Bias in random forest variable importance measures: illustrations, sources and a solution

Bipolar disorder: Construction and analysis of a joint diagnostic model using random forest and feedforward neural networks

Background

To construct a diagnostic model for Bipolar Disorder (BD) depressive phase using peripheral tissue RNA data from patients and combining Random Forest with Feedforward Neural Network methods.

Methods

Datasets GSE23848, GSE39653, and GSE69486 were selected, and differential gene expression analysis was conducted using the limma package in R. Key genes from the differentially expressed genes were identified using the Random Forest method. These key genes' expression levels in each sample were used to train a Feedforward Neural Network model. Techniques like L1 regularization, early stopping, and dropout layers were employed to prevent model overfitting. Model performance was then validated, followed by GO, KEGG, and protein-protein interaction network analyses.

Results

The final model was a Feedforward Neural Network with two hidden layers and two dropout layers, comprising 2345 trainable parameters. Model performance on the validation set, assessed through 1000 bootstrap resampling iterations, demonstrated a specificity of 0.769 (95 % CI 0.571-1.000), sensitivity of 0.818 (95 % CI 0.533-1.000), AUC value of 0.832 (95 % CI 0.642-0.979), and accuracy of 0.792 (95 % CI 0.625-0.958). Enrichment analysis of key genes indicated no significant enrichment in any known pathways.

Conclusion

Key genes with biological significance were identified based on the decrease in Gini coefficient within the Random Forest model. The combined use of Random Forest and Feedforward Neural Network to establish a diagnostic model showed good classification performance in Bipolar Disorder.

Survival analysis for lung cancer patients: A comparison of Cox regression and machine learning models

INTRODUCTION

Survival analysis based on cancer registry data is of paramount importance for monitoring the effectiveness of health care. As new methods arise, the compendium of statistical tools applicable to cancer registry data grows. In recent years, machine learning approaches for survival analysis were developed. The aim of this study is to compare the model performance of the well established Cox regression and novel machine learning approaches on a previously unused dataset.

MATERIAL AND METHODS

The study is based on lung cancer data from the Schleswig-Holstein Cancer Registry. Four survival analysis models are compared: Cox Proportional Hazard Regression (CoxPH) as the most commonly used statistical model, as well as Random Survival Forests (RSF) and two neural network architectures based on the DeepSurv and TabNet approaches. The models are evaluated using the concordance index (C-I), the Brier score and the AUC-ROC score. In addition, to gain more insight in the decision process of the models, we identified the features that have an higher impact on patient survival using permutation feature importance scores and SHAP values.

RESULTS

Using a dataset including the cancer stage established by the Union for International Cancer Control (UICC), the best performing model is the CoxPH (C-I: 0.698±0.005), while using a dataset which includes the tumor size, lymph node and metastasis status (TNM) leads to the RSF as best performing model (C-I: 0.703±0.004). The explainability metrics show that the models rely on the combined UICC stage and the metastasis status in the first place, which corresponds to other studies.

DISCUSSION

The studied methods are highly relevant for epidemiological researchers to create more accurate survival models, which can help physicians make informed decisions about appropriate therapies and management of patients with lung cancer, ultimately improving survival and quality of life.

Prediction of developmental toxic effects of fine particulate matter (PM2.5) water-soluble components via machine learning through observation of PM2.5 from diverse urban areas

The global health implications of fine particulate matter (PM2.5) underscore the imperative need for research into its toxicity and chemical composition. In this study, zebrafish embryos exposed to the water-soluble components of PM2.5 from two cities (Harbin and Hangzhou) with differences in air quality, underwent microscopic examination to identify primary target organs. The Harbin PM2.5 induced dose-dependent organ malformation in zebrafish, indicating a higher level of toxicity than that of the Hangzhou sample. Harbin PM2.5 led to severe deformities such as pericardial edema and a high mortality rate, while the Hangzhou sample exhibited hepatotoxicity, causing delayed yolk sac absorption. The experimental determination of PM2.5 constituents was followed by the application of four algorithms for predictive toxicological assessment. The random forest algorithm correctly predicted each of the effect classes and showed the best performance, suggesting that zebrafish malformation rates were strongly correlated with water-soluble components of PM2.5. Feature selection identified the water-soluble ions F- and Cl- and metallic elements Al, K, Mn, and Be as potential key components affecting zebrafish development. This study provides new insights into the developmental toxicity of PM2.5 and offers a new approach for predicting and exploring the health effects of PM2.5.

Understanding overfitting in random forest for probability estimation: a visualization and simulation study

BACKGROUND

Random forests have become popular for clinical risk prediction modeling. In a case study on predicting ovarian malignancy, we observed training AUCs close to 1. Although this suggests overfitting, performance was competitive on test data. We aimed to understand the behavior of random forests for probability estimation by (1) visualizing data space in three real-world case studies and (2) a simulation study.

METHODS

For the case studies, multinomial risk estimates were visualized using heatmaps in a 2-dimensional subspace. The simulation study included 48 logistic data-generating mechanisms (DGM), varying the predictor distribution, the number of predictors, the correlation between predictors, the true AUC, and the strength of true predictors. For each DGM, 1000 training datasets of size 200 or 4000 with binary outcomes were simulated, and random forest models were trained with minimum node size 2 or 20 using the ranger R package, resulting in 192 scenarios in total. Model performance was evaluated on large test datasets (N = 100,000).

RESULTS

The visualizations suggested that the model learned "spikes of probability" around events in the training set. A cluster of events created a bigger peak or plateau (signal), isolated events local peaks (noise). In the simulation study, median training AUCs were between 0.97 and 1 unless there were 4 binary predictors or 16 binary predictors with a minimum node size of 20. The median discrimination loss, i.e., the difference between the median test AUC and the true AUC, was 0.025 (range 0.00 to 0.13). Median training AUCs had Spearman correlations of around 0.70 with discrimination loss. Median test AUCs were higher with higher events per variable, higher minimum node size, and binary predictors. Median training calibration slopes were always above 1 and were not correlated with median test slopes across scenarios (Spearman correlation - 0.11). Median test slopes were higher with higher true AUC, higher minimum node size, and higher sample size.

CONCLUSIONS

Random forests learn local probability peaks that often yield near perfect training AUCs without strongly affecting AUCs on test data. When the aim is probability estimation, the simulation results go against the common recommendation to use fully grown trees in random forest models.

Antigenic drift and subtype interference shape A(H3N2) epidemic dynamics in the United States

Influenza viruses continually evolve new antigenic variants, through mutations in epitopes of their major surface proteins, hemagglutinin (HA) and neuraminidase (NA). Antigenic drift potentiates the reinfection of previously infected individuals, but the contribution of this process to variability in annual epidemics is not well understood. Here, we link influenza A(H3N2) virus evolution to regional epidemic dynamics in the United States during 1997-2019. We integrate phenotypic measures of HA antigenic drift and sequence-based measures of HA and NA fitness to infer antigenic and genetic distances between viruses circulating in successive seasons. We estimate the magnitude, severity, timing, transmission rate, age-specific patterns, and subtype dominance of each regional outbreak and find that genetic distance based on broad sets of epitope sites is the strongest evolutionary predictor of A(H3N2) virus epidemiology. Increased HA and NA epitope distance between seasons correlates with larger, more intense epidemics, higher transmission, greater A(H3N2) subtype dominance, and a greater proportion of cases in adults relative to children, consistent with increased population susceptibility. Based on random forest models, A(H1N1) incidence impacts A(H3N2) epidemics to a greater extent than viral evolution, suggesting that subtype interference is a major driver of influenza A virus infection ynamics, presumably via heterosubtypic cross-immunity.

Initial evaluation of a personalized advantage index to determine which individuals may benefit from mindfulness-based cognitive therapy for suicide prevention

OBJECTIVE

Develop and evaluate a treatment matching algorithm to predict differential treatment response to Mindfulness-Based Cognitive Therapy for suicide prevention (MBCT-S) versus enhanced treatment-as-usual (eTAU).

METHODS

Analyses used data from Veterans at high-risk for suicide assigned to either MBCT-S (n = 71) or eTAU (n = 69) in a randomized clinical trial. Potential predictors (n = 55) included available demographic, clinical, and neurocognitive variables. Random forest models were used to predict risk of suicidal event (suicidal behaviors, or ideation resulting in hospitalization or emergency department visit) within 12 months following randomization, characterize the prediction, and develop a Personalized Advantage Index (PAI).

RESULTS

A slightly better prediction model emerged for MBCT-S (AUC = 0.70) than eTAU (AUC = 0.63). Important outcome predictors for participants in the MBCT-S arm included PTSD diagnosis, decisional efficiency on a neurocognitive task (Go/No-Go), prior-year mental health residential treatment, and non-suicidal self-injury. Significant predictors for participants in the eTAU arm included past-year acute psychiatric hospitalizations, past-year outpatient psychotherapy visits, past-year suicidal ideation severity, and attentional control (indexed by Stroop task). A moderation analysis showed that fewer suicidal events occurred among those randomized to their PAI-indicated optimal treatment.

CONCLUSIONS

PAI-guided treatment assignment may enhance suicide prevention outcomes. However, prior to real-world application, additional research is required to improve model accuracy and evaluate model generalization.

Child maltreatment in a high adversity context: Associations of age, type and timing of exposure with psychopathology in middle childhood

BACKGROUND

While cumulative childhood maltreatment (CM) has been linked to psychopathological outcomes, recent studies point to the relevance of the type and timing of exposure. The aim of the current study was to better understand their importance beyond the cumulative burden of CM for psychopathological symptoms in middle childhood.

METHODS

A total of N = 341 children (M = 9.92, SD = 1.51) were interviewed to assess trauma load (UCLA - University of California at Los Angeles Event List), exposure to CM (pediMACE - Maltreatment and Abuse Chronology of Exposure - Pediatric Interview) and different outcomes of psychopathology (UCLA Posttraumatic Stress Disorder Reaction Index, Children's Depression Inventory (CDI), Strengths and Difficulties Questionnaire (SDQ). We employed conditioned random forest regression, incorporating type, timing, and cumulative indicators of CM, to assess the importance of each predictor simultaneously.

RESULTS

Exposure to CM (abuse, neglect and cumulative indicators) exhibited a robust association with psychopathological outcomes. Recent abuse and recent neglect showed most robust associations with outcomes, neglect was stronger related to internalizing problems and timing of exposure showed clear associations with diverse pathological outcomes.

CONCLUSION

Beyond the cumulative burden, type and timing of CM show direct and diverse associations to pathological outcomes in middle childhood. Our results highlight the critical importance of early and detailed identification of CM, particularly recent exposure. This finding is valuable for researchers and clinicians, as it can refine diagnostic assessments and pave the way for effective early intervention strategies for affected children.

Random forest machine learning for maize yield and agronomic efficiency prediction in Ghana

Maize (Zea mays) is an important staple crop for food security in Sub-Saharan Africa. However, there is need to increase production to feed a growing population. In Ghana, this is mainly done by increasing acreage with adverse environmental consequences, rather than yield increment per unit area. Accurate prediction of maize yields and nutrient use efficiency in production is critical to making informed decisions toward economic and ecological sustainability. We trained the random forest machine learning algorithm to predict maize yield and agronomic efficiency in Ghana using soil, climate, environment, and management factors, including fertilizer application. We calibrated and evaluated the performance of the random forest machine learning algorithm using a 5 × 10-fold nested cross-validation approach. Data from 482 maize field trials consisting of 3136 georeferenced treatment plots conducted in Ghana from 1991 to 2020 were used to train the algorithm, identify important predictor variables, and quantify the uncertainties associated with the random forest predictions. The mean error, root mean squared error, model efficiency coefficient and 90 % prediction interval coverage probability were calculated. The results obtained on test data demonstrate good prediction performance for yield (MEC = 0.81) and moderate performance for agronomic efficiency (MEC = 0.63, 0.55 and 0.54 for AE-N, AE-P and AE-K, respectively). We found that climatic variables were less important predictors than soil variables for yield prediction, but temperature was of key importance to yield prediction and rainfall to agronomic efficiency. The developed random forest models provided a better understanding of the drivers of maize yield and agronomic efficiency in a tropical climate and an insight towards improving fertilizer recommendations for sustainable maize production and food security in Sub-Saharan Africa.

Knowledge-slanted random forest method for high-dimensional data and small sample size with a feature selection application for gene expression data

The use of prior knowledge in the machine learning framework has been considered a potential tool to handle the curse of dimensionality in genetic and genomics data. Although random forest (RF) represents a flexible non-parametric approach with several advantages, it can provide poor accuracy in high-dimensional settings, mainly in scenarios with small sample sizes. We propose a knowledge-slanted RF that integrates biological networks as prior knowledge into the model to improve its performance and explainability, exemplifying its use for selecting and identifying relevant genes. knowledge-slanted RF is a combination of two stages. First, prior knowledge represented by graphs is translated by running a random walk with restart algorithm to determine the relevance of each gene based on its connection and localization on a protein-protein interaction network. Then, each relevance is used to modify the selection probability to draw a gene as a candidate split-feature in the conventional RF. Experiments in simulated datasets with very small sample sizes ( n ≤ 30 ) comparing knowledge-slanted RF against conventional RF and logistic lasso regression, suggest an improved precision in outcome prediction compared to the other methods. The knowledge-slanted RF was completed with the introduction of a modified version of the Boruta feature selection algorithm. Finally, knowledge-slanted RF identified more relevant biological genes, offering a higher level of explainability for users than conventional RF. These findings were corroborated in one real case to identify relevant genes to calcific aortic valve stenosis.

Random forest analysis and lasso regression outperform traditional methods in identifying missing data auxiliary variables when the MAR mechanism is nonlinear (p.s. Stop using Little's MCAR test)

The selection of auxiliary variables is an important first step in appropriately implementing missing data methods such as full information maximum likelihood (FIML) estimation or multiple imputation. However, practical guidelines and statistical tests for selecting useful auxiliary variables are somewhat lacking, leading to potentially biased estimates. We propose the use of random forest analysis and lasso regression as alternative methods to select auxiliary variables, particularly in situations in which the missing data pattern is nonlinear or otherwise complex (i.e., interactive relationships between variables and missingness). Monte Carlo simulations demonstrate the effectiveness of random forest analysis and lasso regression compared to traditional methods (t-tests, Little's MCAR test, logistic regressions), in terms of both selecting auxiliary variables and the performance of said auxiliary variables when incorporated in an analysis with missing data. Both techniques outperformed traditional methods, providing a promising direction for improvement of practical methods for handling missing data in statistical analyses.

Earth observation data uncover green spaces' role in mental health

The prevalence of mental health disorders, a key disability cause, is linked to demographic and socioeconomic factors. However, limited data exists on mental health and the urban environment. Urbanization exposes populations to environmental stressors, particularly affecting low-middle-income countries with complex urban arrangements. We used remote sensing and census data to investigate potential connections between environmental factors and mental health disorders. Land cover variables were assessed using the European Space Agency (ESA) global WorldCover product at 10 m resolution together with the database of mental health diagnosed cases (n = 5769) from the Brazilian Unified Health System's Department of Informatics (DATASUS) from every health facility of the city of Porto Alegre. The association of mental health data with land cover was established with machine learning algorithms and polynomial regression models. The results suggest that higher trees cover at neighborhood level was associated with better mental health index. A lower mental health index was also found to be associated with an higher Human Development Index. Our results highlight the potential of greenness in the city environment to achieve substantially better mental health outcomes.

Metabolic syndrome predictive modelling in Bangladesh applying machine learning approach

Metabolic syndrome (MetS) is a cluster of interconnected metabolic risk factors, including abdominal obesity, high blood pressure, and elevated fasting blood glucose levels, that result in an increased risk of heart disease and stroke. In this research, we aim to identify the risk factors that have an impact on MetS in the Bangladeshi population. Subsequently, we intend to construct predictive machine learning (ML) models and ultimately, assess the accuracy and reliability of these models. In this particular study, we utilized the ATP III criteria as the basis for evaluating various health parameters from a dataset comprising 8185 participants in Bangladesh. After employing multiple ML algorithms, we identified that 27.8% of the population exhibited a prevalence of MetS. The prevalence of MetS was higher among females, accounting for 58.3% of the cases, compared to males with a prevalence of 41.7%. Initially, we identified the crucial variables using Chi-Square and Random Forest techniques. Subsequently, the obtained optimal variables are employed to train various models including Decision Trees, Random Forests, Support Vector Machines, Extreme Gradient Boosting, K-nearest neighbors, and Logistic Regression. Particularly we employed the ATP III criteria, which utilizes the Waist-to-Height Ratio (WHtR) as an anthropometric index for diagnosing abdominal obesity. Our analysis indicated that Age, SBP, WHtR, FBG, WC, DBP, marital status, HC, TGs, and smoking emerged as the most significant factors when using Chi-Square and Random Forest analyses. However, further investigation is necessary to evaluate its precision as a classification tool and to improve the accuracy of all classifiers for MetS prediction.

PCA, PC-CVA, and Random Forest of GCIB-SIMS Data for the Elucidation of Bacterial Envelope Differences in Antibiotic Resistance Research

Antibiotic resistance can rapidly spread through bacterial populations via bacterial conjugation. The bacterial membrane has an important role in facilitating conjugation, thus investigating the effects on the bacterial membrane caused by conjugative plasmids, antibiotic resistance, and genes involved in conjugation is of interest. Analysis of bacterial membranes was conducted using gas cluster ion beam-secondary ion mass spectrometry (GCIB-SIMS). The complexity of the data means that data analysis is important for the identification of changes in the membrane composition. Preprocessing of data and several analytical methods for identification of changes in bacterial membranes have been investigated. GCIB-SIMS data from Escherichia coli samples were subjected to principal components analysis (PCA), principal components-canonical variate analysis (PC-CVA), and Random Forests (RF) data analysis with the aim of extracting the maximum biological information. The influence of increasing replicate data was assessed, and the effect of diminishing biological variation was studied. Optimized m/z region-specific scaling provided improved clustering, with an increase in biologically significant peaks contributing to the loadings. PC-CVA improved clustering, provided clearer loadings, and benefited from larger data sets collected over several months. RF required larger sample numbers and while showing overlap with the PC-CVA, produced additional peaks of interest. The combination of PC-CVA and RF allowed very subtle differences between bacterial strains and growth conditions to be elucidated for the first time. Specifically, comparative analysis of an E. coli strain with and without the F-plasmid revealed changes in cyclopropanation of fatty acids, where the addition of the F-plasmid led to a reduction in cyclopropanation.

AbMelt: Learning antibody thermostability from molecular dynamics

Antibody thermostability is challenging to predict from sequence and/or structure. This difficulty is likely due to the absence of direct entropic information. Herein, we present AbMelt where we model the inherent flexibility of homologous antibody structures using molecular dynamics simulations at three temperatures and learn the relevant descriptors to predict the temperatures of aggregation (Tagg), melt onset (Tm,on), and melt (Tm). We observed that the radius of gyration deviation of the complementarity determining regions at 400 K is the highest Pearson correlated descriptor with aggregation temperature (rp = -0.68 ± 0.23) and the deviation of internal molecular contacts at 350 K is the highest correlated descriptor with both Tm,on (rp = -0.74 ± 0.04) as well as Tm (rp = -0.69 ± 0.03). Moreover, after descriptor selection and machine learning regression, we predict on a held-out test set containing both internal and public data and achieve robust performance for all endpoints compared with baseline models (Tagg R2 = 0.57 ± 0.11, Tm,on R2 = 0.56 ± 0.01, and Tm R2 = 0.60 ± 0.06). In addition, the robustness of the AbMelt molecular dynamics methodology is demonstrated by only training on <5% of the data and outperforming more traditional machine learning models trained on the entire data set of more than 500 internal antibodies. Users can predict thermostability measurements for antibody variable fragments by collecting descriptors and using AbMelt, which has been made available.

Development of a High-Resolution Indoor Radon Map Using a New Machine Learning-Based Probabilistic Model and German Radon Survey Data

BACKGROUND

Radon is a carcinogenic, radioactive gas that can accumulate indoors and is undetected by human senses. Therefore, accurate knowledge of indoor radon concentration is crucial for assessing radon-related health effects or identifying radon-prone areas.

OBJECTIVES

Indoor radon concentration at the national scale is usually estimated on the basis of extensive measurement campaigns. However, characteristics of the sampled households often differ from the characteristics of the target population owing to the large number of relevant factors that control the indoor radon concentration, such as the availability of geogenic radon or floor level. Furthermore, the sample size usually does not allow estimation with high spatial resolution. We propose a model-based approach that allows a more realistic estimation of indoor radon distribution with a higher spatial resolution than a purely data-based approach.

METHODS

A multistage modeling approach was used by applying a quantile regression forest that uses environmental and building data as predictors to estimate the probability distribution function of indoor radon for each floor level of each residential building in Germany. Based on the estimated probability distribution function, a probabilistic Monte Carlo sampling technique was applied, enabling the combination and population weighting of floor-level predictions. In this way, the uncertainty of the individual predictions is effectively propagated into the estimate of variability at the aggregated level.

RESULTS

The results show an approximate lognormal distribution of indoor radon in dwellings in Germany with an arithmetic mean of 63  Bq / m 3 , a geometric mean of 41  Bq / m 3 , and a 95th percentile of 180  Bq / m 3 . The exceedance probabilities for 100 and 300  Bq / m 3 are 12.5% (10.5 million people affected) and 2.2% (1.9 million people affected), respectively. In large cities, individual indoor radon concentration is generally estimated to be lower than in rural areas, which is due to the different distribution of the population on floor levels.

DISCUSSION

The advantages of our approach are that is yields a) an accurate estimation of indoor radon concentration even if the survey is not fully representative with respect to floor level and radon concentration in soil, and b) an estimate of the indoor radon distribution with a much higher spatial resolution than basic descriptive statistics. https://doi.org/10.1289/EHP14171.

Exploring cranial macromorphoscopic variation and classification accuracy in a South African sample

To date South African forensic anthropologists are only able to successfully apply a metric approach to estimate population affinity when constructing a biological profile from skeletal remains. While a non-metric, or macromorphoscopic approach exists, limited research has been conducted to explore its use in a South African population. This study aimed to explore 17 cranial macromorphoscopic traits to develop improved methodology for the estimation of population affinity among black, white and coloured South Africans and for the method to be compliant with standards of best practice. The trait frequency distributions revealed substantial group variation and overlap, and not a single trait can be considered characteristic of any one population group. Kruskal-Wallis and Dunn's tests demonstrated significant population differences for 13 of the 17 traits. Random forest modelling was used to develop classification models to assess the reliability and accuracy of the traits in identifying population affinity. Overall, the model including all traits obtained a classification accuracy of 79% when assessing population affinity, which is comparable to current craniometric methods. The variable importance indicates that all the traits contributed some information to the model, with the inferior nasal margin, nasal bone contour, and nasal aperture shape ranked the most useful for classification. Thus, this study validates the use of macromorphoscopic traits in a South African sample, and the population-specific data from this study can potentially be incorporated into forensic casework and skeletal analyses in South Africa to improve population affinity estimates.

Exploring the determinants of under-five mortality and morbidity from infectious diseases in Cambodia-a traditional and machine learning approach

Cambodia has made progress in reducing the under-five mortality rate and burden of infectious diseases among children over the last decades. However the determinants of child mortality and morbidity in Cambodia is not well understood, and no recent analysis has been conducted to investigate possible determinants. We applied a multivariable logistical regression model and a conditional random forest to explore possible determinants of under-five mortality and under-five child morbidity from infectious diseases using the most recent Demographic Health Survey in 2021-2022. Our findings show that the majority (58%) of under-five deaths occurred during the neonatal period. Contraceptive use of the mother led to lower odds of under-five mortality (0.51 [95% CI 0.32-0.80], p-value 0.003), while being born fourth or later was associated with increased odds (3.25 [95% CI 1.09-9.66], p-value 0.034). Improved household water source and higher household wealth quintile was associated with lower odds of infectious disease while living in the Great Lake or Coastal region led to increased odds respectively. The odds ratios were consistent with the results from the conditional random forest. The study showcases how closely related child mortality and morbidity due to infectious disease are to broader social development in Cambodia and the importance of accelerating progress in many sectors to end preventable child mortality and morbidity.

Unravelling large-scale patterns and drivers of biodiversity in dry rivers

More than half of the world's rivers dry up periodically, but our understanding of the biological communities in dry riverbeds remains limited. Specifically, the roles of dispersal, environmental filtering and biotic interactions in driving biodiversity in dry rivers are poorly understood. Here, we conduct a large-scale coordinated survey of patterns and drivers of biodiversity in dry riverbeds. We focus on eight major taxa, including microorganisms, invertebrates and plants: Algae, Archaea, Bacteria, Fungi, Protozoa, Arthropods, Nematodes and Streptophyta. We use environmental DNA metabarcoding to assess biodiversity in dry sediments collected over a 1-year period from 84 non-perennial rivers across 19 countries on four continents. Both direct factors, such as nutrient and carbon availability, and indirect factors such as climate influence the local biodiversity of most taxa. Limited resource availability and prolonged dry phases favor oligotrophic microbial taxa. Co-variation among taxa, particularly Bacteria, Fungi, Algae and Protozoa, explain more spatial variation in community composition than dispersal or environmental gradients. This finding suggests that biotic interactions or unmeasured ecological and evolutionary factors may strongly influence communities during dry phases, altering biodiversity responses to global changes.

Exploring Responsiveness to Highly Challenging Balance and Gait Training in Parkinson's Disease

BACKGROUND

Exercise potentially improves gait, balance, and habitual physical activity in Parkinson's disease (PD). However, given the heterogeneous nature of the disease, it is likely that people respond differently to exercise interventions. Factors determining responsiveness to exercise interventions remain unclear.

OBJECTIVES

To address this uncertainty, we explored the responsiveness to our highly challenging balance and gait intervention (HiBalance) in people with PD.

METHODS

Thirty-nine participants with mild-moderate PD who underwent the HiBalance intervention from our randomized controlled trial were included. We defined response in three domains: (1) balance based on Mini-BESTest, (2) gait based on gait velocity, and (3) physical activity based on accelerometry-derived steps per day. In each domain, we explored three responsiveness levels: high, low, or non-responders according to the change from pre- to post-intervention. Separate Random Forests for each responder domain classified these responsiveness levels and identified variable importance.

RESULTS

Only the Random Forest for the balance domain classified all responsiveness levels above the chance level indicated by a Cohen's kappa of "slight" agreement. Variable importance differed among the responsiveness levels. Slow gait velocity indicated high responders in the balance domain but showed low probabilities for low and non-responders. For low and non-responders, fall history or no falls, respectively, were more important.

CONCLUSIONS

Among three responder domains and responsiveness levels, we could moderately classify responders in the balance domain, but not for the gait or physical activity domain. This can guide inclusion criteria for balance-targeted, personalized intervention studies in people with PD.

EXPLANA: A user-friendly workflow for EXPLoratory ANAlysis and feature selection in cross-sectional and longitudinal microbiome studies

Motivation

Longitudinal microbiome studies (LMS) are increasingly common but have analytic challenges including non-independent data requiring mixed-effects models and large amounts of data that motivate exploratory analysis to identify factors related to outcome variables. Although change analysis (i.e. calculating deltas between values at different timepoints) can be powerful, how to best conduct these analyses is not always clear. For example, observational LMS measurements show natural fluctuations, so baseline might not be a reference of primary interest; whereas, for interventional LMS, baseline is a key reference point, often indicating the start of treatment.

Results

To address these challenges, we developed a feature selection workflow for cross-sectional and LMS that supports numerical and categorical data called EXPLANA (EXPLoratory ANAlysis). Machine-learning methods were combined with different types of change calculations and downstream interpretation methods to identify statistically meaningful variables and explain their relationship to outcomes. EXPLANA generates an interactive report that textually and graphically summarizes methods and results. EXPLANA had good performance on simulated data, with an average area under the curve (AUC) of 0.91 (range: 0.79-1.0, SD = 0.05), outperformed an existing tool (AUC: 0.95 vs. 0.56), and identified novel order-dependent categorical feature changes. EXPLANA is broadly applicable and simplifies analytics for identifying features related to outcomes of interest.

Research on CO2-WAG in Thick Reservoirs: Geological Influencing Factors and Random Forest Importance Evaluation

In the development process of thick reservoirs, the impact of various geological factors on the effectiveness of the CO2 water alternating gas (CO2-WAG) flooding technology remains unclear. This paper establishes multiple CO2-WAG flooding models for thick reservoirs to study the effects of sedimentary rhythm, dip angle, matrix permeability, high-permeability streaks (HPS), and barrier layers on the effectiveness of CO2-WAG flooding and then uses the random forest algorithm to rank the importance of these geological factors. The results show that different geological factors have varying degrees of impact on the distribution of water and gas migration and recovery rates during the CO2-WAG flooding process. The ranking of the importance of various factors obtained by reservoir numerical simulations and the random forest algorithm is HPS, sedimentary rhythm, dip angle, matrix permeability, and barrier layers. These research findings will provide effective guidance and a reference for the optimal selection of CO2-WAG flooding schemes for similar thick reservoirs under different geological conditions.

The impact of parental acceptance and childhood maltreatment on mental health and physical pain in Burundian survivors of childhood sexual abuse

BACKGROUND

Parental support has been suggested to mitigate mental and physical consequences following childhood sexual abuse (CSA). However, many CSA survivors experience parental rejection post-CSA.

OBJECTIVE

We aimed to understand the impact of abuse-specific parental acceptance on post-traumatic stress disorder (PTSD) and physical pain in Burundian CSA-survivors. We further assessed the significance of parental acceptance among known risk factors for predicting PTSD.

METHODS, PARTICIPANTS, AND SETTINGS

Participants (N = 131, 80.9 % female, mean age 16.21 years) were recruited via primary health care centers for survivors of sexual violence which survivors approached post-CSA. Survivors reported on PTSD symptoms, daytime/nighttime pain, and adverse childhood experiences in semi-structured interviews. Parental acceptance levels were categorized (acceptance, no acceptance, no contact) for mothers and fathers separately. Kruskal-Wallis tests assessed group differences. Conditional random forests (CRF) evaluated the significance of parental acceptance in predicting PTSD symptom severity.

RESULTS

No significant differences regarding PTSD symptoms and physical pain between levels of maternal acceptance were obtained. Pairwise comparisons revealed significant differences in PTSD symptom severity between paternal acceptance and no acceptance (d = 1.04) and paternal acceptance and no contact (d = 0.81). The CRF identified paternal acceptance as important variable for the prediction of PTSD symptom severity. Even though results were less conclusive, medium effect sizes hint at less pain perception within the paternal acceptance group.

CONCLUSIONS

The results highlight paternal acceptance as a potential risk or protective factor regarding psychological and possibly physical well-being in the aftermath of CSA, even in the context of other known risk factors.

An evaluation of statistical models of microcystin detection in lakes applied forward under varying climate conditions

Algal blooms can threaten human health if cyanotoxins such as microcystin are produced by cyanobacteria. Regularly monitoring microcystin concentrations in recreational waters to inform management action is a tool for protecting public health; however, monitoring cyanotoxins is resource- and time-intensive. Statistical models that identify waterbodies likely to produce microcystin can help guide monitoring efforts, but variability in bloom severity and cyanotoxin production among lakes and years makes prediction challenging. We evaluated the skill of a statistical classification model developed from water quality surveys in one season with low temporal replication but broad spatial coverage to predict if microcystin is likely to be detected in a lake in subsequent years. We used summertime monitoring data from 128 lakes in Iowa (USA) sampled between 2017 and 2021 to build and evaluate a predictive model of microcystin detection as a function of lake physical and chemical attributes, watershed characteristics, zooplankton abundance, and weather. The model built from 2017 data identified pH, total nutrient concentrations, and ecogeographic variables as the best predictors of microcystin detection in this population of lakes. We then applied the 2017 classification model to data collected in subsequent years and found that model skill declined but remained effective at predicting microcystin detection (area under the curve, AUC ≥ 0.7). We assessed if classification skill could be improved by assimilating the previous years' monitoring data into the model, but model skill was only minimally enhanced. Overall, the classification model remained reliable under varying climatic conditions. Finally, we tested if early season observations could be combined with a trained model to provide early warning for late summer microcystin detection, but model skill was low in all years and below the AUC threshold for two years. The results of these modeling exercises support the application of correlative analyses built on single-season sampling data to monitoring decision-making, but similar investigations are needed in other regions to build further evidence for this approach in management application.

Impact of climate change on future flood susceptibility projections under shared socioeconomic pathway scenarios in South Asia using artificial intelligence algorithms

This study investigated the impact of climate change on flood susceptibility in six South Asian countries Afghanistan, Bangladesh, Bhutan, Bharat (India), Nepal, and Pakistan-under two distinct Shared Socioeconomic Pathway (SSP) scenarios: SSP1-2.6 and SSP5-5.8, for 2041-2060 and 2081-2100. To predict flood susceptibility, we employed three artificial intelligence (AI) algorithms: the K-nearest neighbor (KNN), conditional inference random forest (CIRF), and regularized random forest (RRF). Predictions were based on data from 2452 historical flood events, alongside climatic variables measured over monthly, seasonal, and annual timeframes. The innovative aspect of this research is the emphasis on using climatic variables across these progressively condensed timeframes, specifically addressing eight precipitation factors. The performance evaluation, employing the area under the receiver operating characteristic curve (AUC) metric, identified the RRF model as the most accurate, with the highest AUC of 0.94 during the testing phase, followed by the CIRF (AUC = 0.91) and the KNN (AUC = 0.86). An analysis of variable importance highlighted the substantial role of certain climatic factors, namely precipitation in the warmest quarter, annual precipitation, and precipitation during the wettest month, in the modeling of flood susceptibility in South Asia. The resultant flood susceptibility maps demonstrated the influence of climate change scenarios on susceptibility classifications, signalling a dynamic landscape of flood-prone areas over time. The findings revealed variable trends under different climate change scenarios and periods, with marked differences in the percentage of areas classified as having high and very high flood susceptibility. Overall, this study advances our understanding of how climate change affects flood susceptibility in South Asia and offers an essential tool for assessing and managing flood risks in the region.

Validation of Variables for Use in Pediatric Obesity Risk Score Development in Demographically and Racially Diverse United States Cohorts

OBJECTIVE

To evaluate the performance of childhood obesity prediction models in four independent cohorts in the United States, using previously validated variables obtained easily from medical records as measured in different clinical settings.

STUDY DESIGN

Data from four prospective cohorts, Latinx, Eating, and Diabetes; Stress in Pregnancy Study; Project Viva; and Center for the Health Assessment of Mothers and Children of Salinas were used to test childhood obesity risk models and predict childhood obesity by ages 4 through 6, using five clinical variables (maternal age, maternal prepregnancy body mass index, birth weight Z-score, weight-for-age Z-score change, and breastfeeding), derived from a previously validated risk model and as measured in each cohort's clinical setting. Multivariable logistic regression was performed within each cohort, and performance of each model was assessed based on discrimination and predictive accuracy.

RESULTS

The risk models performed well across all four cohorts, achieving excellent discrimination. The area under the receiver operator curve was 0.79 for Center for the Health Assessment of Mothers and Children of Salinas and Project Viva, 0.83 for Stress in Pregnancy Study, and 0.86 for Latinx, Eating, and Diabetes. At a 50th percentile threshold, the sensitivity of the models ranged from 12% to 53%, and specificity was ≥ 90%. The negative predictive values were ≥ 80% for all cohorts, and the positive predictive values ranged from 62% to 86%.

CONCLUSION

All four risk models performed well in each independent and demographically diverse cohort, demonstrating the utility of these five variables for identifying children at high risk for developing early childhood obesity in the United States.

Harnessing Machine Learning to Uncover Hidden Patterns in Azole-Resistant CYP51/ERG11 Proteins

Fungal resistance is a public health concern due to the limited availability of antifungal resources and the complexities associated with treating persistent fungal infections. Azoles are thus far the primary line of defense against fungi. Specifically, azoles inhibit the conversion of lanosterol to ergosterol, producing defective sterols and impairing fluidity in fungal plasmatic membranes. Studies on azole resistance have emphasized specific point mutations in CYP51/ERG11 proteins linked to resistance. Although very insightful, the traditional approach to studying azole resistance is time-consuming and prone to errors during meticulous alignment evaluation. It relies on a reference-based method using a specific protein sequence obtained from a wild-type (WT) phenotype. Therefore, this study introduces a machine learning (ML)-based approach utilizing molecular descriptors representing the physiochemical attributes of CYP51/ERG11 protein isoforms. This approach aims to unravel hidden patterns associated with azole resistance. The results highlight that descriptors related to amino acid composition and their combination of hydrophobicity and hydrophilicity effectively explain the slight differences between the resistant non-wild-type (NWT) and WT (nonresistant) protein sequences. This study underscores the potential of ML to unravel nuanced patterns in CYP51/ERG11 sequences, providing valuable molecular signatures that could inform future endeavors in drug development and computational screening of resistant and nonresistant fungal lineages.

Interpretable Machine Learning-Based Risk Score for Predicting Ten-Year Corneal Graft Survival After Penetrating Keratoplasty and Deep Anterior Lamellar Keratoplasty in Asian Eyes

PURPOSE

To predict 10-year graft survival after deep anterior lamellar keratoplasty (DALK) and penetrating keratoplasty (PK) using a machine learning (ML)-based interpretable risk score.

METHODS

Singapore Corneal Transplant Registry patients (n = 1687) who underwent DALK (n = 524) or PK (n = 1163) for optical indications (excluding endothelial diseases) were followed up for 10 years. Variable importance scores from random survival forests were used to identify variables associated with graft survival. Parsimonious analysis using nested Cox models selected the top factors. An ML-based clinical score generator (AutoScore) converted identified variables into an interpretable risk score. Predictive performance was evaluated using Kaplan-Meier (KM) curves and time-integrated AUC (iAUC) on an independent testing set.

RESULTS

Mean recipient age was 51.8 years, 54.1% were male, and majority were Chinese (60.0%). Surgical indications included corneal scar (46.5%), keratoconus (18.3%), and regraft (16.2%). Five-year and ten-year KM survival was 93.4% and 92.3% for DALK, compared with 67.6% and 56.6% for PK (log-rank P < 0.001). Five factors were identified by ML algorithm as predictors of 10-year graft survival: recipient sex, preoperative visual acuity, choice of procedure, surgical indication, and active inflammation. AutoScore stratified participants into low-risk and high-risk groups-with KM survival of 73.6% and 39.0%, respectively (log-rank P < 0.001). ML analysis outperformed traditional Cox regression in predicting graft survival beyond 5 years (iAUC 0.75 vs. 0.69).

CONCLUSIONS

A combination of ML and traditional techniques identified factors associated with graft failure to derive a clinically interpretable risk score to stratify PK and DALK patients-a technique that may be replicated in other corneal transplant programs.

Survival analysis in breast cancer: evaluating ensemble learning techniques for prediction

Breast cancer is most commonly faced with form of cancer amongst women worldwide. In spite of the fact that the breast cancer research and awareness have gained considerable momentum, there is still no one treatment due to disease heterogeneity. Survival data may be of specific interest in breast cancer studies to understand its dynamic and complex trajectories. This study copes with the most important covariates affecting the disease progression. The study utilizes the German Breast Cancer Study Group 2 (GBSG2) and the Molecular Taxonomy of Breast Cancer International Consortium dataset (METABRIC) datasets. In both datasets, interests lie in relapse of the disease and the time when the relapse happens. The three models, namely the Cox proportional hazards (PH) model, random survival forest (RSF) and conditional inference forest (Cforest) were employed to analyse the breast cancer datasets. The goal of this study is to apply these methods in prediction of breast cancer progression and compare their performances based on two different estimation methods: the bootstrap estimation and the bootstrap .632 estimation. The model performance was evaluated in concordance index (C-index) and prediction error curves (pec) for discrimination. The Cox PH model has a lower C-index and bigger prediction error compared to the RSF and the Cforest approach for both datasets. The analysis results of GBSG2 and METABRIC datasets reveal that the RSF and the Cforest algorithms provide non-parametric alternatives to Cox PH model for estimation of the survival probability of breast cancer patients.

Brain structure correlates of social information use: an exploratory machine learning approach

Introduction

Individual differences in social learning impact many important decisions, from voting behavior to polarization. Prior research has found that there are consistent and stable individual differences in social information use. However, the underlying mechanisms of these individual differences are still poorly understood.

Methods

We used two complementary exploratory machine learning approaches to identify brain volumes related to individual differences in social information use.

Results and discussion

Using lasso regression and random forest regression we were able to capture linear and non-linear brain-behavior relationships. Consistent with previous studies, our results suggest there is a robust positive relationship between the volume of the left pars triangularis and social information use. Moreover, our results largely overlap with common social brain network regions, such as the medial prefrontal cortex, superior temporal sulcus, temporal parietal junction, and anterior cingulate cortex. Besides, our analyses also revealed several novel regions related to individual differences in social information use, such as the postcentral gyrus, the left caudal middle frontal gyrus, the left pallidum, and the entorhinal cortex. Together, these results provide novel insights into the neural mechanisms that underly individual differences in social learning and provide important new leads for future research.

Regulating greenhouse gas dynamics in tidal wetlands: Impacts of salinity gradients and water pollution

Tidal wetlands play a critical role in emitting greenhouse gases (GHGs) into the atmosphere; our understanding of the intricate interplay between natural processes and human activities shaping their biogeochemistry and GHG emissions remains lacking. In this study, we delve into the spatiotemporal dynamics and key drivers of the GHG emissions from five tidal wetlands in the Scheldt Estuary by focusing on the interactive impacts of salinity and water pollution, two factors exhibiting contrasting gradients in this estuarine system: pollution escalates as salinity declines. Our findings reveal a marked escalation in GHG emissions when moving upstream, primarily attributed to increased concentrations of organic matter and nutrients, coupled with reduced levels of dissolved oxygen and pH. These low water quality conditions not only promote methanogenesis and denitrification to produce CH4 and N2O, respectively, but also shift the carbonate equilibria towards releasing more CO2. As a result, the most upstream freshwater wetland was the largest GHG emitter with a global warming potential around 35 to 70 times higher than the other wetlands. When moving seaward along a gradient of decreasing urbanization and increasing salinity, wetlands become less polluted and are characterized by lower concentrations of NO3-, TN and TOC, which induces stronger negative impact of elevated salinity on the GHG emissions from the saline wetlands. Consequently, these meso-to polyhaline wetlands released considerably smaller amounts of GHGs. These findings emphasize the importance of integrating management strategies, such as wetland restoration and pollution prevention, that address both natural salinity gradients and human-induced water pollution to effectively mitigate GHG emissions from tidal wetlands.

Probe-based mass spectrometry approaches for single-cell and single-organelle measurements

Exploring the chemical content of individual cells not only reveals underlying cell-to-cell chemical heterogeneity but is also a key component in understanding how cells combine to form emergent properties of cellular networks and tissues. Recent technological advances in many analytical techniques including mass spectrometry (MS) have improved instrumental limits of detection and laser/ion probe dimensions, allowing the analysis of micron and submicron sized areas. In the case of MS, these improvements combined with MS's broad analyte detection capabilities have enabled the rise of single-cell and single-organelle chemical characterization. As the chemical coverage and throughput of single-cell measurements increase, more advanced statistical and data analysis methods have aided in data visualization and interpretation. This review focuses on secondary ion MS and matrix-assisted laser desorption/ionization MS approaches for single-cell and single-organelle characterization, which is followed by advances in mass spectral data visualization and analysis.

Metabolites Associated With Uremic Symptoms in Patients With CKD: Findings From the Chronic Renal Insufficiency Cohort (CRIC) Study

RATIONALE & OBJECTIVE

The toxins that contribute to uremic symptoms in patients with chronic kidney disease (CKD) are unknown. We sought to apply complementary statistical modeling approaches to data from untargeted plasma metabolomic profiling to identify solutes associated with uremic symptoms in patients with CKD.

STUDY DESIGN

Cross-sectional.

SETTING & PARTICIPANTS

1,761 Chronic Renal Insufficiency Cohort (CRIC) participants with CKD not treated with dialysis.

PREDICTORS

Measurement of 448 known plasma metabolites.

OUTCOMES

The uremic symptoms of fatigue, anorexia, pruritus, nausea, paresthesia, and pain were assessed by single items on the Kidney Disease Quality of Life-36 instrument.

ANALYTICAL APPROACH

Multivariable adjusted linear regression, least absolute shrinkage and selection operator linear regression, and random forest models were used to identify metabolites associated with symptom severity. After adjustment for multiple comparisons, metabolites selected in at least 2 of the 3 modeling approaches were deemed "overall significant."

RESULTS

Participant mean estimated glomerular filtration rate was 43mL/min/1.73m2, with 44% self-identifying as female and 41% as non-Hispanic Black. The prevalence of uremic symptoms ranged from 22% to 55%. We identified 17 metabolites for which a higher level was associated with greater severity of at least one uremic symptom and 9 metabolites inversely associated with uremic symptom severity. Many of these metabolites exhibited at least a moderate correlation with estimated glomerular filtration rate (Pearson's r≥0.5), and some were also associated with the risk of developing kidney failure or death in multivariable adjusted Cox regression models.

LIMITATIONS

Lack of a second independent cohort for external validation of our findings.

CONCLUSIONS

Metabolomic profiling was used to identify multiple solutes associated with uremic symptoms in adults with CKD, but future validation and mechanistic studies are needed.

PLAIN-LANGUAGE SUMMARY

Individuals living with chronic kidney disease (CKD) often experience symptoms related to CKD, traditionally called uremic symptoms. It is likely that CKD results in alterations in the levels of numerous circulating substances that, in turn, cause uremic symptoms; however, the identity of these solutes is not known. In this study, we used metabolomic profiling in patients with CKD to gain insights into the pathophysiology of uremic symptoms. We identified 26 metabolites whose levels were significantly associated with at least one of the symptoms of fatigue, anorexia, itchiness, nausea, paresthesia, and pain. The results of this study lay the groundwork for future research into the biological causes of symptoms in patients with CKD.

Supervised machine learning for the prediction of post-operative clinical outcomes of hip and knee replacements: a review

Prediction models are being increasingly used in the medical field to identify risk factors and possible outcomes. Some of these are presently being used to develop guidelines for improving clinical practice. The application of machine learning (ML), comprising a powerful set of computational tools for analysing data, has been clearly expanding in the role of predictive modelling. This paper reviews the latest developments of supervised ML techniques that have been used to analyse data related to post-operative total hip and knee replacements. The aim was to review the most recent findings of relevant published studies by outlining the methodologies employed (most-widely used supervised ML techniques), data sources, domains, limitations of predictive analytics and the quality of predictions.

A Study on the Robustness and Stability of Explainable Deep Learning in an Imbalanced Setting: The Exploration of the Conformational Space of G Protein-Coupled Receptors

G-protein coupled receptors (GPCRs) are transmembrane proteins that transmit signals from the extracellular environment to the inside of the cells. Their ability to adopt various conformational states, which influence their function, makes them crucial in pharmacoproteomic studies. While many drugs target specific GPCR states to exert their effects-thereby regulating the protein's activity-unraveling the activation pathway remains challenging due to the multitude of intermediate transformations occurring throughout this process, and intrinsically influencing the dynamics of the receptors. In this context, computational modeling, particularly molecular dynamics (MD) simulations, may offer valuable insights into the dynamics and energetics of GPCR transformations, especially when combined with machine learning (ML) methods and techniques for achieving model interpretability for knowledge generation. The current study builds upon previous work in which the layer relevance propagation (LRP) technique was employed to interpret the predictions in a multi-class classification problem concerning the conformational states of the β2-adrenergic (β2AR) receptor from MD simulations. Here, we address the challenges posed by class imbalance and extend previous analyses by evaluating the robustness and stability of deep learning (DL)-based predictions under different imbalance mitigation techniques. By meticulously evaluating explainability and imbalance strategies, we aim to produce reliable and robust insights.

Considerations and targeted approaches to identifying bad actors in exposure mixtures

Variable importance is a key statistical issue in exposure mixtures, as it allows a ranking of exposures as potential targets for intervention, and helps to identify bad actors within a mixture. In settings where mixtures have many constituents or high between-constituent correlations, estimators of importance can be subject to bias or high variance. Current approaches to assessing variable importance have major limitations, including reliance on overly strong or incorrect constraints or assumptions, excessive model extrapolation, or poor interpretability, especially regarding practical significance. We sought to overcome these limitations by applying an established doubly-robust, machine learning-based approach to estimating variable importance in a mixtures context. This method reduces model extrapolation, appropriately controls confounding, and provides both interpretability and model flexibility. We illustrate its use with an evaluation of the relationship between telomere length, a measure of biologic aging, and exposure to a mixture of polychlorinated biphenyls (PCBs), dioxins, and furans among 979 US adults from the National Health and Nutrition Examination Survey (NHANES). In contrast with standard approaches for mixtures, our approach selected PCB 180 and PCB 194 as important contributors to telomere length. We hypothesize that this difference could be due to residual confounding in standard methods that rely on variable selection. Further empirical evaluation of this method is needed, but it is a promising tool in the search for bad actors within a mixture.

Machine Learning Algorithms to Predict Delayed Cerebral Ischemia After Subarachnoid Hemorrhage: A Systematic Review and Meta-analysis

Delayed cerebral ischemia (DCI) is a common and severe complication after subarachnoid hemorrhage (SAH). Logistic regression (LR) is the primary method to predict DCI, but it has low accuracy. This study assessed whether other machine learning (ML) models can predict DCI after SAH more accurately than conventional LR. PubMed, Embase, and Web of Science were systematically searched for studies directly comparing LR and other ML algorithms to forecast DCI in patients with SAH. Our main outcome was the accuracy measurement, represented by sensitivity, specificity, and area under the receiver operating characteristic. In the six studies included, comprising 1828 patients, about 28% (519) developed DCI. For LR models, the pooled sensitivity was 0.71 (95% confidence interval [CI] 0.57-0.84; p < 0.01) and the pooled specificity was 0.63 (95% CI 0.42-0.85; p < 0.01). For ML models, the pooled sensitivity was 0.74 (95% CI 0.61-0.86; p < 0.01) and the pooled specificity was 0.78 (95% CI 0.71-0.86; p = 0.02). Our results suggest that ML algorithms performed better than conventional LR at predicting DCI.Trial Registration: PROSPERO (International Prospective Register of Systematic Reviews) CRD42023441586; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=441586.

Using random forests to uncover the predictive power of distance-varying cell interactions in tumor microenvironments

Tumor microenvironments (TMEs) contain vast amounts of information on patient's cancer through their cellular composition and the spatial distribution of tumor cells and immune cell populations. Exploring variations in TMEs between patient groups, as well as determining the extent to which this information can predict outcomes such as patient survival or treatment success with emerging immunotherapies, is of great interest. Moreover, in the face of a large number of cell interactions to consider, we often wish to identify specific interactions that are useful in making such predictions. We present an approach to achieve these goals based on summarizing spatial relationships in the TME using spatial K functions, and then applying functional data analysis and random forest models to both predict outcomes of interest and identify important spatial relationships. This approach is shown to be effective in simulation experiments at both identifying important spatial interactions while also controlling the false discovery rate. We further used the proposed approach to interrogate two real data sets of Multiplexed Ion Beam Images of TMEs in triple negative breast cancer and lung cancer patients. The methods proposed are publicly available in a companion R package funkycells.

Improving Prediction of Survival and Progression in Metastatic Non-Small Cell Lung Cancer After Immunotherapy Through Machine Learning of Circulating Tumor DNA

PURPOSE

To use modern machine learning approaches to enhance and automate the feature extraction from the longitudinal circulating tumor DNA (ctDNA) data and to improve the prediction of survival and disease progression, risk stratification, and treatment strategies for patients with 1L non-small cell lung cancer (NSCLC).

MATERIALS AND METHODS

Using IMpower150 trial data on patients with untreated metastatic NSCLC treated with atezolizumab and chemotherapies, we developed a machine learning algorithm to extract predictive features from ctDNA kinetics, improving survival and progression prediction. We analyzed kinetic data from 17 ctDNA summary markers, including cell-free DNA concentration, allele frequency, tumor molecules in plasma, and mutation counts.

RESULTS

Three hundred and ninety-eight patients with ctDNA data (206 in training and 192 in validation) were analyzed. Our models outperformed existing workflow using conventional temporal ctDNA features, raising overall survival (OS) concordance index to 0.72 and 0.71 from 0.67 and 0.63 for C3D1 and C4D1, respectively, and substantially improving progression-free survival (PFS) to approximately 0.65 from the previous 0.54-0.58, a 12%-20% increase. Additionally, they enhanced risk stratification for patients with NSCLC, achieving clear OS and PFS separation. Distinct patterns of ctDNA kinetic characteristics (eg, baseline ctDNA markers, depth of ctDNA responses, and timing of ctDNA clearance, etc) were revealed across the risk groups. Rapid and complete ctDNA clearance appears essential for long-term clinical benefit.

CONCLUSION

Our machine learning approach offers a novel tool for analyzing ctDNA kinetics, extracting critical features from longitudinal data, improving our understanding of the link between ctDNA kinetics and progression/mortality risks, and optimizing personalized immunotherapies for 1L NSCLC.

Range and climate niche shifts in European and North American breeding birds

Species respond dynamically to climate change and exhibit time lags. Consequently, species may not occupy their full climatic niche during range shifting. Here, we assessed climate niche tracking during recent range shifts of European and United States (US) birds. Using data from two European bird atlases and from the North American Breeding Bird Survey between the 1980s and 2010s, we analysed range overlap and climate niche overlap based on kernel density estimation. Phylogenetic multiple regression was used to assess the effect of species morphological, ecological and biogeographic traits on range and niche metrics. European birds shifted their ranges north and north-eastwards, US birds westwards. Range unfilling was lower than expected by null models, and niche expansion was more common than niche unfilling. Also, climate niche tracking was generally lower in US birds and poorly explained by species traits. Overall, our results suggest that dispersal limitations were minor in range shifting birds in Europe and the USA while delayed extinctions from unfavourable areas seem more important. Regional differences could be related to differences in land use history and monitoring schemes. Comparative analyses of range and niche shifts provide a useful screening approach for identifying the importance of transient dynamics and time-lagged responses to climate change. This article is part of the theme issue 'Ecological novelty and planetary stewardship: biodiversity dynamics in a transforming biosphere'.

Enhancing mass spectrometry imaging accessibility using convolutional autoencoders for deriving hypoxia-associated peptides from tumors

Mass spectrometry imaging (MSI) allows to study cancer's intratumoral heterogeneity through spatially-resolved peptides, metabolites and lipids. Yet, in biomedical research MSI is rarely used for biomarker discovery. Besides its high dimensionality and multicollinearity, mass spectrometry (MS) technologies typically output mass-to-charge ratio values but not the biochemical compounds of interest. Our framework makes particularly low-abundant signals in MSI more accessible. We utilized convolutional autoencoders to aggregate features associated with tumor hypoxia, a parameter with significant spatial heterogeneity, in cancer xenograft models. We highlight that MSI captures these low-abundant signals and that autoencoders can preserve them in their latent space. The relevance of individual hyperparameters is demonstrated through ablation experiments, and the contribution from original features to latent features is unraveled. Complementing MSI with tandem MS from the same tumor model, multiple hypoxia-associated peptide candidates were derived. Compared to random forests alone, our autoencoder approach yielded more biologically relevant insights for biomarker discovery.

Multimodal imaging and electrophysiological study in the differential diagnosis of rest tremor

Introduction

Distinguishing tremor-dominant Parkinson's disease (tPD) from essential tremor with rest tremor (rET) can be challenging and often requires dopamine imaging. This study aimed to differentiate between these two diseases through a machine learning (ML) approach based on rest tremor (RT) electrophysiological features and structural MRI data.

Methods

We enrolled 72 patients including 40 tPD patients and 32 rET patients, and 45 control subjects (HC). RT electrophysiological features (frequency, amplitude, and phase) were calculated using surface electromyography (sEMG). Several MRI morphometric variables (cortical thickness, surface area, cortical/subcortical volumes, roughness, and mean curvature) were extracted using Freesurfer. ML models based on a tree-based classification algorithm termed XGBoost using MRI and/or electrophysiological data were tested in distinguishing tPD from rET patients.

Results

Both structural MRI and sEMG data showed acceptable performance in distinguishing the two patient groups. Models based on electrophysiological data performed slightly better than those based on MRI data only (mean AUC: 0.92 and 0.87, respectively; p = 0.0071). The top-performing model used a combination of sEMG features (amplitude and phase) and MRI data (cortical volumes, surface area, and mean curvature), reaching AUC: 0.97 ± 0.03 and outperforming models using separately either MRI (p = 0.0001) or EMG data (p = 0.0231). In the best model, the most important feature was the RT phase.

Conclusion

Machine learning models combining electrophysiological and MRI data showed great potential in distinguishing between tPD and rET patients and may serve as biomarkers to support clinicians in the differential diagnosis of rest tremor syndromes in the absence of expensive and invasive diagnostic procedures such as dopamine imaging.

Antigenic drift and subtype interference shape A(H3N2) epidemic dynamics in the United States

Influenza viruses continually evolve new antigenic variants, through mutations in epitopes of their major surface proteins, hemagglutinin (HA) and neuraminidase (NA). Antigenic drift potentiates the reinfection of previously infected individuals, but the contribution of this process to variability in annual epidemics is not well understood. Here we link influenza A(H3N2) virus evolution to regional epidemic dynamics in the United States during 1997-2019. We integrate phenotypic measures of HA antigenic drift and sequence-based measures of HA and NA fitness to infer antigenic and genetic distances between viruses circulating in successive seasons. We estimate the magnitude, severity, timing, transmission rate, age-specific patterns, and subtype dominance of each regional outbreak and find that genetic distance based on broad sets of epitope sites is the strongest evolutionary predictor of A(H3N2) virus epidemiology. Increased HA and NA epitope distance between seasons correlates with larger, more intense epidemics, higher transmission, greater A(H3N2) subtype dominance, and a greater proportion of cases in adults relative to children, consistent with increased population susceptibility. Based on random forest models, A(H1N1) incidence impacts A(H3N2) epidemics to a greater extent than viral evolution, suggesting that subtype interference is a major driver of influenza A virus infection dynamics, presumably via heterosubtypic cross-immunity.

Cherry on Top or Real Need? A Review of Explainable Machine Learning in Kidney Transplantation

Research on solid organ transplantation has taken advantage of the substantial acquisition of medical data and the use of artificial intelligence (AI) and machine learning (ML) to answer diagnostic, prognostic, and therapeutic questions for many years. Nevertheless, despite the question of whether AI models add value to traditional modeling approaches, such as regression models, their "black box" nature is one of the factors that have hindered the translation from research to clinical practice. Several techniques that make such models understandable to humans were developed with the promise of increasing transparency in the support of medical decision-making. These techniques should help AI to close the gap between theory and practice by yielding trust in the model by doctors and patients, allowing model auditing, and facilitating compliance with emergent AI regulations. But is this also happening in the field of kidney transplantation? This review reports the use and explanation of "black box" models to diagnose and predict kidney allograft rejection, delayed graft function, graft failure, and other related outcomes after kidney transplantation. In particular, we emphasize the discussion on the need (or not) to explain ML models for biological discovery and clinical implementation in kidney transplantation. We also discuss promising future research paths for these computational tools.

New insights into the effects of type and timing of childhood maltreatment on brain morphometry

Childhood maltreatment (CM) is known to influence brain development. To obtain a better understanding of related brain alterations, recent research has focused on the influence of the type and timing of CM. We aimed to investigate the association between type and timing of CM and local brain volume. Anatomical magnetic resonance images were collected from 93 participants (79 female/14 male) with a history of CM. CM history was assessed with the German Interview Version of the "Maltreatment and Abuse Chronology of Exposure" scale, "KERF-40 + ". Random forest regressions were performed to assess the impact of CM characteristics on the volume of amygdala, hippocampus and anterior cingulate cortex (ACC). The volume of the left ACC was predicted by neglect at age 3 and 4 and abuse at age 16 in a model including both type and timing of CM. For the right ACC, overall CM severity and duration had the greatest impact on volumetric alterations. Our data point to an influence of CM timing on left ACC volume, which was most pronounced in early childhood and in adolescence. We were not able to replicate previously reported effects of maltreatment type and timing on amygdala and hippocampal volume.

Deep-learning survival analysis for patients with calcific aortic valve disease undergoing valve replacement

Calcification of the aortic valve (CAVDS) is a major cause of aortic stenosis (AS) leading to loss of valve function which requires the substitution by surgical aortic valve replacement (SAVR) or transcatheter aortic valve intervention (TAVI). These procedures are associated with high post-intervention mortality, then the corresponding risk assessment is relevant from a clinical standpoint. This study compares the traditional Cox Proportional Hazard (CPH) against Machine Learning (ML) based methods, such as Deep Learning Survival (DeepSurv) and Random Survival Forest (RSF), to identify variables able to estimate the risk of death one year after the intervention, in patients undergoing either to SAVR or TAVI. We found that with all three approaches the combination of six variables, named albumin, age, BMI, glucose, hypertension, and clonal hemopoiesis of indeterminate potential (CHIP), allows for predicting mortality with a c-index of approximately 80 % . Importantly, we found that the ML models have a better prediction capability, making them as effective for statistical analysis in medicine as most state-of-the-art approaches, with the additional advantage that they may expose non-linear relationships. This study aims to improve the early identification of patients at higher risk of death, who could then benefit from a more appropriate therapeutic intervention.

Socio-economic variables improve accuracy and change spatial predictions in species distribution models

In an era marked by increasing anthropogenic pressure, understanding the relations between human activities and wildlife is crucial for understanding ecological patterns, effective conservation, and management strategies. Here, we explore the potential and usefulness of socio-economic variables in species distribution modelling (SDM), focusing on their impact on the occurrence of wild mammals in Poland. Beyond the environmental factors commonly considered in SDM, like land-use, the study tests the importance of socio-economic characteristics of local human societies, such as age, income, working sector, gender, education, and village characteristics for explaining distribution of diverse mammalian groups, including carnivores, ungulates, rodents, soricids, and bats. The study revealed that incorporating socio-economic variables enhances the predictive power for >60 % of species and overall for most groups, with the exception being carnivores. For all the species combined, among the 10 predictors with highest predictive power, 6 belong to socio-economic group, while for specific species groups, socio-economic variables had similar predictive power as environmental variables. Furthermore, spatial predictions of species occurrence underwent changes when socio-economic variables were included in the model, resulting in a substantial mismatch in spatial predictions of species occurrence between environment-only models and models containing socio-economic variables. We conclude that socio-economic data has potential as useful predictors which increase prediction accuracy of wildlife occurrence and recommend its wider usage. Further, to our knowledge this is a first study on such a big scale for terrestrial mammals which evaluates performance based on presence or absence of socio-economic predictors in the model. We recognise the need for a more comprehensive approach in SDMs and that bridging the gap between human socio-economic dynamics and ecological processes may contribute to the understanding of the factors influencing biodiversity.

Exploring the Association Between Structural Racism and Mental Health: Geospatial and Machine Learning Analysis

BACKGROUND

Structural racism produces mental health disparities. While studies have examined the impact of individual factors such as poverty and education, the collective contribution of these elements, as manifestations of structural racism, has been less explored. Milwaukee County, Wisconsin, with its racial and socioeconomic diversity, provides a unique context for this multifactorial investigation.

OBJECTIVE

This research aimed to delineate the association between structural racism and mental health disparities in Milwaukee County, using a combination of geospatial and deep learning techniques. We used secondary data sets where all data were aggregated and anonymized before being released by federal agencies.

METHODS

We compiled 217 georeferenced explanatory variables across domains, initially deliberately excluding race-based factors to focus on nonracial determinants. This approach was designed to reveal the underlying patterns of risk factors contributing to poor mental health, subsequently reintegrating race to assess the effects of racism quantitatively. The variable selection combined tree-based methods (random forest) and conventional techniques, supported by variance inflation factor and Pearson correlation analysis for multicollinearity mitigation. The geographically weighted random forest model was used to investigate spatial heterogeneity and dependence. Self-organizing maps, combined with K-means clustering, were used to analyze data from Milwaukee communities, focusing on quantifying the impact of structural racism on the prevalence of poor mental health.

RESULTS

While 12 influential factors collectively accounted for 95.11% of the variability in mental health across communities, the top 6 factors-smoking, poverty, insufficient sleep, lack of health insurance, employment, and age-were particularly impactful. Predominantly, African American neighborhoods were disproportionately affected, which is 2.23 times more likely to encounter high-risk clusters for poor mental health.

CONCLUSIONS

The findings demonstrate that structural racism shapes mental health disparities, with Black community members disproportionately impacted. The multifaceted methodological approach underscores the value of integrating geospatial analysis and deep learning to understand complex social determinants of mental health. These insights highlight the need for targeted interventions, addressing both individual and systemic factors to mitigate mental health disparities rooted in structural racism.

Variable Selection Using Bayesian Additive Regression Trees

Variable selection is an important statistical problem. This problem becomes more challenging when the candidate predictors are of mixed type (e.g. continuous and binary) and impact the response variable in nonlinear and/or non-additive ways. In this paper, we review existing variable selection approaches for the Bayesian additive regression trees (BART) model, a nonparametric regression model, which is flexible enough to capture the interactions between predictors and nonlinear relationships with the response. An emphasis of this review is on the ability to identify relevant predictors. We also propose two variable importance measures which can be used in a permutation-based variable selection approach, and a backward variable selection procedure for BART. We introduce these variations as a way of illustrating limitations and opportunities for improving current approaches and assess these via simulations.

Alternations (at) that time: NP versus PP time adjuncts in the history of English

The present paper investigates variation between nominal and prepositional adjuncts of time as in, for example, [on] that day, they left. The main goals are (i) to assess potential changes in the distribution of these variants in the history of English, specifically from Middle English to Late Modern English (1150-1914), and (ii) to test which factors most strongly impact the choice between the two variants, with a focus on the impact of different complexity measures. To address these questions, the paper makes use of data from the Penn-Helsinki Parsed Corpora of Historical English (PPCME2; PPCEME; PPCMBE), explored by means of logistic regression modelling. The results suggest that there is no dramatic, sweeping change in this abstract alternation over time, but that this variation may mainly plays out on lower, noun-specific levels.