Separate Predictive Models Research Articles

Machine learning-based variable selection for clustered credit risk modeling

AbstractSeveral studies have demonstrated the high prediction accuracy of clustered credit risk modeling. In clustered modeling, borrowers are segmented based on their similarities through cluster analysis, and a separate predictive model is developed for each cluster, resulting in increased predictive accuracy. Unambiguously, its effectiveness depends on the quality of the segmentation, which in turn depends primarily on the choice of variables used in the cluster analysis. However, appropriate variable selection for clustering is a major challenge, particularly for high-dimensional data. In the present study, we propose a machine learning-based variable selection method based on theoretical and regulatory considerations. Formally, the most influential risk drivers from a best-in-class machine learning model are identified using Shapley values and employed as clustering variables. Thus, the information of the explanatory variables crucial for the prediction of the dependent variable is already processed during data segmentation, making each individual predictive model more effective. Through a comparative analysis using two real-world credit default datasets, we show that our proposed approach to clustered modeling leads to the highest prediction accuracy among various clustering models.

Prediction of individual total amino acids and free amino acids in Chinese Holstein cows milk using mid-infrared spectroscopy and their phenotypic variability

Establishing a high-throughput detection technology for amino acid (AA) content in milk using mid-infrared (MIR) spectroscopy has profound implications for enhancing nutritional value of milk, identifying superior milk sources, producing specialty dairy products, and expanding Dairy Herd Improvement (DHI) metrics. The aim of this study was to evaluate the effectiveness of MIR spectroscopy in predicting the content of 15 individual total AA (TAAs) and 16 free AA (FAAs) in bovine milk as well as to investigate the major factors affecting the phenotypic variability of AA content. From March 2023 to March 2024, 513 milk samples were collected from 10 Holstein dairy farms in China and analyzed using Bentley spectrometers for MIR measurements. Their TAAs and FAAs concentrations were assessed through an AA autoanalyzer. Separate quantitative prediction models were developed for each AA using partial least squares regression; accuracy of prediction was assessed using Cow-independent external validation (CEV) and Farm-independent external validation (FEV) set. In CEV, the ratio of performance to deviation (RPD) of the TAAs models ranged from 1.45 (Ser) to 2.19 (Leu), while the FAA models ranged from 1.15 (Ser) to 2.44 (Met). In FEV, the RPD of the TAAs models ranged from 0.98 (Met) to 1.76 (Asp, Glu, and Ala), and the FAAs models ranged from 0.33 (Phe) to 1.23 (Asp and Tyr). For farms included in the calibration set, MIR spectroscopy provided a rough quantitative estimation for 4 individual TAAs (Ile, Leu, Glu, and Tyr) and 2 FAAs (Met and His), as well as a qualitative determination for high and low values in 9 individual TAAs (Phe, Met, Val, Lys, Thr, Asp, Ala, His, and Arg). For farms outside the calibration set, MIR spectroscopy could only distinguish between high and low contents for 5 individual TAAs (Glu, Asp, Ala, Leu, and Arg). Phenotypically, the variation pattern in TAAs contents mirrored that of protein, while FAAs did not show a clear trend, though mastitis led to a significant elevation of FAAs in milk (p < 0.05). Overall, the application of MIR spectroscopy can be considered very promising for a low-cost, rapid, large-scale assessment of individual TAAs and FAAs contents in milk. After refinement, some models could potentially be incorporated into DHI, which would greatly benefit the milk production and food industries.

A Model for Fat Content Detection in Walnuts Based on Near-Infrared Spectroscopy

Near-infrared spectroscopy (NIR) is an efficient and accurate method for fat content detection in walnuts. ‘Wen 185’ walnut is grown in large quantities in southern Xinjiang, and its fat content is an important indicator for evaluating the intrinsic quality. The excessive pursuit of yield efficiency, combined with the neglect of quality, agricultural product safety and other factors, has led to the production of poor-quality walnuts. Moreover, research on predicting walnut kernel fat content based on near-infrared spectroscopy technology is rarely reported. Therefore, a technical framework for walnut kernel detection based on near-infrared spectroscopy and the technical standards for ‘Wen 185’ are urgently needed. After first optimizing the initial spectrum data using five preprocessing methods, we established separate prediction models for walnut kernel fat content based on either a back propagation neural network or a support vector regression (SVR) algorithm. The results show that the correction set and validation set coefficients of determination of the walnut kernel fat content prediction model using the back propagation neural network algorithm were 0.86 and 0.89, respectively, with root mean square errors of 1.56 and 1.58, and an RPD value of 2.57; the coefficients of determination for the calibration and validation sets of SVR were 0.90 and 0.83, respectively, with root mean square errors of 1.76 and 1.70, respectively, and an RPD value of 1.70. Thus, near-infrared spectroscopy can be used to establish a foundation for realizing the rapid detection of walnut kernel fat content.

Predicting Survival Status in COVID-19 Patients: Machine Learning Models Development with Ventilator-Related and Biochemical Parameters from Early Stages: A Pilot Study.

Objective: Coronavirus disease 2019 (COVID-19) can cause intubation and ventilatory support due to respiratory failure, and extubation failure increases mortality risk. This study, therefore, aimed to explore the feasibility of using specific biochemical and ventilator parameters to predict survival status among COVID-19 patients by using machine learning. Methods: This study included COVID-19 patients from Taipei Medical University-affiliated hospitals from May 2021 to May 2022. Sequential data on specific biochemical and ventilator parameters from days 0-2, 3-5, and 6-7 were analyzed to explore differences between the surviving (successfully weaned off the ventilator) and non-surviving groups. These data were further used to establish separate survival prediction models using random forest (RF). Results: The surviving group exhibited significantly lower mean C-reactive protein (CRP) levels and mean potential of hydrogen ions levels (pH) levels on days 0-2 compared to the non-surviving group (CRP: non-surviving group: 13.16 ± 5.15 ng/mL, surviving group: 10.23 ± 5.15 ng/mL; pH: non-surviving group: 7.32 ± 0.07, survival group: 7.37 ± 0.07). Regarding the survival prediction performanace, the RF model trained solely with data from days 0-2 outperformed models trained with data from days 3-5 and 6-7. Subsequently, CRP, the partial pressure of carbon dioxide in arterial blood (PaCO2), pH, and the arterial oxygen partial pressure to fractional inspired oxygen (P/F) ratio served as primary indicators in survival prediction in the day 0-2 model. Conclusions: The present developed models confirmed that early biochemical and ventilatory parameters-specifically, CRP levels, pH, PaCO2, and P/F ratio-were key predictors of survival for COVID-19 patients. Assessed during the initial two days, these indicators effectively predicted the likelihood of successful weaning of from ventilators, emphasizing their importance in early management and improved outcomes in COVID-19-related respiratory failure.

Short-Term Wind Speed Prediction Study Based on Variational Mode Decompositions–Sparrow Search Algorithm–Gated Recurrent Units

Improving the accuracy of short-term wind speed predictions is crucial for mitigating the impact on power systems when integrating wind power into an electricity grid. This study developed a hybrid short-term wind speed prediction method, termed VMD–SSA–GRU, by combining variational mode decomposition (VMD) with gated recurrent units (GRUs) and optimizing it using a sparrow search algorithm (SSA). Initially, VMD was used to decompose the wind speed time series into subtime series. After reconstructing these subtime series, a GRU model was employed to establish separate prediction models for each series. Furthermore, an enhanced SSA was proposed to optimize the hyperparameters of the GRU model, which improved the prediction accuracy. Ultimately, the sub-series predictions were aggregated to produce the final wind speed prediction values. The predictive accuracy of this model was validated using the wind speed data measured at a meteorological station near a bridge site. The performance of the VMD–SSA–GRU model was compared with several other hybrid models, including those using wavelet transform, long short-term memory, and other neural networks. Comparably, the RMSE value of the VMD-SSA-GRU model was lower by 25.3%, 60.2%, and 61.7% in comparison to the VMD–SSA–LSTM, VMD–GRU, and VMD–LSTM models, respectively. The experimental results demonstrated that the proposed method achieved higher prediction accuracy than traditional methods.

Key influences on dysglycemia across Fujian's urban-rural divide.

Screening and treatment of dysglycemia (prediabetes and diabetes) represent significant challenges in advancing the Healthy China initiative. Identifying the crucial factors contributing to dysglycemia in urban-rural areas is essential for the implementation of targeted, precise interventions. Data for 26,157 adults in Fujian Province, China, were collected using the Social Factors Special Survey Form through a multi-stage random sampling method, wherein 18 variables contributing to dysglycemia were analyzed with logistic regression and the random forest model. Investigating urban-rural differences and critical factors in dysglycemia prevalence in Fujian, China, with the simultaneous development of separate predictive models for urban and rural areas. The detection rate of dysglycemia among adults was 35.26%, with rates of 34.1% in urban areas and 35.8% in rural areas. Common factors influencing dysglycemia included education, age, BMI, hypertension, and dyslipidemia. For rural residents, higher income (OR = 0.80, 95% CI [0.74, 0.87]), average sleep quality (OR = 0.89, 95% CI [0.80, 0.99]), good sleep quality (OR = 0.89, 95% CI [0.80, 1.00]), and high physical activity (PA) (OR = 0.87, 95% CI [0.79, 0.96]) emerged as protective factors. Conversely, a daily sleep duration over 8 hours (OR = 1.46, 95% CI [1.03, 1.28]) and middle income (OR = 1.12, 95% CI [1.03, 1.22]) were specific risk factors. In urban areas, being male (OR = 1.14, 95% CI [1.02, 1.26]), cohabitation (OR = 1.18, 95% CI [1.02, 1.37]), and central obesity (OR = 1.35, 95% CI [1.19, 1.53]) were identified as unique risk factors. Using logistic regression outcomes, a random forest model was developed to predict dysglycemia, achieving accuracies of 75.35% (rural) and 76.95% (urban) with ROC areas of 0.77 (rural) and 0.75 (urban). This study identifies key factors affecting dysglycemia in urban and rural Fujian residents, including common factors such as education, age, BMI, hypertension, and dyslipidemia. Notably, rural-specific protective factors are higher income and good sleep quality, while urban-specific risk factors include being male and central obesity. These findings support the development of targeted prevention and intervention strategies for dysglycemia, tailored to the unique characteristics of urban and rural populations.

Machine learning models to predict systemic inflammatory response syndrome after percutaneous nephrolithotomy

ObjectiveThe objective of this study was to develop and evaluate the performance of machine learning models for predicting the possibility of systemic inflammatory response syndrome (SIRS) following percutaneous nephrolithotomy (PCNL).MethodsWe retrospectively reviewed the clinical data of 337 patients who received PCNL between May 2020 and June 2022. In our study, 80% of the data were used as the training set, and the remaining data were used as the testing set. Separate prediction models based on the six machine learning algorithms were created using the training set. The predictive performance of each machine learning model was determined by the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity and specificity using the testing set. We used coefficients to interpret the contribution of each variable to the predictive performance.ResultsAmong the six machine learning algorithms, the support vector machine (SVM) delivered the best performance with accuracy of 0.868, AUC of 0.942 (95% CI 0.890–0.994) in the testing set. Further analysis using the SVM model showed that prealbumin contributed the most to the prediction of the outcome, followed by preoperative urine culture, systemic immune-inflammation (SII), neutrophil to lymphocyte ratio (NLR), staghorn stones, fibrinogen, operation time, preoperative urine white blood cell (WBC), preoperative urea nitrogen, hydronephrosis, stone burden, sex and preoperative lymphocyte count.ConclusionMachine learning-based prediction models can accurately predict the possibility of SIRS after PCNL in advance by learning patient clinical data, and should be used to guide surgeons in clinical decision-making.

MRI-based radiomics feature combined with tumor markers to predict TN staging of rectal cancer.

The aim of this study is to evaluate the predictive ability of MRI-based radiomics combined with tumor markers for TN staging in patients with rectal cancer and to develop a prediction model for TN staging. A total of 190 patients with rectal adenocarcinoma who underwent total mesorectal excision at the First Affiliated Hospital of the Air Force Medical University between January 2016 and December 2020 were included in the study. An additional 54 patients from a prospective validation cohort were included between August 2022 and August 2023. Preoperative tumor markers and MRI imaging data were collected from all enrolled patients. The 190 patients were divided into a training cohort (n = 133) and a validation cohort (n = 57). Radiomics features were extracted by outlining the region of interest (ROI) on T2WI sequence images. Feature selection and radiomics score (Rad-score) construction were performed using least absolute shrinkage and selection operator regression analysis (LASSO). The postoperative pathology TNM stage was used to differentiate locally advanced rectal cancer (T3/4 or N1/2) from locally early rectal cancer (T1/2, N0). Logistic regression was used to construct separate prediction models for T stage and N stage. The models' predictive performance was evaluated using DCA curves and calibration curves. The T staging model showed that Rad-score, based on 8 radiomics features, was an independent predictor of T staging. When combined with CEA, tumor diameter, mesoretal fascia (MRF), and extramural venous invasion (EMVI), it effectively differentiated between T1/2 and T3/4 stage rectal cancers in the training cohort (AUC 0.87 [95% CI: 0.81-0.93]). The N-staging model found that Rad-score, based on 10 radiomics features, was an independent predictor of N-staging. When combined with CA19.9, degree of differentiation, and EMVI, it effectively differentiated between N0 and N1/2 stage rectal cancers. The training cohort had an AUC of 0.84 (95% CI: 0.77-0.91). The calibration curves demonstrated good precision between the predicted and actual results. The DCA curves indicated that both sets of predictive models could provide net clinical benefits for diagnosis. MRI-based radiomics features are independent predictors of T staging and N staging. When combined with tumor markers, they have good predictive efficacy for TN staging of rectal cancer.

Prediction of the success of electrical cardioversion in patients with atrial fibrillation using parameters obtained by spectral and vectorcardiographic analysis of the ECG

Abstract Introduction Electrical cardioversion (ECV) still presents a treatment option for atrial fibrillation (AF). Despite the great success rate in the acute setting, its long-term success rate is low. The aim of the study was to find predictors of SR maintenance after ECV using spectral and vectorcardiographic (VCG) analysis of the ECG curve. Methods Consecutive patients with AF referred for elective ECV were prospectively enrolled. The digital ECG recording obtained before the ECV was analyzed using spectral and VCG analysis. All patients had Holter monitoring and clinical follow-up three months after cardioversion to assess SR maintenance. Atrial fibrillation activity was analyzed by spectral analysis, determining the dominant frequency (DF), RI (regularity index) and OI (organization index). QRS complexes were analyzed by vectorcardiography and dXmean, dYmean and dZmean (derivation of VCG signals) parameters were determined. Using logistic regression (LR) with backward parameter reduction, separate and combined predictive models of SR maintenance were calculated based on both the analysis from the VCG data and the spectral analysis data. Additional models with a forward selection of clinical variables were also calculated. The efficiency of the models was evaluated on a separate test subset of the data (N=32). Results A total of 80 patients were enrolled (age 70+10 years, 48 (64%) men, CHADSVASc 3.17+1.5, BMI 29.75kg/m2). At the 3-month follow-up, AF recurrence was present in 36 patients (45%). Using only vectorcardiographic parameters, dXmean (OR 3.07, 95 %CI 1.14- 8.63), p=0.001) and dYmean (OR 0.26, 95% CI 0.09-0.72), p &amp;lt;0.001) best predicted the SR maintenance (AUC of 0.72). Using only spectral analysis parameters, the best predictor was DF (OR 3.07, 95%CI 1.14-8.63), p=0.012) with AUC of 0.66. By combining the results of spectral analysis and vectorcardiography, an AUC of 0.78 was achieved. Among the clinical parameters, the most significant predictors were insufficiently controlled arterial hypertension and previous history of stroke. Conclusion Digital analysis of ECG curves using deeper mathematical models provides a better predictive model than classical clinical and echocardiographic models. The combination of spectral analysis of fibrillation atrial activity and vectorocardiographic analysis of QRS complexes revealed more accurate predictive information compared to both analyses alone.

Integrating IASLC grading and radiomics for predicting postoperative outcomes in stage IA invasive lung adenocarcinoma.

The International Association for the Study of Lung Cancer (IASLC) Pathology Committee introduced a histologic grading system for invasive lung adenocarcinoma (LUAD) in 2020. The IASLC grading system, hinging on the evaluation of predominant and high-grade histologic patterns, has proven to be practical and prognostic for invasive LUAD. However, there are still limitations in evaluating the prognosis of stage IA LUAD. Radiomics may serve as a valuable complement. To establish a model that integrates IASLC grading and radiomics, aimed at predicting the prognosis of stage IA LUAD. We conducted a retrospective analysis of 628 patients diagnosed with stage IA LUAD who underwent surgical resection between January 2015 and December 2018 at our institution. The patients were randomly divided into the training set (n=439) and testing set (n=189) at a ratio of 7:3. Overall survival (OS) and disease-free survival (DFS) were taken as the end points. Radiomics features were obtained by PyRadiomics. Feature selection was performed using the least absolute shrinkage and selection operator (LASSO). The prediction models for OS and DFS were developed using multivariate Cox regression analysis, and the models were visualized through nomogram plots. The model's performance was evaluated using area under the curves (AUC), concordance index (C-index), calibration curves, and survival decision curve analysis (DCA). In total, nine radiomics features were selected for the OS prediction model, and 15 radiomics features were selected for the DFS prediction model. Patients with high radiomics scores were associated with a worse prognosis (p<0.001). We built separate prediction models using radiomics or IASLC alone, as well as a combined prediction model. In the prediction of OS, we observed that the combined model (C-index: 0.812±0.024, 3 years AUC: 0.692, 5 years AUC: 0.792) achieved superior predictive performance than the radiomics (C-index: 0.743±0.038, 3 years AUC: 0.633, 5 years AUC: 0.768) and IASLC grading (C-index: 0.765±0.042, 3 years AUC: 0.658, 5 years AUC: 0.743) models alone. Similar results were obtained in the models for DFS. The combination of radiomics and IASLC pathological grading proves to be an effective approach for predicting the prognosis of stage IA LUAD. This has substantial clinical relevance in guiding treatment decisions for early-stage LUAD.

Predicting survival in patients with myelodysplastic/myeloproliferative neoplasms with SF3B1 mutation and thrombocytosis

We investigated data from 180 consecutive patients with myelodysplastic/myeloproliferative neoplasms with SF3B1 mutation and thrombocytosis (MDS/MPN-SF3B1-T) who were diagnosed according to the 2022 World Health Organization (WHO) classification of myeloid neoplasms to identify covariates associated with survival. At a median follow-up of 48 months (95% confidence interval [CI] 35–61 months), the median survival was 69 months (95% CI 59–79 months). Patients with bone marrow ring sideroblasts (RS) < 15% had shorter median overall survival (OS) than did those with bone marrow RS ≥ 15% (41 months [95% CI 32–50 months] versus 76 months [95% CI 59–93 months]; P < 0.001). According to the univariable analyses of OS, age ≥ 65 years (P < 0.001), hemoglobin concentration (Hb) < 80 g/L (P = 0.090), platelet count (PLT) ≥ 800 × 10E + 9/L (P = 0.087), bone marrow RS < 15% (P < 0.001), the Revised International Prognostic Scoring System (IPSS-R) cytogenetic category intermediate/poor/very poor (P = 0.005), SETBP1 mutation (P = 0.061) and SRSF2 mutation (P < 0.001) were associated with poor survival. Based on variables selected from univariable analyses, two separate survival prediction models, a clinical survival model, and a clinical-molecular survival model, were developed using multivariable analyses with the minimum value of the Akaike information criterion (AIC) to specifically predict outcomes in patients with MDS/MPN-SF3B1-T according to the 2022 WHO classification.

Multi-source domain generalization deep neural network model for predicting energy consumption in multiple office buildings

The effective prediction of building energy consumption can be used to optimize building operating modes and reduce the overall energy consumption and carbon emission of the building. To predict the energy consumption of the same type of building in the same area, it is often necessary to train separate prediction models for different buildings, which is usually costly computationally in order to get better prediction results. Therefore, this study will combine deep neural networks and multi-source domain generalization in an encoder-decoder architecture to train a model that can be directly applied to predict the energy consumption of multiple buildings in same type. In order to validate the predictive effectiveness of the model, it will be tested on real office building energy consumption dataset and different comparative experiments will be designed on the source and target domains. The results show that the constructed multi-source domain generalization model is able to accurately predict the energy consumption trend of different buildings 1 h in advance. It also has some energy consumption prediction ability for the unknow training set of buildings.

Proactive Identification of Patients with Diabetes at Risk of Uncontrolled Outcomes during a Diabetes Management Program: Conceptualization and Development Study Using Machine Learning.

The growth in the capabilities of telehealth have made it possible to identify individuals with a higher risk of uncontrolled diabetes and provide them with targeted support and resources to help them manage their condition. Thus, predictive modeling has emerged as a valuable tool for the advancement of diabetes management. This study aimed to conceptualize and develop a novel machine learning (ML) approach to proactively identify participants enrolled in a remote diabetes monitoring program (RDMP) who were at risk of uncontrolled diabetes at 12 months in the program. Registry data from the Livongo for Diabetes RDMP were used to design separate dynamic predictive ML models to predict participant outcomes at each monthly checkpoint of the participants' program journey (month-n models) from the first day of onboarding (month-0 model) up to the 11th month (month-11 model). A participant's program journey began upon onboarding into the RDMP and monitoring their own blood glucose (BG) levels through the RDMP-provided BG meter. Each participant passed through 12 predicative models through their first year enrolled in the RDMP. Four categories of participant attributes (ie, survey data, BG data, medication fills, and health signals) were used for feature construction. The models were trained using the light gradient boosting machine and underwent hyperparameter tuning. The performance of the models was evaluated using standard metrics, including precision, recall, specificity, the area under the curve, the F1-score, and accuracy. The ML models exhibited strong performance, accurately identifying observable at-risk participants, with recall ranging from 70% to 94% and precision from 40% to 88% across the 12-month program journey. Unobservable at-risk participants also showed promising performance, with recall ranging from 61% to 82% and precision from 42% to 61%. Overall, model performance improved as participants progressed through their program journey, demonstrating the importance of engagement data in predicting long-term clinical outcomes. This study explored the Livongo for Diabetes RDMP participants' temporal and static attributes, identification of diabetes management patterns and characteristics, and their relationship to predict diabetes management outcomes. Proactive targeting ML models accurately identified participants at risk of uncontrolled diabetes with a high level of precision that was generalizable through future years within the RDMP. The ability to identify participants who are at risk at various time points throughout the program journey allows for personalized interventions to improve outcomes. This approach offers significant advancements in the feasibility of large-scale implementation in remote monitoring programs and can help prevent uncontrolled glycemic levels and diabetes-related complications. Future research should include the impact of significant changes that can affect a participant's diabetes management.

Balancing fracture risk versus risk of mortality before fracture among women aged 80 years or older.

Most fractures occur in women aged ≥80 years but competing mortality unrelated to fracture may limit the benefit of osteoporosis drug therapy for some women in late life. Our primary aim was to develop separate prediction models for non-spine fracture (NSF) and mortality before fracture to identify subsets of women with varying fracture versus mortality risks. Separate prediction models were developed for NSF and mortality before NSF for 4895 women aged ≥80 years enrolled in the Study of Osteoporotic Fractures (SOF) or the Health Aging and Body Composition (HABC) study. Proportional hazards models modified to account for competing mortality were used to identify candidate risk factors for each outcome. Predictors associated with NSF or mortality (p < 0.2) were included in separate competing risk models to estimate the cumulative incidence of NSF and mortality before NSF during 5 years of follow-up. This process was repeated to develop separate prediction models for hip fracture and mortality before hip fracture. Significant predictors of NSF (race, total hip BMD, grip strength, prior fracture, falls, and use of selective serotonin reuptake inhibitors, benzodiazepines, or oral/transdermal estrogen) differed from predictors of mortality before NSF (age, walking speed, multimorbidity, weight change, shrinking, smoking, self-rated health, dementia, and use of warfarin). Within nine subsets of women defined by tertiles of risk, 5-year outcomes varied from 28% NSF and 8% mortality in the high-risk NSF/low-risk mortality subset, to 9% NSF and 22% mortality in the low-risk NSF/high-risk mortality subset. Similar results were seen for predictors of hip fracture and mortality before hip fracture. Considerable variation in 5-year competing mortality risk is present among women in late life with similar 5-year NSF risk. Both fracture risk and life expectancy should inform shared clinical decision-making regarding initiation or continuation of osteoporosis drug therapy for women aged ≥80 years.

Abstract B092: Development of predictive models for expression of a tumor specific biomarker and CD3 on H&amp;E digital slides

Abstract Characterization of a patient’s tumor microenvironment is fundamental to advancing translational strategies in immuno-oncology. Histopathological evaluation of tissue slides from patients can provide this invaluable information, informing disease heterogeneity, tumor contexture, and target expression - all important to identifying patients more likely to benefit from therapy. However, availability of sufficient tissue is often a challenge to produce such a comprehensive tumor characterization. Measurement of target expression using an H&amp;E slide could greatly reduce the usage of tissue for IHC making sections available for other investigative purposes. Here, an AI-driven predictive model was developed on H&amp;E digital slides to simulate the expression of a tumor specific biomarker and CD3 across 4 different tumor types. Four epithelial tumor types (TNBC, NSCLC, ovarian, and cervical cancer) consisted of 400 H&amp;E slides, 400 CD3 slides and 400 tumor specific biomarker IHC slides (100 cases per tumor) were used for this study. The Bio-AI Predict-X platform was utilized for the optimization of separate tumor models using pathology annotations as basis for development. The network output included tumor tiles as well as the tiles from adjacent stromal/normal tissue. Models were optimized until an AUC &amp;gt; 0.8 was achieved and prediction results were approved by the combined pathology teams. Predict-X was used for IHC marker quantification on a tile basis for both CD3 and the tumor specific biomarker slides with significant correlation to manual counts (p&amp;lt;0.001, r&amp;gt;0.8). Images from each case were co-registered so that ground truth for each tile with the IHC count was used in the predictive model. Tiles from tumor and corresponding adjacent stroma/normal tissue from all cases were subdivided into 3 groups for predictive model development. A training, validation, and test set were used, the latter of which was not used for development, only as a final assessment of model performance. A pan-tumor predictive model was developed for CD3 IHC on TNBC, cervical and ovarian cancers. The test set AUC values on cases from all 4 tumors were &amp;gt;0.7. Since the tumor specific biomarker stain was specific to tumor morphology, 4 separate predictive models were developed - one for each of the indications used with a test AUC &amp;gt;0.86. These findings suggest that deep learning can be used as a complementary method to prescreen H&amp;E-stained images to enhance the detection rate of the tumor specific biomarker and CD3 positivity in patient tumors. Additional work will involve the optimization of these models for implementation in a clinical setting. Citation Format: Alan Jerusalmi, Krishna Bairavi, Ayushi Shah, Tom Chittenden, Mike Bonham, Chung-Wein Lee, Brandon Higgs, Anantharaman Muthuswamy. Development of predictive models for expression of a tumor specific biomarker and CD3 on H&amp;E digital slides [abstract]. In: Proceedings of the AACR Special Conference on Ovarian Cancer; 2023 Oct 5-7; Boston, Massachusetts. Philadelphia (PA): AACR; Cancer Res 2024;84(5 Suppl_2):Abstract nr B092.

Throughput Prediction of 5G Network Based on Trace Similarity for Adaptive Video

Predicting throughput is essential to reduce latency in time-critical services like video streaming, which constitutes a significant portion of mobile network traffic. The video player continuously monitors network throughput during playback and adjusts the video quality according to the network conditions. This means that the quality of the video depends on the player’s ability to predict network throughput accurately, which can be challenging in the unpredictable environment of mobile networks. To improve the prediction accuracy, we grouped the throughput trace into clusters taking into account the similarity of their mean and variance. Once we distinguished the similar trace fragments, we built a separate LSTM predictive model for each cluster. For the experiment, we used traffic captured from 5G networks generated by individual user equipment (UE) in fixed and mobile scenarios. Our results show that the prior grouping of the network traces improved the prediction compared to the global model operating on the whole trace.

Multi-source AdaBoost with cross-weight method for virtual metrology in semiconductor manufacturing

In the context of Industry 4.0, many production scenarios utilise sensors to monitor manufacturing processes, resulting in massive data that can be leveraged to build machine-learning models to improve production efficiency and quality. In semiconductor manufacturing, where many sensors are employed to monitor the production process, the resulting multi-sensor data poses challenges for constructing a virtual metrology (VM) model to assess wafer quality. Furthermore, building separate prediction models for each sensor can result in a loss of redundancy present in the multi-sensor data. Therefore, this paper proposes a multi-source AdaBoost with a cross-weight sampling method for predicting the critical dimension required in VM. Compared with the existing AdaBoost algorithm in regression, the proposed method adapts well to multi-source data by incorporating a cross-weight sampling distribution. Thus, when updating the sampling weight distribution, the impact of both individual and multi-source data is considered to re-sample high-quality observations. Based on the numerical results from the synthetic datasets and case study in semiconductor manufacturing, the proposed method outperforms benchmark methods. Additionally, owing to the redundancy of multi-sensor data, the proposed method demonstrates robust performance regardless of the noise level in the semiconductor VM data.

Prediction models to evaluate baking quality instruments for commercial wheat flour

AbstractBackground and ObjectivesLoaf volume is the main indicator of wheat flour quality, but test baking has major limitations. Here, prediction models were used to evaluate which methodology best captured the baking quality in Swedish commercial wheat flour and if the chemical composition of flour increased prediction accuracy.FindingsFlour type (e.g., winter vs. spring wheat) affected prediction model results significantly. Thus, separate prediction models should be developed for each flour type. Combining data from alveograph, farinograph, and glutomatic tests with protein and damaged starch gave the best prediction results. The main loaf volume predictors were dough strength for winter wheat, stability for spring wheat, and extensibility for flour blends. The composition of protein and arabinoxylan influenced several quality parameters but did not improve loaf volume predictions.ConclusionsBest predictions were obtained for winter wheat. Spring wheat and flour blend models contained only one latent variable, indicating that protein content was the main determinant for loaf volume in these samples.Significance and NoveltyThis study is one of few using prediction models to evaluate instrument suitability to determine loaf volume. Instruments suitable for predicting quality were determined for commercial winter wheat flour, which is the main product of Swedish mills.

Baseline Toxicity Model to Identify the Specific and Nonspecific Effects of Per- and Polyfluoroalkyl Substances in Cell-Based Bioassays.

High-throughput screening is a strategy to identify potential adverse outcome pathways (AOP) for thousands of per- and polyfluoroalkyl substances (PFAS) if the specific effects can be distinguished from nonspecific effects. We hypothesize that baseline toxicity may serve as a reference to determine the specificity of the cell responses. Baseline toxicity is the minimum (cyto)toxicity caused by the accumulation of chemicals in cell membranes, which disturbs their structure and function. A mass balance model linking the critical membrane concentration for baseline toxicity to nominal (i.e., dosed) concentrations of PFAS in cell-based bioassays yielded separate baseline toxicity prediction models for anionic and neutral PFAS, which were based on liposome-water distribution ratios as the sole model descriptors. The specificity of cell responses to 30 PFAS on six target effects (activation of peroxisome proliferator-activated receptor (PPAR) gamma, aryl hydrocarbon receptor, oxidative stress response, and neurotoxicity in own experiments, and literature data for activation of several PPARs and the estrogen receptor) were assessed by comparing effective concentrations to predicted baseline toxic concentrations. HFPO-DA, HFPO-DA-AS, and PFMOAA showed high specificity on PPARs, which provides information on key events in AOPs relevant to PFAS. However, PFAS were of low specificity in the other experimentally evaluated assays and others from the literature. Even if PFAS are not highly specific for certain defined targets but disturb many toxicity pathways with low potency, such effects are toxicologically relevant, especially for hydrophobic PFAS and because PFAS are highly persistent and cause chronic effects. This implicates a heightened need for the risk assessment of PFAS mixtures because nonspecific effects behave concentration-additive in mixtures.

Prediction of Body Weight by Using PCA-Supported Gradient Boosting and Random Forest Algorithms in Water Buffaloes (Bubalus bubalis) Reared in South-Eastern Mexico.

This study aims to use advanced machine learning techniques supported by Principal Component Analysis (PCA) to estimate body weight (BW) in buffalos raised in southeastern Mexico and compare their performance. The first stage of the current study consists of body measurements and the process of determining the most informative variables using PCA, a dimension reduction method. This process reduces the data size by eliminating the complex structure of the model and provides a faster and more effective learning process. As a second stage, two separate prediction models were developed with Gradient Boosting and Random Forest algorithms, using the principal components obtained from the data set reduced by PCA. The performances of both models were compared using R2, RMSE and MAE metrics, and showed that the Gradient Boosting model achieved a better prediction performance with a higher R2 value and lower error rates than the Random Forest model. In conclusion, PCA-supported modeling applications can provide more reliable results, and the Gradient Boosting algorithm is superior to Random Forest in this context. The current study demonstrates the potential use of machine learning approaches in estimating body weight in water buffalos, and will support sustainable animal husbandry by contributing to decision making processes in the field of animal science.