Explaining a series of models by propagating Shapley values

Explainable PET-based intratumoral and peritumoral machine learning model for predicting visceral pleural invasion in clinical-stage IA non-small cell lung cancer: A two-center study

AIM

The aim of this study was to develop a PET-based machine learning model for predicting visceral pleural invasion (VPI) in patients with clinical stage IA non-small cell lung cancer.

MATERIALS AND METHODS

A total of 294 patients and 69 patients from two institutions who underwent the 18F-FDG-PET scan were retrospectively analyzed. We extracted PET-based radiomics features from the gross tumor volume (GTV) and gross tumor volume incorporating peritumoral 4, 8 and 12 mm regions (GPTV4, GPTV8, GPTV12), respectively. Then four models were respectively established by using machine learning algorithms. The performance of the models was assessed by the receiver operating characteristic (ROC) curve and decision curve analyses (DCA). Shapley additive explanation (SHAP) was employed to explain the machine learning (ML) models and visualize variable predictions.

RESULTS

Compared with GTV, GPTV4, and GPTV12 radiomics models, the radiomics model based on GPTV8 using random forest (RF) among the 10 features demonstrated better prediction performance, with the AUC of 0.879, 0.846, and 0.745 in the training, internal validation, and external validation sets, respectively. The results of the SHAP method showed that the GLRLM_ShortRunLowGreyLevel Emphasis features were the most important factors in VPI. At the patient level, SHAP force plots provided a deep understanding for predicting VPI.

CONCLUSION

The PET-based intratumoral and peritumoral model based on machine learning offers an innovative tool for preoperative prediction of VPI in patients with lung adenocarcinoma. By employing the SHAP method, clinicians may gain a clearer insight into the factors contributing to VPI, which could enhance clinical decision-making of prognosis assessment.

Development and validation of an interpretable machine learning model for predicting hyperuricemia risk: Based on environmental chemical exposure

Hyperuricemia is a global health concern, with environmental chemicals as risk factors. This study used data of multiple environmental chemical exposures from the 2011-2012 cycle of the National Health and Nutrition Examination Survey (NHANES) to develop an interpretable machine learning model for hyperuricemia risk prediction. The least absolute shrinkage and selection operator (LASSO) regression method was employed to select relevant variables. The dataset was split into training (80 %) and test (20 %) sets and six machine learning models were constructed, including Random Forest (RF), Gaussian Naive Bayes (GNB), Light Gradient Boosting (LGB), Extreme Gradient Boosting (XGB), Adaptive Boosting Classifier (AB), and Support Vector Machine (SVM). Our study identified a hyperuricemia prevalence of 20.58 % in the 2011-2012 NHANES cycle, which was consistent with previous studies. The XGB model exhibited optimal performance, achieving the highest AUC (0.806, 95 % CI: 0.768-0.845), balanced accuracy (0.762; 95 % CI: 0.721-0.802), F1 value (0585; 95 % CI: 0.535-0.635), as well as the lowest Brier score (0.133; 95 % CI:0.122-0.144). Estimated glomerular filtration rate (eGFR), body mass index (BMI), cobalt (Co), mono-(2-ethyl)-hexyl phthalate (MEHP), mono-(3-carboxypropyl) phthalate (MCPP), mono-(2-ethyl-5-hydroxyhexyl) phthalate (MEHHP), 2-hydroxynaphthalene (OHNa2) were identified as the key factors contributing to the predictive model. The results of Shapley additive explanations and partial dependence plots indicated that hyperuricemia was positively associated with MCPP, MEHHP, and OHNa2, while negatively associated with Co and MEHP. This study is the first to predict the risk of hyperuricemia based on multiple environmental chemical exposures using a machine learning model.

Machine learning approaches enable the discovery of therapeutics across domains

Multi-modal datasets have grown exponentially in the last decade. This has created an enormous demand for machine learning models that can predict complex outcomes by leveraging cellular, molecular, and humoral profiles. Corresponding inference of mechanisms can help to uncover new therapeutic targets. Here, we discuss how biological principles guide the design of predictive models and how interpretable machine learning can lead to novel mechanistic insights. We provide descriptions of multiple learning techniques and how suited they are to domain adaptations. Finally, we talk about broad learning capabilities of foundation models on large datasets and whether they can be used to provide meaningful inference about biological datasets.

Assessing the association of multi-environmental chemical exposures on metabolic syndrome: A machine learning approach

Metabolic syndrome (MetS) is a major global public health concern due to its rising prevalence and association with increased risks of cardiovascular disease and type 2 diabetes. Emerging evidence suggests that environmental chemical exposures may play a significant role in the development of MetS by disrupting metabolic pathways. This study used data from 2,960 participants in the Korean National Environmental Health Survey (KoNEHS) cycle 4 (2018-2020) to examine associations between environmental exposures and MetS risk through machine learning (ML) approaches. Eight ML algorithms were applied, with the multilayer perceptron (MLP) and random forest (RF) models identified as optimal predictors. The MLP achieved an AUC of 0.79, and the RF achieved the highest F1 score of 0.82. Both models highlighted PFOA and PFOS, alongside age and BMI, as key predictors. SHapley Additive exPlanations (SHAP) and partial dependence plots (PDP) revealed both linear and nonlinear exposure-response patterns, suggesting threshold effects for key chemicals. These findings underscore the importance of incorporating environmental exposures into MetS risk assessments. The ML models provided robust predictive performance and novel insights into chemical and metabolic interactions, advocating for regulatory measures to reduce harmful exposures and integrate environmental factors into MetS prevention strategies.

Predicting treatment response and prognosis of immune checkpoint inhibitors-based combination therapy in advanced hepatocellular carcinoma using a longitudinal CT-based radiomics model: a multicenter study

BACKGROUND

Identifying effective predictive strategies to assess the response of immune checkpoint inhibitors (ICIs)-based combination therapy in advanced hepatocellular carcinoma (HCC) is crucial. This study presents a new longitudinal CT-based radiomics model to predict treatment response and prognosis in advanced HCC patients undergoing ICIs-based combination therapy.

METHODS

Longitudinal CT images were collected before and during the treatment for HCC patients across three institutions from January 2019 to April 2022. A total of 1316 radiomic features were extracted from arterial and portal venous phase abdominal CT images for each patient. A model called Longitudinal Whole-liver CT-based Radiomics (LWCTR) was developed to categorize patients into responders or non-responders using radiomic features and clinical information through support vector machine (SVM) classifiers. The area under the curve (AUC) was used as the performance metric and subsequently applied for risk stratification and prognostic assessment. The Shapley Additive explanations (SHAP) method was used to calculate the Shapley value, which explains the contribution of each feature in the SVM model to the prediction.

RESULTS

This study included 395 eligible participants, with a median age of 57 years (IQR 51-66), comprising 344 males and 51 females. The LWCTR model performed well in predicting treatment response, achieving an AUC of 0.883 (95% confidence interval [CI] 0.881-0.888) in the training cohort, 0.876 (0.858-0.895) in the internal validation cohort, and 0.875 (0.860-0.887) in the external test cohort. The Rad-Nomo model, integrating the LWCTR model's prediction score (Rad-score) with the modified Response Evaluation Criteria in Solid Tumors (mRECIST), demonstrated strong prognostic performance. It achieved time-dependent AUC values of 0.902, 0.823, and 0.850 at 1, 2, and 3 years in the internal validation cohort and 0.893, 0.848, and 0.762 at the same intervals in the external test cohort.

CONCLUSION

The proposed LWCTR model performs well in predicting treatment response and prognosis in patients with HCC receiving ICIs-based combination therapy, potentially contributing to personalized and timely treatment decisions.

A quantitatively interpretable model for Alzheimer's disease prediction using deep counterfactuals

Deep learning (DL) for predicting Alzheimer's disease (AD) has provided timely intervention in disease progression yet still demands attentive interpretability to explain how their DL models make definitive decisions. Counterfactual reasoning has recently gained increasing attention in medical research because of its ability to provide a refined visual explanatory map. However, such visual explanatory maps based on visual inspection alone are insufficient unless we intuitively demonstrate their medical or neuroscientific validity via quantitative features. In this study, we synthesize the counterfactual-labeled structural MRIs using our proposed framework and transform it into a gray matter density map to measure its volumetric changes over the parcellated region of interest (ROI). We also devised a lightweight linear classifier to boost the effectiveness of constructed ROIs, promoted quantitative interpretation, and achieved comparable predictive performance to DL methods. Throughout this, our framework produces an "AD-relatedness index" for each ROI. It offers an intuitive understanding of brain status for an individual patient and across patient groups concerning AD progression.

Targeted Detection of 76 Carnitine Indicators Combined with a Machine Learning Algorithm Based on HPLC-MS/MS in the Diagnosis of Rheumatoid Arthritis

BACKGROUND/OBJECTIVES

Early diagnosis and treatment of rheumatoid arthritis (RA) are essential to reducing disability. However, the diagnostic criteria remain unclear, relying on clinical symptoms and blood markers.

METHODS

Using high-performance liquid chromatography-mass spectrometry (HPLC-MS/MS) targeted detection, we evaluated 76 carnitine indicators (55 carnitines and 21 corresponding ratios) in the serum of patients with RA to investigate the role of carnitine in RA. A total of 359 patients (207 patients with RA and 152 healthy controls) were included in the study. Screening involved three methods and integrated 76 carnitine indicators and 128 clinical indicators to identify candidate markers to establish a theoretical basis for RA diagnosis and new therapeutic targets. The diagnostic model derived from the screened markers was validated using three machine learning algorithms.

RESULTS

The model was refined using eight candidate indicators (C0, C10:1, LYMPH, platelet distribution width, anti-keratin antibody, glucose, urobilinogen, and erythrocyte sedimentation rate (ESR)). The receiver operating characteristic curve, sensitivity, specificity, and accuracy of the V8 model obtained from the training set were >0.948, 79.46%, 92.99%, and 89.18%, whereas those of the test set were >0.925, 78.89%, 89.22%, and 85.87%, respectively. Twenty-four carnitines were identified as risk factors of RA, with three significantly correlating with ESR, four with anti-cyclic citrullinated peptide antibody activity, two with C-reactive protein, five with immunoglobulin-G, eight with immunoglobulin-A levels, and eleven with immunoglobulin-M levels.

CONCLUSIONS

Carnitine is integral in the progression of RA. The diagnostic model developed shows excellent diagnostic capacity, improving early detection and enabling timely intervention to minimize disability associated with RA.

Development and Validation of Machine Learning Models for Outcome Prediction in Patients with Poor-Grade Aneurysmal Subarachnoid Hemorrhage Following Endovascular Treatment

Background

Endovascular treatment (EVT) has been recommended as a superior modality for the treatment of intracranial aneurysm. However, there still exists a worse percentage of poor functional outcome in patients with poor-grade aneurysmal subarachnoid hemorrhage (aSAH) undergoing EVT. Therefore, it is urgently needed to investigate the risk factors and develop a critical decision model in the subtype of such patients.

Methods

We extracted the target variables from an ongoing registry cohort study, PROSAH-MPC, which was conducted in multiple centers in China. We randomly assigned these patients to training and validation cohorts with a ratio of 7:3. Univariate and multivariate logistic regressions were performed to find the potential factors, and then nine machine learning models and a stack ensemble model were developed with optimized variables. The performance of these models was evaluated through several indicators, including area under the receiver operating characteristic curve (AUC-ROC). We further use Shapley Additive Explanations (SHAP) methods for the distribution of feature visualization based on the optimal models.

Results

A total of 226 eligible patients with poor-grade aSAH undergoing EVT were enrolled, while 89 (39.4%) has a poor 12-month outcome. Age (Adjusted OR [aOR], 1.08; 95% CI: 1.03-1.13; p = 0.002), subarachnoid hemorrhage volume (aOR, 1.02; 95% CI: 1.00-1.05; p = 0.033), World Federation of Neurosurgical Societies grade (WFNS) (aOR, 2.03; 95% CI: 1.05-3.93; p = 0.035), and Hunt-Hess grade (aOR, 2.36; 95% CI: 1.13-4.93; p = 0.022) were identified as the independent risk factors of the poor outcome. Then, the prediction models developed have revealed that LightGBM algorithm has a superior performance with an AUC-ROC value of 0.842 in the validation cohort, while the SHAP results showed that age is the most important risk factor affecting functional outcomes.

Conclusion

The LightGBM model holds immense potential in facilitating risk stratification for poor-grade aSAH patients undergoing endovascular treatment who are at risk of adverse outcomes, thereby enhancing clinical decision-making processes.

Trial Registration

PROSAH-MPC. NCT05738083. Registered 16 November 2022 - Retrospectively registered, https://clinicaltrials.gov/study/NCT05738083.

Comparing Explanations of Molecular Machine Learning Models Generated with Different Methods for the Calculation of Shapley Values

Feature attribution methods from explainable artificial intelligence (XAI) provide explanations of machine learning models by quantifying feature importance for predictions of test instances. While features determining individual predictions have frequently been identified in machine learning applications, the consistency of feature importance-based explanations of machine learning models using different attribution methods has not been thoroughly investigated. We have systematically compared model explanations in molecular machine learning. Therefore, a test system of highly accurate compound activity predictions for different targets using different machine learning methods was generated. For these predictions, explanations were computed using methodological variants of the Shapley value formalism, a popular feature attribution approach in machine learning adapted from game theory. Predictions of each model were assessed using a model-agnostic and model-specific Shapley value-based method. The resulting feature importance distributions were characterized and compared by a global statistical analysis using diverse measures. Unexpectedly, methodological variants for Shapley value calculations yielded distinct feature importance distributions for highly accurate predictions. There was only little agreement between alternative model explanations. Our findings suggest that feature importance-based explanations of machine learning predictions should include an assessment of consistency using alternative methods.

System for Predicting Neurological Outcomes Following Cardiac Arrest Based on Clinical Predictors Using a Machine Learning Method: The Neurological Outcomes After Cardiac Arrest Method

BACKGROUND

A multimodal approach may prove effective for predicting clinical outcomes following cardiac arrest (CA). We aimed to develop a practical predictive model that incorporates clinical factors related to CA and multiple prognostic tests using machine learning methods.

METHODS

The neurological outcomes after CA (NOCA) method for predicting poor outcomes were developed using data from 390 patients with CA between May 2018 and June 2023. The outcome was poor neurological outcome, defined as a Cerebral Performance Category score of 3-5 at discharge. We analyzed 31 variables describing the circumstances at CA, demographics, comorbidities, and prognostic studies. The prognostic method was developed based on an extreme gradient-boosting algorithm with threefold cross-validation and hyperparameter optimization. The performance of the predictive model was evaluated using the receiver operating characteristic curve analysis and calculating the area under the curve (AUC).

RESULTS

Of the 390 total patients (mean age 64.2 years; 71.3% male), 235 (60.3%) experienced poor outcomes at discharge. We selected variables to predict poor neurological outcomes using least absolute shrinkage and selection operator regression. The Glasgow Coma Scale-M (best motor response), electroencephalographic features, the neurological pupil index, time from CA to return of spontaneous circulation, and brain imaging were found to be important key parameters in the NOCA score. The AUC of the NOCA method was 0.965 (95% confidence interval 0.941-0.976).

CONCLUSIONS

The NOCA score represents a simple method for predicting neurological outcomes, with good performance in patients with CA, using a machine learning analysis that incorporates widely available variables.

CaXML: Chemistry-informed machine learning explains mutual changes between protein conformations and calcium ions in calcium-binding proteins using structural and topological features

Proteins' flexibility is a feature in communicating changes in cell signaling instigated by binding with secondary messengers, such as calcium ions, associated with the coordination of muscle contraction, neurotransmitter release, and gene expression. When binding with the disordered parts of a protein, calcium ions must balance their charge states with the shape of calcium-binding proteins and their versatile pool of partners depending on the circumstances they transmit. Accurately determining the ionic charges of those ions is essential for understanding their role in such processes. However, it is unclear whether the limited experimental data available can be effectively used to train models to accurately predict the charges of calcium-binding protein variants. Here, we developed a chemistry-informed, machine-learning algorithm that implements a game theoretic approach to explain the output of a machine-learning model without the prerequisite of an excessively large database for high-performance prediction of atomic charges. We used the ab initio electronic structure data representing calcium ions and the structures of the disordered segments of calcium-binding peptides with surrounding water molecules to train several explainable models. Network theory was used to extract the topological features of atomic interactions in the structurally complex data dictated by the coordination chemistry of a calcium ion, a potent indicator of its charge state in protein. Our design created a computational tool of CaXML, which provided a framework of explainable machine learning model to annotate ionic charges of calcium ions in calcium-binding proteins in response to the chemical changes in an environment. Our framework will provide new insights into protein design for engineering functionality based on the limited size of scientific data in a genome space.

CellPhenoX: An eXplainable Cell-specific machine learning method to predict clinical Phenotypes using single-cell multi-omics

Single-cell technologies have enhanced our knowledge of molecular and cellular heterogeneity underlying disease. As the scale of single-cell datasets expands, linking cell-level phenotypic alterations with clinical outcomes becomes increasingly challenging. To address this, we introduce CellPhenoX, an eXplainable machine learning method to identify cell-specific phenotypes that influence clinical outcomes. CellPhenoX integrates classification models, explainable AI techniques, and a statistical framework to generate interpretable, cell-specific scores that uncover cell populations associated with relevant clinical phenotypes and interaction effects. We demonstrated the performance of CellPhenoX across diverse single-cell designs, including simulations, binary disease-control comparisons, and multi-class studies. Notably, CellPhenoX identified an activated monocyte phenotype in COVID-19, with expansion correlated with disease severity after adjusting for covariates and interactive effects. It also uncovered an inflammation-associated gradient in fibroblasts from ulcerative colitis. We anticipate that CellPhenoX holds the potential to detect clinically relevant phenotypic changes in single-cell data with multiple sources of variation, paving the way for translating single-cell findings into clinical impact.

Optimizing Hydrogen Production in the Co-Gasification Process: Comparison of Explainable Regression Models Using Shapley Additive Explanations

The co-gasification of biomass and plastic waste offers a promising solution for producing hydrogen-rich syngas, addressing the rising demand for cleaner energy. However, optimizing this complex process to maximize hydrogen yield remains challenging, particularly when balancing diverse feedstocks and improving process efficiency. While machine learning (ML) has shown significant potential in simulating and optimizing such processes, there is no clear consensus on the most effective regression models for co-gasification, especially with limited experimental data. Additionally, the interpretability of these models is a key concern. This study aims to bridge these gaps through two primary objectives: (1) modeling the co-gasification process using seven different ML algorithms, and (2) developing a framework for evaluating model interpretability, ultimately identifying the most suitable model for process optimization. A comprehensive set of experiments was conducted across three key dimensions, generalization ability, predictive accuracy, and interpretability, to thoroughly assess the models. Support Vector Regression (SVR) exhibited superior performance, achieving the highest coefficient of determination (R2) of 0.86. SVR outperformed other models in capturing non-linear dependencies and demonstrated effective overfitting mitigation. This study further highlights the limitations of other ML models, emphasizing the importance of regularization and hyperparameter tuning in improving model stability. By integrating Shapley Additive Explanations (SHAP) into model evaluation, this work is the first to provide detailed insights into feature importance and demonstrate the operational feasibility of ML models for industrial-scale hydrogen production in the co-gasification process. The findings contribute to the development of a robust framework for optimizing co-gasification, supporting the advancement of sustainable energy technologies and the reduction of greenhouse gas (GHG) emissions.

A multimodal deep-learning model based on multichannel CT radiomics for predicting pathological grade of bladder cancer

OBJECTIVE

To construct a predictive model using deep-learning radiomics and clinical risk factors for assessing the preoperative histopathological grade of bladder cancer according to computed tomography (CT) images.

METHODS

A retrospective analysis was conducted involving 201 bladder cancer patients with definite pathological grading results after surgical excision at the organization between January 2019 and June 2023. The cohort was classified into a test set of 81 cases and a training set of 120 cases. Hand-crafted radiomics (HCR) and features derived from deep-learning (DL) were obtained from computed tomography (CT) images. The research builds a prediction model using 12 machine-learning classifiers, which integrate HCR, DL features, and clinical data. Model performance was estimated utilizing decision-curve analysis (DCA), the area under the curve (AUC), and calibration curves.

RESULTS

Among the classifiers tested, the logistic regression model that combined DL and HCR characteristics demonstrated the finest performance. The AUC values were 0.912 (training set) and 0.777 (test set). The AUC values of clinical model achieved 0.850 (training set) and 0.804 (test set). The AUC values of the combined model were 0.933 (training set) and 0.824 (test set), outperforming both the clinical and HCR-only models.

CONCLUSION

The CT-based combined model demonstrated considerable diagnostic capability in differentiating high-grade from low-grade bladder cancer, serving as a valuable noninvasive instrument for preoperative pathological evaluation.

Brain clocks capture diversity and disparities in aging and dementia across geographically diverse populations

Brain clocks, which quantify discrepancies between brain age and chronological age, hold promise for understanding brain health and disease. However, the impact of diversity (including geographical, socioeconomic, sociodemographic, sex and neurodegeneration) on the brain-age gap is unknown. We analyzed datasets from 5,306 participants across 15 countries (7 Latin American and Caribbean countries (LAC) and 8 non-LAC countries). Based on higher-order interactions, we developed a brain-age gap deep learning architecture for functional magnetic resonance imaging (2,953) and electroencephalography (2,353). The datasets comprised healthy controls and individuals with mild cognitive impairment, Alzheimer disease and behavioral variant frontotemporal dementia. LAC models evidenced older brain ages (functional magnetic resonance imaging: mean directional error = 5.60, root mean square error (r.m.s.e.) = 11.91; electroencephalography: mean directional error = 5.34, r.m.s.e. = 9.82) associated with frontoposterior networks compared with non-LAC models. Structural socioeconomic inequality, pollution and health disparities were influential predictors of increased brain-age gaps, especially in LAC (R² = 0.37, F² = 0.59, r.m.s.e. = 6.9). An ascending brain-age gap from healthy controls to mild cognitive impairment to Alzheimer disease was found. In LAC, we observed larger brain-age gaps in females in control and Alzheimer disease groups compared with the respective males. The results were not explained by variations in signal quality, demographics or acquisition methods. These findings provide a quantitative framework capturing the diversity of accelerated brain aging.

Enhancing type 2 diabetes mellitus prediction by integrating metabolomics and tree-based boosting approaches

Background

Type 2 diabetes mellitus (T2DM) is a global health problem characterized by insulin resistance and hyperglycemia. Early detection and accurate prediction of T2DM is crucial for effective management and prevention. This study explores the integration of machine learning (ML) and explainable artificial intelligence (XAI) approaches based on metabolomics panel data to identify biomarkers and develop predictive models for T2DM.

Methods

Metabolomics data from T2DM (n = 31) and healthy controls (n = 34) were analyzed for biomarker discovery (mostly amino acids, fatty acids, and purines) and T2DM prediction. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO) regression to enhance the model's accuracy and interpretability. Advanced three tree-based ML algorithms (KTBoost: Kernel-Tree Boosting; XGBoost: eXtreme Gradient Boosting; NGBoost: Natural Gradient Boosting) were employed to predict T2DM using these biomarkers. The SHapley Additive exPlanations (SHAP) method was used to explain the effects of metabolomics biomarkers on the prediction of the model.

Results

The study identified multiple metabolites associated with T2DM, where LASSO feature selection highlighted important biomarkers. KTBoost [Accuracy: 0.938; CI: (0.880-0.997), Sensitivity: 0.971; CI: (0.847-0.999), Area under the Curve (AUC): 0.965; CI: (0.937-0.994)] demonstrated its effectiveness in using complex metabolomics data for T2DM prediction and achieved better performance than other models. According to KTBoost's SHAP, high levels of phenylactate (pla) and taurine metabolites, as well as low concentrations of cysteine, laspartate, and lcysteate, are strongly associated with the presence of T2DM.

Conclusion

The integration of metabolomics profiling and XAI offers a promising approach to predicting T2DM. The use of tree-based algorithms, in particular KTBoost, provides a robust framework for analyzing complex datasets and improves the prediction accuracy of T2DM onset. Future research should focus on validating these biomarkers and models in larger, more diverse populations to solidify their clinical utility.

Trends in the prevalence of osteoporosis and effects of heavy metal exposure using interpretable machine learning

There is limited evidence that heavy metals exposure contributes to osteoporosis. Multi-parameter scoring machine learning (ML) techniques were developed using National Health and Nutrition Examination Survey data to predict osteoporosis based on heavy metal exposure levels. For generating an optimal predictive model for osteoporosis, 12 ML models were used. Identification was carried out using the model that performed the best. For interpretation of models, Shapley additive explanation (SHAP) methods and partial dependence plots (PDP) were integrated into a pipeline and incorporated into the ML pipeline. By regressing osteoporosis on survey cycles, logistic regression was used to evaluate linear trends in osteoporosis over time. For the purpose of training and validating predictive models, 5745 eligible participants were randomly selected into training and testing set. It was evident from the results that the gradient boosting decision tree model performed the best among the predictive models, attributing to an accuracy rate of 89.40 % in the testing set. Based on the model results, the area under the curve and F1 score were 0.88 and 0.39, respectively. As a result of the SHAP analysis, urinary Co, urinary Tu, blood Cd, and urinary Hg levels were identified as the most influential factors influencing osteoporosis. Urinary Co (0.20-6.10 μg/mg creatinine), urinary Tu (0.06-1.93 μg/mg creatinine), blood Cd (0.07-0.50 μg/L), and urinary Hg (0.06-0.75 μg/mg creatinine) levels displayed a distinctive upward trend with risk of osteoporosis as values increased. Our analysis revealed that urinary Co, urinary Tu, blood Cd, and urinary Hg played a significant role in predictability.

Decoding micro-electrocorticographic signals by using explainable 3D convolutional neural network to predict finger movements

BACKGROUND

Electroencephalography (EEG) and electrocorticography (ECoG) recordings have been used to decode finger movements by analyzing brain activity. Traditional methods focused on single bandpass power changes for movement decoding, utilizing machine learning models requiring manual feature extraction.

NEW METHOD

This study introduces a 3D convolutional neural network (3D-CNN) model to decode finger movements using ECoG data. The model employs adaptive, explainable AI (xAI) techniques to interpret the physiological relevance of brain signals. ECoG signals from epilepsy patients during awake craniotomy were processed to extract power spectral density across multiple frequency bands. These data formed a 3D matrix used to train the 3D-CNN to predict finger trajectories.

RESULTS

The 3D-CNN model showed significant accuracy in predicting finger movements, with root-mean-square error (RMSE) values of 0.26-0.38 for single finger movements and 0.20-0.24 for combined movements. Explainable AI techniques, Grad-CAM and SHAP, identified the high gamma (HG) band as crucial for movement prediction, showing specific cortical regions involved in different finger movements. These findings highlighted the physiological significance of the HG band in motor control.

COMPARISON WITH EXISTING METHODS

The 3D-CNN model outperformed traditional machine learning approaches by effectively capturing spatial and temporal patterns in ECoG data. The use of xAI techniques provided clearer insights into the model's decision-making process, unlike the "black box" nature of standard deep learning models.

CONCLUSIONS

The proposed 3D-CNN model, combined with xAI methods, enhances the decoding accuracy of finger movements from ECoG data. This approach offers a more efficient and interpretable solution for brain-computer interface (BCI) applications, emphasizing the HG band's role in motor control.

Length-scale study in deep learning prediction for non-small cell lung cancer brain metastasis

Deep learning-assisted digital pathology has demonstrated the potential to profoundly impact clinical practice, even surpassing human pathologists in performance. However, as deep neural network (DNN) architectures grow in size and complexity, their explainability decreases, posing challenges in interpreting pathology features for broader clinical insights into physiological diseases. To better assess the interpretability of digital microscopic images and guide future microscopic system design, we developed a novel method to study the predictive feature length-scale that underpins a DNN's predictive power. We applied this method to analyze a DNN's capability in predicting brain metastasis from early-stage non-small-cell lung cancer biopsy slides. This study quantifies DNN's attention for brain metastasis prediction, targeting features at both the cellular scale and tissue scale in H&E-stained histological whole slide images. At the cellular scale, the predictive power of DNNs progressively increases with higher resolution and significantly decreases when the resolvable feature length exceeds 5 microns. Additionally, DNN uses more macro-scale features associated with tissue architecture and is optimized when assessing visual fields greater than 41 microns. Our study computes the length-scale requirements for optimal DNN learning on digital whole-slide microscopic images, holding the promise to guide future optical microscope designs in pathology applications and facilitating downstream deep learning analysis.

Quick and effective approximation of in silico saturation mutagenesis experiments with first-order taylor expansion

To understand the decision process of genomic sequence-to-function models, explainable AI algorithms determine the importance of each nucleotide in a given input sequence to the model's predictions and enable discovery of cis-regulatory motifs for gene regulation. The most commonly applied method is in silico saturation mutagenesis (ISM) because its per-nucleotide importance scores can be intuitively understood as the computational counterpart to in vivo saturation mutagenesis experiments. While ISM is highly interpretable, it is computationally challenging to perform for many sequences, and becomes prohibitive as the length of the input sequences and size of the model grows. Here, we use the first-order Taylor approximation to approximate ISM values from the model's gradient, which reduces its computation cost to a single forward pass for an input sequence. We show that the Taylor ISM (TISM) approximation is robust across different model ablations, random initializations, training parameters, and dataset sizes.

Flexible use of conserved motif vocabularies constrains genome access in cell type evolution

Cell types evolve into a hierarchy with related types grouped into families. How cell type diversification is constrained by the stable separation between families over vast evolutionary times remains unknown. Here, integrating single-nucleus multiomic sequencing and deep learning, we show that hundreds of sequence features (motifs) divide into distinct sets associated with accessible genomes of specific cell type families. This division is conserved across highly divergent, early-branching animals including flatworms and cnidarians. While specific interactions between motifs delineate cell type relationships within families, surprisingly, these interactions are not conserved between species. Consistently, while deep learning models trained on one species can predict accessibility of other species' sequences, their predictions frequently rely on distinct, but synonymous, motif combinations. We propose that long-term stability of cell type families is maintained through genome access specified by conserved motif sets, or 'vocabularies', whereas cell types diversify through flexible use of motifs within each set.

Explainable and visualizable machine learning models to predict biochemical recurrence of prostate cancer

PURPOSE

Machine learning (ML) models presented an excellent performance in the prognosis prediction. However, the black box characteristic of ML models limited the clinical applications. Here, we aimed to establish explainable and visualizable ML models to predict biochemical recurrence (BCR) of prostate cancer (PCa).

MATERIALS AND METHODS

A total of 647 PCa patients were retrospectively evaluated. Clinical parameters were identified using LASSO regression. Then, cohort was split into training and validation datasets with a ratio of 0.75:0.25 and BCR-related features were included in Cox regression and five ML algorithm to construct BCR prediction models. The clinical utility of each model was evaluated by concordance index (C-index) values and decision curve analyses (DCA). Besides, Shapley Additive Explanation (SHAP) values were used to explain the features in the models.

RESULTS

We identified 11 BCR-related features using LASSO regression, then establishing five ML-based models, including random survival forest (RSF), survival support vector machine (SSVM), survival Tree (sTree), gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and a Cox regression model, C-index were 0.846 (95%CI 0.796-0.894), 0.774 (95%CI 0.712-0.834), 0.757 (95%CI 0.694-0.818), 0.820 (95%CI 0.765-0.869), 0.793 (95%CI 0.735-0.852), and 0.807 (95%CI 0.753-0.858), respectively. The DCA showed that RSF model had significant advantages over all models. In interpretability of ML models, the SHAP value demonstrated the tangible contribution of each feature in RSF model.

CONCLUSIONS

Our score system provide reference for the identification for BCR, and the crafting of a framework for making therapeutic decisions for PCa on a personalized basis.

Prediction of Mycobacterium tuberculosis cell wall permeability using machine learning methods

Tuberculosis (TB) caused by the bacteria Mycobacterium tuberculosis (M. tb), continues to pose a significant worldwide health threat. The advent of drug-resistant strains of the disease highlights the critical need for novel treatments. The unique cell wall of M. tb provides an extra layer of protection for the bacteria and hence only compounds that can penetrate this barrier can reach their targets within the bacterial cell wall. The creation of a reliable machine learning (ML) model to predict the mycobacterial cell wall permeability of small molecules is presented in this work and four ML algorithms, including Random Forest, Support Vector Machines (SVM), k-nearest Neighbour (k-NN) and Logistic Regression were trained on a dataset of 5368 compounds. RDKit and Mordred toolkits were used to calculate features. To determine the most effective model, various performance metrics were used such as accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve. The best-performing model was further refined with hyperparameter tuning and tenfold cross-validation. The SVM model with filtering outperformed the other machine learning models and demonstrated 80.26% and 81.13% accuracy on the test and validation datasets, respectively. The study also provided insights into the molecular descriptors that play the most important role in predicting the ability of a molecule to pass the M. tb cell wall, which could guide future compound design. The model is available at https://github.com/PGlab-NIPER/MTB_Permeability .

Deciphering the Feature Representation of Deep Neural Networks for High-Performance AI

AI driven by deep learning is transforming many aspects of science and technology. The enormous success of deep learning stems from its unique capability of extracting essential features from Big Data for decision-making. However, the feature extraction and hidden representations in deep neural networks (DNNs) remain inexplicable, primarily because of lack of technical tools to comprehend and interrogate the feature space data. The main hurdle here is that the feature data are often noisy in nature, complex in structure, and huge in size and dimensionality, making it intractable for existing techniques to analyze the data reliably. In this work, we develop a computational framework named contrastive feature analysis (CFA) to facilitate the exploration of the DNN feature space and improve the performance of AI. By utilizing the interaction relations among the features and incorporating a novel data-driven kernel formation strategy into the feature analysis pipeline, CFA mitigates the limitations of traditional approaches and provides an urgently needed solution for the analysis of feature space data. The technique allows feature data exploration in unsupervised, semi-supervised and supervised formats to address different needs of downstream applications. The potential of CFA and its applications for pruning of neural network architectures are demonstrated using several state-of-the-art networks and well-annotated datasets across different disciplines.

Enhanced nitrous oxide emission factors due to climate change increase the mitigation challenge in the agricultural sector

Effective nitrogen fertilizer management is crucial for reducing nitrous oxide (N2O) emissions while ensuring food security within planetary boundaries. However, climate change might also interact with management practices to alter N2O emission and emission factors (EFs), adding further uncertainties to estimating mitigation potentials. Here, we developed a new hybrid modeling framework that integrates a machine learning model with an ensemble of eight process-based models to project EFs under different climate and nitrogen policy scenarios. Our findings reveal that EFs are dynamically modulated by environmental changes, including climate, soil properties, and nitrogen management practices. Under low-ambition nitrogen regulation policies, EF would increase from 1.18%-1.22% in 2010 to 1.27%-1.34% by 2050, representing a relative increase of 4.4%-11.4% and exceeding the IPCC tier-1 EF of 1%. This trend is particularly pronounced in tropical and subtropical regions with high nitrogen inputs, where EFs could increase by 0.14%-0.35% (relative increase of 11.9%-17%). In contrast, high-ambition policies have the potential to mitigate the increases in EF caused by climate change, possibly leading to slight decreases in EFs. Furthermore, our results demonstrate that global EFs are expected to continue rising due to warming and regional drying-wetting cycles, even in the absence of changes in nitrogen management practices. This asymmetrical influence of nitrogen fertilizers on EFs, driven by climate change, underscores the urgent need for immediate N2O emission reductions and further assessments of mitigation potentials. This hybrid modeling framework offers a computationally efficient approach to projecting future N2O emissions across various climate, soil, and nitrogen management scenarios, facilitating socio-economic assessments and policy-making efforts.

Leading role of Saharan dust on tropical cyclone rainfall in the Atlantic Basin

Tropical cyclone rainfall (TCR) extensively affects coastal communities, primarily through inland flooding. The impact of global climate changes on TCR is complex and debatable. This study uses an XGBoost machine learning model with 19-year meteorological data and hourly satellite precipitation observations to predict TCR for individual storms. The model identifies dust optical depth (DOD) as a key predictor that enhances performance evidently. The model also uncovers a nonlinear and boomerang-shape relationship between Saharan dust and TCR, with a TCR peak at 0.06 DOD and a sharp decrease thereafter. This indicates a shift from microphysical enhancement to radiative suppression at high dust concentrations. The model also highlights meaningful correlations between TCR and meteorological factors like sea surface temperature and equivalent potential temperature near storm cores. These findings illustrate the effectiveness of machine learning in predicting TCR and understanding its driving factors and physical mechanisms.

Chemistry-informed Machine Learning Explains Calcium-binding Proteins' Fuzzy Shape for Communicating Changes in the Atomic States of Calcium Ions

Proteins' fuzziness are features for communicating changes in cell signaling instigated by binding with secondary messengers, such as calcium ions, associated with the coordination of muscle contraction, neurotransmitter release, and gene expression. Binding with the disordered parts of a protein, calcium ions must balance their charge states with the shape of calcium-binding proteins and their versatile pool of partners depending on the circumstances they transmit, but it is unclear whether the limited experimental data available can be used to train models to accurately predict the charges of calcium-binding protein variants. Here, we developed a chemistry-informed, machine-learning algorithm that implements a game theoretic approach to explain the output of a machine-learning model without the prerequisite of an excessively large database for high-performance prediction of atomic charges. We used the ab initio electronic structure data representing calcium ions and the structures of the disordered segments of calcium-binding peptides with surrounding water molecules to train several explainable models. Network theory was used to extract the topological features of atomic interactions in the structurally complex data dictated by the coordination chemistry of a calcium ion, a potent indicator of its charge state in protein. With our designs, we provided a framework of explainable machine learning model to annotate atomic charges of calcium ions in calcium-binding proteins with domain knowledge in response to the chemical changes in an environment based on the limited size of scientific data in a genome space.

Prospector Heads: Generalized Feature Attribution for Large Models & Data

Feature attribution, the ability to localize regions of the input data that are relevant for classification, is an important capability for ML models in scientific and biomedical domains. Current methods for feature attribution, which rely on "explaining" the predictions of end-to-end classifiers, suffer from imprecise feature localization and are inadequate for use with small sample sizes and high-dimensional datasets due to computational challenges. We introduce prospector heads, an efficient and interpretable alternative to explanation-based attribution methods that can be applied to any encoder and any data modality. Prospector heads generalize across modalities through experiments on sequences (text), images (pathology), and graphs (protein structures), outperforming baseline attribution methods by up to 26.3 points in mean localization AUPRC. We also demonstrate how prospector heads enable improved interpretation and discovery of class-specific patterns in input data. Through their high performance, flexibility, and generalizability, prospectors provide a framework for improving trust and transparency for ML models in complex domains.

Interpretable machine learning identifies metabolites associated with glomerular filtration rate in type 2 diabetes patients

Objective

The co-occurrence of kidney disease in patients with type 2 diabetes (T2D) is a major public health challenge. Although early detection and intervention can prevent or slow down the progression, the commonly used estimated glomerular filtration rate (eGFR) based on serum creatinine may be influenced by factors unrelated to kidney function. Therefore, there is a need to identify novel biomarkers that can more accurately assess renal function in T2D patients. In this study, we employed an interpretable machine-learning framework to identify plasma metabolomic features associated with GFR in T2D patients.

Methods

We retrieved 1626 patients with type 2 diabetes (T2D) in Liaoning Medical University First Affiliated Hospital (LMUFAH) as a development cohort and 716 T2D patients in Second Affiliated Hospital of Dalian Medical University (SAHDMU) as an external validation cohort. The metabolite features were screened by the orthogonal partial least squares discriminant analysis (OPLS-DA). We compared machine learning prediction methods, including logistic regression (LR), support vector machine (SVM), random forest (RF), and eXtreme Gradient Boosting (XGBoost). The Shapley Additive exPlanations (SHAP) were used to explain the optimal model.

Results

For T2D patients, compared with the normal or elevated eGFR group, glutarylcarnitine (C5DC) and decanoylcarnitine (C10) were significantly elevated in GFR mild reduction group, and citrulline and 9 acylcarnitines were also elevated significantly (FDR<0.05, FC > 1.2 and VIP > 1) in moderate or severe reduction group. The XGBoost model with metabolites had the best performance: in the internal validate dataset (AUROC=0.90, AUPRC=0.65, BS=0.064) and external validate cohort (AUROC=0.970, AUPRC=0.857, BS=0.046). Through the SHAP method, we found that C5DC higher than 0.1μmol/L, Cit higher than 26 μmol/L, triglyceride higher than 2 mmol/L, age greater than 65 years old, and duration of T2D more than 10 years were associated with reduced GFR.

Conclusion

Elevated plasma levels of citrulline and a panel of acylcarnitines were associated with reduced GFR in T2D patients, independent of other conventional risk factors.

Research on Active Safety Situation of Road Passenger Transportation Enterprises: Evaluation, Prediction, and Analysis

The road passenger transportation enterprise is a complex system, requiring a clear understanding of their active safety situation (ASS), trends, and influencing factors. This facilitates transportation authorities to promptly receive signals and take effective measures. Through exploratory factor analysis and confirmatory factor analysis, we delved into potential factors for evaluating ASS and extracted an ASS index. To predict obtaining a higher ASS information rate, we compared multiple time series models, including GRU (gated recurrent unit), LSTM (long short-term memory), ARIMA, Prophet, Conv_LSTM, and TCN (temporal convolutional network). This paper proposed the WDA-DBN (water drop algorithm-Deep Belief Network) model and employed DEEPSHAP to identify factors with higher ASS information content. TCN and GRU performed well in the prediction. Compared to the other models, WDA-DBN exhibited the best performance in terms of MSE and MAE. Overall, deep learning models outperform econometric models in terms of information processing. The total time spent processing alarms positively influences ASS, while variables such as fatigue driving occurrences, abnormal driving occurrences, and nighttime driving alarm occurrences have a negative impact on ASS.

Machine Learning-Based Prediction of Suicidal Thinking in Adolescents by Derivation and Validation in 3 Independent Worldwide Cohorts: Algorithm Development and Validation Study

BACKGROUND

Suicide is the second-leading cause of death among adolescents and is associated with clusters of suicides. Despite numerous studies on this preventable cause of death, the focus has primarily been on single nations and traditional statistical methods.

OBJECTIVE

This study aims to develop a predictive model for adolescent suicidal thinking using multinational data sets and machine learning (ML).

METHODS

We used data from the Korea Youth Risk Behavior Web-based Survey with 566,875 adolescents aged between 13 and 18 years and conducted external validation using the Youth Risk Behavior Survey with 103,874 adolescents and Norway's University National General Survey with 19,574 adolescents. Several tree-based ML models were developed, and feature importance and Shapley additive explanations values were analyzed to identify risk factors for adolescent suicidal thinking.

RESULTS

When trained on the Korea Youth Risk Behavior Web-based Survey data from South Korea with a 95% CI, the XGBoost model reported an area under the receiver operating characteristic (AUROC) curve of 90.06% (95% CI 89.97-90.16), displaying superior performance compared to other models. For external validation using the Youth Risk Behavior Survey data from the United States and the University National General Survey from Norway, the XGBoost model achieved AUROCs of 83.09% and 81.27%, respectively. Across all data sets, XGBoost consistently outperformed the other models with the highest AUROC score, and was selected as the optimal model. In terms of predictors of suicidal thinking, feelings of sadness and despair were the most influential, accounting for 57.4% of the impact, followed by stress status at 19.8%. This was followed by age (5.7%), household income (4%), academic achievement (3.4%), sex (2.1%), and others, which contributed less than 2% each.

CONCLUSIONS

This study used ML by integrating diverse data sets from 3 countries to address adolescent suicide. The findings highlight the important role of emotional health indicators in predicting suicidal thinking among adolescents. Specifically, sadness and despair were identified as the most significant predictors, followed by stressful conditions and age. These findings emphasize the critical need for early diagnosis and prevention of mental health issues during adolescence.

Conventional and unconventional T-cell responses contribute to the prediction of clinical outcome and causative bacterial pathogen in sepsis patients

Sepsis is characterized by a dysfunctional host response to infection culminating in life-threatening organ failure that requires complex patient management and rapid intervention. Timely diagnosis of the underlying cause of sepsis is crucial, and identifying those at risk of complications and death is imperative for triaging treatment and resource allocation. Here, we explored the potential of explainable machine learning models to predict mortality and causative pathogen in sepsis patients. By using a modelling pipeline employing multiple feature selection algorithms, we demonstrate the feasibility of identifying integrative patterns from clinical parameters, plasma biomarkers, and extensive phenotyping of blood immune cells. While no single variable had sufficient predictive power, models that combined five and more features showed a macro area under the curve (AUC) of 0.85 to predict 90-day mortality after sepsis diagnosis, and a macro AUC of 0.86 to discriminate between Gram-positive and Gram-negative bacterial infections. Parameters associated with the cellular immune response contributed the most to models predictive of 90-day mortality, most notably, the proportion of T cells among PBMCs, together with expression of CXCR3 by CD4+ T cells and CD25 by mucosal-associated invariant T (MAIT) cells. Frequencies of Vδ2+ γδ T cells had the most profound impact on the prediction of Gram-negative infections, alongside other T-cell-related variables and total neutrophil count. Overall, our findings highlight the added value of measuring the proportion and activation patterns of conventional and unconventional T cells in the blood of sepsis patients in combination with other immunological, biochemical, and clinical parameters.

Machine learning for predicting cognitive deficits using auditory and demographic factors

IMPORTANCE

Predicting neurocognitive deficits using complex auditory assessments could change how cognitive dysfunction is identified, and monitored over time. Detecting cognitive impairment in people living with HIV (PLWH) is important for early intervention, especially in low- to middle-income countries where most cases exist. Auditory tests relate to neurocognitive test results, but the incremental predictive capability beyond demographic factors is unknown.

OBJECTIVE

Use machine learning to predict neurocognitive deficits, using auditory tests and demographic factors.

SETTING

The Infectious Disease Center in Dar es Salaam, Tanzania.

PARTICIPANTS

Participants were 939 Tanzanian individuals from Dar es Salaam living with and without HIV who were part of a longitudinal study. Patients who had only one visit, a positive history of ear drainage, concussion, significant noise or chemical exposure, neurological disease, mental illness, or exposure to ototoxic antibiotics (e.g., gentamycin), or chemotherapy were excluded. This provided 478 participants (349 PLWH, 129 HIV-negative). Participant data were randomized to training and test sets for machine learning.

MAIN OUTCOME(S) AND MEASURE(S)

The main outcome was whether auditory variables combined with relevant demographic variables could predict neurocognitive dysfunction (defined as a score of <26 on the Kiswahili Montreal Cognitive Assessment) better than demographic factors alone. The performance of predictive machine learning algorithms was primarily evaluated using the area under the receiver operational characteristic curve. Secondary metrics for evaluation included F1 scores, accuracies, and the Youden's indices for the algorithms.

RESULTS

The percentage of individuals with cognitive deficits was 36.2% (139 PLWH and 34 HIV-negative). The Gaussian and kernel naïve Bayes classifiers were the most predictive algorithms for neurocognitive impairment. Algorithms trained with auditory variables had average area under the curve values of 0.91 and 0.87, F1 scores (metric for precision and recall) of 0.81 and 0.76, and average accuracies of 86.3% and 81.9% respectively. Algorithms trained without auditory variables as features were statistically worse (p < .001) in both the primary measure of area under the curve (0.82/0.78) and the secondary measure of accuracy (72.3%/74.5%) for the Gaussian and kernel algorithms respectively.

CONCLUSIONS AND RELEVANCE

Auditory variables improved the prediction of cognitive function. Since auditory tests are easy-to-administer and often naturalistic tasks, they may offer objective measures or predictors of neurocognitive performance suitable for many global settings. Further research and development into using machine learning algorithms for predicting cognitive outcomes should be pursued.

MixEHR-SurG: A joint proportional hazard and guided topic model for inferring mortality-associated topics from electronic health records

Survival models can help medical practitioners to evaluate the prognostic importance of clinical variables to patient outcomes such as mortality or hospital readmission and subsequently design personalized treatment regimes. Electronic Health Records (EHRs) hold the promise for large-scale survival analysis based on systematically recorded clinical features for each patient. However, existing survival models either do not scale to high dimensional and multi-modal EHR data or are difficult to interpret. In this study, we present a supervised topic model called MixEHR-SurG to simultaneously integrate heterogeneous EHR data and model survival hazard. Our contributions are three-folds: (1) integrating EHR topic inference with Cox proportional hazards likelihood; (2) integrating patient-specific topic hyperparameters using the PheCode concepts such that each topic can be identified with exactly one PheCode-associated phenotype; (3) multi-modal survival topic inference. This leads to a highly interpretable survival topic model that can infer PheCode-specific phenotype topics associated with patient mortality. We evaluated MixEHR-SurG using a simulated dataset and two real-world EHR datasets: the Quebec Congenital Heart Disease (CHD) data consisting of 8211 subjects with 75,187 outpatient claim records of 1767 unique ICD codes; the MIMIC-III consisting of 1458 subjects with multi-modal EHR records. Compared to the baselines, MixEHR-SurG achieved a superior dynamic AUROC for mortality prediction, with a mean AUROC score of 0.89 in the simulation dataset and a mean AUROC of 0.645 on the CHD dataset. Qualitatively, MixEHR-SurG associates severe cardiac conditions with high mortality risk among the CHD patients after the first heart failure hospitalization and critical brain injuries with increased mortality among the MIMIC-III patients after their ICU discharge. Together, the integration of the Cox proportional hazards model and EHR topic inference in MixEHR-SurG not only leads to competitive mortality prediction but also meaningful phenotype topics for in-depth survival analysis. The software is available at GitHub: https://github.com/li-lab-mcgill/MixEHR-SurG.

Ensemble Machine Learning for Predicting 90-Day Outcomes and Analyzing Risk Factors in Acute Kidney Injury Requiring Dialysis

Purpose

Our objectives were to (1) employ ensemble machine learning algorithms utilizing real-world clinical data to predict 90-day prognosis, including dialysis dependence and mortality, following the first hospitalized dialysis and (2) identify the significant factors associated with overall outcomes.

Patients and Methods

We identified hospitalized patients with Acute kidney injury requiring dialysis (AKI-D) from a dataset of the Taipei Medical University Clinical Research Database (TMUCRD) from January 2008 to December 2020. The extracted data comprise demographics, comorbidities, medications, and laboratory parameters. Ensemble machine learning models were developed utilizing real-world clinical data through the Google Cloud Platform.

Results

The Study Analyzed 1080 Patients in the Dialysis-Dependent Module, Out of Which 616 Received Regular Dialysis After 90 Days. Our Ensemble Model, Consisting of 25 Feedforward Neural Network Models, Demonstrated the Best Performance with an Auroc of 0.846. We Identified the Baseline Creatinine Value, Assessed at Least 90 Days Before the Initial Dialysis, as the Most Crucial Factor. We selected 2358 patients, 984 of whom were deceased after 90 days, for the survival module. The ensemble model, comprising 15 feedforward neural network models and 10 gradient-boosted decision tree models, achieved superior performance with an AUROC of 0.865. The pre-dialysis creatinine value, tested within 90 days prior to the initial dialysis, was identified as the most significant factor.

Conclusion

Ensemble machine learning models outperform logistic regression models in predicting outcomes of AKI-D, compared to existing literature. Our study, which includes a large sample size from three different hospitals, supports the significance of the creatinine value tested before the first hospitalized dialysis in determining overall prognosis. Healthcare providers could benefit from utilizing our validated prediction model to improve clinical decision-making and enhance patient care for the high-risk population.

Explainable Deep-Learning Prediction for Brain-Computer Interfaces Supported Lower Extremity Motor Gains Based on Multistate Fusion

Predicting the potential for recovery of motor function in stroke patients who undergo specific rehabilitation treatments is an important and major challenge. Recently, electroencephalography (EEG) has shown potential in helping to determine the relationship between cortical neural activity and motor recovery. EEG recorded in different states could more accurately predict motor recovery than single-state recordings. Here, we design a multi-state (combining eyes closed, EC, and eyes open, EO) fusion neural network for predicting the motor recovery of patients with stroke after EEG-brain-computer-interface (BCI) rehabilitation training and use an explainable deep learning method to identify the most important features of EEG power spectral density and functional connectivity contributing to prediction. The prediction accuracy of the multi-states fusion network was 82%, significantly improved compared with a single-state model. The neural network explanation result demonstrated the important region and frequency oscillation bands. Specifically, in those two states, power spectral density and functional connectivity were shown as the regions and bands related to motor recovery in frontal, central, and occipital. Moreover, the motor recovery relation in bands, the power spectrum density shows the bands at delta and alpha bands. The functional connectivity shows the delta, theta, and alpha bands in the EC state; delta, theta, and beta mid at the EO state are related to motor recovery. Multi-state fusion neural networks, which combine multiple states of EEG signals into a single network, can increase the accuracy of predicting motor recovery after BCI training, and reveal the underlying mechanisms of motor recovery in brain activity.

MLe-KCNQ2: An Artificial Intelligence Model for the Prognosis of Missense KCNQ2 Gene Variants

Despite the increasing availability of genomic data and enhanced data analysis procedures, predicting the severity of associated diseases remains elusive in the absence of clinical descriptors. To address this challenge, we have focused on the KV7.2 voltage-gated potassium channel gene (KCNQ2), known for its link to developmental delays and various epilepsies, including self-limited benign familial neonatal epilepsy and epileptic encephalopathy. Genome-wide tools often exhibit a tendency to overestimate deleterious mutations, frequently overlooking tolerated variants, and lack the capacity to discriminate variant severity. This study introduces a novel approach by evaluating multiple machine learning (ML) protocols and descriptors. The combination of genomic information with a novel Variant Frequency Index (VFI) builds a robust foundation for constructing reliable gene-specific ML models. The ensemble model, MLe-KCNQ2, formed through logistic regression, support vector machine, random forest and gradient boosting algorithms, achieves specificity and sensitivity values surpassing 0.95 (AUC-ROC > 0.98). The ensemble MLe-KCNQ2 model also categorizes pathogenic mutations as benign or severe, with an area under the receiver operating characteristic curve (AUC-ROC) above 0.67. This study not only presents a transferable methodology for accurately classifying KCNQ2 missense variants, but also provides valuable insights for clinical counseling and aids in the determination of variant severity. The research context emphasizes the necessity of precise variant classification, especially for genes like KCNQ2, contributing to the broader understanding of gene-specific challenges in the field of genomic research. The MLe-KCNQ2 model stands as a promising tool for enhancing clinical decision making and prognosis in the realm of KCNQ2-related pathologies.

Interpreting Stroke-Impaired Electromyography Patterns through Explainable Artificial Intelligence

Electromyography (EMG) proves invaluable myoelectric manifestation in identifying neuromuscular alterations resulting from ischemic strokes, serving as a potential marker for diagnostics of gait impairments caused by ischemia. This study aims to develop an interpretable machine learning (ML) framework capable of distinguishing between the myoelectric patterns of stroke patients and those of healthy individuals through Explainable Artificial Intelligence (XAI) techniques. The research included 48 stroke patients (average age 70.6 years, 65% male) undergoing treatment at a rehabilitation center, alongside 75 healthy adults (average age 76.3 years, 32% male) as the control group. EMG signals were recorded from wearable devices positioned on the bicep femoris and lateral gastrocnemius muscles of both lower limbs during indoor ground walking in a gait laboratory. Boosting ML techniques were deployed to identify stroke-related gait impairments using EMG gait features. Furthermore, we employed XAI techniques, such as Shapley Additive Explanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and Anchors to interpret the role of EMG variables in the stroke-prediction models. Among the ML models assessed, the GBoost model demonstrated the highest classification performance (AUROC: 0.94) during cross-validation with the training dataset, and it also overperformed (AUROC: 0.92, accuracy: 85.26%) when evaluated using the testing EMG dataset. Through SHAP and LIME analyses, the study identified that EMG spectral features contributing to distinguishing the stroke group from the control group were associated with the right bicep femoris and lateral gastrocnemius muscles. This interpretable EMG-based stroke prediction model holds promise as an objective tool for predicting post-stroke gait impairments. Its potential application could greatly assist in managing post-stroke rehabilitation by providing reliable EMG biomarkers and address potential gait impairment in individuals recovering from ischemic stroke.

Association between co-exposure to phenols, phthalates, and polycyclic aromatic hydrocarbons with the risk of frailty

The phenomenon of population aging has brought forth the challenge of frailty. Nevertheless, the contribution of environmental exposure to frailty remains ambiguous. Our objective was to investigate the association between phenols, phthalates (PAEs), and polycyclic aromatic hydrocarbons (PAHs) with frailty. We constructed a 48-item frailty index using data from the National Health and Nutrition Examination Survey (NHANES). The exposure levels of 20 organic contaminants were obtained from the survey circle between 2005 and 2016. The association between individual organic contaminants and the frailty index was assessed using negative binomial regression models. The combined effect of organic contaminants was examined using weighted quantile sum (WQS) regression. Dose-response patterns were modeled using generalized additive models (GAMs). Additionally, an interpretable machine learning approach was employed to develop a predictive model for the frailty index. A total of 1566 participants were included in the analysis. Positive associations were observed between exposure to MIB, P02, ECP, MBP, MHH, MOH, MZP, MC1, and P01 with the frailty index. WQS regression analysis revealed a significant increase in the frailty index with higher levels of the mixture of organic contaminants (aOR, 1.12; 95% CI, 1.05-1.20; p < 0.001), with MIB, ECP, COP, MBP, P02, and P01 identified as the major contributors. Dose-response relationships were observed between MIB, ECP, MBP, P02, and P01 exposure with an increased risk of frailty (both with p < 0.05). The developed predictive model based on organic contaminants exposure demonstrated high performance, with an R2 of 0.9634 and 0.9611 in the training and testing sets, respectively. Furthermore, the predictive model suggested potential synergistic effects in the MIB-MBP and P01-P02 pairs. Taken together, these findings suggest a significant association between exposure to phthalates and PAHs with an increased susceptibility to frailty.

Advancing microbiome research with machine learning: key findings from the ML4Microbiome COST action

The rapid development of machine learning (ML) techniques has opened up the data-dense field of microbiome research for novel therapeutic, diagnostic, and prognostic applications targeting a wide range of disorders, which could substantially improve healthcare practices in the era of precision medicine. However, several challenges must be addressed to exploit the benefits of ML in this field fully. In particular, there is a need to establish "gold standard" protocols for conducting ML analysis experiments and improve interactions between microbiome researchers and ML experts. The Machine Learning Techniques in Human Microbiome Studies (ML4Microbiome) COST Action CA18131 is a European network established in 2019 to promote collaboration between discovery-oriented microbiome researchers and data-driven ML experts to optimize and standardize ML approaches for microbiome analysis. This perspective paper presents the key achievements of ML4Microbiome, which include identifying predictive and discriminatory 'omics' features, improving repeatability and comparability, developing automation procedures, and defining priority areas for the novel development of ML methods targeting the microbiome. The insights gained from ML4Microbiome will help to maximize the potential of ML in microbiome research and pave the way for new and improved healthcare practices.

Effects of heavy metal exposure on hypertension: A machine learning modeling approach

Heavy metal exposure is a common risk factor for hypertension. To develop an interpretable predictive machine learning (ML) model for hypertension based on levels of heavy metal exposure, data from the NHANES (2003-2016) were employed. Random forest (RF), support vector machine (SVM), decision tree (DT), multilayer perceptron (MLP), ridge regression (RR), AdaBoost (AB), gradient boosting decision tree (GBDT), voting classifier (VC), and K-nearest neighbour (KNN) algorithms were utilized to generate an optimal predictive model for hypertension. Three interpretable methods, the permutation feature importance analysis, partial dependence plot (PDP), and Shapley additive explanations (SHAP) methods, were integrated into a pipeline and embedded in ML for model interpretation. A total of 9005 eligible individuals were randomly allocated into two distinct sets for predictive model training and validation. The results showed that among the predictive models, the RF model demonstrated the highest performance, achieving an accuracy rate of 77.40% in the validation set. The AUC and F1 score for the model were 0.84 and 0.76, respectively. Blood Pb, urinary Cd, urinary Tl, and urinary Co levels were identified as the main influencers of hypertension, and their contribution weights were 0.0504 ± 0.0482, 0.0389 ± 0.0256, 0.0307 ± 0.0179, and 0.0296 ± 0.0162, respectively. Blood Pb (0.55-2.93 μg/dL) and urinary Cd (0.06-0.15 μg/L) levels exhibited the most pronounced upwards trend with the risk of hypertension within a specific value range, while urinary Tl (0.06-0.26 μg/L) and urinary Co (0.02-0.32 μg/L) levels demonstrated a declining trend with hypertension. The findings on the synergistic effects indicated that Pb and Cd were the primary determinants of hypertension. Our findings underscore the predictive value of heavy metals for hypertension. By utilizing interpretable methods, we discerned that Pb, Cd, Tl, and Co emerged as noteworthy contributors within the predictive model.

Artificial intelligence, machine learning, and deep learning in liver transplantation

Liver transplantation (LT) is a life-saving treatment for individuals with end-stage liver disease. The management of LT recipients is complex, predominantly because of the need to consider demographic, clinical, laboratory, pathology, imaging, and omics data in the development of an appropriate treatment plan. Current methods to collate clinical information are susceptible to some degree of subjectivity; thus, clinical decision-making in LT could benefit from the data-driven approach offered by artificial intelligence (AI). Machine learning and deep learning could be applied in both the pre- and post-LT settings. Some examples of AI applications pre-transplant include optimising transplant candidacy decision-making and donor-recipient matching to reduce waitlist mortality and improve post-transplant outcomes. In the post-LT setting, AI could help guide the management of LT recipients, particularly by predicting patient and graft survival, along with identifying risk factors for disease recurrence and other associated complications. Although AI shows promise in medicine, there are limitations to its clinical deployment which include dataset imbalances for model training, data privacy issues, and a lack of available research practices to benchmark model performance in the real world. Overall, AI tools have the potential to enhance personalised clinical decision-making, especially in the context of liver transplant medicine.

The predictors of death within 1 year in acute ischemic stroke patients based on machine learning

OBJECTIVE

To explore the predictors of death in acute ischemic stroke (AIS) patients within 1 year based on machine learning (ML) algorithms.

METHODS

This study retrospectively analyzed the clinical data of patients hospitalized and diagnosed with AIS in the Second Affiliated Hospital of Xuzhou Medical University between August 2017 and July 2019. The patients were randomly divided into training and validation sets at a ratio of 7:3, and the clinical characteristic variables of the patients were screened using univariate and multivariate logistics regression. Six ML algorithms, including logistic regression (LR), gradient boosting machine (GBM), extreme gradient boosting (XGB), random forest (RF), decision tree (DT), and naive Bayes classifier (NBC), were applied to develop models to predict death in AIS patients within 1 year. During training, a 10-fold cross-validation approach was used to validate the training set internally, and the models were interpreted using important ranking and the SHapley Additive exPlanations (SHAP) principle. The validation set was used to externally validate the models. Ultimately, the highest-performing model was selected to build a web-based calculator.

RESULTS

Multivariate logistic regression analysis revealed that C-reactive protein (CRP), homocysteine (HCY) levels, stroke severity (SS), and the number of stroke lesions (NOS) were independent risk factors for death within 1 year in patients with AIS. The area under the curve value of the XGB model was 0.846, which was the highest among the six ML algorithms. Therefore, we built an ML network calculator (https://mlmedicine-de-stroke-de-stroke-m5pijk.streamlitapp.com/) based on XGB to predict death in AIS patients within 1 year.

CONCLUSIONS

The network calculator based on the XGB model developed in this study can help clinicians make more personalized and rational clinical decisions.

Deep Learning in Population Genetics

Population genetics is transitioning into a data-driven discipline thanks to the availability of large-scale genomic data and the need to study increasingly complex evolutionary scenarios. With likelihood and Bayesian approaches becoming either intractable or computationally unfeasible, machine learning, and in particular deep learning, algorithms are emerging as popular techniques for population genetic inferences. These approaches rely on algorithms that learn non-linear relationships between the input data and the model parameters being estimated through representation learning from training data sets. Deep learning algorithms currently employed in the field comprise discriminative and generative models with fully connected, convolutional, or recurrent layers. Additionally, a wide range of powerful simulators to generate training data under complex scenarios are now available. The application of deep learning to empirical data sets mostly replicates previous findings of demography reconstruction and signals of natural selection in model organisms. To showcase the feasibility of deep learning to tackle new challenges, we designed a branched architecture to detect signals of recent balancing selection from temporal haplotypic data, which exhibited good predictive performance on simulated data. Investigations on the interpretability of neural networks, their robustness to uncertain training data, and creative representation of population genetic data, will provide further opportunities for technological advancements in the field.

Disease classification for whole-blood DNA methylation: Meta-analysis, missing values imputation, and XAI

BACKGROUND

DNA methylation has a significant effect on gene expression and can be associated with various diseases. Meta-analysis of available DNA methylation datasets requires development of a specific workflow for joint data processing.

RESULTS

We propose a comprehensive approach of combined DNA methylation datasets to classify controls and patients. The solution includes data harmonization, construction of machine learning classification models, dimensionality reduction of models, imputation of missing values, and explanation of model predictions by explainable artificial intelligence (XAI) algorithms. We show that harmonization can improve classification accuracy by up to 20% when preprocessing methods of the training and test datasets are different. The best accuracy results were obtained with tree ensembles, reaching above 95% for Parkinson's disease. Dimensionality reduction can substantially decrease the number of features, without detriment to the classification accuracy. The best imputation methods achieve almost the same classification accuracy for data with missing values as for the original data. XAI approaches have allowed us to explain model predictions from both populational and individual perspectives.

CONCLUSIONS

We propose a methodologically valid and comprehensive approach to the classification of healthy individuals and patients with various diseases based on whole-blood DNA methylation data using Parkinson's disease and schizophrenia as examples. The proposed algorithm works better for the former pathology, characterized by a complex set of symptoms. It allows to solve data harmonization problems for meta-analysis of many different datasets, impute missing values, and build classification models of small dimensionality.