Multiple Cross-validation Research Articles

AIGD-PFT: the first AI-driven global daily gap-free 4 km phytoplankton functional type data product from 1998 to 2023

Abstract. Long time series of spatiotemporally continuous phytoplankton functional type (PFT) data are essential for understanding marine ecosystems and global biogeochemical cycles as well as for effective marine management. In this study, we integrated artificial intelligence (AI) technology with multisource marine big data to develop a spatial–temporal–ecological ensemble model based on deep learning (STEE-DL). This model generated the first AI-driven global daily gap-free 4 km PFT chlorophyll a concentration product from 1998 to 2023 (AIGD-PFT). The AIGD-PFT significantly enhances the accuracy and spatiotemporal coverage of quantifying eight major PFTs: diatoms, dinoflagellates, haptophytes, pelagophytes, cryptophytes, green algae, prokaryotes, and Prochlorococcus. The model input encompasses (1) physical oceanographic, biogeochemical, and spatiotemporal information and (2) ocean colour data (OC-CCI v6.0) that have been gap-filled using a discrete cosine transform–penalized least squares (DCT-PLS) approach. The STEE-DL model utilizes an ensemble strategy with 100 residual neural network (ResNet) models, applying Monte Carlo and bootstrapping methods to estimate the optimal PFT chlorophyll a concentration and assess the model uncertainty through ensemble means and standard deviations. The model's performance was validated using multiple cross-validation strategies – random, spatial-block, and temporal-block methods – combined with in situ data, demonstrating STEE-DL's robustness and generalization capability. The daily updates and seamless nature of the AIGD-PFT data product capture the complex dynamics of coastal regions effectively. Finally, through a comparative analysis using a triple-collocation analysis (TCA) approach, the competitive advantages of the AIGD-PFT data product over existing products were validated. The complete product dataset (1998–2023) can be freely downloaded from https://doi.org/10.11888/RemoteSen.tpdc.301164 (Zhang and Shen, 2024a).

A 10 km daily-level ultraviolet-radiation-predicting dataset based on machine learning models in China from 2005 to 2020

Abstract. Ultraviolet (UV) radiation is closely related to health; however, limited measurements have hindered further investigation of its health effects in China. Machine learning algorithms have been widely used to predict environmental factors with high accuracy, but a limited number of studies have implemented it for UV radiation. The main aim of this study is to develop a UV radiation prediction model using the random forest approach and predict the UV radiation with a daily and 10 km resolution in mainland China from 2005 to 2020. The model was developed with multiple predictors, such as UV radiation data from satellites as independent variables and ground UV radiation measurements from monitoring stations as the dependent variable. Missing satellite-based UV radiation data were obtained using the 3 d moving average method. The model performance was evaluated using multiple cross-validation (CV) methods. The overall R2 and root mean square error between measured and predicted UV radiation from model development and model 10-fold CV were 0.97 and 15.64 W m−2 and 0.83 and 37.44 W m−2 at the daily level, respectively. The model that incorporated erythemal daily dose (EDD) retrieved from the Ozone Monitoring Instrument (OMI) had a higher prediction accuracy than that without it. Based on predictions of UV radiation at the daily level, 10 km spatial resolution, and nearly 100 % spatiotemporal coverage, we found that UV radiation increased by 4.20 %, PM2.5 levels decreased by 48.51 %, and O3 levels increased by 22.70 % from 2013–2020, suggesting a potential correlation among these environmental factors. The uneven spatial distribution of UV radiation was associated with factors such as latitude, elevation, meteorological factors, and season. The eastern areas of China pose a higher risk due to both high population density and high UV radiation intensity. Using a machine learning algorithm, this study generated a gridded UV radiation dataset with extensive spatiotemporal coverage, which can be utilized for future health-related research. This dataset is freely available at https://doi.org/10.5281/zenodo.10884591 (Jiang et al., 2024).

Machine Learning-Based Algorithm for the Early Prediction of Postoperative Hypocalcemia Risk After Thyroidectomy.

We used machine learning to develop and validate a multivariable algorithm allowing the accurate and early prediction of postoperative hypocalcemia risk. Postoperative hypocalcemia is frequent after total thyroidectomy. An early and accurate individualized prediction of the risk of hypocalcemia could guide the selective prescription of calcium supplementation only to patients most likely to present with hypocalcemia after total thyroidectomy. This retrospective study enrolled all patients undergoing total thyroidectomy in a single referral center between November 2019 and March 2022 (derivation cohort) and April 2022 and September 2022 (validation cohort). The primary study outcome was postoperative hypocalcemia (serum calcium under 80mg/L). Exposures were multiple clinical and biological variables prospectively collected and analyzed with various machine learning methods to develop and validate a multivariable prediction algorithm. Among 610/118 participants in the derivation/validation cohorts, 100 (16.4%)/26 (22%) presented postoperative hypocalcemia. The most accurate prediction algorithm was obtained with random forest and combined intraoperative parathyroid hormone measurements with 3 clinical variables (age, sex, and body mass index) to calculate a postoperative hypocalcemia risk for each patient. After multiple cross-validation, the area under the receiver operative characteristic curve was 0.902 (0.829-0.970) in the derivation cohort, and 0.928 (95% CI: 0.86; 0.97) in the validation cohort. Postoperative hypocalcemia risk values of 7% (low threshold) and 20% (high threshold) had, respectively, a sensitivity of 92%, a negative likelihood ratio of 0.11, a specificity of 90%, and a positive of 7.6 for the prediction of postoperative hypocalcemia. Using machine learning, we developed and validated a simple multivariable model that allowed the accurate prediction of postoperative hypocalcemia. The resulting algorithm could be used at the point of care to guide clinical management after total thyroidectomy.

Developing a Machine Learning Algorithm for Improved Management of Congestive Heart Failure Patients in the Emergency Department

Background and aim: Congestive heart failure is a prevalent and serious condition that poses significant challenges in the emergency department setting. Prompt and accurate management of congestive heart failure patients is crucial for improving outcomes and optimizing resource utilization. This study aims to address these challenges by developing a machine learning algorithm and comparing it to a traditional logistic regression model that can assist in the triage, resource allocation, and long-term prognostication of congestive heart failure patients. Methods: In this investigation, we used the MIMIC-III database, a publicly accessible resource containing patient data from ICU settings. Traditional logistic regression, along with the robust XGBoost and random forest algorithms, was harnessed to construct predictive models. These models were built using a range of pretreatment clinical variables. To pinpoint the most pertinent features, we carried out a univariate analysis. Ensuring robust performance and broad applicability, we adopted a nested cross-validation approach. This method enhances the precision and validation of our models by implementing multiple cross-validation iterations. Results: The performance of machine learning algorithms was assessed using the area under the receiver operating characteristic curve (AUC). Notably, the random forest algorithm, despite having lower performance among the machine learning models still demonstrated significantly higher AUC than traditional logistic regression. The AUC for the XGBoost was 0.99, random forest 0.98, while traditional logistic regression was 0.57. The most important pretreatment variables associated with congestive heart failure include total bilirubin, creatine kinase, international normalized ratio (INR), sodium, age, creatinine, potassium, gender, alkaline phosphatase, and platelets. Conclusion: Machine learning techniques utilizing multiple pretreatment clinical variables outperform traditional logistic regression in aiding the triage, resource allocation, and long-term prognostication of congestive heart failure patients in the intensive care unit setting using MIMIC III data.

Determination of organic contaminants in L'Albufera Natural Park using microporous polyethylene tube passive samplers: An environmental risk assessment

L'Albufera Natural Park (Valencia, Spain) is a protected wetland of international significance that provides critical habitats to endemic and threatened bird and plant species. This study aims to use multiple cross-validation techniques to generate an accurate estimation of the environmental risk of organic contaminants (OCs) in an internationally important coastal wetland, to identify compounds of concern and their potential sources and risk factors. Microporous polyethylene tube (MPT) passive samplers were deployed at 12 locations across L'Albufera Natural Park with concurrent grab samples collected. A subset of MPT samplers were also analysed by an additional laboratory in Australia to widen the range of contaminants and assess interlaboratory reproducibility of results. Forty-three pesticides, 20 pharmaceuticals and personal care products (PPCPs), 20 per-and polyfluoroalkyl substances (PFAS) and 4 organophosphorus flame retardants (OPFRs) were detected in the MPT samplers. The fungicides tebuconazole and difenoconazole were detected at the highest concentrations in passive samplers (maximum concentrations, 153 ng sampler−1 and 106 ng sampler−1, respectively). Several other pesticides were detected in all locations (mean concentrations >1 ng sampler−1). The compounds fenamiphos, propyzamide, difenoconazole, propiconazole, metsulfuron methyl, sodium bis (perfluorohexyl) phosphinate (6:6 PFPiA), 6:2 fluorotelomer sulfonamide alkylbetaine (6:2 FTAB), 6:2 fluorotelomersulfonate (6:2 FTS), citalopram desmethyl and citalopram were reported in the wetland for the first time. Spatial distribution analysis revealed higher pesticide concentrations in the North of L'Albufera. A risk quotient (RQ) analysis showed that ibuprofen is of concern in the area. Overall, the MPT sampling approach is promising as a risk assessment tool for better understanding the transport and fate of OCs in protected areas.

Feasibility study on using house-tree-person drawings for automatic analysis of depression

Major depression is a severe psychological disorder typically diagnosed using scale tests and through the subjective assessment of medical professionals. Along with the continuous development of machine learning techniques, computer technology has been increasingly employed to identify depression in recent years. Traditional methods of automatic depression recognition rely on using the patient’s physiological data, such as facial expressions, voice, electroencephalography (EEG), and magnetic resonance imaging (MRI) as input. However, the acquisition cost of these data is relatively high, making it unsuitable for large-scale depression screening. Thus, we explore the possibility of utilizing a house-tree-person (HTP) drawing to automatically detect major depression without requiring the patient’s physiological data. The dataset we used for this study consisted of 309 drawings depicting individuals at risk of major depression and 290 drawings depicting individuals without depression risk. We classified the eight features extracted from HTP sketches using four machine-learning models and used multiple cross-validations to calculate recognition rates. The best classification accuracy rate among these models reached 97.2%. Additionally, we conducted ablation experiments to analyze the association between features and information on depression pathology. The results of Wilcoxon rank-sum tests showed that seven of the eight features significantly differed between the major depression group and the regular group. We demonstrated significant differences in HTP drawings between patients with severe depression and everyday individuals, and using HTP sketches to identify depression automatically is feasible, providing a new approach for automatic identification and large-scale screening of depression.

Assessing facial weakness in myasthenia gravis with facial recognition software and deep learning.

Myasthenia gravis (MG) is an autoimmune disease leading to fatigable muscle weakness. Extra-ocular and bulbar muscles are most commonly affected. We aimed to investigate whether facial weakness can be quantified automatically and used for diagnosis and disease monitoring. In this cross-sectional study, we analyzed video recordings of 70 MG patients and 69 healthy controls (HC) with two different methods. Facial weakness was first quantified with facial expression recognition software. Subsequently, a deep learning (DL) computer model was trained for the classification of diagnosis and disease severity using multiple cross-validations on videos of 50 patients and 50 controls. Results were validated using unseen videos of 20 MG patients and 19 HC. Expression of anger (p = 0.026), fear (p = 0.003), and happiness (p < 0.001) was significantly decreased in MG compared to HC. Specific patterns of decreased facial movement were detectable in each emotion. Results of the DL model for diagnosis were as follows: area under the curve (AUC) of the receiver operator curve 0.75 (95% CI 0.65-0.85), sensitivity 0.76, specificity 0.76, and accuracy 76%. For disease severity: AUC 0.75 (95% CI 0.60-0.90), sensitivity 0.93, specificity 0.63, and accuracy 80%. Results of validation, diagnosis: AUC 0.82 (95% CI: 0.67-0.97), sensitivity 1.0, specificity 0.74, and accuracy 87%. For disease severity: AUC 0.88 (95% CI: 0.67-1.0), sensitivity 1.0, specificity 0.86, and accuracy 94%. Patterns of facial weakness can be detected with facial recognition software. Second, this study delivers a 'proof of concept' for a DL model that can distinguish MG from HC and classifies disease severity.

Marine big data-driven ensemble learning for estimating global phytoplankton group composition over two decades (1997–2020)

Accurate monitoring of the spatial-temporal distribution and variability of phytoplankton group (PG) composition is of vital importance in better understanding of marine ecosystem dynamics and biogeochemical cycles. While existing bio-optical algorithms provide valuable information, relying solely on satellite ocean color data remains insufficient to obtain high-precision retrieval of PG due to the intricate nature of the bio-optical signal and PG composition itself. An interdisciplinary approach combining advancements in machine learning with big data from ocean observations and simulations offers a promising avenue for more accurate quantification of PG composition. In this study, an ensemble learning approach, called the spatial-temporal-ecological ensemble (STEE) model, is developed to construct a robust prediction model for eight distinct phytoplankton groups (i.e., Diatoms, Dinoflagellates, Haptophytes, Pelagophytes, Cryptophytes, Green Algae, Prokaryotes, and Prochlorococcus). The proposed method introduces multiple data simultaneously: ocean color, physical oceanographic, biogeochemical, and spatial and temporal information. An ensemble strategy is applied to increase the performance of the model by merging three advanced machine-learning algorithms. The combined validation of multiple cross-validation (CV) strategies (i.e., standard, spatial block, and temporal block CVs) shows that the proposed STEE model has superior robustness and generalization ability. In addition, the analysis shows a high degree of concordance between the independent datasets and the modeled estimations for long-time series sites, indicating that the STEE model is capable of effectively monitoring long-term trends in phytoplankton group composition. Finally, the proposed model was utilized to retrieve global monthly phytoplankton group products (STEE-PG) over an extended period (September 1997 to May 2020), and comparisons demonstrated better rationality of spatio-temporal distribution than existing satellite-derived phytoplankton group products. Hence, this new model comprehensively integrates all kinds of observation data and yields long-term global PG products with high accuracy, which will enhance our understanding of the response of marine ecosystems to environmental and climate change.

Resting-state EEG Predicts Cognitive Impairment in Parkinson’s Disease (P6-11.015)

<h3>Objective:</h3> To develop and evaluate an EEG-based assay to predict cognitive functions in Parkinson’s disease (PD) using resting-state EEG data. <h3>Background:</h3> Cognitive dysfunction is a prominent feature of PD. PD related cognitive impairment is usually diagnosed by neuropsychological testing; however, these tests are complex, cannot capture cognitive fluctuations in PD patients, and are not readily compatible with repeated measurements, real-time feedback, or closed-loop neuromodulation therapies. <h3>Design/Methods:</h3> We recorded a few minutes of resting-state EEG data from a few electrodes. Our approach optimized a data-driven, machine-learning technique called ‘Linear predictive coding EEG Algorithms for Parkinson’s disease’ (LEAD-PD) to predict cognitive measures from 100 PD and 49 control participants. We compared our EEG-based cognitive index with the Montreal Cognitive Assessment (MOCA), as well as with different cognitive domains from NIH Toolbox, using multiple cross-validation schemes and randomization tests. We also investigated the relationship between them with multiple regression models, compared our approach with traditional EEG-based features and measured robustness with data truncation tests. <h3>Results:</h3> We found a strong correlation (rho = 0.69, <i>p</i> value &lt; 0.001) between our proposed LEAD-PD cognitive index and the MOCA score using data from only 8 EEG electrodes. LEAD-PD index also strongly predicted other cognitive domains from the NIH toolbox (Dimensional Change Card Sort, Flanker Attention and Inhibitory Control, Picture Vocabulary Task, Pattern Completion Speed Processing, and the Picture Sequence Memory tests). Our approach was consistently more predictive than traditional EEG spectral analysis, yielded 80.44% accuracy (0.90 AUC, 79.52 % sensitivity, and 81.14% specificity) in detecting cognitive impairment, and extended to both PD and control participants. The relationship between our approach and the MOCA was nonlinear with an R² of 0.46 for quadratic regression models. <h3>Conclusions:</h3> LEAD-PD approach using resting-state EEG is computationally-efficient and capable of providing continuous real-time control and feedback, facilitating the diagnosis and monitoring of cognitive impairment in PD. <b>Disclosure:</b> An immediate family member of Dr. Uc has received personal compensation in the range of $500-$4,999 for serving as an officer or member of the Board of Directors for American Board of Pediatrics. Dr. Uc has received personal compensation in the range of $500-$4,999 for serving as an Editor, Associate Editor, or Editorial Advisory Board Member for Elsevier. The institution of Dr. Uc has received research support from Department of Veterans Affairs. The institution of Dr. Uc has received research support from Department of Defense. The institution of Dr. Uc has received research support from Michael J. Fox Foundation. The institution of an immediate family member of Dr. Uc has received research support from NIH. Dr. Anjum has nothing to disclose. Dr. Dasgupta has nothing to disclose. Dr. Narayanan has nothing to disclose.

Harmonizing across datasets to improve the transferability of drug combination prediction

Combination treatment has multiple advantages over traditional monotherapy in clinics, thus becoming a target of interest for many high-throughput screening (HTS) studies, which enables the development of machine learning models predicting the response of new drug combinations. However, most existing models have been tested only within a single study, and these models cannot generalize across different datasets due to significantly variable experimental settings. Here, we thoroughly assessed the transferability issue of single-study-derived models on new datasets. More importantly, we propose a method to overcome the experimental variability by harmonizing dose–response curves of different studies. Our method improves the prediction performance of machine learning models by 184% and 1367% compared to the baseline models in intra-study and inter-study predictions, respectively, and shows consistent improvement in multiple cross-validation settings. Our study addresses the crucial question of the transferability in drug combination predictions, which is fundamental for such models to be extrapolated to new drug combination discovery and clinical applications that are de facto different datasets.

Predicting COVID-19 Severity Integrating RNA-Seq Data Using Machine Learning Techniques

Abstract: A fundamental challenge in the fight against COVID -19 is the development of reliable and accurate tools to predict disease progression in a patient. This information can be extremely useful in distinguishing hospitalized patients at higher risk for needing UCI from patients with low severity. How SARS-CoV-2 infection will evolve is still unclear. Methods: A novel pipeline was developed that can integrate RNA-Seq data from different databases to obtain a genetic biomarker COVID -19 severity index using an artificial intelligence algorithm. Our pipeline ensures robustness through multiple cross-validation processes in different steps. Results: CD93, RPS24, PSCA, and CD300E were identified as a COVID -19 severity gene signature. Furthermore, using the obtained gene signature, an effective multi-class classifier capable of discriminating between control, outpatient, inpatient, and ICU COVID -19 patients was optimized, achieving an accuracy of 97.5%. Conclusion: In summary, during this research, a new intelligent pipeline was implemented with the goal of developing a specific gene signature that can detect the severity of patients suffering COVID -19. Our approach to clinical decision support systems achieved excellent results, even when processing unseen samples. Our system can be of great clinical utility for the strategy of planning, organizing and managing human and material resources, as well as for automatically classifying the severity of patients affected by COVID -19.

Risk-Adjustment of Return to Work After Medical Rehabilitation: Methodical Advancements in Quality Assurance of the German Pension Insurance

Besides the quality of life, patients' return to work is one of the most important treatment results of medical rehabilitation paid by the German Pension Insurance. In order to be able to use the return to work as a quality indicator for medical rehabilitation, a risk adjustment strategy for pre-existing characteristics of patients, rehabilitation departments and labour markets had to be developed. Multiple regression analyses and cross validation were used to develop a risk adjustment strategy, which mathematically compensates the influence of confounders and thus allows for appropriate comparisons between rehabilitation departments regarding patients' return to work after medical rehabilitation. Under the inclusion of experts, the number of employment days in the first and second year after medical rehabilitation were chosen as an appropriate operationalization of return to work. Methodological challenges in the development of the risk adjustment strategy were the identification of a suitable regression method for the distribution of the dependent variable, modelling the multilevel structure of the data appropriately and selecting relevant confounders for return to work. A user-friendly way of communicating the results was developed. The fractional logit regression was chosen as an appropriate regression method to model the U-shaped distribution of the employment days. Low intraclass correlations indicate that the multilevel structure of the data (cross-classified labour market regions and rehabilitation departments) is statistically negligible. Potential confounding factors were theoretically preselected (medical experts were involved for medical parameters) and tested for their prognostic relevance in each indication area using backwards selection. Cross validations proved the risk adjustment strategy to be stable. Adjustment results were displayed in a user-friendly report, including the users' perspective (focus groups and interviews). The developed risk adjustment strategy allows for adequate comparisons between rehabilitation departments and thus enables a quality assessment of treatment results. Methodological challenges, decisions and limitations are discussed in details throughout this paper.

DoubleSG-DTA: Deep Learning for Drug Discovery: Case Study on the Non-Small Cell Lung Cancer with EGFRT790M Mutation

Drug-targeted therapies are promising approaches to treating tumors, and research on receptor-ligand interactions for discovering high-affinity targeted drugs has been accelerating drug development. This study presents a mechanism-driven deep learning-based computational model to learn double drug sequences, protein sequences, and drug graphs to project drug-target affinities (DTAs), which was termed the DoubleSG-DTA. We deployed lightweight graph isomorphism networks to aggregate drug graph representations and discriminate between molecular structures, and stacked multilayer squeeze-and-excitation networks to selectively enhance spatial features of drug and protein sequences. What is more, cross-multi-head attentions were constructed to further model the non-covalent molecular docking behavior. The multiple cross-validation experimental evaluations on various datasets indicated that DoubleSG-DTA consistently outperformed all previously reported works. To showcase the value of DoubleSG-DTA, we applied it to generate promising hit compounds of Non-Small Cell Lung Cancer harboring EGFRT790M mutation from natural products, which were consistent with reported laboratory studies. Afterward, we further investigated the interpretability of the graph-based "black box" model and highlighted the active structures that contributed the most. DoubleSG-DTA thus provides a powerful and interpretable framework that extrapolates for potential chemicals to modulate the systemic response to disease.

Optimal deep brain stimulation sites and networks for stimulation of the fornix in Alzheimer’s disease

Deep brain stimulation (DBS) to the fornix is an investigational treatment for patients with mild Alzheimer’s Disease. Outcomes from randomized clinical trials have shown that cognitive function improved in some patients but deteriorated in others. This could be explained by variance in electrode placement leading to differential engagement of neural circuits. To investigate this, we performed a post-hoc analysis on a multi-center cohort of 46 patients with DBS to the fornix (NCT00658125, NCT01608061). Using normative structural and functional connectivity data, we found that stimulation of the circuit of Papez and stria terminalis robustly associated with cognitive improvement (R = 0.53, p < 0.001). On a local level, the optimal stimulation site resided at the direct interface between these structures (R = 0.48, p < 0.001). Finally, modulating specific distributed brain networks related to memory accounted for optimal outcomes (R = 0.48, p < 0.001). Findings were robust to multiple cross-validation designs and may define an optimal network target that could refine DBS surgery and programming.

Advanced modeling method for quantifying cumulative subjective fatigue in mid-air interaction

Interaction in mid-air can be fatiguing. A model-based method to quantify cumulative subjective fatigue for such interaction was recently introduced in HCI research. This model separates muscle units into three states: active (MA) fatigued (MF) or rested (MR) and defines transition rules between states. This method demonstrated promising accuracy in predicting subjective fatigue accumulated in mid-air pointing tasks. In this paper, we introduce an improved model that additionally captures the variations of the maximum arm strength based on arm postures and adds linearly-varying model parameters based on current muscle strength. To validate the applicability and capabilities of the new model, we tested its performance in various mid-air interaction conditions, including mid-air pointing/docking tasks, with shorter and longer rest and task periods, and a long-term evaluation with individual participants. We present results from multiple cross-validations and comparisons against the previous model and identify that our new model predicts fatigue more accurately. Our modeling approach showed a 42.5% reduction in fatigue estimation error when the longitudinal experiment data is used for an individual participant’s fatigue. Finally, we discuss the applicability and capabilities of our new approach.

De Novo Prediction of Drug Targets and Candidates by Chemical Similarity-Guided Network-Based Inference.

Identifying drug–target interactions is a crucial step in discovering novel drugs and for drug repositioning. Network-based methods have shown great potential thanks to the straightforward integration of information from different sources and the possibility of extracting novel information from the graph topology. However, despite recent advances, there is still an urgent need for efficient and robust prediction methods. Here, we present SimSpread, a novel method that combines network-based inference with chemical similarity. This method employs a tripartite drug–drug–target network constructed from protein–ligand interaction annotations and drug–drug chemical similarity on which a resource-spreading algorithm predicts potential biological targets for both known or failed drugs and novel compounds. We describe small molecules as vectors of similarity indices to other compounds, thereby providing a flexible means to explore diverse molecular representations. We show that our proposed method achieves high prediction performance through multiple cross-validation and time-split validation procedures over a series of datasets. In addition, we demonstrate that our method performed a balanced exploration of both chemical ligand space (scaffold hopping) and biological target space (target hopping). Our results suggest robust and balanced performance, and our method may be useful for predicting drug targets, virtual screening, and drug repositioning.

Abstract 1613: Systematic screening for genomic vulnerabilities as drug targets for mesothelioma

Abstract Mesothelioma is a rare, lethal cancer caused by asbestos. Despite recent advances molecularly stratified therapy is lacking, in part due to the challenge of targeting the preponderance of tumor suppressors. Recent efforts to select therapy from our group have focused on poly(ADP-ribose) polymerase (PARP) and cyclin-dependent kinase 4/6 (CDK4/6) however the molecular determinants of response to therapy are unknown. To address this knowledge gap, we have undertaken high-throughput screening of exome sequenced primary mesothelioma cell lines alongside haploid cell lines harboring CRISPR knockouts of key, positively selected drivers BAP1, NF2 and MTAP. This strategy should enable us to identify robust and repeated drug-gene interactions using ensemble machine learning methods across a drug panel of drugs, both novel and in the clinic. Our platform encompasses multiple cross validation with both primary mesothelioma explants and clinical trial drug-response data. Patients were consented for extended pleurectomy decortication (EPD) and tissue sampled, disaggregated and cultured for no more than four passages to establish cell lines. Immunophenotyping and whole exome sequencing was conducted to validate their mesothelioma origins across an initial panel of 20 cell lines comprising primary cell lines established at the NKI (Netherlands) and Cardiff University (United Kingdom). To further characterize these, we investigated the response of the cell lines to 58 clinically approved cancer drugs using Imagen’s 3-day PredictRx assay. In addition, cell lines were treated with PARP inhibitors, olaparib, rucaparib, niraparib and talazoparib and incubated for 7 days. Cell count and cell death endpoints were captured in addition to phospho-H2AX immunofluorescence. Mesothelioma patient cells displayed differential sensitivity to targeted and non-targeted drugs. There were more nuanced responses amongst the targeted drugs, where for example epithelioid subtypes showed a greater sensitivity to mTOR, VEGFR and HDAC inhibitors in the 3-day assay. Compared with wild type cells, the haploid knockout cells were broadly 3-fold more sensitive to PARP inhibitors; mesothelioma primary cell lines showed exquisite sensitivity to talazoparib (IC50 &amp;lt; 0.1 µM). Multiple cross validation of our results across exome sequenced primary mesothelioma explants and patient tumors (PARP Phase II) is ongoing (data to be presented), leveraging machine learning based prediction of sensitivity, to be further validated in an upcoming PARP randomized trial. Citation Format: Gareth J. Griffiths, Dominic I. James, Beth Dyson, Immaculate M. Nalubowa, Apostolos Nakas, Paul Baas, Zsuzsanna Tabi, Min Zhang, Qianqian Sun, Daniel Faulkner, Nada Nusrat, Essa Y. Baitei, Joanna Dzialo, Alan Dawson, Charlotte Poile, Jake Spicer, Aleksandra Bzura, Jin-Li Luo, Hongji Yang, Cathy Richards, Dean A. Fennell. Systematic screening for genomic vulnerabilities as drug targets for mesothelioma [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1613.

MOOC Dropout Prediction Based on Multidimensional Time-Series Data

Massive open online courses have attracted millions of learners worldwide with flexible learning options. However, online learning differs from offline education in that the lack of communicative feedback is a drawback that magnifies high dropout rates. The analysis and prediction of student’s online learning process can help teachers find the students with dropout tendencies in time and provide additional help. Previous studies have shown that analyzing learning behaviors at different time scales leads to different prediction results. In addition, noise in the time-series data of student behavior can also interfere with the prediction results. To address these issues, we propose a dropout prediction model that combines a multiscale fully convolutional network and a variational information bottleneck. The model extracts multiscale features of student behavior time-series data by constructing a multiscale full convolutional network and then uses a variational information bottleneck to suppress the effect of noise on the prediction results. This study conducted multiple cross-validation experiments on KDD CUP 2015 data set. The results showed that the proposed method achieved the best performance compared to the baseline method.

Multiclass stand-alone and ensemble machine learning algorithms utilised to classify soils based on their physico-chemical characteristics

This study has provided an approach to classify soil using machine learning. Multiclass elements of stand-alone machine learning algorithms (i.e. logistic regression (LR) and artificial neural network (ANN)), decision tree ensembles (i.e. decision forest (DF) and decision jungle (DJ)), and meta-ensemble models (i.e. stacking ensemble (SE) and voting ensemble (VE)) were used to classify soils based on their intrinsic physico-chemical properties. Also, the multiclass prediction was carried out across multiple cross-validation (CV) methods, i.e. train validation split (TVS), k -fold cross-validation (KFCV), and Monte Carlo cross-validation (MCCV). Results indicated that the soils' clay fraction (CF) had the most influence on the multiclass prediction of natural soils' plasticity while specific surface and carbonate content (CC) possessed the least within the nature of the dataset used in this study. Stand-alone machine learning models (LR and ANN) produced relatively less accurate predictive performance (accuracy of 0.45, average precision of 0.5, and average recall of 0.44) compared to tree-based models (accuracy of 0.68, average precision of 0.71, and recall rate of 0.68), while the meta-ensembles (SE and VE) outperformed (accuracy of 0.75, average precision of 0.74, and average recall rate of 0.72) all the models utilised for multiclass classification. Sensitivity analysis of the meta-ensembles proved their capacities to discriminate between soil classes across the methods of CV considered. Machine learning training and validation using MCCV and KFCV methods enabled better prediction while also ensuring that the dataset was not overfitted by the machine learning models. Further confirmation of this phenomenon was depicted by the continuous rise of the cumulative lift curve (LC) of the best performing models when using the MCCV technique. Overall, this study demonstrated that soil's physico-chemical properties do have a direct influence on plastic behaviour and, therefore, can be relied upon to classify soils.

Adverse Drug Event Prediction Using Noisy Literature-Derived Knowledge Graphs: Algorithm Development and Validation.

BackgroundAdverse drug events (ADEs) are unintended side effects of drugs that cause substantial clinical and economic burdens globally. Not all ADEs are discovered during clinical trials; therefore, postmarketing surveillance, called pharmacovigilance, is routinely conducted to find unknown ADEs. A wealth of information, which facilitates ADE discovery, lies in the growing body of biomedical literature. Knowledge graphs (KGs) encode information from the literature, where the vertices and the edges represent clinical concepts and their relations, respectively. The scale and unstructured form of the literature necessitates the use of natural language processing (NLP) to automatically create such KGs. Previous studies have demonstrated the utility of such literature-derived KGs in ADE prediction. Through unsupervised learning of the representations (features) of clinical concepts from the KG, which are used in machine learning models, state-of-the-art results for ADE prediction were obtained on benchmark data sets.ObjectiveDue to the use of NLP to infer literature-derived KGs, there is noise in the form of false positive (erroneous) and false negative (absent) nodes and edges. Previous representation learning methods do not account for such inaccuracies in the graph. NLP algorithms can quantify the confidence in their inference of extracted concepts and relations from the literature. Our hypothesis, which motivates this work, is that by using such confidence scores during representation learning, the learned embeddings would yield better features for ADE prediction models.MethodsWe developed methods to use these confidence scores on two well-known representation learning methods—DeepWalk and Translating Embeddings for Modeling Multi-relational Data (TransE)—to develop their weighted versions: Weighted DeepWalk and Weighted TransE. These methods were used to learn representations from a large literature-derived KG, the Semantic MEDLINE Database, which contains more than 93 million clinical relations. They were compared with Embedding of Semantic Predications, which, to our knowledge, is the best reported representation learning method using the Semantic MEDLINE Database with state-of-the-art results for ADE prediction. Representations learned from different methods were used (separately) as features of drugs and diseases to build classification models for ADE prediction using benchmark data sets. The methods were compared rigorously over multiple cross-validation settings.ResultsThe weighted versions we designed were able to learn representations that yielded more accurate predictive models than the corresponding unweighted versions of both DeepWalk and TransE, as well as Embedding of Semantic Predications, in our experiments. There were performance improvements of up to 5.75% in the F1-score and 8.4% in the area under the receiver operating characteristic curve value, thus advancing the state of the art in ADE prediction from literature-derived KGs.ConclusionsOur classification models can be used to aid pharmacovigilance teams in detecting potentially new ADEs. Our experiments demonstrate the importance of modeling inaccuracies in the inferred KGs for representation learning.