.780, AUPRC 0.440, sensitivity 0.611, specificity 0.855, F1-score 0.458). Among SARC-CXR model features, predicted low ALM from chest X-ray was the most important predictor of sarcopenia based on SHapley Additive exPlanations values. Higher estimation uncertainty of HGS contributed to elevate the predicted risk of sarcopenia. In internal test set, SARC-CXR score showed better discriminatory performance than SARC-F score (AUROC 0.813 vs. 0.691, P=0.029). Chest X-ray-based deep leaning model improved detection of sarcopenia, which merits further investigation.

Interpretable Machine Learning for Inpatient COVID-19 Mortality Risk Assessments: Diabetes Mellitus Exclusive Interplay.

People with diabetes mellitus (DM) are at elevated risk of in-hospital mortality from coronavirus disease-2019 (COVID-19). This vulnerability has spurred efforts to pinpoint distinctive characteristics of COVID-19 patients with DM. In this context, the present article develops ML models equipped with interpretation modules for inpatient mortality risk assessments of COVID-19 patients with DM. To this end, a cohort of 156 hospitalised COVID-19 patients with pre-existing DM is studied. For creating risk assessment platforms, this work explores a pool of historical, on-admission, and during-admission data that are DM-related or, according to preliminary investigations, are exclusively attributed to the COVID-19 susceptibility of DM patients. First, a set of careful pre-modelling steps are executed on the clinical data, including cleaning, pre-processing, subdivision, and feature elimination. Subsequently, standard machine learning (ML) modelling analysis is performed on the cured data. Initially, a classifier is tasked with forecasting COVID-19 fatality from selected features. The model undergoes thorough evaluation analysis. The results achieved substantiate the efficacy of the undertaken data curation and modelling steps. Afterwards, SHapley Additive exPlanations (SHAP) technique is assigned to interpret the generated mortality risk prediction model by rating the predictors' global and local influence on the model's outputs. These interpretations advance the comprehensibility of the analysis by explaining the formation of outcomes and, in this way, foster the adoption of the proposed methodologies. Next, a clustering algorithm demarcates patients into four separate groups based on their SHAP values, providing a practical risk stratification method. Finally, a re-evaluation analysis is performed to verify the robustness of the proposed framework.

Abstract 13139: Explainable Machine Learning Model for Prediction of Cardiovascular Disease With Routinely Collected Clinical Data Combined With Publicly Available Data on Environmental Risk Factors

Introduction: Little is known about the performance benefit of combining these data with publicly available environmental risk factors for CVD prediction models. Hypothesis: We aimed to test whether routinely collected clinical data (age, sex, systolic blood pressure, total cholesterol, cigarette smoking) in combination with data on environmental risk factors could improve the performance of CVD prediction using an explainable machine learning (ML) technique. Methods: We identified individuals without previous history of CVD from the National Health Insurance Service, 2002-2015 in the Republic of Korea, linked to publicly available data on annual average of fine particulate matter (PM 10 ) exposure and urban green space coverage according to the administrative code for residence in each individual. Random forest (RF) model and SHapley Additive exPlanations (SHAP) value were used for prediction for newly diagnosed CVD and explanation for contribution of each component to the model output, respectively. Performance of the prediction models were evaluated using area under the curve (AUC). Results: Among 151,936 individuals included, there were 2,837 subsequent CVD events. The AUCs for the RF model with the clinical data only and the model with environmental risk factors (high annual average PM 10 and low UGS coverage) in addition to the clinical data were 0.731 and 0.733, respectively. The SHAP values showed that adding these environmental risk factors did not have significant impact on the model output (SHAP summary plot). However, the clinical data comprised of traditional CVD risk factors had notable contribution to the performance of prediction model. Conclusions: This study showed that adding publicly available data on environmental risk factors to the routinely collected clinical data had only marginal improvement in the prediction of CVD outcomes.

Spec guidance for engineering design based on data mining and neural networks

Recently, many studies on product design have been utilizing online data. They analyze user-generated online data and draw design implications. However, most of them provide customers’ tendency for feature categories rather than spec ranges for sub-features, which are crucial in industrial applications. This paper proposes an approach based on data mining and neural networks to extract spec guidance for engineering design from online data. First, product sub-features are extracted from online data, and customer choice sets are constructed. Next, a neural network choice model is trained based on these choice sets. Finally, the model is interpreted by SHAP (SHapley Additive exPlanations). In the final stage, this study proposes a method for analyzing the obtained SHAP values to draw new design implications. The suggested approach was tested on smartphone review data, and the result provides a set of recommended spec values for each sub-feature. The resultant spec guidance can help companies design a product with spec configuration preferred by customers.

An Interpretable Machine Learning Approach for Hepatitis B Diagnosis

Hepatitis B is a potentially deadly liver infection caused by the hepatitis B virus. It is a serious public health problem globally. Substantial efforts have been made to apply machine learning in detecting the virus. However, the application of model interpretability is limited in the existing literature. Model interpretability makes it easier for humans to understand and trust the machine-learning model. Therefore, in this study, we used SHapley Additive exPlanations (SHAP), a game-based theoretical approach to explain and visualize the predictions of machine learning models applied for hepatitis B diagnosis. The algorithms used in building the models include decision tree, logistic regression, support vector machines, random forest, adaptive boosting (AdaBoost), and extreme gradient boosting (XGBoost), and they achieved balanced accuracies of 75%, 82%, 75%, 86%, 92%, and 90%, respectively. Meanwhile, the SHAP values showed that bilirubin is the most significant feature contributing to a higher mortality rate. Consequently, older patients are more likely to die with elevated bilirubin levels. The outcome of this study can aid health practitioners and health policymakers in explaining the result of machine learning models for health-related problems.

Development of an interpretable machine learning model associated with heavy metals’ exposure to identify coronary heart disease among US adults via SHAP: Findings of the US NHANES from 2003 to 2018

Limited information is available on the links between heavy metals' exposure and coronary heart disease (CHD). We aim to establish an efficient and explainable machine learning (ML) model that associates heavy metals' exposure with CHD identification. Our datasets for investigating the associations between heavy metals and CHD were sourced from the US National Health and Nutrition Examination Survey (US NHANES, 2003–2018). Five ML models were established to identify CHD by heavy metals' exposure. Further, 11 discrimination characteristics were used to test the strength of the models. The optimally performing model was selected for identification. Finally, the SHapley Additive exPlanations (SHAP) tool was used for interpreting the features to visualize the selected model's decision-making capacity. In total, 12,554 participants were eligible for this study. The best performing random forest classifier (RF) based on 13 heavy metals to identify CHD was chosen (AUC: 0.827; 95%CI: 0.777–0.877; accuracy: 95.9%). SHAP values indicated that cesium (1.62), thallium (1.17), antimony (1.63), dimethylarsonic acid (0.91), barium (0.76), arsenous acid (0.79), total arsenic (0.01) in urine, and lead (3.58) and cadmium (4.66) in blood positively contributed to the model, while cobalt (−0.15), cadmium (−2.93), and uranium (−0.13) in urine negatively contributed to the model. The RF model was efficient, accurate, and robust in identifying an association between heavy metals' exposure and CHD among US NHANES 2003–2018 participants. Cesium, thallium, antimony, dimethylarsonic acid, barium, arsenous acid, and total arsenic in urine, and lead and cadmium in blood show positive relationships with CHD, while cobalt, cadmium, and uranium in urine show negative relationships with CHD.

New onset delirium prediction using machine learning and long short-term memory (LSTM) in electronic health record.

To develop and test an accurate deep learning model for predicting new onset delirium in hospitalized adult patients. Using electronic health record (EHR) data extracted from a large academic medical center, we developed a model combining long short-term memory (LSTM) and machine learning to predict new onset delirium and compared its performance with machine-learning-only models (logistic regression, random forest, support vector machine, neural network, and LightGBM). The labels of models were confusion assessment method (CAM) assessments. We evaluated models on a hold-out dataset. We calculated Shapley additive explanations (SHAP) measures to gauge the feature impact on the model. A total of 331489 CAM assessments with 896 features from 34035 patients were included. The LightGBM model achieved the best performance (AUC 0.927 [0.924, 0.929] and F1 0.626 [0.618, 0.634]) among the machine learning models. When combined with the LSTM model, the final model's performance improved significantly (P = .001) with AUC 0.952 [0.950, 0.955] and F1 0.759 [0.755, 0.765]. The precision value of the combined model improved from 0.497 to 0.751 with a fixed recall of 0.8. Using the mean absolute SHAP values, we identified the top 20 features, including age, heart rate, Richmond Agitation-Sedation Scale score, Morse fall risk score, pulse, respiratory rate, and level of care. Leveraging LSTM to capture temporal trends and combining it with the LightGBM model can significantly improve the prediction of new onset delirium, providing an algorithmic basis for the subsequent development of clinical decision support tools for proactive delirium interventions.

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.

Evidence for residential building retrofitting practices using explainable AI and socio-demographic data

Extensive retrofits and effective policy measures are needed to meet the ambitious climate goals, particularly in the UK, with the EU’s oldest residential building stock. Researchers must investigate the factors influencing retrofits to enable effective and targeted policy measures. To date, however, there is a lack of holistically large-scale quantitative studies accounting for such factors. At the same time, great potential is seen in data-driven solutions and the use of explainable artificial intelligence (XAI). We address this research gap by combining supervised machine learning with XAI employing a three-stage approach: First, we consolidate datasets of Energy Performance Certificates from England and Wales from which we extract conducted retrofits, house prices, and socio-demographic information. Second, we apply an eXtreme Gradient Boosting (XGBoost) model that predicts whether a building has been retrofitted or not. Lastly, we use SHapley Additive exPlanations values (SHAP) as an XAI technique to identify the key factors and relationships that influence the implementation of retrofits. We succeed in substantiating results previously obtained in qualitative or small-scale studies and also find that retrofit-related policies already implemented in regional cases, such as the ”Better Homes for Yorkshire” initiative, can successfully achieve large-scale success through replication in other regions. Further, our results suggest the implementation of income-based CO2 taxes as a reasonable and easy-to-implement policy measure.

Turnaround time prediction for clinical chemistry samples using machine learning.

Turnaround time (TAT) is an essential performance indicator of a medical diagnostic laboratory. Accurate TAT prediction is crucial for taking timely action in case of prolonged TAT and is important for efficient organization of healthcare. The objective was todevelop a model to accurately predict TAT, focusing on the automated pre-analytical and analytical phase. A total of 90,543 clinical chemistry samples from Erasmus MC were included and 39 features were analyzed, including priority level and workload in the different stages upon sample arrival. PyCaret was used toevaluate and compare multiple regression models, including the Extra Trees (ET) Regressor, Ridge Regression and K Neighbors Regressor, to determine the best model forTAT prediction. The relative residual and SHAP (SHapley Additive exPlanations) values were plotted formodel evaluation. The regression-tree-based method ET Regressor performed best with an R2 of 0.63, a mean absolute error of2.42min and a mean absolute percentage error of 7.35%,where the average TAT was 30.09min. Of the test setsamples, 77% had a relative residual error of at most 10%. SHAP value analysis indicated that TAT was mainly influenced by the workload in pre-analysis upon sample arrival and the number of modules visited. Accurate TAT predictions were attained with the ET Regressor and features with the biggest impact on TAT were identified, enabling the laboratory to take timely action in case of prolonged TAT and helping healthcare providers to improve planning of scarce resources to increase healthcare efficiency.

Climatic water balance forecasting with machine learning and deep learning models over Bangladesh

AbstractUnderstanding the impact of input variables on black‐box machine learning and deep learning models is necessary. Therefore, this study proposed SHapley Additive exPlanations (SHAP) values to address the problem of the interpretability of the output of the support vector machine (SVM), random forest (RF), convolutional neural network (CNN), and long short‐term memory (LSTM) models to forecast climatic water balance (CWB) within 1–3 months lead‐time. The current study uses two Koppen–Geiger climate zones over Bangladesh: the humid subtropical climate with dry winter and hot summer (Cwa) and the tropical climate (Af‐Am). Monthly antecedent CWB, potential evapotranspiration (PET), convective available potential energy (CAPE), relative humidity (RH), air temperature (TEM), El Niño–Southern Oscillation (ENSO), Indian Ocean Dipole (IOD), and North Atlantic oscillation (NAO) are used as the inputs to the proposed model. SHAP values show that CWB (current month) for both climate zones has the maximum impact on the prediction. The effects of atmospheric variables and ocean–atmospheric teleconnections on CWB forecasts over Bangladesh are low. Forecasting results show that the SVM model shows the best performance for CWB forecasts in the Cwa and Af‐Am climate zones in terms of antecedent CWB as input. And this model effectively decreases the negative effect of increasing forecast lead‐time. Since the proposed model can predict CWB in 3‐months lead‐time, it would help policy‐makers and practitioners to reduce the drought and flood impacts in the future by adopting better preparedness plans.