SHapley Additive exPlanations Values Research Articles

Deep learning based data-driven model for detecting time-delay water quality indicators of wastewater treatment plant influent

Rapid and accurate detection of time-delayed water quality indicators (WQIs) is the key to achieving fast feedback regulation of wastewater treatment plants (WWTPs) that enables its energy-efficient operation and high tolerance towards shock sewage loads. However, advanced oxidation methods are costly, and data-driven modeling methods based on traditional machine learning algorithms for detecting time-delayed WQIs have limited detection accuracy. This work develops deep learning models based on long short-term memory (LSTM) neural networks to detect time-delayed WQIs in WWTPs intake accurately. The lack of interpretability of the deep learning models hampers the optimization of the developed LSTM models in applications. Therefore, a global sensitivity analysis (GSA) based on Shapley additive explanations (SHAP) is performed to quantify the contribution of the input indicators to detection results of the developed LSTM models. The direct contributions provide the basis for optimizing the input indicators to achieve more cost-effective modeling detection. In the case study, the developed LSTM models achieved good accuracy (R2 of 0.9141, 0.9239, and 0.9040, respectively) in detecting chemical oxygen demand, total nitrogen, and total phosphorus in the influent of a WWTP, outperforming the four types of baseline models. According to the SHAP values, the contributions of dissolved oxygen, turbidity, and ammonia nitrogen to the above detection targets are always in the top third of all input indicators, which are more outstanding than meteorological indicators. Removing the indicator with the smallest SHAP value reduces the build and run costs of the models with minimal loss of detection accuracy. Combining deep learning and GSA to detect WWTPs influent is a novel and effective attempt. This attempt provides a more sustainable solution for rapid and accurate detection of time-delayed WQIs, which drives WWTPs' operation in an intelligent, clean, and safe direction.

PPSW-SHAP: Towards Interpretable Cell Classification Using Tree-Based SHAP Image Decomposition and Restoration for High-Throughput Bright-Field Imaging.

Advancements in high-throughput microscopy imaging have transformed cell analytics, enabling functionally relevant, rapid, and in-depth bioanalytics with Artificial Intelligence (AI) as a powerful driving force in cell therapy (CT) manufacturing. High-content microscopy screening often suffers from systematic noise, such as uneven illumination or vignetting artifacts, which can result in false-negative findings in AI models. Traditionally, AI models have been expected to learn to deal with these artifacts, but success in an inductive framework depends on sufficient training examples. To address this challenge, we propose a two-fold approach: (1) reducing noise through an image decomposition and restoration technique called the Periodic Plus Smooth Wavelet transform (PPSW) and (2) developing an interpretable machine learning (ML) platform using tree-based Shapley Additive exPlanations (SHAP) to enhance end-user understanding. By correcting artifacts during pre-processing, we lower the inductive learning load on the AI and improve end-user acceptance through a more interpretable heuristic approach to problem solving. Using a dataset of human Mesenchymal Stem Cells (MSCs) cultured under diverse density and media environment conditions, we demonstrate supervised clustering with mean SHAP values, derived from the 'DFT Modulus' applied to the decomposition of bright-field images, in the trained tree-based ML model. Our innovative ML framework offers end-to-end interpretability, leading to improved precision in cell characterization during CT manufacturing.

Machine learning algorithms identify hypokalaemia risk in people with hypertension in the United States National Health and Nutrition Examination Survey 1999–2018

Background Hypokalaemia is a side-effect of diuretics. We aimed to use machine learning to identify features predicting hypokalaemia risk in hypertensive patients. Methods Participants with hypertension in the United States National Health and Nutrition Examination Survey 1999–2018 were included for analysis. To select the most suitable algorithm, we tested and evaluated five machine learning algorithms commonly employed in epidemiological studies: Logistic Regression, k-Nearest Neighbor, Random Forest, Recursive Partitioning and Regression Trees, and eXtreme Gradient Boosting. These algorithms were accessed using a set of 38 screened features. We then selected the key hypokalaemia-associated features in the hypertension group and their cardiovascular diseases (CVD) subgroup using the SHapley Additive exPlanations (SHAP) values. Using SHAP values, the key features and their impact pattern on hypokalaemia risk were determined. Results A total of 25,326 hypertensive participants were included for analysis, of whom 4,511 had known CVD. The Random Forest algorithm had the highest AUROC (hypertension dataset: 0.73 [95%CI, 0.71–0.76]; CVD subgroup: 0.72 [95%CI, 0.66–0.78]). Moreover, the nomogram based on the top twelve key features screened by random forest retained good performance: age, sex, race, poverty income ratio, body mass index, systolic and diastolic blood pressure, non-potassium-sparing diuretics use and duration, renin-angiotensin blockers use and duration, and CVD history in hypertension dataset; while in CVD subgroup, the additional key features were comorbid diabetes, education level, smoking status, and use of bronchodilators. Conclusion Our predictive model based on the random forest algorithm performed best among the tested and evaluated five algorithms. Hypokalaemia-associated key features have been identified in hypertensive patients and the subgroup with CVD. These findings from machine learning facilitate the development of artificial intelligence to highlight hypokalaemia risk in hypertension patients. Key messages: Our predictive model based on the random forest algorithm performed best among the tested and evaluated five algorithms, and hypokalemia-associated key features have been identified in hypertensive patients and the subgroup with cardiovascular disease. The nomogram we developed including twelve key features might be useful and applied in primary clinical consultations to identify the hypertensive patients at risk of hypokalaemia. These findings from machine learning facilitate the development of artificial intelligence to highlight hypokalaemia risk in hypertension patients

Modeling land susceptibility to wind erosion hazards using LASSO regression and graph convolutional networks

Predicting land susceptibility to wind erosion is necessary to mitigate the negative impacts of erosion on soil fertility, ecosystems, and human health. This study is the first attempt to model wind erosion hazards through the application of a novel approach, the graph convolutional networks (GCNs), as deep learning models with Monte Carlo dropout. This approach is applied to Semnan Province in arid central Iran, an area vulnerable to dust storms and climate change. We mapped 15 potential factors controlling wind erosion, including climatic variables, soil characteristics, lithology, vegetation cover, land use, and a digital elevation model (DEM), and then applied the least absolute shrinkage and selection operator (LASSO) regression to discriminate the most important factors. We constructed a predictive model by randomly selecting 70% and 30% of the pixels, as training and validation datasets, respectively, focusing on locations with severe wind erosion on the inventory map. The current LASSO regression identified eight out of the 15 features (four soil property categories, vegetation cover, land use, wind speed, and evaporation) as the most important factors controlling wind erosion in Semnan Province. These factors were adopted into the GCN model, which estimated that 15.5%, 19.8%, 33.2%, and 31.4% of the total area is characterized by low, moderate, high, and very high susceptibility to wind erosion, respectively. The area under curve (AUC) and SHapley Additive exPlanations (SHAP) of game theory were applied to assess the performance and interpretability of GCN output, respectively. The AUC values for training and validation datasets were estimated at 97.2% and 97.25%, respectively, indicating excellent model prediction. SHAP values ranged between −0.3 and 0.4, while SHAP analyses revealed that the coarse clastic component, vegetation cover, and land use were the most effective features of the GCN output. Our results suggest that this novel suite of methods is highly recommended for future spatial prediction of wind erosion hazards in other arid environments around the globe.

Artificial Intelligence (AI) based approach for identifying biomarkers associated with neutralizing antidrug antibodies to Factor VIII used in Hemophilia A treatment

Abstract Lack of clinically validated markers associated with the development of neutralizing antibodies (inhibitors) to the replacement protein-drug, Factor VIII (FVIII) is a major challenge in managing Hemophilia A (HA). By utilizing Machine Learning (ML) and Explainable AI (XAI) we identified and ranked relevant variables (biomarkers) that could be predictive for developing inhibitors to FVIII drugs in hemophilia A patients. The dataset used for the study was derived from the My Life Our Future (MLOF) repository and included variables such as age, race, ethnicity, mutations in the F8 gene and the Human Leucocyte Antigen Class II (HLA-II) type. We computed the number of foreign FVIII derived peptides, based on the alignment of the endogenous FVIII and infused drug sequences, and the foreign-peptide HLA-II molecule binding affinity (using the NetMHCIIpan algorithm). The complete dataset with the added variables was trained and validated using multiple ML classification models to identify the top performing model, which was chosen for further processing with XAI via SHAP (SHapley Additive exPlanations) to identify the variables critical for the prediction of FVIII inhibitor development in a hemophilia A patients. The top five variables for predicting inhibitor development based on SHAP values are: (i) the baseline activity of the FVIII protein, (ii) mean affinity of all foreign peptides for HLA DRB 3, 4, & 5 alleles, (iii) mean affinity of all foreign peptides for HLA DRB1 alleles, (iv) the maximum affinity among all foreign peptides for HLA DRB1 alleles, and (v) F8 mutation type. This research was supported by funds from the Hemostasis Branch/Division of Plasma Protein 344 Therapeutics/Office of Tissues and Advanced Therapies/Center for Biologics Evaluation and Research 345 of the U.S. Food and Drug Administration and in part by an appointment to the Research Participation 346 Program at the Center for Biologics Evaluation and Research administered by the Oak Ridge Institute 347 for Science and Education through an interagency agreement between the U.S. Department of Energy 348 and the U.S. Food and Drug Administration. The MLOF program was developed as a partnership 349 between NHF, ATHN, Bloodworks Northwest, and Bioverativ and supported financially by 350 Bioverativ, NHF, Bloodworks Northwest, and ATHN.

The prediction of distant metastasis risk for male breast cancer patients based on an interpretable machine learning model

ObjectivesThis research was designed to compare the ability of different machine learning (ML) models and nomogram to predict distant metastasis in male breast cancer (MBC) patients and to interpret the optimal ML model by SHapley Additive exPlanations (SHAP) framework.MethodsFour powerful ML models were developed using data from male breast cancer (MBC) patients in the SEER database between 2010 and 2015 and MBC patients from our hospital between 2010 and 2020. The area under curve (AUC) and Brier score were used to assess the capacity of different models. The Delong test was applied to compare the performance of the models. Univariable and multivariable analysis were conducted using logistic regression.ResultsOf 2351 patients were analyzed; 168 (7.1%) had distant metastasis (M1); 117 (5.0%) had bone metastasis, and 71 (3.0%) had lung metastasis. The median age at diagnosis is 68.0 years old. Most patients did not receive radiotherapy (1723, 73.3%) or chemotherapy (1447, 61.5%). The XGB model was the best ML model for predicting M1 in MBC patients. It showed the largest AUC value in the tenfold cross validation (AUC:0.884; SD:0.02), training (AUC:0.907; 95% CI: 0.899—0.917), testing (AUC:0.827; 95% CI: 0.802—0.857) and external validation (AUC:0.754; 95% CI: 0.739—0.771) sets. It also showed powerful ability in the prediction of bone metastasis (AUC: 0.880, 95% CI: 0.856—0.903 in the training set; AUC: 0.823, 95% CI:0.790—0.848 in the test set; AUC: 0.747, 95% CI: 0.727—0.764 in the external validation set) and lung metastasis (AUC: 0.906, 95% CI: 0.877—0.928 in training set; AUC: 0.859, 95% CI: 0.816—0.891 in the test set; AUC: 0.756, 95% CI: 0.732—0.777 in the external validation set). The AUC value of the XGB model was larger than that of nomogram in the training (0.907 vs 0.802) and external validation (0.754 vs 0.706) sets.ConclusionsThe XGB model is a better predictor of distant metastasis among MBC patients than other ML models and nomogram; furthermore, the XGB model is a powerful model for predicting bone and lung metastasis. Combining with SHAP values, it could help doctors intuitively understand the impact of each variable on outcome.

Potential of Coupling Metaheuristics-Optimized-XGBoost and SHAP in Revealing PAHs Environmental Fate.

Polycyclic aromatic hydrocarbons (PAHs) refer to a group of several hundred compounds, among which 16 are identified as priority pollutants, due to their adverse health effects, frequency of occurrence, and potential for human exposure. This study is focused on benzo(a)pyrene, being considered an indicator of exposure to a PAH carcinogenic mixture. For this purpose, we have applied the XGBoost model to a two-year database of pollutant concentrations and meteorological parameters, with the aim to identify the factors which were mostly associated with the observed benzo(a)pyrene concentrations and to describe types of environments that supported the interactions between benzo(a)pyrene and other polluting species. The pollutant data were collected at the energy industry center in Serbia, in the vicinity of coal mining areas and power stations, where the observed benzo(a)pyrene maximum concentration for a study period reached 43.7 ngm-3. The metaheuristics algorithm has been used to optimize the XGBoost hyperparameters, and the results have been compared to the results of XGBoost models tuned by eight other cutting-edge metaheuristics algorithms. The best-produced model was later on interpreted by applying Shapley Additive exPlanations (SHAP). As indicated by mean absolute SHAP values, the temperature at the surface, arsenic, PM10, and total nitrogen oxide (NOx) concentrations appear to be the major factors affecting benzo(a)pyrene concentrations and its environmental fate.

Supervised learning applied to classifying fallers versus non-fallers among older adults with cancer

IntroductionSupervised machine learning approaches are increasingly used to analyze clinical data, including in geriatric oncology. This study presents a machine learning approach to understand falls in a cohort of older adults with advanced cancer starting chemotherapy, including fall prediction and identification of contributing factors. Materials and MethodsThis secondary analysis of prospectively collected data from the GAP 70+ Trial (NCT02054741; PI: Mohile) enrolled patients aged ≥70 with advanced cancer and ≥ 1 geriatric assessment domain impairment who planned to start a new cancer treatment regimen. Of ≥2000 baseline variables (“features”) collected, 73 were selected based on clinical judgment. Machine learning models to predict falls at three months were developed, optimized, and tested using data from 522 patients. A custom data preprocessing pipeline was implemented to prepare data for analysis. Both undersampling and oversampling techniques were applied to balance the outcome measure. Ensemble feature selection was applied to identify and select the most relevant features. Four models (logistic regression [LR], k-nearest neighbor [kNN], random forest [RF], and MultiLayer Perceptron [MLP]) were trained and subsequently tested on a holdout set. Receiver operating characteristic (ROC) curves were generated and area under the curve (AUC) was calculated for each model. SHapley Additive exPlanations (SHAP) values were utilized to further understand individual feature contributions to observed predictions. ResultsBased on the ensemble feature selection algorithm, the top eight features were selected for inclusion in the final models. Selected features aligned with clinical intuition and prior literature. The LR, kNN, and RF models performed equivalently well in predicting falls in the test set, with AUC values 0.66–0.67, and the MLP model showed AUC 0.75. Ensemble feature selection resulted in improved AUC values compared to using LASSO alone. SHAP values, a model-agnostic technique, revealed logical associations between selected features and model predictions. DiscussionMachine learning techniques can augment hypothesis-driven research, including in older adults for whom randomized trial data are limited. Interpretable machine learning is particularly important, as understanding which features impact predictions is a critical aspect of decision-making and intervention. Clinicians should understand the philosophy, strengths, and limitations of a machine learning approach applied to patient data.

Explainable Machine-Learning Predictions for Peak Ground Acceleration

Peak ground acceleration (PGA) prediction is of great significance in the seismic design of engineering structures. Machine learning is a new method to predict PGA and does have some advantages. To establish explainable prediction models of PGA, 3104 sets of uphole and downhole seismic records collected by the KiK-net in Japan were used. The feature combinations that make the models perform best were selected through feature selection. The peak bedrock acceleration (PBA), the predominant frequency (FP), the depth of the soil when the shear wave velocity reaches 800 m/s (D800), and the bedrock shear wave velocity (Bedrock Vs) were used as inputs to predict the PGA. The XGBoost (eXtreme Gradient Boosting), random forest, and decision tree models were established, and the prediction results were compared with the numerical simulation results The influence between the input features and the model prediction results were analyzed with the SHAP (SHapley Additive exPlanations) value. The results show that the R2 of the training dataset and testing dataset reach up to 0.945 and 0.915, respectively. On different site classifications and different PGA intervals, the prediction results of the XGBoost model are better than the random forest model and the decision tree model. Even if a non-integrated algorithm (decision tree model) is used, its prediction effect is better than the numerical simulation methods. The SHAP values of the three machine learning models have the same distribution and densities, and the influence of each feature on the prediction results is consistent with the existing empirical data, which shows the rationality of the machine learning models and provides reliable support for the prediction results.

Interpretable & Explainable Machine Learning for Ultrasonic Defect Sizing.

Despite its popularity in literature, there are few examples of machine learning (ML) being used for industrial nondestructive evaluation (NDE) applications. A significant barrier is the 'black box' nature of most ML algorithms. This paper aims to improve the interpretability and explainability of ML for ultrasonic NDE by presenting a novel dimensionality reduction method: Gaussian feature approximation (GFA). GFA involves fitting a 2D elliptical Gaussian function an ultrasonic image and storing the seven parameters that describe each Gaussian. These seven parameters can then be used as inputs to data analysis methods such as the defect sizing neural network presented in this paper. GFA is applied to ultrasonic defect sizing for inline pipe inspection as an example application. This approach is compared to sizing with the same neural network, and two other dimensionality reduction methods (the parameters of 6 dB drop boxes and principal component analysis), as well as a convolutional neural network applied to raw ultrasonic images. Of the dimensionality reduction methods tested, GFA features produce the closest sizing accuracy to sizing from the raw images, with only a 23% increase in RMSE, despite a 96.5% reduction in the dimensionality of the input data. Implementing ML with GFA is implicitly more interpretable than doing so with principal component analysis or raw images as inputs, and gives significantly more sizing accuracy than 6 dB drop boxes. Shapley additive explanations (SHAP) are used to calculate how each feature contributes to the prediction of an individual defect's length. Analysis of SHAP values demonstrates that the GFA-based neural network proposed displays many of the same relationships between defect indications and their predicted size as occur in traditional NDE sizing methods.

Machine learning algorithm to predict mortality in critically ill patients with sepsis-associated acute kidney injury

This study aimed to establish and validate a machine learning (ML) model for predicting in-hospital mortality in patients with sepsis-associated acute kidney injury (SA-AKI). This study collected data on SA-AKI patients from 2008 to 2019 using the Medical Information Mart for Intensive Care IV. After employing Lasso regression for feature selection, six ML approaches were used to build the model. The optimal model was chosen based on precision and area under curve (AUC). In addition, the best model was interpreted using SHapley Additive exPlanations (SHAP) values and Local Interpretable Model-Agnostic Explanations (LIME) algorithms. There were 8129 sepsis patients eligible for participation; the median age was 68.7 (interquartile range: 57.2–79.6) years, and 57.9% (4708/8129) were male. After selection, 24 of the 44 clinical characteristics gathered after intensive care unit admission remained linked with prognosis and were utilized developing ML models. Among the six models developed, the eXtreme Gradient Boosting (XGBoost) model had the highest AUC, at 0.794. According to the SHAP values, the sequential organ failure assessment score, respiration, simplified acute physiology score II, and age were the four most influential variables in the XGBoost model. Individualized forecasts were clarified using the LIME algorithm. We built and verified ML models that excel in early mortality risk prediction in SA-AKI and the XGBoost model performed best.

Predicting future depressive episodes from resting-state fMRI with generative embedding

After a first episode of major depressive disorder (MDD), there is substantial risk for a long-term remitting-relapsing course. Prevention and early interventions are thus critically important. Various studies have examined the feasibility of detecting at-risk individuals based on out-of-sample predictions about the future occurrence of depression. However, functional magnetic resonance imaging (fMRI) has received very little attention for this purpose so far.Here, we explored the utility of generative models (i.e. different dynamic causal models, DCMs) as well as functional connectivity (FC) for predicting future episodes of depression in never-depressed adults, using a large dataset (N = 906) of task-free ("resting state") fMRI data from the UK Biobank (UKB). Connectivity analyses were conducted using timeseries from pre-computed spatially independent components of different dimensionalities. Over a three-year period, 50% of selected participants showed indications of at least one depressive episode, while the other 50% did not. Using nested cross-validation for training and a held-out test set (80/20 split), we systematically examined the combination of 8 connectivity feature sets and 17 classifiers. We found that a generative embedding procedure based on combining regression DCM (rDCM) with a support vector machine (SVM) enabled the best predictions, both on the training set (0.63 accuracy, 0.66 area under the curve, AUC) and the test set (0.62 accuracy, 0.64 AUC; p < 0.001). However, on the test set, rDCM was only slightly superior to predictions based on FC (0.59 accuracy, 0.61 AUC). Interpreting model predictions based on SHAP (SHapley Additive exPlanations) values suggested that the most predictive connections were widely distributed and not confined to specific networks. Overall, our analyses suggest (i) ways of improving future fMRI-based generative embedding approaches for the early detection of individuals at-risk for depression and that (ii) achieving accuracies of clinical utility may require combination of fMRI with other data modalities.

Attitudes of Different Religions Toward Surrogacy: Analysis of 11 Countries' Situation Using Machine Learning Approach and Artificial Neural Networks.

Individuals may develop different attitudes on bioethics in general and reproductive ethics in particular, due to the effects of different sociocultural environments. Individuals' attitudes toward surrogacy are affected positively or negatively depending on religious and cultural environments. This study was conducted to determine and compare the attitudes of different religions toward surrogacy. This study is cross-sectional and collected from individuals living in Turkey, India, Iran, the Turkish Republic of Northern Cyprus, Madagascar, Nepal, Nigeria, Pakistan, Mexico, England, and Japan between May 2022 and December 2022. The study was conducted with individuals belonging to Islam, Christianity, Hinduism, Buddhism, and Atheism. The study was conducted with 1177 individuals from different religions who agreed to participate in the study by snowball sampling method. The introductory Information Form and "Attitude Questionnaire Toward Surrogacy" were used as data collection tools. R programming language 4.1.3 was used for regression analysis with machine learning approach and artificial neural networks, and SPSS-25 was used for other statistical analyses. There was a significant difference between the total mean score of the individuals' Attitudes toward Surrogacy Questionnaire and their religious beliefs (p < 0.05). When the results of the analysis of the regression model with the dummy variable, which was carried out with the aim of revealing the effects of religious belief on the attitude toward surrogacy, are examined, statistical estimates of the regression model show that the model is significant and usable F(4,1172) = 5.005, p = 0.001). It explains 1.7% of the total variance of the level of religious belief's attitude toward surrogacy. In the regression model, when the t-test results regarding the significance of the regression coefficient are examined, among the participants, it was determined that the mean score of those who believed in Islam (t = - 3.827, p < 0.001) and those who believed in Christianity (t = - 2.548, p < 0.001) was lower than the mean score of those who believed in Hinduism (Constant) (p < 0.05). Individuals' attitudes toward surrogacy differ according to their religion. The best performing algorithm for the prediction model was random forest (RF) regression. The contributions of the variables to the model were calculated with Shapley values (Shapley Additive Explanations (SHAP)). The SHAP values of the variables in the best performing model were examined to avoid bias in terms of comparison in the performance criterion. SHAP values (Shapley Additive Explanations) show the contribution or importance of each variable in the estimation of the model. It is determined that the most important variable that should be in the model to predict the Attitude Toward Surrogacy Survey variable is the Nationality variable. It is recommended that studies on attitudes toward surrogacy should be conducted by taking religious and cultural values into consideration.

Generative models as computational assays for psychiatry

AbstractPsychiatry faces fundamental challenges with regard to mechanistically guided differential diagnosis, as well as prediction of clinical trajectories and treatment response. This has motivated novel approaches that aim to develop “computational assays” for inferring patient-specific disease processes from neuroimaging data, which can then be incorporated into decision making in everyday clinical practice. Such computational assays are often based on generative models, which describe how measured data may be caused by a particular mechanism. Combining generative models with machine learning allows translating the inferences from computational assays into patient-specific predictions, an approach referred to as generative embedding.Here, I illustrate the clinical potential of generative embedding for the exemplary case of a generative model of whole-brain effective (directed) connectivity: regression DCM (rDCM). First, I introduce rDCM to the audience and highlight its relevance for understanding the pathophysiology of psychiatric disorders. I then provide an initial demonstration of the clinical utility of rDCM. Specifically, we assessed the ability of rDCM for predicting future episodes of depression in never-depressed adults, using a large dataset (N=906) of resting-state fMRI data from the UK Biobank. Over a 3-year period, half of the participants showed indications of at least one depressive episode, while the other half did not. Using nested cross-validation for training and a held-out test set (80/20 split), we found that a generative embedding procedure based on rDCM in combination with a support vector machine enables statistically significant predictions of future depressive episodes, both on the training (accuracy: 0.63, area under the curve (AUC): 0.66, p&lt;0.001) and test set (accuracy: 0.62, AUC: 0.64, p&lt;0.001). Interpreting model predictions based on SHAP (Shapley Additive exPlanations) values suggested that the most predictive connections were widely distributed and not confined to specific networks.In summary, generative models of brain connectivity in general, and rDCM in particular, show initial promise to serve as computational assays for psychiatry. Our analyses suggest that (i) fMRI-based generative embedding approaches have some capacity for early detection of individuals at-risk for depression and (ii) achieving accuracies of clinical utility may require combination of fMRI with other data modalities.Disclosure of InterestNone Declared

Analyzing the operation reliability of aeroengine using Quick Access Recorder flight data

Aeroengine operation reliability (AOR) estimation is important for stakeholders in operating, monitoring, designing, and improving. The Quick Access Recorder (QAR) flight data and failure rate of aeroengine are utilized to analyze AOR. Considering the uncertainty in AOR assessment, a Bayesian neural network (BNN) is trained to evaluate and forecast AOR based on aeroengine status data within a confidence interval. Further, to quantify the degree of each feature on AOR, Shapley Additive ex-Planations (SHAP) values are calculated based on the Light gradient boosting machine (LightGBM) to study the degree and direction of influence feature on AOR. In this study, it is revealed that (i) AOR is closely related to the airplane flight stages; and (ii) after training with eight flights and validation with two flights data from QAR data, BNN can achieve AOR analysis and prediction within a certain confidence interval while obtaining aeroengine state data; and (iii) the feature importance and influence direction are quantified by SHAP values, it demonstrates the sensitive factors in AOR analysis. Based QAR data, this study provide an AOR analysis framework to improve the operation and design, which has the potential to support aeroengine real-time status monitoring and health management.

To Establish an Early Prediction Model for Acute Respiratory Distress Syndrome in Severe Acute Pancreatitis Using Machine Learning Algorithm.

To develop binary and quaternary classification prediction models in patients with severe acute pancreatitis (SAP) using machine learning methods, so that doctors can evaluate the risk of patients with acute respiratory distress syndrome (ARDS) and severe ARDS at an early stage. A retrospective study was conducted on SAP patients hospitalized in our hospital from August 2017 to August 2022. Logical Regression (LR), Random Forest (RF), Support Vector Machine (SVM), Decision Tree (DT), and eXtreme Gradient Boosting (XGB) were used to build the binary classification prediction model of ARDS. Shapley Additive explanations (SHAP) values were used to interpret the machine learning model, and the model was optimized according to the interpretability results of SHAP values. Combined with the optimized characteristic variables, four-class classification models, including RF, SVM, DT, XGB, and Artificial Neural Network (ANN), were constructed to predict mild, moderate, and severe ARDS, and the prediction effects of each model were compared. The XGB model showed the best effect (AUC = 0.84) in the prediction of binary classification (ARDS or non-ARDS). According to SHAP values, the prediction model of ARDS severity was constructed with four characteristic variables (PaO2/FiO2, APACHE II, SOFA, AMY). Among them, the overall prediction accuracy of ANN is 86%, which is the best. Machine learning has a good effect in predicting the occurrence and severity of ARDS in SAP patients. It can also provide a valuable tool for doctors to make clinical decisions.

Operation reliability monitoring towards fault diagnosis of airplane hydraulic system using Quick Access Recorder flight data

Hydraulic system operation reliability (HSOR) can evaluate time series state reliability for hydraulic system fault diagnosis and provide condition based maintenance decisions. The quick access recorder (QAR) flight data and normal values of the hydraulic system are utilized to analyze time series HSOR by calculating the operation reliability index. Considering the relationship of the hydraulic subsystem among the components, hydraulic components Bayesian Network is constructed to analyze time series HSOR. Furthermore, the sensitivity of HSOR features to fault location is assessed using categorical boosting (CatBoost) and Shapley Additive ex-Planations values. Through the analysis of two flights hydraulic system QAR datasets, it is revealed that (a) HSOR can accurately monitor the time series operating states of the hydraulic system; and (b) with demonstrating two illustrative case, the HSOR values and features sensitivity analysis can be a useful reference for the fault diagnosis and location of the airplane hydraulic system. The study intends to develop a practical reference approach for hydraulic system fault diagnosis and location using QAR data.

Prediction of Suicidal Behaviors in the Middle-aged Population: Machine Learning Analyses of UK Biobank.

Suicidal behaviors, including suicide deaths and attempts, are major public health concerns. However, previous suicide models required a huge amount of input features, resulting in limited applicability in clinical practice. We aimed to construct applicable models (ie, with limited features) for short- and long-term suicidal behavior prediction. We further validated these models among individuals with different genetic risks of suicide. Based on the prospective cohort of UK Biobank, we included 223 (0.06%) eligible cases of suicide attempts or deaths, according to hospital inpatient or death register data within 1 year from baseline and randomly selected 4460 (1.18%) controls (1:20) without such records. We similarly identified 833 (0.22%) cases of suicidal behaviors 1 to 6 years from baseline and 16,660 (4.42%) corresponding controls. Based on 143 input features, mainly including sociodemographic, environmental, and psychosocial factors; medical history; and polygenic risk scores (PRS) for suicidality, we applied a bagged balanced light gradient-boosting machine (LightGBM) with stratified 10-fold cross-validation and grid-search to construct the full prediction models for suicide attempts or deaths within 1 year or between 1 and 6 years. The Shapley Additive Explanations (SHAP) approach was used to quantify the importance of input features, and the top 20 features with the highest SHAP values were selected to train the applicable models. The external validity of the established models was assessed among 50,310 individuals who participated in UK Biobank repeated assessments both overall and by the level of PRS for suicidality. Individuals with suicidal behaviors were on average 56 years old, with equal sex distribution. The application of these full models in the external validation data set demonstrated good model performance, with the area under the receiver operating characteristic (AUROC) curves of 0.919 and 0.892 within 1 year and between 1 and 6 years, respectively. Importantly, the applicable models with the top 20 most important features showed comparable external-validated performance (AUROC curves of 0.901 and 0.885) as the full models, based on which we found that individuals in the top quintile of predicted risk accounted for 91.7% (n=11) and 80.7% (n=25) of all suicidality cases within 1 year and during 1 to 6 years, respectively. We further obtained comparable prediction accuracy when applying these models to subpopulations with different genetic susceptibilities to suicidality. For example, for the 1-year risk prediction, the AUROC curves were 0.907 and 0.885 for the high (>2nd tertile of PRS) and low (<1st) genetic susceptibilities groups, respectively. We established applicable machine learning-based models for predicting both the short- and long-term risk of suicidality with high accuracy across populations of varying genetic risk for suicide, highlighting a cost-effective method of identifying individuals with a high risk of suicidality.

Explainability of High Energy Physics events classification using SHAP

Complex machine learning models have been fundamental for achieving accurate results regarding events classification in High Energy Physics (HEP). However, these complex models or black-box systems lack transparency and interpretability. In this work, we use the SHapley Additive exPlanations (SHAP) method for explaining the output of two event machine learning classifiers, based on eXtreme Gradient Boost (XGBoost) and deep neural networks (DNN). We compute SHAP values to interpret the results and analyze the importance of individual features, and the experiments show that SHAP method has high potential for understanding complex machine learning model in the context of high energy physics.

Machine learning-based approach reveals essential features for simplified TSPO PET quantification in ischemic stroke patients

IntroductionNeuroinflammation evaluation after acute ischemic stroke is a promising option for selecting an appropriate post-stroke treatment strategy. To assess neuroinflammation in vivo, translocator protein PET (TSPO PET) can be used. However, the gold standard TSPO PET quantification method includes a 90 min scan and continuous arterial blood sampling, which is challenging to perform on a routine basis. In this work, we determine what information is required for a simplified quantification approach using a machine learning algorithm. Materials and MethodsWe analyzed data from 18 patients with ischemic stroke who received 0–90 min [18F]GE-180 PET as well as T1-weigted (T1w), FLAIR, and arterial spin labeling (ASL) MRI scans. During PET scans, five manual venous blood samples at 5, 15, 30, 60, and 85 min post injection (p.i.) were drawn, and plasma activity concentration was measured. Total distribution volume (VT) was calculated using Logan plot with the full dynamic PET and an image-derived input function (IDIF) from the carotid arteries. IDIF was scaled by a calibration factor derived from all the measured plasma activity concentrations. The calculated VT values were used for training a random forest regressor. As input features for the model, we used three late PET frames (60–70, 70–80, and 80–90 min p.i.), the ASL image reflecting perfusion, the voxel coordinates, the lesion mask, and the five plasma activity concentrations. The algorithm was validated with the leave-one-out approach. To estimate the impact of the individual features on the algorithm’s performance, we used Shapley Additive Explanations (SHAP). Having determined that the three late PET frames and the plasma activity concentrations were the most important features, we tested a simplified quantification approach consisting of dividing a late PET frame by a plasma activity concentration. All the combinations of frames/samples were compared by means of concordance correlation coefficient and Bland-Altman plots. ResultsWhen using all the input features, the algorithm predicted VT values with high accuracy (87.8 ± 8.3%) for both lesion and non-lesion voxels. The SHAP values demonstrated high impact of the late PET frames (60–70, 70–80, and 80–90 min p.i.) and plasma activity concentrations on the VT prediction, while the influence of the ASL-derived perfusion, voxel coordinates, and the lesion mask was low. Among all the combinations of the late PET frames and plasma activity concentrations, the 70–80 min p.i. frame divided by the 30 min p.i. plasma sample produced the closest VT estimate in the ischemic lesion. ConclusionReliable TSPO PET quantification is achievable by using a single late PET frame divided by a late blood sample activity concentration.