Leave-one-out Cross-validation Research Articles

The Determination of Ethanol Levels in Facial Freshener Using the NIR Spectroscopy and Chemometric Method

A facial freshener, also known as toner, is a cosmetic product that is commonly used to invigorate the face after a busy day. Ethanol serves as a key component in toner, serving multiple purposes such as being a solvent, preservative, and antimicrobial agent. However, it's important to note that toner formulated for normal skin types typically contain ethanol in small concentrations, adhering to a limit of not more than 10%. Therefore, this study aims to determine ethanol levels in toner using the NIR spectroscopy and chemometric techniques. The NIR spectra of the simulated samples were correlated with ethanol concentration using chemometric calibration model. The calibration models used were partial least square (PLS), principal component regression (PCR), and support vector regression (SVR). The calibration model was validated by leave one out cross validation (LOOCV) as well as the external validation, and the precision and accuracy of the method was evaluated. Among the calibration models, the PLS model exhibited the best performance, yielding an impressive R2 0.9976; with an RMSEC value of 0.4364 and RMSECV value of 0.4704. The internal validation yield R2 value more than 0.99 and RMSE of less than 0,4198. Furthermore, external validation showed the R2 and RMSEP value of 0.989 and 0.920 respectively. The %recovery and RSD value were 101.2% and 0.129%. Comparing ethanol measurements obtained through the NIR chemometric method with those obtained using gas chromatography as the reference method, no significant difference was observed at a 95% confidence levels, as indicated by a significance value of 0.231.

Harnessing acoustic speech parameters to decipher amyloid status in individuals with mild cognitive impairment.

Alzheimer's disease (AD) is a neurodegenerative condition characterized by a gradual decline in cognitive functions. Currently, there are no effective treatments for AD, underscoring the importance of identifying individuals in the preclinical stages of mild cognitive impairment (MCI) to enable early interventions. Among the neuropathological events associated with the onset of the disease is the accumulation of amyloid protein in the brain, which correlates with decreased levels of Aβ42 peptide in the cerebrospinal fluid (CSF). Consequently, the development of non-invasive, low-cost, and easy-to-administer proxies for detecting Aβ42 positivity in CSF becomes particularly valuable. A promising approach to achieve this is spontaneous speech analysis, which combined with machine learning (ML) techniques, has proven highly useful in AD. In this study, we examined the relationship between amyloid status in CSF and acoustic features derived from the description of the Cookie Theft picture in MCI patients from a memory clinic. The cohort consisted of fifty-two patients with MCI (mean age 73 years, 65% female, and 57% positive amyloid status). Eighty-eight acoustic parameters were extracted from voice recordings using the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS), and several ML models were used to classify the amyloid status. Furthermore, interpretability techniques were employed to examine the influence of input variables on the determination of amyloid-positive status. The best model, based on acoustic variables, achieved an accuracy of 75% with an area under the curve (AUC) of 0.79 in the prediction of amyloid status evaluated by bootstrapping and Leave-One-Out Cross Validation (LOOCV), outperforming conventional neuropsychological tests (AUC = 0.66). Our results showed that the automated analysis of voice recordings derived from spontaneous speech tests offers valuable insights into AD biomarkers during the preclinical stages. These findings introduce novel possibilities for the use of digital biomarkers to identify subjects at high risk of developing AD.

An assessment of Sentinel‐1 synthetic aperture radar, geophysical and topographical covariates for estimating topsoil particle‐size fractions

AbstractData derived from synthetic aperture radar (SAR) are widely employed to predict soil properties, particularly soil moisture and soil carbon content. However, few studies address the use of microwave sensors for soil texture retrieval and those that do are typically constrained to bare soil conditions. Here, we test two statistical modelling approaches—linear (with and without interaction terms) and tree‐based models, namely compositional linear regression model (LRM) and random forest (RF)—and both nongeophysical (e.g., surface soil moisture, topographic, etc) and geophysical‐based (electromagnetic, magnetic and radiometric) covariates to estimate soil texture (sand %, silt % and clay %), using microwave remote sensing data (ESA Sentinel‐1). The statistical models evaluated explicitly consider the compositional nature of soil texture and were evaluated with leave‐one‐out cross‐validation (LOOCV). Our findings indicate that both modelling approaches yielded better estimates when fitted without the geophysical covariates. Based on the Nash–Sutcliffe efficiency coefficient (NSE), LRM slightly outperformed RF, with NSE values for sand, silt and clay of 0.94, 0.62 and 0.46, respectively; for RF, the NSE values were 0.93, 0.59 and 0.44. When interaction terms were included, RF was found to outperform LRM. The inclusion of interactions in the LRM resulted in a decrease in NSE value and an increase in the size of the residuals. Findings also indicate that the use of radar‐derived variables (e.g., VV, VH, RVI) alone was not able to predict soil particle size without the aid of other covariates. Our findings highlight the importance of explicitly considering the compositional nature of soil texture information in statistical analysis and regression modelling. As part of the continued assessment of microwave remote sensing data (e.g., ESA Sentinel‐1) for predicting topsoil particle size, we intend to test surface scattering information derived from the dual‐polarimetric decomposition technique and integrate that predictor into the models in order to deal with the effects of vegetation cover on topsoil backscattering.

Comparasion of Error Rate Prediction Methods of C4.5 Algorithm for Balanced Data

C45 is a highly effective decision tree algorithm widely used for classification purposes. Compared to CHAID, Cart, and ID3, C4.5 generates decision trees that are easier to understand and does so in a faster manner. This is due to C4.5 selecting attributes based on their information content during each stage of the process. After generating the decision tree model, its performance needs to be evaluated. One commonly used method is the prediction error rate, which assesses the model's performance. The prediction error rate consists of two approaches: the train error rate, which employs the same data for both building and testing the model, potentially leading to overfitting, and the test error rate, which divides the data into training and testing sets. The test error rate includes cross validation techniques such as Leave One Out Cross Validation (LOOCV), Hold Out (HO), and k-folds cross validation. Considering these factors, this research focuses on comparing the three cross-validation methods for predicting error rates applied to the C4.5 algorithm. The study utilizes artificially generated data with a normal distribution, including univariate, bivariate, and multivariate datasets with various combinations of mean differences and correlations. Different correlation structures are applied between two relevant variables and between relevant and irrelevant variables in the bivariate and multivariate data, including three correlation levels: no correlation, moderate correlation, and high correlation. This research findings that k-folds cross validation is the most suitable cross validation method to apply to C4.5.

Comparison of Error Rate Prediction Methods in Classification Modeling with Classification and Regression Tree (CART) Methods for Balanced Data

CART (Classification and Regression Tree) is one of the classification algorithms in the decision tree method. The model formed in CART is a tree consisting of root nodes, internal nodes, and terminal nodes. After the model is formed, it is necessary to calculate the accuracy of the model. The aims is to see the performance of the model. The accuracy of this model can be done by calculating the predicted error rate in the model. The error rate prediction method works by dividing the data into training data and testing data. There are three methods in the error rate prediction method, such as Leave One Out Cross Validation (LOOCV), Hold Out (HO), and K-Fold Cross Validation. These methods have different performance in dividing data into training data and testing data, so there are advantages and disadvantages to each method. Therefore, a comparison was made for the three error rate prediction methods with the aim of determining the appropriate method for the CART algorithm. This comparison was made by considering several factors, for instance variations in the mean, number of variables, and correlations in normal distributed random data. The results of the comparison will be observed using a boxplot by looking at the median error rate and the lowest variance. The results of this study indicate that the K-Fold Cross Validation has the median error rate and the lowest variance, so the most suitable error prediction method used for the CART method is the K-Fold Cross Validation method.

Comparison of Error Rate Prediction Methods in Classification Modeling with the CHAID Method for Imbalanced Data

CHAID (Chi-Square Automatic Interaction Detection) is one of the classification algorithms in the decision tree method. The classification results are displayed in the form of a tree diagram model. After the model is formed, it is necessary to calculate the accuracy of the model. The aims is to see the performance of the model. The accuracy of this model can be done by calculating the predicted error rate in the model. There are three methods, such as Leave one out cross-validation (LOOCV), Hold-out, and K-fold cross-validation. These methods have different performances in dividing data into training and testing data, so each method has advantages and disadvantages. Imbalanced data is data that has a different number of class observations. In the CHAID method, imbalanced data affects the prediction results. When the data is increasingly imbalanced the prediction result will approach the number of minority classes. Therefore, a comparison was made for the three error rate prediction methods to determine the appropriate method for the CHAID method in imbalanced data. This research is included in experimental research and uses simulated data from the results of generating data in RStudio. This comparison was made by considering several factors, for the marginal opportunity matrix, different correlations, and several observation ratios. The results of the comparison will be observed using a boxplot by looking at the median error rate and the lowest variance. This research finds that K-fold cross-validation is the most suitable error rate prediction method applied to the CHAID method for imbalanced data.

Ocular condition prognosis in Keratoconus patients after corneal ring implantation using artificial neural networks

The common disorder, Keratoconus (KC), is distinguished by cumulative corneal slimming and steepening. The corneal ring implantation has become a successful surgical procedure to correct the KC patient’s vision. The determination of suitable patients for the surgery alternative is among the paramount concerns of ophthalmologists. To reduce the burden on them and enhance the treatment, this research aims to previse the ocular condition of KC patients after the corneal ring implantation. It focuses on predicting post-surgical corneal topographic indices and visual characteristics. This study applied an efficacious artificial neural network approach to foretell the aforementioned ocular features of KC subjects 6 and 12 months after implanting KeraRing and MyoRing based on the accumulated data. The datasets are composed of sufficient numbers of corneal topographic maps and visual characteristics recorded from KC patients before and after implanting the rings. The visual characteristics under study are uncorrected visual acuity (UCVA), sphere (SPH), astigmatism (Ast), astigmatism orientation (Axe), and best corrected visual acuity (BCVA). In addition, the statistical data of multiple KC subjects were registered, including three effective indices of corneal topography (i.e., Ast, K-reading, and pachymetry) pre- and post-ring embedding. The outcomes represent the contribution of practical training of the introduced models to the estimation of ocular features of KC subjects following the implantation. The corneal topographic indices and visual characteristics were estimated with mean errors of 7.29% and 8.60%, respectively. Further, the errors of 6.82% and 7.65% were respectively realized for the visual characteristics and corneal topographic indices while assessing the predictions by the leave-one-out cross-validation (LOOCV) procedure. The results confirm the great potential of neural networks to guide ophthalmologists in choosing appropriate surgical candidates and their specific intracorneal rings by predicting post-implantation ocular features.

Neighborhood-based inference and restricted Boltzmann machine for small molecule-miRNA associations prediction.

A growing number of experiments have shown that microRNAs (miRNAs) can be used as target of small molecules (SMs) to regulate gene expression for treating diseases. Therefore, identifying SM-related miRNAs is helpful for the treatment of diseases in the domain of medical investigation. This article presents a new computational model, called NIRBMSMMA (neighborhood-based inference (NI) and restricted Boltzmann machine (RBM)), which we developed to identify potential small molecule-miRNA associations (NIRBMSMMA). First, grounded on known SM-miRNAs associations, SM similarity and miRNA similarity, NI was used to predict score of an unknown SM-miRNA pair by reckoning the sum of known associations between neighbors of the SM (miRNA) and the miRNA (SM). Second, utilizing a two-layered generative stochastic artificial neural network, RBM was used to predict SM-miRNA association by learning potential probability distribution from known SM-miRNA associations. At last, an ensemble learning model was conducted to combine NI and RBM for identifying potential SM-miRNA associations. Furthermore, we conducted global leave one out cross validation (LOOCV), miRNA-fixed LOOCV, SM-fixed LOOCV and five-fold cross validation to assess performance of NIRBMSMMA based on three datasets. Results showed that NIRBMSMMA obtained areas under the curve (AUC) of 0.9912, 0.9875, 0.8376 and 0.9898 ±0.0009 under global LOOCV, miRNA-fixed LOOCV, SM-fixed LOOCV and five-fold cross validation based on dataset 1, respectively. For dataset 2, the AUCs are 0.8645, 0.8720, 0.7066 and 0.8547 ±0.0046 in turn. For dataset 3, the AUCs are 0.9884, 0.9802, 0.8239 and 0.9870 ±0.0015 in turn. Also, we conducted case studies to further assess the predictive performance of NIRBMSMMA. These results illustrated the proposed model is a useful tool in predicting potential SM-miRNA associations.

Land Use Regression Models for Particle Number Concentration and Black Carbon in Lanzhou, Northwest of China

It is necessary to predict the spatial variation in particle number concentration (PNC) and black carbon (BC) because they are considered air pollutants associated with traffic and many diseases. In this study, land use regression (LUR) models for PNC and BC were developed based on a mobile monitoring campaign in January 2020 in Lanzhou, and the performance of models was evaluated with hold-out validation (HV) and leave-one-out cross-validation (LOOCV). The results show that the adjusted R2 of the LUR models for PNC and BC are 0.51 and 0.53, respectively. The R2 of HV and LOOCV are 0.43 and 0.44, respectively, for the PNC model and 0.42 and 0.50, respectively, for the BC model. The performances of the LUR models are of a moderate level. The spatial distribution of the predicted PNC is related to the distance from water bodies. The high PNC is related to industrial pollution. The BC concentration decreases from south to north. High BC concentrations are associated with freight distribution centres and coal-fired power plants. The range of PNC particle sizes in this study is larger than in most studies. As one of few studies in Lanzhou to develop LUR models of air pollutants, it is important to accurately estimate pollutant concentrations to improve air quality and provide health benefits for residents.

A new integrated framework for the identification of potential virus-drug associations.

With the increasingly serious problem of antiviral drug resistance, drug repurposing offers a time-efficient and cost-effective way to find potential therapeutic agents for disease. Computational models have the ability to quickly predict potential reusable drug candidates to treat diseases. In this study, two matrix decomposition-based methods, i.e., Matrix Decomposition with Heterogeneous Graph Inference (MDHGI) and Bounded Nuclear Norm Regularization (BNNR), were integrated to predict anti-viral drugs. Moreover, global leave-one-out cross-validation (LOOCV), local LOOCV, and 5-fold cross-validation were implemented to evaluate the performance of the proposed model based on datasets of DrugVirus that consist of 933 known associations between 175 drugs and 95 viruses. The results showed that the area under the receiver operating characteristics curve (AUC) of global LOOCV and local LOOCV are 0.9035 and 0.8786, respectively. The average AUC and the standard deviation of the 5-fold cross-validation for DrugVirus datasets are 0.8856 ± 0.0032. We further implemented cross-validation based on MDAD and aBiofilm, respectively, to evaluate the performance of the model. In particle, MDAD (aBiofilm) dataset contains 2,470 (2,884) known associations between 1,373 (1,470) drugs and 173 (140) microbes. In addition, two types of case studies were carried out further to verify the effectiveness of the model based on the DrugVirus and MDAD datasets. The results of the case studies supported the effectiveness of MHBVDA in identifying potential virus-drug associations as well as predicting potential drugs for new microbes.

Machine Learning and Quantum Calculation for Predicting Yield in Cu-Catalyzed P-H Reactions.

The paper discussed the use of machine learning (ML) and quantum chemistry calculations to predict the transition state and yield of copper-catalyzed P-H insertion reactions. By analyzing a dataset of 120 experimental data points, the transition state was determined using density functional theory (DFT). ML algorithms were then applied to analyze 16 descriptors derived from the quantum chemical transition state to predict the product yield. Among the algorithms studied, the Support Vector Machine (SVM) achieved the highest prediction accuracy of 97%, with over 80% correlation in Leave-One-Out Cross-Validation (LOOCV). Sensitivity analysis was performed on each descriptor, and a comprehensive investigation of the reaction mechanism was conducted to better understand the transition state characteristics. Finally, the ML model was used to predict reaction plans for experimental design, demonstrating strong predictive performance in subsequent experimental validation.

The Application of Adaptive Minimum Match k-Nearest Neighbors to Identify At-Risk Students in Health Professions Education.

When learners fail to reach milestones, educators often wonder if any warning signs could have allowed them to intervene sooner. Machine learning can predict which students are at risk for failing a high-stakes certification examination. If predictions can be made well before the examination, educators can meaningfully intervene before students take the examination to reduce their chances of failing. The authors used already-collected, first-year student assessment data from 5 cohorts in a single Master of Physician Assistant Studies program to implement an "adaptive minimum match" version of the k-nearest neighbors algorithm using changing numbers of neighbors to predict each student's future examination scores on the Physician Assistant National Certifying Exam (PANCE). Validation occurred in 2 ways by using leave-one-out cross-validation (LOOCV) and by evaluating predictions in a new cohort. "Adaptive minimum match" version of the k-nearest neighbors algorithm achieved an accuracy of 93% in LOOCV. "Adaptive minimum match" version of the k-nearest neighbors algorithm generates a predicted PANCE score for each student one year before they take the examination. Students are classified into extra support, optional extra support, or no extra support categories. Then, one year remains to provide appropriate support to each category of student. Predictive analytics can identify at-risk students who might need additional support or remediation before high-stakes certification examinations. Educators can use the included methods and code to generate predicted test outcomes for students. The authors recommend that educators use predictive modeling responsibly and transparently, as one of many tools used to support students. More research is needed to test alternative machine learning methods across a variety of educational programs.

Low-Flow (7-Day, 10-Year) Classical Statistical and Improved Machine Learning Estimation Methodologies

Water resource managers require accurate estimates of the 7-day, 10-year low flow (7Q10) of streams for many reasons, including protecting aquatic species, designing wastewater treatment plants, and calculating municipal water availability. StreamStats, a publicly available web application developed by the United States Geologic Survey that is commonly used by resource managers for estimating the 7Q10 in states where it is available, utilizes state-by-state, locally calibrated regression equations for estimation. This paper expands StreamStats’ methodology and improves 7Q10 estimation by developing a more regionally applicable and generalized methodology for 7Q10 estimation. In addition to classical methodologies, namely multiple linear regression (MLR) and multiple linear regression in log space (LTLR), three promising machine learning algorithms, random forest (RF) decision trees, neural networks (NN), and generalized additive models (GAM), are tested to determine if more advanced statistical methods offer improved estimation. For illustrative purposes, this methodology is applied to and verified for the full range of unimpaired, gaged basins in both the northeast and mid-Atlantic hydrologic regions of the United States (with basin sizes ranging from 2–1419 mi2) using leave-one-out cross-validation (LOOCV). Pearson’s correlation coefficient (R2), root mean square error (RMSE), Kling–Gupta Efficiency (KGE), and Nash–Sutcliffe Efficiency (NSE) are used to evaluate the performance of each method. Results suggest that each method provides varying results based on basin size, with RF displaying the smallest average RMSE (5.85) across all ranges of basin sizes.

A CT Radiomics Analysis of the Adrenal Masses: Can We Discriminate Lipid-poor Adenomas from the Pheochromocytoma and Malignant Masses?

The aim of the study is to demonstrate a non-invasive alternative method to aid the decision making process in the management of adrenal masses. Lipid-poor adenomas constitute 30% of all adrenal adenomas. When discovered incidentally, additional dynamic adrenal examinations are required to differentiate them from an adrenal malignancy or pheochromocytoma. In this retrospective study, we aimed to discriminate lipid-poor adenomas from other lipidpoor adrenal masses by using radiomics analysis in single contrast phase CT scans. A total of 38 histologically proven lipid-poor adenomas (Group 1) and 38 cases of pheochromocytoma or malignant adrenal mass (Group 2) were included in this retrospective study. Lesions were segmented volumetrically by two independent authors, and a total of 63 sizes, shapes, and first- and second-order parameters were calculated. Among these parameters, a logit-fit model was produced by using 6 parameters selected by the LASSO (least absolute shrinkage and selection operator) regression. The model was cross-validated with LOOCV (leave-one-out crossvalidation) and 1000-bootstrap sampling. A random forest model was also generated in order to use all parameters without the risk of multicollinearity. This model was examined with the nested crossvalidation method. Sensitivity, specificity, accuracy and AUC were calculated in test sets as 84.2%, 81.6%, 82.9% and 0.829 in the logit fit model and 91%, 80%, 82.8% and 0.975 in the RF model, respectively. Predictive models based on radiomics analysis using single-phase contrast-enhanced CT can help characterize adrenal lesions.

Leave-one-ion-out cross-validation for assisting in developing robust QSPR models of ionic liquids

Precisely determining thermodynamic properties such as density (ρ) and viscosity (η) at variable temperatures and pressures is necessary in the chemical process design of ionic liquids (ILs). Quantitative structure–property relationship (QSPR) models offer a quick and accurate route for obtaining the thermodynamic properties of ILs under variable environmental conditions. However, QSPR models employing the leave-one-out cross-validation (LOO-CV) method are often hindered by a “pseudo-high” robustness, weakened stability, and unbalanced data point distribution when used for determining the ρ and η of ILs. In this work, we propose a rigorous evaluation method for QSPR models, called leave-one-ion-out cross-validation (LOIO-CV), to evaluate the thermodynamic properties of ILs. By balancing the distribution of data points in ILs, two f(T,P,I)-QSPR models are developed with norm indexes (I) to predict ρ and η of ILs at variable temperatures and pressures. Our LOIO-CV method enhances the stability of QSPR models when predicting the properties of ILs with novel cations and anions, which is crucial for the data-driven design of new ILs.

Hyperspectral estimation of maize (Zea mays L.) yield loss under lodging stress

The frequency and intensity of maize (Zea mays L.) yield disturbance caused by lodging stress are getting higher and higher, so it is of great significance to take effective methods to monitor the yield loss of lodging maize. This study aimed to explore the ability of hyperspectral technology to estimate maize yield loss under lodging. The lodging control experiment of maize was carried out, the changes of canopy hyperspectral and per unit yield loss of maize in multiple growth stages were analyzed. First, the successive projections algorithm (SPA) and recursive feature elimination (RFE) methods were used to select yield-sensitive wavelengths from original canopy spectrum (OCS) to achieve dimensionality reduction. Then, the fractional-order differential (FOD) transform was applied for the canopy spectrum, and the RFE method was used to select the optimal wavelength combination. Finally, the models of estimating per unit yield of maize with different lodging periods and days after lodging (DAL) were constructed by using the optimal wavelength combination, and the model was verified by leave-one-out cross-validation (LOOCV) method. It was found that the more serious the lodging severity, the greater the maize per unit yield loss. On 1, 7, 14, 21 DAL at vegetative tasseling (VT) stage and 1, 7 DAL at reproductive milk (R3) stage, the model accuracy of RFE method was 5.26%, 17.39%, 9.46%, 6.41%, 20.37% and 11.11% higher than that of SPA method. The estimation accuracy of lodging maize yield model was improved by using FOD method and RFE method (FOD-RFE). The R2 of the maize yield model of 1.4-order, 0.9-order, 1.8-order, 0.6-order on 1, 7, 14 and 21 DAL at VT stage reached 0.90, 0.89, 0.92, 0.93, which were 12.5%, 11.11%, 13.58% and 12.05% higher than OCS-RFE. The R2 of the maize yield model of 1.5-order and 1.5-order on 1 and 7 DAL at R3 stage were 0.84 and 0.87, which were 29.23% and 24.29% higher than OCS-RFE. Therefore, fractional-order differential transform and spectral feature selection could improve the estimation accuracy of maize yield under lodging stress. The hyperspectral technology could be used to quickly estimate maize yield loss under lodging stress.

New Score Models for Assessing Disease Activity in Crohn's Disease Based on Bowel Ultrasound and Biomarkers: Ideal Surrogates for Endoscopy or Imaging.

Disease activity evaluation is important in Crohn's disease (CD). We aimed to establish new disease activity indices for CD based on noninvasive parameters. The data of 110 CD patients were retrospectively analyzed. Parameters from bowel ultrasound and biomarkers were measured to select the variables included in the models by univariate analysis. Logistic regression analysis was performed to predict mucosal and transmural activities defined by ileocolonoscopy or computed tomography enterography (CTE), respectively. The models' performance was measured by the area under the receiver operating characteristic (ROC) curve (AUC). Leave-one-out cross validation (LOOCV) was applied to adjust for overconfidence in the newly established score models. To predict mucosal activity, erythrocyte sedimentation rate (ESR) and (LimG × BWT)-SUM (the sum of the product of Limberg grade [LimG] and bowel wall thickness [BWT] of each bowel segment) were selected for Model A, and the equation was A=2 × ESR + 9.3 × (LimG × BWT)-SUM. The AUC of ROC, sensitivity, and specificity were 0.927, 89.8%, and 86.4%, respectively. The AUC of the ROC curve verified by LOOCV was 0.913. To predict transmural activity, albumin (ALB) and LimG-SUM (the sum of the LimG of all the bowel segments) were selected for Model B, which was established as B=-1.3 × ALB + 1.7 × LimG-SUM. The AUC of ROC, sensitivity, and specificity were 0.851, 78.0%, and 84.2%, respectively. The AUC of the ROC curve verified by LOOCV was 0.833. Nomograms were developed for two score models. New score models based on noninvasive parameters established in this study showed good abilities in detecting active disease and performed well in the validation phase.

Interpretable machine learning assisted multi-objective optimization design for small molecule hole transport materials

Small molecule hole transport materials (SM-HTMs) play an important role in the development of perovskite solar cells (PSCs) due to the simple structure and excellent photoelectric performance. Herein, we established interpretable models for designing materials with high hole mobility (μ) and absolute hardness (η). The μ and η models were constructed by Gradient Boosting Regression (GBR) due to the better performances. The coefficient of determination (R2) for leave-one-out cross validation (LOOCV) and the test set are 0.86 and 0.86 in μ model, 0.84 and 0.89 in η model, respectively. SHapley Additive exPlanations (SHAP) analysis reveals that higher μ are more likely to be obtained with fewer C–O and N-O pairs at topological distance 9, and molecules with high degree of cyclization and less polar fragments might have higher η. New groups were designed based on the strategies and used to generate 785 virtual samples. Finally, 35 promising candidates (μ > 10−1.54 cm2 V−1 s−1 and η > 2.71 eV) with high synthesizability were screened out using high-throughput screening and pattern recognition. Our work provides an effective way to accelerate the design of prominent SM-HTMs.

Enhanced Kriging leave-one-out cross-validation in improving model estimation and optimization

Leave-one-out cross-validation (LOOCV) is a widely used technique in model estimation and selection of the Kriging surrogate model for engineering problems, such as structural optimization and reliability analysis. However, the traditional LOOCV method has some disadvantages in terms of accuracy and efficiency. This paper proposes an enhanced-LOOCV method that incorporates hyperparameters from the Kriging model based on the complete training dataset (i.e., the complete Kriging model) into the LOOCV error calculation. By keeping the model hyperparameters in LOOCV consistent with the complete Kriging model, it reduces the number of hyperparameter optimizations and significantly increases the accuracy and efficiency of the LOOCV process. Additionally, a decremental calculation is proposed to reduce the computational costs of the correlation matrix inversion without sacrificing accuracy, resulting in an improved time complexity of the traditional LOOCV from O(n4) to O(n3). Experimental results with thirty test functions verify that the enhanced-LOOCV has better estimation performance than the Kriging model with significantly higher efficiency compared to the traditional LOOCV. Numerical experiments and an engineering case in optimization demonstrate that the enhanced-LOOCV can reduce the number of samples needed for infilling in the Kriging model, and it is more suitable for expensive optimizations in engineering.

Enhanced Differentiation of Wild and Feeding Civet Coffee Using Near-Infrared Spectroscopy with Various Sample Pretreatments and Chemometric Approaches

Civet coffee is the world’s most expensive and rarest coffee bean. Indonesia was the first country to be identified as the origin of civet coffee. First, it is produced spontaneously by collecting civet feces from coffee plantations near the forest. Due to limited stock, farmers began cultivating civets to obtain safe supplies of civet coffee. Based on this, civet coffee can be divided into two types: wild and fed. A combination of spectroscopy and chemometrics can be used to evaluate authenticity with high speed and precision. In this study, seven samples from different regions were analyzed using NIR Spectroscopy with various preparations: unroasted, roasted, unground, and ground. The spectroscopic data were combined with unsupervised exploratory methods (hierarchical cluster analysis (HCA) and principal component analysis (PCA)) and supervised classification methods (support vector machine (SVM) and random forest (RF)). The HCA results showed a trend between roasted and unroasted beans; meanwhile, the PCA showed a trend based on coffee bean regions. Combining the SVM with leave-one-out-cross-validation (LOOCV) successfully differentiated 57.14% in all sample groups (unground, ground, unroasted, unroasted–unground, and roasted–unground), 78.57% in roasted, 92.86% in roasted–ground, and 100% in unroasted–ground. However, using the Boruta filter, the accuracy increased to 89.29% for all samples, to 85.71% for unground and unroasted–unground, and 100% for roasted, unroasted–ground, and roasted–ground. Ultimately, RF successfully differentiated 100% of all grouped samples. In general, roasting and grinding the samples before analysis improved the accuracy of differentiating between wild and feeding civet coffee using NIR Spectroscopy.