SHapley Additive exPlanations Research Articles

Explainable highway performance degradation prediction model based on LSTM

With the dense and huge highway network in China, the highway maintenance management system has become the most concerned issue for Chinese highway managers in recent years. Therefore, based on the highway network in Guizhou Province of China, this paper established the semi-rigid asphalt pavement performance multi-output Long Short-Term Memory (LSTM) prediction model with regional applicability, including Pavement Surface Condition Index (PCI), Pavement Rutting Depth Index (RDI), Pavement Riding Quality Index (RQI) and Pavement Skidding Resistance Index (SRI) and the hyperparameters of the model were optimized by Bayesian optimization algorithm and the casual analysis of the model was conducted based on SHapley Additive exPlanations (SHAP), providing a reference for China's highway network-level preventive maintenance decision-making. The results prove that the established model can effectively predict the performance of the following year through the data of the previous two years. The established model has a better comprehensive prediction effect in comparison to similar multi-output models. The average coefficient of determination (R2) of the four prediction indexes can reach 0.823. Besides, the prediction effect among the indexes is stable and multiple pavement performance indexes can be effectively and reliably predicted at the same time. In addition, the contribution of input feature variables to output is quantified based on SHAP. According to quantified the contribution, the association relationship of the neural network model is set-valued mapped to the causality relationship in the post hoc. The input feature variables with the main contribution are extracted for causal analysis. The causal analysis shows that the degradation of the four pavement performances all have obvious time memory characteristics and have significant differences, but their pavement performance degradation is the result of nonlinear changes caused by the coupling effects of traffic load, road age, climate, and pavement structure.

Explainable artificial intelligence for spectroscopy data: a review.

Explainable artificial intelligence (XAI) has gained significant attention in various domains, including natural and medical image analysis. However, its application in spectroscopy remains relatively unexplored. This systematic review aims to fill this gap by providing a comprehensive overview of the current landscape of XAI in spectroscopy and identifying potential benefits and challenges associated with its implementation. Following the PRISMA guideline 2020, we conducted a systematic search across major journal databases, resulting in 259 initial search results. After removing duplicates and applying inclusion and exclusion criteria, 21 scientific studies were included in this review. Notably, most of the studies focused on using XAI methods for spectral data analysis, emphasizing identifying significant spectral bands rather than specific intensity peaks. Among the most utilized AI techniques were SHapley Additive exPlanations (SHAP), masking methods inspired by Local Interpretable Model-agnostic Explanations (LIME), and Class Activation Mapping (CAM). These methods were favored due to their model-agnostic nature and ease of use, enabling interpretable explanations without modifying the original models. Future research should propose new methods and explore the adaptation of other XAI employed in other domains to better suit the unique characteristics of spectroscopic data.

Explaining transfer learning models for the detection of COVID-19 on X-ray lung images

Amidst the coronavirus disease 2019 (COVID-19) pandemic, researchers are exploring innovative approaches to enhance diagnostic accuracy. One avenue is utilizing deep learning models to analyze lung X-ray images for COVID-19 diagnosis, complementing existing tests like reverse transcription polymerase chain reaction (RT-PCR). However, trusting these models, often viewed as black boxes, presents a challenge. To address this, six explainable artificial intelligence (XAI) techniques: local interpretable model agnostic explanations (LIME), Shapley additive explanations (SHAP), integrated gradients, smooth-grad, gradient-weighted class activation mapping (Grad-CAM), and Layer-CAM are applied to interpret four transfer learning models. These models: VGG16, ResNet50, InceptionV3, and DenseNet121 are analyzed to understand their workings and the rationale behind their predictions. Validating the results with medical experts poses difficulties due to time and resource constraints, alongside the scarcity of annotated X-ray datasets. To address this, a voting mechanism employing different XAI methods across various models is proposed. This approach highlights regions of lung infection, potentially reducing individual model biases stemming from their structures. If successful, this research could pave the way for an automated system for annotating infection regions, bolstering confidence in predictions and aiding in the development of more effective diagnostic tools for COVID-19.

SHAP based predictive modeling for 1 year all-cause readmission risk in elderly heart failure patients: feature selection and model interpretation

Heart failure (HF) is a significant global public health concern with a high readmission rate, posing a serious threat to the health of the elderly population. While several studies have used machine learning (ML) to develop all-cause readmission risk prediction models for elderly patients with HF, few have integrated ML-selected features with those chosen by human experts to assess HF patients readmission. A retrospective analysis of 8396 elderly HF patients hospitalized at the Affiliated Hospital of North Sichuan Medical College from January 1, 2018 to December 31, 2021 was conducted. Variables selected by XGBoost, LASSO regression, and random forest constituted the machine group, while the human expert group comprised variables chosen by two experienced cardiovascular professors. The variables selected by both groups were combined to form a human–machine collaboration group. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) method was used to elucidate the importance of each predictive feature, explain the impact of individual features on the model, and provide visual representation. A total of 73 features were included for model development. The human–machine collaboration model, utilizing CatBoost, achieved an AUC of 0.83617, an F1-score of 0.73521, and a Brier score of 0.16536 on the validation set. This model demonstrated superior predictive performance compared to those created solely by human experts or machine. The SHAP plot was then used to visually display the feature analysis of the human–machine collaboration model, revealing HGB, NT-proBNP, smoking history, NYHA classification, and LVEF as the 5 most important features. This study indicate that the human–machine collaboration model outperforms those relying solely on human expert selection or machine algorithm at predicting all-cause readmission in elderly HF patients. The application of the SHAP method enhanced the interpretability of the model outcomes, aiding clinicians in accurately pinpointing risk factors associated with HF readmission. This advancement enables the formulation of tailored treatment strategies, offering a more personalized approach to patient care.

Ultra-low cycle fatigue life prediction of stainless steel based on transfer learning guided artificial neural network

Determining ultra-low cycle fatigue (ULCF) life of stainless steel typically involves laborious and time-consuming tests. While machine learning offers an efficient solution for fatigue life prediction, the inherent demand for sufficient training data remains unaddressed. To solve this challenge, a novel method to predict ULCF life using small datasets was proposed by combining artificial neural network (ANN) and transfer learning (TL). A TL-guided ANN (TLNN) framework was developed, wherein the knowledge gained from a pre-trained ANN model for structural steels was transferred to ULCF life prediction of stainless steel. The TLNN model exhibits enhanced predictive capabilities for stainless steel under small amounts of training datasets, with mean R2 (coefficient of determination) value of 0.88 and average MAPE (mean absolute percentage error) value of 25.52%. In comparison, the directly trained ANN model shows lower performance, possessing that the average R2 and MAPE are 0.69 and 47.58%, respectively. We pioneeringly explored and found that adding synthetic data enhancement had no positive effect on the predictive ability of the TLNN model. Furthermore, the dependence of predictive capacity by the TLNN model on the number of training data availability was studied. The prediction process of the TLNN model was interpreted by SHAP (SHapley Additive exPlanations) method. In general, the developed TLNN model can efficiently predict the ULCF life of stainless steel with high accuracy and reduce the cost of consecutive fatigue property evaluation.

Predicting Tracheostomy Need on Admission to the Intensive Care Unit-A Multicenter Machine Learning Analysis.

It is difficult to predict which mechanically ventilated patients will ultimately require a tracheostomy which further predisposes them to unnecessary spontaneous breathing trials, additional time on the ventilator, increased costs, and further ventilation-related complications such as subglottic stenosis. In this study, we aimed to develop a machine learning tool to predict which patients need a tracheostomy at the onset of admission to the intensive care unit (ICU). Retrospective Cohort Study. Multicenter Study of 335 Intensive Care Units between 2014 and 2015. The eICU Collaborative Research Database (eICU-CRD) was utilized to obtain the patient cohort. Inclusion criteria included: (1) Age >18 years and (2) ICU admission requiring mechanical ventilation. The primary outcome of interest included tracheostomy assessed via a binary classification model. Models included logistic regression (LR), random forest (RF), and Extreme Gradient Boosting (XGBoost). Of 38,508 invasively mechanically ventilated patients, 1605 patients underwent a tracheostomy. The XGBoost, RF, and LR models had fair performances at an AUROC 0.794, 0.780, and 0.775 respectively. Limiting the XGBoost model to 20 features out of 331, a minimal reduction in performance was observed with an AUROC of 0.778. Using Shapley Additive Explanations, the top features were an admission diagnosis of pneumonia or sepsis and comorbidity of chronic respiratory failure. Our machine learning model accurately predicts the probability that a patient will eventually require a tracheostomy upon ICU admission, and upon prospective validation, we have the potential to institute earlier interventions and reduce the complications of prolonged ventilation.

Non-Linear Impact of Economic Performance on Social Equity in Rail Transit Station Areas

Rail transit station areas (RSAs) are heralded as a transformative approach to urban planning, emphasizing the integration of transportation, housing, and commercial development to foster sustainable and inclusive cities. This study presents a comprehensive exploration of the interplay between transit-oriented development (TOD) economic performance and social equity in RSAs, employing advanced methodologies, like eXtreme Gradient Boosting (XGBoost) and SHapley Additive exPlanations (SHAPs), to decipher the complex relationships between TOD characteristics and social equity outcomes. Focused on Dalian’s urban center, this study integrates diverse datasets, including mobile location, geospatial, and economic price data, to construct a nuanced analysis framework within the NPE (node–place–economic) model. The results indicate that economic factors significantly impact overall social equity, particularly influencing key variables, such as weekday and weekend commuter population densities. Local explanatory plots reveal that economic performance variables associated with transportation development exhibit a broad non-linear impact on social equity in RSAs. This study advances equitable urban development through TOD by stressing the importance of factoring in multiple variables in RSA planning. This approach is vital for creating synergistic effects, fostering equitable spatial planning, and has both theoretical and practical benefits for improving residents’ well-being.

Revolutionizing engineered cementitious composite materials (ECC): the impact of XGBoost-SHAP analysis on polyvinyl alcohol (PVA) based ECC predictions

This study integrates previous experimental data and employs machine learning (ML) methods, including Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and eXtreme Gradient Boosting (XGBoost), to predict the compressive strength (CS) and tensile strength (TS) of engineered cementitious composites (ECC). XGBoost emerged as the superior model among the four ML models, providing an interpretable and highly accurate predictive framework. To optimize the model performance, hyperparameter tuning using a fivefold cross-validation approach with the data divided into 80% training and 20% testing subsets. The Shapley Additive Explanations (SHAP) algorithm was also employed to reveal the impact of important features, such as the water/binder ratio, fly ash content, and water reducer dosage, on the model’s predictions and their interrelationships. The XGBoost demonstrates the most exemplary performance, as reflected in the R2 values of 0.92 and 0.97 for CS and TS testing, respectively. The SHAP analysis provided insights into the impact of individual features on CS and TS, shedding light on how specific characteristics influence the predictive accuracy of these properties. This highly accurate prediction model uncovers insights into correlated features, aids in creating new mix designs of ECC, and supports global efforts toward a low-carbon future in the construction industry by reducing carbon emissions.

A Machine Learning Model for Predicting In-Hospital Mortality in Chinese Patients With ST-Segment Elevation Myocardial Infarction: Findings From the China Myocardial Infarction Registry.

Machine learning (ML) risk prediction models, although much more accurate than traditional statistical methods, are inconvenient to use in clinical practice due to their nontransparency and requirement of a large number of input variables. We aimed to develop a precise, explainable, and flexible ML model to predict the risk of in-hospital mortality in patients with ST-segment elevation myocardial infarction (STEMI). This study recruited 18,744 patients enrolled in the 2013 China Acute Myocardial Infarction (CAMI) registry and 12,018 patients from the China Patient-Centered Evaluative Assessment of Cardiac Events (PEACE)-Retrospective Acute Myocardial Infarction Study. The Extreme Gradient Boosting (XGBoost) model was derived from 9616 patients in the CAMI registry (2014, 89 variables) with 5-fold cross-validation and validated on both the 9125 patients in the CAMI registry (89 variables) and the independent China PEACE cohort (10 variables). The Shapley Additive Explanations (SHAP) approach was employed to interpret the complex relationships embedded in the proposed model. In the XGBoost model for predicting all-cause in-hospital mortality, the variables with the top 8 most important scores were age, left ventricular ejection fraction, Killip class, heart rate, creatinine, blood glucose, white blood cell count, and use of angiotensin-converting enzyme inhibitors (ACEIs) and angiotensin II receptor blockers (ARBs). The area under the curve (AUC) on the CAMI validation set was 0.896 (95% CI 0.884-0.909), significantly higher than the previous models. The AUC for the Global Registry of Acute Coronary Events (GRACE) model was 0.809 (95% CI 0.790-0.828), and for the TIMI model, it was 0.782 (95% CI 0.763-0.800). Despite the China PEACE validation set only having 10 available variables, the AUC reached 0.840 (0.829-0.852), showing a substantial improvement to the GRACE (0.762, 95% CI 0.748-0.776) and TIMI (0.789, 95% CI 0.776-0.803) scores. Several novel and nonlinear relationships were discovered between patients' characteristics and in-hospital mortality, including a U-shape pattern of high-density lipoprotein cholesterol (HDL-C). The proposed ML risk prediction model was highly accurate in predicting in-hospital mortality. Its flexible and explainable characteristics make the model convenient to use in clinical practice and could help guide patient management. ClinicalTrials.gov NCT01874691; https://clinicaltrials.gov/study/NCT01874691.

Biochar-based persulfate activation: Rate constant prediction, key variables identification, and system optimization

Biochar has emerged as a promising sustainable catalyst for persulfate-based advanced oxidation processes (PS-AOPs), offering synergy between radical and non-radical pathways for superior mineralization. To streamline the traditionally time-consuming and costly biochar selection process, artificial intelligence was employed for a more efficient approach. In this study, an extreme gradient boosting (XGB) model was used to predict the pseudo-first-order kinetic rate constant (k) of a PS-AOP. Important features were analyzed, and optimal feature values were obtained using particle swarm optimization (PSO). The dataset comprised 671 data points with 18 features extracted from peer-reviewed publications. To address data sparsity and imbalance, machine learning techniques and stratified data splits were utilized, achieving high model performance (R2 = 0.96, RMSE = 0.035). Feature importance analysis using XGB, permutation feature importance, and SHapley Additive exPlanations highlighted the importance of a higher biochar oxygen-to‑carbon ratio and encouraged metal modification. The presence of oxygen functional groups, particularly carbonyl and graphitized biochar structures, was emphasized in kinetic performance. PSO results recommended optimized biochar preparation parameters and operating conditions to enhance the k value, aligning with existing literature. This study predicts kinetic outcomes under preset conditions and guides practical biochar selection, demonstrating its relevance and potential in real-world applications.

The Relationship Between Metal Exposure and HPV Infection: Evidence from Explainable Machine Learning Methods.

HPV is a ubiquitous pathogen implicated in cervical and other cancers. Although vaccines are available, they do not encompass all subtypes. Meanwhile, metal exposure may elevate the risk of HPV infection and amplify its carcinogenic potential, but studies to further elucidate this relationship are insufficient.This study entailed a cross-sectional analysis utilizing data from the National Health and Nutrition Examination Survey (NHANES) 2007-2016. The study sample comprised 2765 women. Multivariate logistic regression was employed to examine the association between single metal exposure and HPV infection, weighted quantile sum (WQS) regression was utilized for assessing the mixed metal exposure effect, and the XGBoost + SHapley Additive exPlanations (SHAP) to evaluate the contribution of metal exposure in HPV infection.Multivariate logistic regression analysis indicated that elevated Co concentration was inversely associated with HPV infection (OR 0.891; 95% CI 0.814-0.975), while elevated Pb concentration correlated with an increased HPV infection (OR 1.176; 95% CI 1.074-1.287). Regression analysis of the WQS for mixed metal exposure suggested that the WQS index was potentially linked to an increased likelihood of HPV infection in the positive direction (OR 1.249; 95% CI 1.052-1.482), with no significant association observed in the negative direction (OR 0.852; 95% CI 0.713-1.017). SHAP analysis prioritized the importance of characteristics: number of sexual partners, marital status, poverty-to-income ratio (PIR), Co, Pb, and alcohol consumption.Exposure to Pb was associated with an increase in the incidence of HPV infection, whereas Co exposure demonstrated an inverse relationship. The composite exposure to multiple metals showed a positive association with the prevalence of HPV infection. These findings indicate that exposure to metals could potentially escalate the prevalence of HPV infection.

The inversion of arid-coastal cultivated soil salinity using explainable machine learning and Sentinel-2

The escalating salinization of cultivated soil poses a significant threat to the ecological environment. It is imperative to establish a monitoring system and mitigate the spread of salinization in arid and coastal areas through remote sensing, incorporating high-precision cross-regional models for soil salt content inversion. This study focuses on typical saline-alkali soils in arid and coastal regions of China. Using Sentinel 2 data (including 6 bands and 27 spectral indices), along with soil texture, moisture content, temperature, precipitation, and digital elevation model (DEM) data to establish an arid-coastal salinity inversion model. Variable selection methods such as pearson correlation coefficient (PCC), variable importance in projection (VIP), gray relational analysis (GRA), and gradient boosting machine (GBM) were used, while using 9 models including adaptive boosting (Adaboost), extremely randomized trees (ERT), and light gradient boosting machine (LightGBM). The best model was further elucidated using the Shapley additive explanations method. Results indicate that the common sensitive characteristic variables of arid-coastal areas were spectral indices and soil properties in PCC, the spectral variable bands and indices in VIP, and all variables in GRA and GBM. The best inversion model GBM-ERT (R2 = 0.91, RMSE = 1.06) in arid areas exhibited higher accuracy than the best inversion model GBM-Adaboost (R2 = 0.77, RMSE = 1.74) in coastal areas. The arid-coastal inversion model PCC-LightGBM demonstrated the best inversion performance (R2 = 0.64, RMSE = 2.29) and simulation performance in arid (R2 = 0.67) and coastal areas (R2 = 0.63). Dead fuel index (DFI) had the most significant impact on model prediction (0.89) and the second ratio index (RI2) contributed the highest relative importance (18 %) to the model. Our analysis indicates that the arid-coastal model of PCC-LightGBM established using common characteristic variables, can effectively monitor large-scale soil salinity.

Relation between fault characteristic frequencies and local interpretability shapley additive explanations for continuous machine health monitoring

Recently, the Shapley additive explanations models have been extensively studied to enhance explainability of artificial intelligence algorithms, while most of them simply use Shapley additive explanations to rank or measure the importance of different features. In this study, a novel methodology that studies the relation between fault characteristic frequencies and Shapley values generated by local interpretability Shapley additive explanations for machine health monitoring is proposed. Firstly, a simulation model is introduced to generate vibration signals at different health conditions and their spectral amplitudes transformed from Fourier transform are used to investigate the relationship between fault characteristic frequencies and local interpretability Shapley values. It is interestingly found that Shapley values can be used to locate fault characteristic frequencies. Moreover, most of them have negative values in a normal stage and have positive values in an abnormal stage. Based on this finding and Shapley additive explanations, a health indicator construction methodology is proposed to continuously monitor incipient machine faults. Subsequently, an automatic signal filtering method is proposed to remove and eliminate burrs and noise in Shapley values so that fault characteristic frequencies can be clearly revealed by Shapley values for physical fault diagnosis. Two run-to-failure cases are conducted to demonstrate the effectiveness of the proposed methodology and then the superiority of this study is demonstrated by comparing with existing methods for health indicator construction and fault diagnosis, including sparsity parameters, Hjorth parameters, and fast Kurtogram. Comparison results show that the proposed health indicator is more sensitive to the time of incipient fault initiation and interpretable fault diagnosis based on Shapley values has a robust performance. This study first sheds a light on the relationship between fault characteristic frequencies and Shapley values under the scenario of continuous machine health monitoring and seamlessly guides applicants to realize Shapley additive explanations based incipient fault detection and diagnosis.

Development and Validation of an Interpretable Machine Learning Model for Early Prognosis Prediction in ICU Patients with Malignant Tumors and Hyperkalemia.

This study aims to develop and validate a machine learning (ML) predictive model for assessing mortality in patients with malignant tumors and hyperkalemia (MTH). We extracted data on patients with MTH from the Medical Information Mart for Intensive Care-IV, version 2.2 (MIMIC-IV v2.2) database. The dataset was split into a training set (75%) and a validation set (25%). We used the Least Absolute Shrinkage and Selection Operator (LASSO) regression to identify potential predictors, which included clinical laboratory indicators and vital signs. Pearson correlation analysis tested the correlation between predictors. In-hospital death was the prediction target. The Area Under the Curve (AUC) and accuracy of the training and validation sets of 7 ML algorithms were compared, and the optimal 1 was selected to develop the model. The calibration curve was used to evaluate the prediction accuracy of the model further. SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) enhanced model interpretability. 496 patients with MTH in the Intensive Care Unit (ICU) were included. After screening, 17 clinical features were included in the construction of the ML model, and the Pearson correlation coefficient was <0.8, indicating that the correlation between the clinical features was small. eXtreme Gradient Boosting (XGBoost) outperformed other algorithms, achieving perfect scores in the training set (accuracy: 1.000, AUC: 1.000) and high scores in the validation set (accuracy: 0.734, AUC: 0.733). The calibration curves indicated good predictive calibration of the model. SHAP analysis identified the top 8 predictive factors: urine output, mean heart rate, maximum urea nitrogen, minimum oxygen saturation, minimum mean blood pressure, maximum total bilirubin, mean respiratory rate, and minimum pH. In addition, SHAP and LIME performed in-depth individual case analyses. This study demonstrates the effectiveness of ML methods in predicting mortality risk in ICU patients with MTH. It highlights the importance of predictors like urine output and mean heart rate. SHAP and LIME significantly enhanced the model's interpretability.

Predicting depression in patients with heart failure based on a stacking model.

There is a lack of literature discussing the utilization of the stacking ensemble algorithm for predicting depression in patients with heart failure (HF). To create a stacking model for predicting depression in patients with HF. This study analyzed data on 1084 HF patients from the National Health and Nutrition Examination Survey database spanning from 2005 to 2018. Through univariate analysis and the use of an artificial neural network algorithm, predictors significantly linked to depression were identified. These predictors were utilized to create a stacking model employing tree-based learners. The performances of both the individual models and the stacking model were assessed by using the test dataset. Furthermore, the SHapley additive exPlanations (SHAP) model was applied to interpret the stacking model. The models included five predictors. Among these models, the stacking model demonstrated the highest performance, achieving an area under the curve of 0.77 (95%CI: 0.71-0.84), a sensitivity of 0.71, and a specificity of 0.68. The calibration curve supported the reliability of the models, and decision curve analysis confirmed their clinical value. The SHAP plot demonstrated that age had the most significant impact on the stacking model's output. The stacking model demonstrated strong predictive performance. Clinicians can utilize this model to identify high-risk depression patients with HF, thus enabling early provision of psychological interventions.

Explainable artificial intelligence (XAI) in finance: a systematic literature review

As the range of decisions made by Artificial Intelligence (AI) expands, the need for Explainable AI (XAI) becomes increasingly critical. The reasoning behind the specific outcomes of complex and opaque financial models requires a thorough justification to improve risk assessment, minimise the loss of trust, and promote a more resilient and trustworthy financial ecosystem. This Systematic Literature Review (SLR) identifies 138 relevant articles from 2005 to 2022 and highlights empirical examples demonstrating XAI's potential benefits in the financial industry. We classified the articles according to the financial tasks addressed by AI using XAI, the variation in XAI methods between applications and tasks, and the development and application of new XAI methods. The most popular financial tasks addressed by the AI using XAI were credit management, stock price predictions, and fraud detection. The three most commonly employed AI black-box techniques in finance whose explainability was evaluated were Artificial Neural Networks (ANN), Extreme Gradient Boosting (XGBoost), and Random Forest. Most of the examined publications utilise feature importance, Shapley additive explanations (SHAP), and rule-based methods. In addition, they employ explainability frameworks that integrate multiple XAI techniques. We also concisely define the existing challenges, requirements, and unresolved issues in applying XAI in the financial sector.

High-beta disruption prediction study on HL-2A with instance-based transfer learning

Unmitigated disruptions pose a much more serious threat when large-scale tokamaks are operating in the high performance regime. Machine learning based disruption predictors can exhibit impressive performance. However, their effectiveness is based on a substantial amount of training data. In future reactors, obtaining a substantial amount of disruption data in high performance regimes without risking damage to the machine is highly improbable. Using machine learning to develop disruption predictors on data from the low performance regime and transfer them to the high performance regime is an effective solution for a large reactor-sized tokamak like ITER and beyond. In this study, a number of models are trained using different subsets of data from the HL-2A tokamak experiment. A SHapley Additive exPlanations (SHAP) analysis is executed on the models, revealing that there are different, even contradicting, patterns between different performance regimes. Thus, simply mixing data among different performance regimes will not yield optimal results. Based on this analysis, we propose an instance-based transfer learning technique which trains the model using a dataset generated with an optimized strategy. The strategy involves instance and feature selection based on the physics behind differences in high- and low-performance discharges, as revealed by SHAP model analysis. The TrAdaBoost technique significantly improved the model performance from 0.78 BA (balanced accuracy) to 0.86 BA with a few high-performance operation data.

Machine learning-enabled prediction of prolonged length of stay in hospital after surgery for tuberculosis spondylitis patients with unbalanced data: a novel approach using explainable artificial intelligence (XAI)

BackgroundTuberculosis spondylitis (TS), commonly known as Pott’s disease, is a severe type of skeletal tuberculosis that typically requires surgical treatment. However, this treatment option has led to an increase in healthcare costs due to prolonged hospital stays (PLOS). Therefore, identifying risk factors associated with extended PLOS is necessary. In this research, we intended to develop an interpretable machine learning model that could predict extended PLOS, which can provide valuable insights for treatments and a web-based application was implemented.MethodsWe obtained patient data from the spine surgery department at our hospital. Extended postoperative length of stay (PLOS) refers to a hospitalization duration equal to or exceeding the 75th percentile following spine surgery. To identify relevant variables, we employed several approaches, such as the least absolute shrinkage and selection operator (LASSO), recursive feature elimination (RFE) based on support vector machine classification (SVC), correlation analysis, and permutation importance value. Several models using implemented and some of them are ensembled using soft voting techniques. Models were constructed using grid search with nested cross-validation. The performance of each algorithm was assessed through various metrics, including the AUC value (area under the curve of receiver operating characteristics) and the Brier Score. Model interpretation involved utilizing methods such as Shapley additive explanations (SHAP), the Gini Impurity Index, permutation importance, and local interpretable model-agnostic explanations (LIME). Furthermore, to facilitate the practical application of the model, a web-based interface was developed and deployed.ResultsThe study included a cohort of 580 patients and 11 features include (CRP, transfusions, infusion volume, blood loss, X-ray bone bridge, X-ray osteophyte, CT-vertebral destruction, CT-paravertebral abscess, MRI-paravertebral abscess, MRI-epidural abscess, postoperative drainage) were selected. Most of the classifiers showed better performance, where the XGBoost model has a higher AUC value (0.86) and lower Brier Score (0.126). The XGBoost model was chosen as the optimal model. The results obtained from the calibration and decision curve analysis (DCA) plots demonstrate that XGBoost has achieved promising performance. After conducting tenfold cross-validation, the XGBoost model demonstrated a mean AUC of 0.85 ± 0.09. SHAP and LIME were used to display the variables’ contributions to the predicted value. The stacked bar plots indicated that infusion volume was the primary contributor, as determined by Gini, permutation importance (PFI), and the LIME algorithm.ConclusionsOur methods not only effectively predicted extended PLOS but also identified risk factors that can be utilized for future treatments. The XGBoost model developed in this study is easily accessible through the deployed web application and can aid in clinical research.

Advancing Tau PET Quantification in Alzheimer Disease with Machine Learning: Introducing THETA, a Novel Tau Summary Measure.

Alzheimer disease (AD) exhibits spatially heterogeneous 3- or 4-repeat tau deposition across participants. Our overall goal was to develop an automated method to quantify the heterogeneous burden of tau deposition into a single number that would be clinically useful. Methods: We used tau PET scans from 3 independent cohorts: the Mayo Clinic Study of Aging and Alzheimer's Disease Research Center (Mayo, n = 1,290), the Alzheimer's Disease Neuroimaging Initiative (ADNI, n = 831), and the Open Access Series of Imaging Studies (OASIS-3, n = 430). A machine learning binary classification model was trained on Mayo data and validated on ADNI and OASIS-3 with the goal of predicting visual tau positivity (as determined by 3 raters following Food and Drug Administration criteria for 18F-flortaucipir). The machine learning model used region-specific SUV ratios scaled to cerebellar crus uptake. We estimated feature contributions based on an artificial intelligence-explainable method (Shapley additive explanations) and formulated a global tau summary measure, Tau Heterogeneity Evaluation in Alzheimer's Disease (THETA) score, using SUV ratios and Shapley additive explanations for each participant. We compared the performance of THETA with that of commonly used meta-regions of interest (ROIs) using the Mini-Mental State Examination, the Clinical Dementia Rating-Sum of Boxes, clinical diagnosis, and histopathologic staging. Results: The model achieved a balanced accuracy of 95% on the Mayo test set and at least 87% on the validation sets. It classified tau-positive and -negative participants with an AUC of 1.00, 0.96, and 0.94 on the Mayo, ADNI, and OASIS-3 cohorts, respectively. Across all cohorts, THETA showed a better correlation with the Mini-Mental State Examination and the Clinical Dementia Rating-Sum of Boxes (ρ ≥ 0.45, P < 0.05) than did meta-ROIs (ρ < 0.44, P < 0.05) and discriminated between participants who were cognitively unimpaired and those who had mild cognitive impairment with an effect size of 10.09, compared with an effect size of 3.08 for meta-ROIs. Conclusion: Our proposed approach identifies positive tau PET scans and provides a quantitative summary measure, THETA, that effectively captures heterogeneous tau deposition observed in AD. The application of THETA for quantifying tau PET in AD exhibits great potential.

Machine learning for the prediction of 1-year mortality in patients with sepsis-associated acute kidney injury

IntroductionSepsis-associated acute kidney injury (SA-AKI) is strongly associated with poor prognosis. We aimed to build a machine learning (ML)-based clinical model to predict 1-year mortality in patients with SA-AKI.MethodsSix ML algorithms were included to perform model fitting. Feature selection was based on the feature importance evaluated by the SHapley Additive exPlanations (SHAP) values. Area under the receiver operating characteristic curve (AUROC) was used to evaluate the discriminatory ability of the prediction model. Calibration curve and Brier score were employed to assess the calibrated ability. Our ML-based prediction models were validated both internally and externally.ResultsA total of 12,750 patients with SA-AKI and 55 features were included to build the prediction models. We identified the top 10 predictors including age, ICU stay and GCS score based on the feature importance. Among the six ML algorithms, the CatBoost showed the best prediction performance with an AUROC of 0.813 and Brier score of 0.119. In the external validation set, the predictive value remained favorable (AUROC = 0.784).ConclusionIn this study, we developed and validated a ML-based prediction model based on 10 commonly used clinical features which could accurately and early identify the individuals at high-risk of long-term mortality in patients with SA-AKI.