Highly efficient stacking ensemble learning model for automated keratoconus screening

This paper mainly shares a successful Artificial Intelligence for IT Operations (AIOps) solution we have built for a cloud storage array to deal with unbalanced hard disk failure data and predict disk failure. Firstly, we preprocessed the unbalanced disk data to filter out irrelevant raw data. Based on SMART (Self-Monitoring, Analysis, and Reporting Technology) attributes, we extracted 14 preliminary attributes as training features for disk failure prediction. Secondly, we used a feature extraction library called Tsfresh and 16 extraction methods to regenerate more than 1500 features. To accelerate machine learning process, we used the Benjamini Yekutieli procedure with a significance test to select the most relevant features. Since a single predictive model no longer performs sufficiently well on the unbalanced dataset, we finally input the prediction results calculated by three algorithms(XGBoost classification, LSTM classification, and XGBoost regression) as new features input of a stacking ensemble learning model that can generate more stable and accurate prediction results. The experimental results showed that the proposed stacking ensemble learning model can accurately predict the disk failure and necessity of disk replacement 0 to 14 days, 14 to 42 days and more days in advance.

Weather forecasting is an important factor affecting production and life. With the development of technology, weather forecasting methods such as weather map forecasting, numerical weather forecasting, and quantitative forecasting methods have emerged. However, these traditional data analysis methods have shortcomings such as incomplete analysis, insufficient objectivity, inability to quantitatively predict the weather, and low prediction accuracy. The import of neural networks into the field of weather forecasting helps to alleviate the above shortcomings and improve the accuracy of weather forecasting. In this paper, the LSTM network and the CNN network are used to predict the sand-dust storm weather. In order to improve the prediction performance, we use the Stacking integration algorithm to fuse the LSTM and CNN models. To improve the experimental scientific and comprehensive, using fully connected network and support vector machine as meta-classifiers, two LSTM-CNN integrated sand-dust storm prediction models are established. At last, the above integrated model is used in the prediction of sand-dust storms in Inner Mongolia. The experimental results show that compared with a single LSTM or CNN model, the Stacking ensemble model has different degrees of improvement in model evaluation indicators such as accuracy, precision, recall, and f1-score. The Stacking ensemble model uses the fully connected network model as the meta-classifier is even better. These prove that the Stacking ensemble algorithm improves the sand-dust storm classification effect and generalization ability of a single neural network to a certain extent.

Photodetectors are essential components in modern optoelectronic technologies, yet experimental characterization of nanomaterial-based devices is often time-consuming and resource-intensive. To address this challenge, this study presents a stacking ensemble learning approach to predict the performance of bismuth-doped graphene quantum dots-based photodetectors under illumination levels ranging from 22 to 110 mW/mm². Four boosting algorithms—Adaptive Boosting, Gradient Boosting, Extreme Gradient Boosting, and Categorical Boosting—were trained on datasets obtained under dark, 22, 66, and 110 mW/mm², while 44 and 88 mW/mm² data were reserved for testing. A stacking ensemble learning model further enhanced prediction accuracy. The final model achieved a coefficient of determination of 0.9874 and a mean absolute error of 0.1840 at 88 mW/mm², effectively predicting the logarithmic current–voltage characteristics. The model also estimated key photodetector metrics, including sensitivity (1589.27), responsivity (2.389 mA/W), and specific detectivity (1.16 × 10¹⁰ Jones). This study explores the use of a stacking ensemble of four boosting algorithms to model the performance of Bi-GQD/p-Si photodetectors across different illumination levels, offering a data-driven alternative to traditional characterization.

To evaluate the diagnostic ability of topographic and tomographic indices with Pentacam and Sirius as well as biomechanical parameters with Corvis ST for the detection of clinical and subclinical forms of keratoconus (KCN). In this prospective diagnostic test study, 70 patients with clinical KCN, 79 patients with abnormal findings in topography and tomography maps with no evidence on clinical examination (subclinical KCN), and 68 normal control subjects were enrolled. The accuracy of topographic, tomographic, and biomechanical parameters was evaluated using the area under the receiver operating characteristic curve (AUC) and cross-validation analysis. The Delong method was used for comparing AUCs. In distinguishing KCN from normal, all parameters showed statistically significant differences between the two groups (P<0.001). Indices with the perfect diagnostic ability (AUC≥0.999) were Sirius KCN vertex of back (KVb), Pentacam random forest index (PRFI), Pentacam index of height decentration (IHD), and Corvis integrated tomographic/biomechanical index (TBI). In distinguishing subclinical KCN from normal, Sirius symmetry index of back (SIb; AUC=0.908), Pentacam inferior-superior difference (IS) value (AUC=0.862), PRFI (AUC=0.847), and Corvis TBI (AUC=0.820) performed best. There were no significant differences between the highest AUCs within keratoconic groups (DeLong, P>0.05). In clinical KCN, all topographic, tomographic, and biomechanical indices have acceptable outcomes in terms of sensitivity and specificity. However, in differentiating subclinical forms of KCN from normal corneas, curvature-based parameters (SIb and IS value) followed by integrated indices (PRFI and TBI) are the most powerful tools for early detection of KCN.

Cheese-making traits in dairy cattle are important to the dairy industry but are difficult to measure at the individual level because there are limitations on collecting phenotypic information. Mid-infrared spectroscopy has its advantages, but it can only be used during monthly milk recordings. Recently, in-line devices for real-time analysis of milk quality have been developed. The AfiLab recording system (Afimilk) offers significant benefits as phenotypes can be collected from each cow at each milking session. The objective of this study was to assess the potential of integrating AfiLab real-time milk analyzer measures with the stacking ensemble learning technique using heterogeneous base learners for the in-line daily monitoring of cheese-making traits in Holstein cattle with a view to developing a precision livestock farming system for monitoring the technological quality of milk. Data and samples for wet-laboratory analyses were collected from 499 Holstein cows belonging to 2 farms where the AfiLab system was installed. The traits of concern were 9 milk coagulation traits [3 milk coagulation properties (MCP), and 6 curd firming traits (CFt)], and 7 cheese-making traits [3 cheese yield (CY) traits, and 4 milk nutrient recovery in the curd (REC) traits]. The near-infrared AfiLab spectral data and on-farm information (days in milk and parity) were used to assess the predictive ability of different statistical methods [elastic net (EN), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), and artificial neural network (ANN)] across different cross-validation scenarios. These statistical methods were considered the base learners, which were then combined in a stacking ensemble learning. Results indicate that including information on the cows (days in milk and parity) in the AfiLab infrared prediction increased its accuracy by 10.3% for traditional MCP, 13.8% for curd firming, 9.8% for CY, and 11.2% for REC traits compared with those obtained from near-infrared AfiLab alone. The statistical approaches exhibited high prediction accuracies (R2) averaged across the cross-validation scenarios for traditional MCP (0.58 for ANN, 0.55 for EN and GBM, 0.52 for XGBoost, and 0.62 for stacking ensemble), CFt (0.55 for ANN, 0.54 for EN and GBM, 0.53 for XGBoost, and 0.61 for stacking ensemble), and similar R2 averages for CY and REC (0.55 for ANN, 0.54 for EN and GBM, 0.53 for XGBoost, and 0.61 for stacking ensemble). The ANN approach was more accurate than the other base learners (EN, GBM, and XGBoost) and improved accuracy across cross-validation scenarios on average by 7% for traditional MCP, 5% for CFt, 8% for CY, and 7% for REC. The stacking ensemble method improved prediction accuracy by 3% to 31% for traditional MCP, 2% to 26% for CFt, 1% to 38% for CY traits, and 2% to 27% for REC traits compared with the base learners. The prediction accuracies of the different approaches evaluated tended to decrease from the 10-fold cross-validation to the independent validation scenario, although there was a smaller reduction in prediction accuracy with the stacking ensemble learning technique across all the cross-validation scenarios. Our results show that combining in-line on-farm information with stacking ensemble machine learning represents an effective alternative for obtaining robust daily predictions of milk cheese-making traits.

Objectives:To evaluate tomographic and topographic parameters in subclinical and clinical keratoconus eyes by comparing them with normal eyes in a young Caucasian population.Materials and Methods:This cross-sectional study included 88 normal eyes (control group), bilateral data from the preclinical stage of 24 progressive keratoconus eyes (bilateral subclinical keratoconus group), 40 fellow eyes of patients with unilateral keratoconus (fellow eyes group) and 97 eyes with mild keratoconus (clinical keratoconus group). Topographic and tomographic data, data from enhanced elevation maps and keratoconus indices were measured in all study eyes using Scheimpflug tomography. Receiver operating characteristic (ROC) curve analysis was used to assess individual parameters to discriminate eyes of patients with subclinical and clinical keratoconus from control eyes. The sensitivity and specificity of the main effective parameters were evaluated and optimal cut-off points were identified to differentiate subclinical keratoconus and keratoconus from normal corneas.Results:Comparison of all subclinical and clinical keratoconus eyes from the normal group revealed significant differences in most diagnostic parameters. The ROC curve analysis showed high overall predictive accuracy of several Pentacam parameters (overall D value, anterior and posterior elevations and difference elevations, pachymetry progression index, index of surface variance, index of height decentration and keratoconus index) in discriminating ectatic corneas from normal ones. These outcomes were proportionally less pronounced in all subclinical keratoconus eyes than in the clinical keratoconus eyes. Pachymetric readings were progressively lower in the bilateral subclinical keratoconus eyes and sensitivity and specificity of the analyzed tomographic and topographic parameters were higher than the fellow eyes group when differentiating subclinical keratoconus from healthy corneas.Conclusion:Scheimpflug tomography parameters such as D value, elevation parameters, progression index and several surface indices can effectively differentiate keratoconus from normal corneas in a Caucasian population. Nevertheless, a combination of different data is required to distinguish subclinical keratoconus.

Polycystic ovary syndrome (PCOS) is the most frequent endocrinological anomaly in reproductive women that causes persistent hormonal secretion disruption, leading to the formation of numerous cysts within the ovaries and serious health complications. But the real-world clinical detection technique for PCOS is very critical since the accuracy of interpretations being substantially dependent on the physician's expertise. Thus, an artificially intelligent PCOS prediction model might be a feasible additional technique to the error prone and time-consuming diagnostic technique. In this study, a modified ensemble machine learning (ML) classification approach is proposed utilizing state-of-the-art stacking technique for PCOS identification with patients' symptom data; employing five traditional ML models as base learners and then one bagging or boosting ensemble ML model as the meta-learner of the stacked model. Furthermore, three distinct types of feature selection strategies are applied to pick different sets of features with varied numbers and combinations of attributes. To evaluate and explore the dominant features necessary for predicting PCOS, the proposed technique with five variety of models and other ten types of classifiers is trained, tested and assessed utilizing different feature sets. As outcomes, the proposed stacking ensemble technique significantly enhances the accuracy in comparison to the other existing ML based techniques in case of all varieties of feature sets. However, among various models investigated to categorize PCOS and non-PCOS patients, the stacking ensemble model with ‘Gradient Boosting’ classifier as meta learner outperforms others with 95.7% accuracy while utilizing the top 25 features selected using Principal Component Analysis (PCA) feature selection technique.

IDH wild-type glioblastoma (GBM) intrinsic subtypes have been linked to different molecular landscapes and outcomes. Accurate prediction of molecular subtypes of GBM is very important to guide clinical diagnosis and treatment. Leveraging machine learning technology to improve the subtype classification was considered a robust strategy. Several single machine learning models have been developed to predict survival or stratify patients. An ensemble learning strategy combines several basic learners to boost model performance. However, it still lacked a robust stacking ensemble learning model with high accuracy in clinical practice. Here, we developed a novel integrative stacking ensemble model framework (ecGBMsub) for improving IDH wild-type GBM molecular subtype classification. In the framework, nine single models with the best hyperparameters were fitted based on extrachromosomal circular DNA (eccDNA) molecular profiling. Then, the top five optimal single models were selected as base models. By randomly combining the five optimal base models, 26 different combinations were finally generated. Nine different meta-models with the best hyperparameters were fitted based on the prediction results of 26 different combinations, resulting in 234 different stacked ensemble models. All models in ecGBMsub were comprehensively evaluated and compared. Finally, the stacking ensemble model named "XGBoost.Enet-stacking-Enet" was chosen as the optimal model in the ecGBMsub framework. A user-friendly web tool was developed to facilitate accessibility to the XGBoost.Enet-stacking-Enet models (https://lizesheng20190820.shinyapps.io/ecGBMsub/).

Outage duration prediction under typhoon disaster with stacking ensemble learning

Objective: To evaluate the diagnostic efficacy of stress-strain index (SSI) for different stages or degrees of keratoconus and changes of SSI and stiffness parameter A1 (SPA1) after corneal collagen cross-linking (CXL) surgery. Methods: Cross-sectional study and retrospective case series study. Ninety-four patients (113 eyes) diagnosed as clinical keratoconus (CKC) in Qingdao Eye Hospital from July 2019 to August 2021 were enrolled in the CKC group, including 69 males and 25 females, aged (20.82±4.53) years, and further divided into subgroups of mild (35 patients, 36 eyes), moderate (36 patients, 40 eyes) and severe (33 patients, 37 eyes) CKC. Fifty-six unaffected eyes of monocular keratoconus patients were enrolled in the subclinical keratoconus (SKC) group. Ninety-one healthy subjects (91 eyes) were recruited as the control group. All subjects were examined by Pentacam topography and Corvis ST measurements to obtain mean keratometry, maximal keratometry, deformation amplitude (DA) ratio at 2 mm, integrated radius (IR), Ambrósio's relational thickness to the horizontal profile, corneal central thickness, SPA1 and SSI for comparison. Forty-eight CKC patients (65 eyes) underwent CXL surgery, and the above parameters were recorded before and 3, 6 and 12 months after operation. Data were analyzed by the ANOVA test, Kruskal-Wallis H test, paired sample test, receiver operating characteristic curves and Pearson correlation. Results: The value of SPA1 in the SKC group accounted for 85.53% (87.92±12.38 vs. 102.79±11.74; t=-6.614, P<0.001) compared with the control group, but the value of SSI had no difference in the two groups (t=0.105, P=0.916). The value of SPA1 in the CKC group accounted for 52.87% (54.35±14.70 vs. 102.79±11.74; t=25.985, P<0.001) compared with the control group. The value of SSI in the CKC group accounted for 67.96% (0.70±0.14 vs. 1.03±0.14; t=-15.305, P<0.001) compared with the control group. The more severe the disease was, the smaller the SPA1 and SSI values were 64.27±12.12, 55.22±12.23, 43.75±12.33; 0.78±0.14, 0.71±0.11, 0.61±0.09, and there were significant statistical differences among groups (mild vs. moderate, mild vs. severe, moderate vs. severe; SPA1: t=3.257, -7.249, -4.159; all P<0.001. SSI: t=2.383, 5.065, 2.798; P=0.018,<0.001,=0.006). Receiver operating characteristic analysis showed that SPA1 had good diagnostic efficiency for subclinical patients [area under curve (AUC)=0.802], while the SSI had no diagnostic value (P=0.802). SPA1 had better diagnostic efficiency than the SSI for keratoconus in different stages, especially in the mild CKC and SKC groups (AUC: 0.914 vs. 0.847). The SSI had a significant positive correlation with SPA1 and a significant negative correlation with DA ratio and IR in the control, SKC and CKC groups (r=0.278, 0.368, 0.550; r=-0.346, -0.462, -0.547; r=-0.612, -0.591, -0.718; P<0.01). For patients who received CXL, maximal keratometry decreased significantly at 6 and 12 months postoperatively (t=4.029, 3.633; all P<0.001), whereas SPA1 increased significantly (t=-3.960, -4.500; all P<0.001). However, the SSI only increased significantly at 3 months (t=-2.577, P=0.012) and returned to the preoperative level at 6 and 12 months postoperatively, with no statistical difference compared with the preoperative level (t=-0.544, -0.257; P=0.589, 0.798). Conclusions: While there was no significant change in the SSI of SKC, the SSI of CKC decreased, and the more severe the disease was, the smaller the value was. The SSI was significantly and consistently correlated with DA ratio, IR and SPA1. The SSI compared with SPA1 had a lower degree of identification in different stages and degrees of keratoconus. The consistency of SPA1 with clinical effects after CXL surgery was higher than that of the SSI parameter.

Fourier Analysis on Irregular Corneal Astigmatism Using Optical Coherence Tomography in Various Severity Stages of Keratoconus

In the steam turbine reheating system, maintaining the high-pressure heater liquid level is crucial to prevent equipment damage and safety accidents. High liquid levels may cause steam to enter the feed system, while low liquid levels may lead to overheating and damage to the heater tube bundle. In this study, the stacking ensemble learning model for combined data decomposition is proposed for high-precision liquid level prediction and early warning. Initially, outliers in the high-pressure heater liquid level data from the DSC system are processed. The Adaptive Local Iterative Filtering (ALIF) algorithm is used for the first data decomposition, followed by Empirical Mode Decomposition (EMD) for a second decomposition. Time-domain features such as summation, maximum value, skewness, and peak value are extracted to construct reconstructed data. Random Forest (RF) reduces the dimensionality of the reconstructed data. The prediction model combines Deep Belief Network (DBN), Deep Neural Networks (DNN), Light Gradient Boosting Machine (LGBM), and eXtreme Gradient Boosting (xGBoost) as base models, with Echo State Network (ESN) as the meta-learner. Finally, based on historical data, warning level values are set, and the system outputs warnings when the actual liquid level significantly deviates from the predicted value.

Soil heavy metal pollution threatens agricultural safety and human health, with Cd exceeding standards being the most common problem in contaminated farmland. The development of hyperspectral remote sensing technology has provided a novel methodology of quickly and non-destructively monitoring heavy metal contamination in soil. This study aims to explore the potential of an interpretable Stacking ensemble learning model for the estimation of soil Cd contamination in farmland hyperspectral data. We assume that this method can improve the modeling accuracy. We chose Zhangjiagang City, Jiangsu Province, China, as the study area. We gathered soil samples from wheat fields and analyzed soil spectral data and Cd level in the lab. First, we pre-processed the spectra utilizing fractional-order derivative (FOD) and standard normal variate (SNV) transforms to highlight the spectral features. Second, we applied the competitive adaptive reweighted sampling (CARS) feature selection algorithm to identify the significant wavelengths correlated with soil Cd content. Then, we constructed and compared the estimation accuracy of multiple machine learning models and a Stacking ensemble learning method and utilized the Optuna method for hyperparameter optimization. Ultimately, the SHAP method was used to shed light on the model’s decision-making process. The results show that (1) FOD can further highlight the spectral features, thereby strengthening the correlation between soil Cd content and wavelength; (2) the CARS algorithm extracted 3.4–6.8% of the feature wavelengths from the full spectrum, and most of them were the wavelengths with high correlation with soil Cd; (3) the optimal estimation precision was achieved using the FOD1.5-SNV spectral pre-processing combined with the Stacking model (R2 = 0.77, RMSE = 0.05 mg/kg, RPD = 2.07), and the model effectively quantitatively estimated soil Cd contamination; and (4) SHAP further revealed the contribution of each base model and characteristic wavelengths in the Stacking modeling process. This research confirms the advantages of the interpretable Stacking model in hyperspectral estimation of Cd contamination in farmland wheat soil. Furthermore, it offers a foundational reference for the future implementation of quantitative and non-destructive regional monitoring of heavy metal contamination in farmland soil.

To evaluate corneal thickness and volume in subclinical and clinical keratoconus in Asian population with the aim of discriminating between normal and ectatic cornea. Eyes were placed into one of the following three groups: normal, subclinical, and mild-moderate keratoconus. Pentacam Scheimpflug imaging (Oculus Inc., Wetzlar, Germany) was performed for each participant to record thinnest corneal thickness, central corneal thickness, corneal volume (CV), peripheral corneal thickness (PCT) and percentage thickness increase (PTI) at 2, 4, 6, and 8 mm. The data were exported to SPSS for statistical analysis. Subjects comprised 52 normal, 15 subclinical keratoconus, and 32 mild-moderate clinical keratoconus eyes. Our results indicated that corneal thickness (CT) distribution, PTI, and CV in normal eyes were significantly different compared with subclinical and clinical keratoconus (P < .05). Overall, subclinical group exhibited lower CT distribution and volume, and higher PTI in comparison with normal eyes. However, they showed higher CT distribution and volume, and lower PTI compared with keratoconus group. In addition, there was a smaller change in PCT and PTI from the thinnest point of the cornea to the periphery. The results of the present study indicate that CT parameters and CV were significantly different in normal versus subclinical group and in normal versus keratoconus group. These findings could help clinicians to better discriminate between normal and ectatic cornea.

To evaluate and compare the topographic and topometric parameters, thickness profile data, and data from enhanced elevation maps of thin non-keratoconic, subclinical keratoconic, and mild keratoconic corneas with the Pentacam Scheimpflug corneal tomography and to study the usefulness of different parameters to differentiate keratoconus from topographically normal thin corneas. The study included 30 eyes with subclinical keratoconus, 30 eyes with mild-stage keratoconus, and 54 healthy eyes with minimal pachymetry ≤500 µm, with a mean age of 21.19 ± 2.97, 21.75 ± 1.93, and 21.5 ± 2.95 years, respectively. The area under the receiver operating characteristic curves was used to analyze the diagnostic significance of the Pentacam parameters. The anterior and posterior corneal elevations, pachymetric progression, the percentage of thickness increase measurements, overall D value, and topometric indices were statistically significantly higher in subclinical and mild keratoconic corneas than in normal eyes with thin cornea (p < 0.05). All these parameters had sufficient strength (area under the receiver operating characteristic curves >0.90) to differentiate clinical keratoconus. Posterior elevation showed the excellent area under the receiver operating characteristic curves with 100% sensitivity and 100% specificity for this purpose. However, among all parameters studied, the anterior elevation (0.935) showed the excellent area under the receiver operating characteristic curves to differentiate subclinical keratoconus, followed by posterior elevation (0.897), index of height decentration (0.887), and D value (0.882). The parameters derived from the Scheimpflug device, such as corneal elevations and overall D value, can effectively differentiate subclinical and clinical keratoconus from non-keratoconic thin cornea eyes. However, the specificity levels of these parameters were relatively limited in the diagnosis of subclinical keratoconus.