Elastic Net Model Research Articles

Demystifying the Results of RTOG 0617: Identification of Dose Sensitive Cardiac Subregions Associated With Overall Survival

IntroductionThe RTOG 0617 trial presented a worse survival for patients with lung cancer treated in the high-dose (74 Gy) arm. In multivariable models, radiation level and whole-heart volumetric dose parameters were associated with survival. In this work, we consider heart subregions to explain the observed survival difference between radiation levels. MethodsVoxel-based analysis identified anatomical regions where the dose was associated with survival. Bootstrapping clinical and dosimetric variables into an elastic net model selected variables associated with survival. Multivariable Cox regression survival models assessed the significance of dose to the heart subregion, compared with whole heart v5 and v30. Finally, the trial outcome was assessed after propensity score matching of patients on lung dose, heart subregion dose, and tumor volume. ResultsA total of 458 patients were eligible for voxel-based analysis. A region of significance (p < 0.001) was identified in the base of the heart. Bootstrapping selected mean lung dose, radiation level, log tumor volume, and heart region dose. The multivariable Cox model exhibited dose to the heart region (p = 0.02), and tumor volume (p = 0.03) were significantly associated with survival, and radiation level was not significant (p = 0.07). The models exhibited that whole heart v5 and v30 were not associated with survival, with radiation level being significant (p < 0.05). In the matched cohort, no significant survival difference was seen between radiation levels. ConclusionsDose to the base of the heart is associated with overall survival, partly removing the radiation level effect, and explaining that worse survival in the high-dose arm is owing, in part, to the heart subregion dose. By defining a heart avoidance region, future dose escalation trials may be feasible.

Predicting Patient Satisfaction With Medications for Treating Opioid Use Disorder: Case Study Applying Natural Language Processing to Reviews of Methadone and Buprenorphine/Naloxone on Health-Related Social Media.

Medication-assisted treatment (MAT) is an effective method for treating opioid use disorder (OUD), which combines behavioral therapies with one of three Food and Drug Administration-approved medications: methadone, buprenorphine, and naloxone. While MAT has been shown to be effective initially, there is a need for more information from the patient perspective about the satisfaction with medications. Existing research focuses on patient satisfaction with the entirety of the treatment, making it difficult to determine the unique role of medication and overlooking the views of those who may lack access to treatment due to being uninsured or concerns over stigma. Studies focusing on patients' perspectives are also limited by the lack of scales that can efficiently collect self-reports across domains of concerns. A broad survey of patients' viewpoints can be obtained through social media and drug review forums, which are then assessed using automated methods to discover factors associated with medication satisfaction. Because the text is unstructured, it may contain a mix of formal and informal language. The primary aim of this study was to use natural language processing methods on text posted on health-related social media to detect patients' satisfaction with two well-studied OUD medications: methadone and buprenorphine/naloxone. We collected 4353 patient reviews of methadone and buprenorphine/naloxone from 2008 to 2021 posted on WebMD and Drugs.com. To build our predictive models for detecting patient satisfaction, we first employed different analyses to build four input feature sets using the vectorized text, topic models, duration of treatment, and biomedical concepts by applying MetaMap. We then developed six prediction models: logistic regression, Elastic Net, least absolute shrinkage and selection operator, random forest classifier, Ridge classifier, and extreme gradient boosting to predict patients' satisfaction. Lastly, we compared the prediction models' performance over different feature sets. Topics discovered included oral sensation, side effects, insurance, and doctor visits. Biomedical concepts included symptoms, drugs, and illnesses. The F-score of the predictive models across all methods ranged from 89.9% to 90.8%. The Ridge classifier model, a regression-based method, outperformed the other models. Assessment of patients' satisfaction with opioid dependency treatment medication can be predicted using automated text analysis. Adding biomedical concepts such as symptoms, drug name, and illness, along with the duration of treatment and topic models, had the most benefits for improving the prediction performance of the Elastic Net model compared to other models. Some of the factors associated with patient satisfaction overlap with domains covered in medication satisfaction scales (eg, side effects) and qualitative patient reports (eg, doctors' visits), while others (insurance) are overlooked, thereby underscoring the value added from processing text on online health forums to better understand patient adherence.

MiRNAs as non-invasive biomarkers in the serum of Oral Squamous Cell Carcinoma (OSCC) and Oral Potentially Malignant Disorder (OPMD) patients

ObjectiveCell-free microRNAs have shown differential levels in the serum of individuals under disease conditions suggesting its potential to act as biomarkers. A population specific miRNA signature in oral cancer is reported in different studies. We aim to identify a set of serum specific miRNAs that may differentiate oral cancer, oral pre-malignant conditions from the healthy individuals. DesignWe investigated the levels of 24 miRNAs in the serum of 47 Oral squamous cell carcinoma (OSCC) patients, 20 patients with Oral potentially malignant disorders (OPMD) and 42 healthy controls from Eastern India. Small RNAs were isolated from serum samples followed by cDNA synthesis. Levels of miRNAs were determined using qRT-PCR. The sources of serum specific miRNAs were evaluated using GTEx-RNAseq and TCGA-HNSCC database. ResultsFive miRNAs, miR-483–5p, miR-31–5p, Let-7b-5p, miR-486–5p and miR-30e-5p showed significant elevation in OSCC patients. An Elastic-Net model with 4 miRNAs classified OSCC from healthy controls with 80 % sensitivity, 64.3 % specificity, and 72.4 % accuracy. Mir-483–5p and miR-31–5p was significantly overexpressed in OSCC tissues as well as significantly higher in the serum of Leukoplakia and Verrucous carcinoma patients suggesting their potential as early disease markers. MiR-483–5p showed a consistent elevated level in the serum/plasma of oral cancer patients across different population and was found to be tumour specific while, the rest of the miRNAs showed variable results across different studies. ConclusionsOur study suggested that the serum miRNAs in oral cancer and pre-malignant disorder conditions can be used as a non-invasive marker for screening of these oral conditions.

County-level analysis on occupation and ecological determinants of child abuse and neglect rates employing elastic net regression

BackgroundOccupation is a known determinant of worker physical and behavioral health risk, yet most previous studies have focused on unemployment, underemployment, and job satisfaction to understand child maltreatment risk. ObjectiveThis county-level study (n = 278) investigated the association between occupation and child maltreatment rates and community well-being in California, Colorado, Minnesota, Oregon, and New Mexico. Participants and settingStates were selected due to having comparable, publicly available county-level data on substantiated child abuse and neglect rates within a five-year span between 2015 and 2020. MethodsUsing US Census Bureau American Community Survey data, we collected percentages of the employed population among 13 occupations. Five additional community health indicators came from the County Health Rankings and Roadmaps. Elastic net linear regression was used for variable selection and because of explanatory variables' interrelationships. Linear regression was used to model individual industries positively associated with child abuse rates. ResultsThe elastic net model selected ten important variables in explaining child maltreatment rates. Important occupational sectors were agriculture, forestry, fishing (AFF), manufacturing, wholesale, retail, finance, and education. Important community indicators included housing, injury deaths, and poor mental health days. Only AFF and retail showed greater child abuse rates with increasing percentages of the workforce in these occupations in unadjusted models (AFF: β = 0.03 SE = 0.01, p = 0.02; Retail: β = 0.09 SE = 0.04, p = 0.02). ConclusionsOur findings suggest group-level effects of counties with a larger AFF and retail presence experiencing higher child maltreatment rates. Given that numerous prior studies of county economies note the strong associations of certain employment types with cultural attitudes, educational opportunities, regional biases, and other unmeasured variables, future studies should incorporate individual level data in a multilevel framework.

Machine learning to predict foodborne salmonellosis outbreaks based on genome characteristics and meteorological trends

Several studies have shown a correlation between outbreaks of Salmonella enterica and meteorological trends, especially related to temperature and precipitation. Additionally, current studies based on outbreaks are performed on data for the species Salmonella enterica, without considering its intra-species and genetic heterogeneity. In this study, we analyzed the effect of differential gene expression and a suite of meteorological factors on salmonellosis outbreak scale (typified by case numbers) using a combination of machine learning and count-based modeling methods. Elastic Net regularization model was used to identify significant genes from a Salmonella pan-genome, and a multi-variable Poisson regression developed to fit the individual and mixed effects data. The best-fit Elastic Net model (α = 0.50; λ = 2.18) identified 53 significant gene features. The final multi-variable Poisson regression model (χ2 = 5748.22; pseudo R2 = 0.669; probability >χ2 = 0) identified 127 significant predictor terms (p < 0.10), comprising 45 gene-only predictors, average temperature, average precipitation, and average snowfall, and 79 gene-meteorological interaction terms. The significant genes ranged in functionality from cellular signaling and transport, virulence, metabolism, and stress response, and included gene variables not considered as significant by the baseline model. This study presents a holistic approach towards evaluating multiple data sources (such as genomic and environmental data) to predict outbreak scale, which could help in revising the estimates for human health risk.

Arithmetic Optimization with Ensemble Deep Transfer Learning Based燤elanoma Classification

Melanoma is a skin disease with high mortality rate while early diagnoses of the disease can increase the survival chances of patients. It is challenging to automatically diagnose melanoma from dermoscopic skin samples. Computer-Aided Diagnostic (CAD) tool saves time and effort in diagnosing melanoma compared to existing medical approaches. In this background, there is a need exists to design an automated classification model for melanoma that can utilize deep and rich feature datasets of an image for disease classification. The current study develops an Intelligent Arithmetic Optimization with Ensemble Deep Transfer Learning Based Melanoma Classification (IAOEDTT-MC) model. The proposed IAOEDTT-MC model focuses on identification and classification of melanoma from dermoscopic images. To accomplish this, IAOEDTT-MC model applies image preprocessing at the initial stage in which Gabor Filtering (GF) technique is utilized. In addition, U-Net segmentation approach is employed to segment the lesion regions in dermoscopic images. Besides, an ensemble of DL models including ResNet50 and ElasticNet models is applied in this study. Moreover, AO algorithm with Gated Recurrent Unit (GRU) method is utilized for identification and classification of melanoma. The proposed IAOEDTT-MC method was experimentally validated with the help of benchmark datasets and the proposed model attained maximum accuracy of 92.09% on ISIC 2017 dataset.

Understanding district metered area level leakage using explainable machine learning

Understanding the various interrelated effects that result in leakage is vital to the effort to reduce it. This paper aims to understand, at the district metered area (DMA) level, the relationship between leakage and static characteristics of a DMA, i.e. without considering pressure or flow. The characteristics used include the number of pipes and connections, total DMA volume and network density, as well as pipe diameter, length, age, and material statistics. Leakage, especially background and unreported leakage, can be difficult to accurately quantify. Here, the Average Weekly Minimum Night Flow (AWM) over the last 5 years is used as a proxy for leakage. While this may include some legitimate demand, it is generally assumed that minimum night flow, strongly correlates with leakage. A data-driven case study on over 800 real DMAs from UK networks is conducted. Two regression models, a decision tree model and an elastic net linear regression model, are created to predict the AWM of unseen DMAs. Reasonable accuracy was achieved, considering pressure is not an included feature, and the models are investigated for the most prominent features related to leakage.

Structural brain network deviations predict recovery after traumatic brain injury

ObjectiveTraumatic brain injury results in diffuse axonal injury and the ensuing maladaptive alterations in network function are associated with incomplete recovery and persistent disability. Despite the importance of axonal injury as an endophenotype in TBI, there is no biomarker that can measure the aggregate and region-specific burden of axonal injury. Normative modeling is an emerging quantitative case-control technique that can capture region-specific and aggregate deviations in brain networks at the individual patient level. Our objective was to apply normative modeling in TBI to study deviations in brain networks after primarily complicated mild TBI and study its relationship with other validated measures of injury severity, burden of post-TBI symptoms, and functional impairment. MethodWe analyzed 70 T1-weighted and diffusion-weighted MRIs longitudinally collected from 35 individuals with primarily complicated mild TBI during the subacute and chronic post-injury periods. Each individual underwent longitudinal blood sampling to characterize blood protein biomarkers of axonal and glial injury and assessment of post-injury recovery in the subacute and chronic periods. By comparing the MRI data of individual TBI participants with 35 uninjured controls, we estimated the longitudinal change in structural brain network deviations. We compared network deviation with independent measures of acute intracranial injury estimated from head CT and blood protein biomarkers. Using elastic net regression models, we identified brain regions in which deviations present in the subacute period predict chronic post-TBI symptoms and functional status. ResultsPost-injury structural network deviation was significantly higher than controls in both subacute and chronic periods, associated with an acute CT lesion and subacute blood levels of glial fibrillary acid protein (r = 0.5, p = 0.008) and neurofilament light (r = 0.41, p = 0.02). Longitudinal change in network deviation associated with change in functional outcome status (r = -0.51, p = 0.003) and post-concussive symptoms (BSI: r = 0.46, p = 0.03; RPQ: r = 0.46, p = 0.02). The brain regions where the node deviation index measured in the subacute period predicted chronic TBI symptoms and functional status corresponded to areas known to be susceptible to neurotrauma. ConclusionNormative modeling can capture structural network deviations, which may be useful in estimating the aggregate and region-specific burden of network changes induced by TAI. If validated in larger studies, structural network deviation scores could be useful for enrichment of clinical trials of targeted TAI-directed therapies.

Geomechanical brittleness index prediction for the Marcellus shale exploiting well-log attributes

Geomechanical brittleness of tight formations calculated from Poisson's Ratio with Young's Modulus requires laboratory testing of cores and/or estimation from shear wave acoustic logs. In shale exploitation budgets are not available to provide those measurements in many wells and it is desirable to estimate a geomechanical brittleness index (BIg) from a basic, often limited, suite of available well logs. BIg helps field operators to select the most prospective sections of tight rocks that are likely to benefit most from fracture stimulation. A machine-learning (ML) approach that combines a few recorded well logs with derivative and volatility attributes calculated from that recorded data is applied to two wellbores that penetrate the Middle Devonian formations (Appalachian Basin). This technique is novel as it has not been previously applied to predict BIg. The results demonstrate that ML models involving just two or three recorded well logs plus selected attributes as input variables provide BIg predictions that are as accurate as ML models that involve three and five recorded well logs as input variables, respectively. Moreover, the ML models involving attributes are as generalizable as ML models based only on multiple recorded well logs when applying models trained with data from one well to predict BIg from well log data in another nearby well. Extreme gradient boosting and elastic net models provide the lowest BIg prediction errors when the trained models are applied to log data from another well. Feature importance analysis identifies that the recorded compression sonic log, bulk density and gamma ray, in that order, are the most influential features in the BIg prediction models applied to the two Marcellus Shale wells. However, two of calculated well-log attributes (moving averages of the well-log derivative and volatility) are particularly effective at improving the BIg predictions when used in combination with just those three recorded well logs. The technique is worth testing in other shale basins.

206. Using Genome-Wide Association Analysis of Pneumococcal Isolates to Identify Novel Gene Variants in Central Nervous System Infection

Abstract Background Streptococcus pneumoniae is the leading cause of meningitis in children and is associated with significant morbidity and mortality. Despite conjugate pneumococcal vaccines that have otherwise been successful in reducing invasive disease caused by vaccine serotypes, the incidence of pneumococcal meningitis remains relatively unchanged. Using genome-wide association, we identified sequence elements associated with pneumococcal central nervous system (CNS) infection. Methods From 1996 to 2018, 366 clinical pneumococcal isolates were collected from pediatric patients and paired with clinical metadata. All isolates underwent whole-genome sequencing and in silico serotyping. Fifty-four clinical isolates were from children with CNS infection. Sequence element enrichment analysis was performed using elastic net and linear mixed modeling statistical methods to find genetic variants associated with CNS infection. We identified associated genes by mapping statistically associated variants to a serotype 19F reference S. pneumoniae genome. Results CNS isolates spanned 22 unique serotypes; the most common were 22F (n = 6), 7F (n = 5), 18C (n = 4), and 35B (n = 4). We identified two intergenic sequence variants and one intragenic sequence variant significantly associated with CNS infection (p &amp;lt; 1.48 × 10-7). The top intragenic candidate in our analysis was located within cpsC, a gene involved in regulation of capsule synthesis. The CNS infection-associated cpsC variant was found in serotypes 22F (n = 1), 18C (n = 4), and 14 (n = 2). Conclusion The diversity of pneumococcal isolates associated with CNS infection suggests the importance of genetic determinants beyond capsular serotype. We identified a number of novel microbial genetic variants significantly associated with pneumococcal CNS infection across serotypes, including a coding mutation in cpsC. Functional characterization of cpsC variants is needed to inform mechanisms contributing to pneumococcal CNS virulence. Disclosures Anne J. Blaschke, MD, PhD, BioFire Diagnostics/Biomerieux: Advisor/Consultant|BioFire Diagnostics/Biomerieux: Grant/Research Support|BioFire Diagnostics/Biomerieux: IP licensed to BioFire Diagnostics through the University of Utah and royalties received through the University of Utah related to the FilmArray|Merck: Advisor/Consultant.

Baseline gene expression profiling determines long-term benefit to programmed cell death protein 1 axis blockade

Treatment with immune checkpoint inhibitors has altered the course of malignant melanoma, with approximately half of the patients with advanced disease surviving for more than 5 years after diagnosis. Currently, there are no biomarker methods for predicting outcome from immunotherapy. Here, we obtained transcriptomic information from a total of 105 baseline tumor samples comprising two cohorts of patients with advanced melanoma treated with programmed cell death protein 1 (PD-1)-based immunotherapies. Gene expression profiles were correlated with progression-free survival (PFS) within consecutive clinical benefit intervals (i.e., 6, 12, 18, and 24 months). Elastic net binomial regression models with cross validation were utilized to compare the predictive value of distinct genes across time. Lasso regression was used to generate a signature predicting long-term benefit (LTB), defined as patients who remain alive and free of disease progression at 24 months post treatment initiation. We show that baseline gene expression profiles were consistently able to predict long-term immunotherapy outcomes with high accuracy. The predictive value of different genes fluctuated across consecutive clinical benefit intervals, with a distinct set of genes defining benefit at 24 months compared to earlier outcomes. A 12-gene signature was able to predict LTB following anti-PD-1 therapy with an area under the curve (AUC) equal to 0.92 and 0.74 in the training and validation set, respectively. Evaluation of LTB, via a unique signature may complement objective response classification and characterize the logistics of sustained antitumor immune responses.

An epigenetic DNA methylation clock for age estimates in Indo-Pacific bottlenose dolphins (Tursiops aduncus).

Knowledge of an animal's chronological age is crucial for understanding and predicting population demographics, survival and reproduction, but accurate age determination for many wild animals remains challenging. Previous methods to estimate age require invasive procedures, such as tooth extraction to analyse growth layers, which are difficult to carry out with large, mobile animals such as cetaceans. However, recent advances in epigenetic methods have opened new avenues for precise age determination. These 'epigenetic clocks' present a less invasive alternative and can provide age estimates with unprecedented accuracy. Here, we present a species-specific epigenetic clock based on skin tissue samples for a population of Indo-Pacific bottlenose dolphins (Tursiops aduncus) in Shark Bay, Western Australia. We measured methylation levels at 37,492 cytosine-guanine sites (CpG sites) in 165 samples using the mammalian methylation array. Chronological age estimates with an accuracy of ±1 year were available for 68 animals as part of a long-term behavioral study of this population. Using these samples with known age, we built an elastic net model with Leave-One-Out-Cross-Validation, which retained 43 CpG sites, providing an r= 0.86 and median absolute age error (MAE)=2.1 years (5% of maximum age). This model was more accurate for our data than the previously published methylation clock based on skin samples of common bottlenose dolphins (T. truncatus: r= 0.83, MAE=2.2) and the multi-species odontocete methylation clock (r= 0.68, MAE=6.8), highlighting that species-specific clocks can have superior performance over those of multi-species assemblages. We further developed an epigenetic sex estimator, predicting sex with 100% accuracy. As age and sex are critical parameters for the study of animal populations, this clock and sex estimator will provide a useful tool for extracting life history information from skin samples rather than long-term observational data for free-ranging Indo-Pacific bottlenose dolphins worldwide.

Predicting soil depth in a humid tropical watershed: A comparative analysis of best-fit regression and geospatial models

As the climate changes, water scarcity is becoming a major global concern. The accurate prediction of soil depth is critical for hydrological modeling under different climate change scenarios for the effective management of humid tropical watersheds. This study determined the best-fit assumption-based (i.e., ordinary least square (OLS)) and assumption-relaxed (partial least squared, quantile, and elastic net) regression models, and compared them with geospatial models (ordinary kriging, universal kriging, regression kriging, and cokriging), for predicting soil depth in a 50 × 150 m humid tropical watershed. Soil depth, apparent electrical soil conductivity (ECa) at two depths of exploration (ECas, 0–0.5 m, and ECad, 0–1.6 m using a DUALEM-1S EC meter), slope gradient, and soil physical and chemical properties measured in-field or using soil samples (0–0.2 m), were determined at each sampling location. Multivariate regression models explained 86–89% of the variability in soil depth and had a comparative root mean square error (RMSE), Lin’s concordance correlation coefficient (LCCC), ratio of performance to deviation (RPD), and prediction error rate (PER) to corresponding ECad univariate models. ECad was a significant predictor in all models except the elastic net regression model. The cokriging model was superior to all regression and geospatial models with the lowest RMSE (0.04) and PER (10 %), and highest LCCC (0.99) and RPD (6.71), and can be used to accurately predict soil depth across the watershed, thereby improving hydrological models for watershed management in this and similar vulnerable humid tropical watersheds. Most interestingly, our results suggest that OLS models are in some cases robust enough to handle violations of their assumptions, implying that data transformation may not always be required for OLS regression. Therefore, it is recommended that the traditional workflow for OLS regression include a validation check using an assumption-relaxed model, the outcome of which would determine the necessity of data transformation.

Cognitive Predictors of Perception and Adaptation to Dysarthric Speech in Young Adult Listeners.

Although recruitment of cognitive-linguistic resources to support dysarthric speech perception and adaptation is presumed by theoretical accounts of effortful listening and supported by cross-disciplinary empirical findings, prospective relationships have received limited attention in the disordered speech literature. This study aimed to examine the predictive relationships between cognitive-linguistic parameters and intelligibility outcomes associated with familiarization with dysarthric speech in young adult listeners. A cohort of 156 listener participants between the ages of 18 and 50 years completed a three-phase perceptual training protocol (pretest, training, and posttest) with one of three speakers with dysarthria. Additionally, listeners completed the National Institutes of Health Toolbox Cognition Battery to obtain measures of the following cognitive-linguistic constructs: working memory, inhibitory control of attention, cognitive flexibility, processing speed, and vocabulary knowledge. Elastic net regression models revealed that select cognitive-linguistic measures and their two-way interactions predicted both initial intelligibility and intelligibility improvement of dysarthric speech. While some consistency across models was shown, unique constellations of select cognitive factors and their interactions predicted initial intelligibility and intelligibility improvement of the three different speakers with dysarthria. Current findings extend empirical support for theoretical models of speech perception in adverse listening conditions to dysarthric speech signals. Although predictive relationships were complex, vocabulary knowledge, working memory, and cognitive flexibility often emerged as important variables across the models.

Longitudinal tau accumulation interacts with cortical thinning to predict steeper memory decline in cognitively normal older adults

AbstractBackgroundCross‐sectional measures of tau pathology, a hallmark of Alzheimer’s disease (AD), have been associated with brain structure and memory performance in cognitively normal older adults (OA). Fewer studies have explored the downstream effects of longitudinal tau accumulation in OA. Here, we investigated relationships between tau accumulation, atrophy, and memory performance change over time.MethodSeventy cognitively normal OA (age = 77 ± 5 years; Table1) underwent longitudinal flortaucipir (FTP) PET, structural MRI scanning, and neuropsychological assessment. Annual change slopes for entorhinal cortex (EC) and a temporal meta‐ROI partial‐volume corrected FTP, cortical thickness, and an episodic memory (EM) composite were estimated using linear mixed effects models. Linear methods (Pearson correlation, multiple regression, mixed effects models) were used to relate regional tau slopes to atrophy and EM change. We used elastic net regression to select the best predictors of EM slope, including all FTP and structural ROIs, age, education, beta‐amyloid burden, and sex. Voxel‐wise FTP slope images were also generated to explore associations with atrophy and EM slope across the cortex.ResultHigher FTP slope in EC and meta‐ROI significantly predicted EM decline (β= ‐.26 p=.03 &amp; β= ‐.38 p=.001, respectively; Fig.1). FTP slope did not significantly predict longitudinal atrophy in EC or meta‐ROI (ps&gt;.10). A multiple regression with predictors selected via an elastic net model (Table2) showed a significant interaction between meta‐ROI FTP slope and EC cortical thickness slope such that tau accumulation and EM decline were more strongly related in OA with a higher rate of entorhinal cortical thinning (β=.23 p=.02; Fig.2). Voxel‐wise models with tau slope and memory showed interesting findings in the temporal lobe but clusters did not reach statistical significance. A one‐sample t‐test showed significant tau accumulation in frontal, parietal, and temporal lobe regions (Fig.3).ConclusionEC and meta‐ROI tau accumulation predicts memory decline in cognitively normal OA. EC atrophy moderated the effect of meta‐ROI tau accumulation on memory decline, suggesting a synergic effect of tau pathology and structural atrophy on memory decline in early AD. These novel findings provide insights into mechanisms of memory decline in typical aging and early AD.

Approaches for the short-term prediction of natural daily streamflows using hybrid machine learning enhanced with grey wolf optimization

ABSTRACT This paper presents the development of hybrid machine learning models to forecast the natural flows of water bodies. Five models were considered under the analysis: extreme gradient boosting (XGB), extreme learning machine (ELM), support vector regression (SVR), elastic net linear model (EN), and multivariate adaptive regression splines (MARS). The grey wolf optimization algorithm (GWO) optimized all of the models’ internal parameters. A feature selection approach was embedded in the hybrid model to reduce the number of input variables. The hybrid model performed the forecasts considering one, three, five, and seven days ahead on data collected from Cahora Bassa dam, Mozambique. In the experiments conducted in this paper, XGB outperformed EN, ELM, MARS, and SVR, presenting lower prediction error and uncertainty. The proposed XGB model arises as an alternative to help with flow prediction, which is crucial for hydroelectric power plant activity.

A machine learning framework with uncertainty‐driven neuroimaging modality selection for data‐efficient prediction of Alzheimer’s Disease

AbstractBackgroundEarly prediction of the AD risk for individuals with MCI has important clinical value. Machine Learning (ML) models integrating multi‐modality neuroimaging datasets have shown great promise. Existing approaches either require all the modalities or suffer from performance loss with fewer modalities. We proposed a data‐efficient ML framework, namely the Uncertainty‐driven Modality Selection (UMoS) framework, to sequentially add modalities for each patient on an as‐needed basis while at the same time ensuring high prediction accuracy.MethodThe dataset included 1319 T1‐MRI scans from MCI patients in ADNI and among these 612 additionally had amyloid‐PET. Regional volumetric and thickness measures were computed using FreeSurfer v7.1 from MRI and regional SUVR measures were computed using our in‐house pipeline from amyloid‐PET. MRI is used in the standard of care while amyloid‐PET has less accessibility and is more expensive. Thus, the goal of UMoS was to save a patient from needing PET if doing so did not hurt prediction accuracy. To achieve this, we first trained two elastic net logistic regression models to predict the risk of conversion to AD for MCI patients: one with MRI only; the other one additionally including amyloid‐PET. Uncertainty quantification (UQ) was performed to quantify the predictive uncertainty of the MRI model. Under the UMoS framework, each individual was first predicted using the MRI model. If the predictive uncertainty was high, the individual would be predicted using the MRI+PET model. The uncertainty threshold was selected by balancing the saving of PET in the patient cohort and the prediction accuracy.ResultUMoS achieved 0.851 Area Under the Curve (AUC) with 46.2% patients predicted by the MRI‐only model and 53.8% by the MRI+PET model. If all patients were predicted by the MRI model, the AUC was 0.794. If all patients were predicted by the MRI +PET model, the AUC was 0.875.ConclusionWe proposed an ML framework, UMoS, that used predictive uncertainty of the MRI model to drive the decision that if less‐accessible amyloid‐PET was needed for predicting MCI conversion to AD. This framework allows an individualized approach to make accurate prediction based on the data available and leverage additional data only when needed.

Dementia prediction in the general population using clinically accessible variables: a proof‐of‐concept study using machine learning. The AGES‐Reykjavik Study

AbstractBackgroundEarly identification of dementia is crucial for prompt intervention and better outcomes for high‐risk individuals in the general population. The use of machine learning in dementia prediction has allowed for highly accurate models that could aid in early classification; however, they have focused on expensive predictors. Exploring predictors that are more accessible is crucial for the possible widespread use of machine learning models in clinical practice.MethodData from 4,793 individuals without dementia or mild cognitive impairment at baseline were included from the population‐based AGES‐Reykjavik Study. Cognitive, biometric, and MRI assessments (total: 64 variables) were collected at baseline, with follow‐up of incident dementia diagnoses for a maximum of 12 years. Elastic net regression, random forest, support vector machine, Naïve Bayes, logistic regression, and elastic net Cox regression (for time‐to‐event analysis) were explored as possible algorithms. Model 1 was fit using all variables, model 2 after feature selection using the Boruta package, and model 3 without neuroimaging markers (clinically accessible model). Ten‐fold cross‐validation, repeated ten times, was implemented during training. Upsampling was used to account for imbalanced data. Tuning parameters were optimized for recalibration automatically using the caret package.ResultDuring training, the Model 2 elastic net regression had the highest AUC [0.78; 95% CI: 0.76‐0.80], sensitivity [72; 95% CI: 68‐75], and specificity [72; 95% CI: 70‐73]. For Model 3, excluding MRI markers, the AUC remained high [AUC 0.75; 95% CI: 0.73‐0.77]. Similar results were found in our test data for Model 3 [AUC 0.74; 95% CI: 0.71‐0.77]; thus, the risk of overfitting was low. Time‐to‐event analysis using elastic net Cox models showed similar discrimination [c‐statistic 0.79] during testing. The most important variables based on Boruta selection included the Activities of Daily Living, presence of APOE e4 allele, memory functioning, and sex.ConclusionSupervised machine learning could be used to identify individuals at high‐risk for dementia in the general population using easily accessible variables. As dementia becomes a leading problem in developing countries, this clinically accessible model could be explored for use in these areas for better identification of individuals at risk in the community. Further external validation is needed.

Illumination levels affect the prediction of soil organic carbon using smartphone-based digital images

Soil organic carbon (SOC) analyses based on wet chemistry are labor-intensive and time-consuming and smartphone-based digital images have been used as a rapid, and low-cost alternative for the assessment of SOC content. We explored the effect of four illumination levels (300, 600, 900, and 1,200 lx) on the prediction accuracy of SOC using smartphone-based digital images. A total of 611 surfaces (0–20 cm) soil samples were collected from under pasture, agriculture, and in a former lakebed. Under controlled conditions, digital images were captured using an Apple iPhone 11 under four illumination levels. A total of 5 color coordinates (RGB, HSV, CIE L*a*b*, CIE L*u*v*, and CIE XYZ) were extracted from the images for each illumination level, and these were used to predict SOC by means of Lasso, elastic net (en), ridge, random forest (rf), and support vector machine (svm) models. The R, G, B, S, V, L*, a*, b*, u*, v*, X, Y, and Z color parameters were significantly affected (P < 0.05) by the illumination levels. The 300 lx illumination level had a fair prediction accuracy for SOC (r2 = 0.42 to 0.70) using random forest. The a*, R, and H color parameters were the best variables based on the 300 lx dataset for the prediction of SOC. We conclude that illumination levels significantly affected the prediction accuracy of SOC; lower illumination levels improved the prediction accuracy of SOC content.

Application of elastic net in quantitative analysis of major elements using Martian laser-induced breakdown spectroscopy datasets

Multiple sets of laser-induced breakdown spectroscopy (LIBS) instruments, including ChemCam (Curiosity), SuperCam (Perseverance), and the Mars Surface Composition Detector (MarSCoDe, Zhurong), are working currently on the surface of Mars, a booming scene of LIBS technology application to planetary exploration. One of the primary challenges faced by those LIBS instruments is accurate and quantitative chemical composition determination of unexplored geological targets. Elastic net is a linear regression model that combines ridge regression and lasso model, which inherits the sparseness of lasso and the stability of ridge. In this work, we investigated the spectral features selected by elastic net model, the model performance and application to over 23,000 LIBS points. Selected features exhibit significant emission lines that are attributed to a certain single element and a few lines attributed to other related elements. These features enhance the interpretability of the model and weaken the matrix effect in LIBS quantitative analysis. In comparison, the results of elastic net are comparable and may be slightly better in some cases than other common linear regression models. For igneous calibration targets (norite, picrite, and shergottite), the predictions of elastic net are closer to the values measured in Earth laboratory than the final multivariate oxide composition values predicted by ChemCam team. Finally, we predicted the oxide abundances of 8 major elements from Mars LIBS spectra of over 23,000 points using elastic net. The predicted values of the elastic net models are highly correlated with those from the ChemCam team model. These results indicate that elastic net is viable for quantitative analysis Mars LIBS spectra. We propose that elastic net is an important candidate model for the quantitative analysis of the Zhurong MarSCoDe LIBS spectra.