ML Model Research Articles

Global Neural Network Potential with Explicit Many-Body Functions for Improved Descriptions of Complex Potential Energy Surface.

The high dimensional machine learning potential (MLP) that has developed rapidly in the past decade represents a giant step forward in large-scale atomic simulation for complex systems. The long-range interaction and the poor description of chemical reactions are typical problems of high dimensional MLP, which are mainly caused by the poor structure discrimination of the atom-centered ML model. Herein, we propose a low-cost neural-network-based MLP architecture for fitting global potential energy surface data, namely, G-MBNN, that can offer improved energy and force resolution on a complex potential energy surface. In G-MBNN, a set of many-body energy terms based on the local atomic environment are explicitly included in computing the total energy─the total energy of the system is written as the sum of atomic energy and many-body energy contributions. These extra many-body energy terms are computationally low-cost and, importantly, can provide easy access to delicate energy terms in complex systems such as very short repulsion, long-range attractions, and sensitive angular-dependent covalent interactions. We implement G-MBNN in the LASP code and demonstrate the improved accuracy of the new framework in representative systems, including ternary-element energy materials LiCoOx, TiO2 with defects, and a series of organic reactions.

Machine learning aided evaluation and design based on polymer membrane materials

With the acceleration of the global modern industrialisation process and the increasingly serious environmental problems, the development of low energy consumption, high efficiency electrochemical energy conversion equipment and separation system has become a research hotspot in the scientific and industrial circles. At present, machine learning has become an important research method to explore and expand two-dimensional material family. Traditional experimental and computational methods have low fault tolerance when studying two-dimensional materials, which requires a lot of time and research and development costs. Machine learning, due to its powerful data processing capability and flexible algorithm model, can help reduce the time and cost of discovering and understanding two-dimensional materials, and can effectively predict and expand two-dimensional material systems based on data and explore their potential for experimental synthesis and application. This paper will focus on the methods of machine learning, the exploration of machine learning in 2D material design and synthesis, and the exploration of machine learning in 2D material properties and applications. Finally, this paper uses ML algorithm to test the synthesised polymer. The experimental data points and prediction data points have relatively good consistency with each other, which indicates that ML model can be used as a prediction tool to identify the undeveloped polymer for gas separation.

Machine learning approach for the detection of vitamin D level: a comparative study

BackgroundAfter the World Health Organization declared the COVID-19 pandemic, the role of Vitamin D has become even more critical for people worldwide. The most accurate way to define vitamin D level is 25-hydroxy vitamin D(25-OH-D) blood test. However, this blood test is not always feasible. Most data sets used in health science research usually contain highly correlated features, which is referred to as multicollinearity problem. This problem can lead to misleading results and overfitting problems in the ML training process. Therefore, the proposed study aims to determine a clinically acceptable ML model for the detection of the vitamin D status of the North Cyprus adult participants accurately, without the need to determine 25-OH-D level, taking into account the multicollinearity problem.MethodThe study was conducted with 481 observations who applied voluntarily to Internal Medicine Department at NEU Hospital. The classification performance of four conventional supervised ML models, namely, Ordinal logistic regression(OLR), Elastic-net ordinal regression(ENOR), Support Vector Machine(SVM), and Random Forest (RF) was compared. The comparative analysis is performed regarding the model's sensitivity to the participant’s metabolic syndrome(MtS)'positive status, hyper-parameter tuning, sensitivities to the size of training data, and the classification performance of the models.ResultsDue to the presence of multicollinearity, the findings showed that the performance of the SVM(RBF) is obviously negatively affected when the test is examined. Moreover, it can be obviously detected that RF is more robust than other models when the variations in the size of training data are examined. This experiment's result showed that the selected RF and ENOR showed better performances than the other two models when the size of training samples was reduced. Since the multicollinearity is more severe in the small samples, it can be concluded that RF and ENOR are not affected by the presence of the multicollinearity problem. The comparative analysis revealed that the RF classifier performed better and was more robust than the other proposed models in terms of accuracy (0.94), specificity (0.96), sensitivity or recall (0.94), precision (0.95), F1-score (0.95), and Cohen's kappa (0.90).ConclusionIt is evident that the RF achieved better than the SVM(RBF), ENOR, and OLR. These comparison findings will be applied to develop a Vitamin D level intelligent detection system for being used in routine clinical, biochemical tests, and lifestyle characteristics of individuals to decrease the cost and time of vitamin D level detection.

Antibiotic resistance genes prevalence prediction and interpretation in beaches affected by urban wastewater discharge

BackgroundThe annual death toll of over 1.2 million worldwide is attributed to infections caused by resistant bacteria, driven by the significant impact of antibiotic misuse and overuse in spreading these bacteria and their associated antibiotic resistance genes (ARGs). While limited data suggest the presence of ARGs in beach environments, efficient prediction tools are needed for monitoring and detecting ARGs to ensure public health safety. This study aims to develop interpretable machine learning methods for predicting ARGs in beach waters, addressing the challenge of black-box models and enhancing our understanding of their internal mechanisms. MethodsIn this study, we systematically collected beach water samples and subsequently isolated bacteria from these samples using various differential and selective media supplemented with different antibiotics. Resistance profiles of bacteria were determined by using Kirby-Bauer disk diffusion method. Further, ARGs were enumerated by using the quantitative polymerase chain reaction (qPCR) to detect and quantify ARGs. The obtained qPCR data and hydro-meteorological were used to create an ML model with high prediction performance and we further used two explainable artificial intelligence (xAI) model-agnostic interpretation methods to describe the internal behavior of ML model. ResultsUsing qPCR, we detected blaCTX−M, blaNDM, blaCMY, blaOXA, blatetX, blasul1, and blaaac(6′-Ib-cr) in the beach waters. Further, we developed ML prediction models for blaaac(6′-Ib-cr), blasul1, and blatetX using the hydro-metrological and qPCR-derived data and the models demonstrated strong performance, with R2 values of 0.957, 0.997, and 0.976, respectively. ConclusionsOur findings show that environmental factors, such as water temperature, precipitation, and tide, are among the important predictors of the abundance of resistance genes at beaches.

Identifying biomarkers for tDCS treatment response in Alzheimer's disease patients: a machine learning approach using resting-state EEG classification.

Transcranial direct current stimulation (tDCS) is a promising treatment for Alzheimer's Disease (AD). However, identifying objective biomarkers that can predict brain stimulation efficacy, remains a challenge. The primary aim of this investigation is to delineate the cerebral regions implicated in AD, taking into account the existing lacuna in comprehension of these regions. In pursuit of this objective, we have employed a supervised machine learning algorithm to prognosticate the neurophysiological outcomes resultant from the confluence of tDCS therapy plus cognitive intervention within both the cohort of responders and non-responders to antecedent tDCS treatment, stratified on the basis of antecedent cognitive outcomes. The data were obtained through an interventional trial. The study recorded high-resolution electroencephalography (EEG) in 70 AD patients and analyzed spectral power density during a 6 min resting period with eyes open focusing on a fixed point. The cognitive response was assessed using the AD Assessment Scale-Cognitive Subscale. The training process was carried out through a Random Forest classifier, and the dataset was partitioned into K equally-partitioned subsamples. The model was iterated k times using K-1 subsamples as the training bench and the remaining subsample as validation data for testing the model. A clinical discriminating EEG biomarkers (features) was found. The ML model identified four brain regions that best predict the response to tDCS associated with cognitive intervention in AD patients. These regions included the channels: FC1, F8, CP5, Oz, and F7. These findings suggest that resting-state EEG features can provide valuable information on the likelihood of cognitive response to tDCS plus cognitive intervention in AD patients. The identified brain regions may serve as potential biomarkers for predicting treatment response and maybe guide a patient-centered strategy. https://classic.clinicaltrials.gov/ct2/show/NCT02772185?term=NCT02772185&draw=2&rank=1, identifier ID: NCT02772185.

End-to end decision support system for sleep apnea detection and Apnea-Hypopnea Index calculation using hybrid feature vector and Machine learning

Sleep apnea is a disease that occurs due to the decrease in oxygen saturation in the blood and directly affects people's lives. Detection of sleep apnea is crucial for assessing sleep quality. It is also an important parameter in the diagnosis of various other diseases (diabetes, chronic kidney disease, depression, and cardiological diseases). Recent studies show that detection of sleep apnea can be done via signal processing, especially EEG and ECG signals. However, the detection accuracy needs to be improved. In this paper, a ML model is used for the detection of sleep apnea using 19 static sensor data and 2 dynamic data (Sleep score and Arousal). The sensor data is recorded as a discrete signal and the sleep process is divided into 4.8 M segments. In this work, 19 different sensor data sets were recorded with polysomnography (PSG). These data sets have been used to perform sleep scoring. Then, arousal status marking is done. Model training was carried out with the feature vector consisting of 21 data obtained. Tests were performed with eight different machine learning techniques on a unique dataset consisting of 113 patients. After all, it was automatically determined whether people were diseased (a kind of apnea) or healthy. The proposed model had an average accuracy of 97.27%, while the recall, precision, and f-score values were 99.18%, 95.32%, and 97.20%, respectively. After all, the model that less feature engineering, less complex classification model, higher dataset usage, and higher classification performance has been revealed.

Enhancing Health and Wellness: Deep Learning Powered Exercise Recognition and Tracking

Abstract: This project aims to develop a React-based web application that enables users to track, analyze, manage, and schedule their workout routines. Additionally, the application incorporates a machine learning model that utilizes the user’s web camera to track their movements and count reps, sets, and calories burned. This innovative feature provides users with a more accurate representation of their workout progress and empowers them to make data-driven decisions regarding their fitness goals. To achieve this, we utilized various technologies such as React for the frontend, Node.js for the backend, and MongoDB for the database. We also integrated TensorFlow.js, an open-source machine learning library, to build the ML model that powers the application’s tracking and analysis capabilities. Furthermore, we designed the user interface using Bootstrap with Vanilla CSS to ensure a responsive and user-friendly experience. Through this project, our primary objective was to address the limitations of existing workout tracking applications that heavily rely on subjective and often inaccurate user input. Our solution offers a more objective and efficient approach to tracking fitness progress. We firmly believe that this technology has the potential to revolutionize the fitness industry. Overall, our project showcases the power of machine learning and its ability to enhance people’s lives in meaningful ways.

A Prospective Study of Machine Learning-Assisted Radiotherapy Planning for Patients Receiving 54 Gy to the Brain

Radiotherapy (RT) planning is presently a semi-manual, iterative, labor-intensive process which may result in unnecessary variation in plan quality. To improve treatment plan quality and decrease RT planning time, we conducted a prospective, blinded study to compare machine learning-assisted planning with conventional manual planning for patients receiving 54 Gy in 30 fractions for a primary brain tumor. From January 31, 2022 to January 10, 2023, 40 patients receiving 54 Gy for primary CNS tumors were prospectively enrolled (median age 50 years, range 4-78 years). Patients underwent standard CT/MR simulation and target/OAR delineation by the treating radiation oncologist. Each patient had one ML plan and 1-2 manual RT plans created by different planners. The reviewing oncologist was blinded to planning method by removing optimization and IMRT/VMAT beam arrangement details from all plans, which were then rated based on clinical acceptability, target coverage, OAR sparing, conformity, and dose-fall off. One preferred plan was chosen and used for clinical treatment. A total of 115 plans for 40 patients were evaluated: 40 ML plans (35% of all plans), and 75 manual plans (65% of all plans; 5 and 35 patients had 1 and 2 manual plans created, respectively). ML plans required a mean planning time of 65 min as compared to 107 min for manual plans, with a mean time savings of 41 min per patient (paired t-test p = 0.002). 97% of ML plans (95% confidence interval [CI] 85-100) and 96% of manual plans (95% CI 87-99) were designated clinically acceptable by the treating radiation oncologist. While ML-assisted plans represented 35% of plans evaluated, they were chosen as preferred for clinical treatment in 43% of cases (17/40, 95% CI 29-58, p = 0.32). Median doses to the brain (10.8 Gy vs. 11.3 Gy, Wilcoxon rank-sum p = 0.012) and brain minus PTV (9.2 Gy vs 10.0 Gy, Wilcoxon rank-sum p = 0.009) were lower with ML planning versus manual planning, respectively. Doses to other structures, including hippocampi, cochlea, pituitary and hypothalamus were not statistically different. In this prospective study with blinded oncologist evaluation, ML-assisted RT planning for primary CNS tumors was faster than manual planning, and produced a very high rate of acceptable plans with similar or superior OAR sparing. Future work will be undertaken to iteratively refine the ML model using the preferred cases from this study.

Development and validation of a machine learning-based nomogram for predicting HLA-B27 expression

BackgroundHLA-B27 positivity is normal in patients undergoing rheumatic diseases. The diagnosis of many diseases requires an HLA-B27 examination.MethodsThis study screened totally 1503 patients who underwent HLA-B27 examination, liver/kidney function tests, and complete blood routine examination in First Affiliated Hospital of Guangxi Medical University. The training cohort included 509 cases with HLA-B27 positivity whereas 611 with HLA-B27 negativity. In addition, validation cohort included 147 cases with HLA-B27 positivity whereas 236 with HLA-B27 negativity. In this study, 3 ML approaches, namely, LASSO, support vector machine (SVM) recursive feature elimination and random forest, were adopted for screening feature variables. Subsequently, to acquire the prediction model, the intersection was selected. Finally, differences among 148 cases with HLA-B27 positivity and negativity suffering from ankylosing spondylitis (AS) were investigated.ResultsSix factors, namely red blood cell count, human major compatibility complex, mean platelet volume, albumin/globulin ratio (ALB/GLB), prealbumin, and bicarbonate radical, were chosen with the aim of constructing the diagnostic nomogram using ML methods. For training queue, nomogram curve exhibited the value of area under the curve (AUC) of 0.8254496, and C-value of the model was 0.825. Moreover, nomogram C-value of the validation queue was 0.853, and the AUC value was 0.852675. Furthermore, a significant decrease in the ALB/GLB was noted among cases with HLA-B27 positivity and AS cases.ConclusionTo conclude, the proposed ML model can effectively predict HLA-B27 and help doctors in the diagnosis of various immune diseases.

A machine learning-based study of sCO2 cycle precooler's design and performance with straight and zigzag channels

This research addresses a gap by investigating pumping power reduction through efficient channel geometries, an area underexplored due to complexities in pseudo-critical flow and heat transfer assessment. The study evaluates straight and zigzag channeled printed circuit heat exchangers' potential as sCO2-BC precoolers. A comprehensive approach integrates CFD, machine learning ML, in-house codes for heat exchanger (RPDAC) design and cycle simulation (CDPC) and multi-objective genetic algorithm (MOGA) techniques. CFD assesses thermohydraulic performance and informs an ML model, which guides RPDAC design. MOGA identifies optimal designs minimizing precooler size and maximizing cycle performance. PDAC findings reveal that straight channel precoolers can halve pumping power compared to zigzag channels, albeit at the cost of a threefold increase in precooler length. Optimization results emphasize that peak efficiency is attainable through straight-channel designs, despite their size being up to eight times larger than zigzag-channel counterparts. Concurrently, precooler designs featuring zigzag channels, aligned with compressor inlet temperatures of 305 K to 306 K and channel Reynolds numbers spanning 17,000 to 18,000, strike a balance between cycle efficiency and precooler size. While cycle efficiency in the proposed zigzag designs slightly lags behind that of optimal straight channel precoolers, the recommended configuration presents a threefold reduction in size.

The Rossby Normal Mode as a Physical Linkage in a Machine Learning Forecast Model for the SST and SSH of South China Sea Deep Basin

AbstractMachine learning (ML) has been widely applied in Earth sciences, but its black‐box nature still remains. Existing techniques for interpretable ML developed by deep learning researchers are not satisfactory in understanding physical problems. In this study, our objective is to establish the physical linkage between features extracted from ML model and results derived from physical equations that describe the problem. This approach aims to provide a more comprehensible way of understanding an ML model. We select a less complex ML forecasting model, consisting of traditional statistical algorithms, which effectively forecast sea surface temperature anomaly (SSTA) and sea level anomaly (SLA) of the South China Sea (SCS) in 30 days ahead. Here, we focus on the SLA prediction and detect the physical mechanism of model capability to predict SCS SLA. Previous study has identified Rossby normal modes as the physical mechanism for long‐lived eddies in SCS in previous study. We here demonstrate the relationship between the ML model and Rossby normal modes in SCS in terms of temporal variations and spatial distributions of water mass. The model's skill in predicting Rossby normal modes partially explains its skill in predicting SLA, thereby providing a more transparent and interpretable basis for forecasts in SCS.

Data-driven machine learning models for quick prediction of the Stokes shift of organic fluorescent materials

Organic fluorescent materials are widely used in various fields, including OLEDs, organic solar cells, and bio-imaging. However, designing and synthesizing new fluorescent organic materials with desirable properties for specific applications require knowledge of the chemical and physical properties of previously studied molecules. One critical property of fluorescent organic compounds is the Stokes shift, which is usually measured experimentally and is known to be time-consuming. Time-dependent density functional theory (TD-DFT) has been used to predict Stokes shifts, but its computational costs restrict the screening of fluorescent organic materials. To address this challenge, we propose a machine learning model based on an ensemble learning approach called LightGBM algorithm, to predict the Stokes shift of organic fluorescent compounds. Based on 15,987 sets of Stokes shift data processed with molecular fingerprints, our ML models show satisfactory results. The squared correlation coefficient (R2), mean absolute error (MAE), and root mean square error (RMSE) of the independent test set for the optimal model are 0.86, 12.27 nm, and 19.16 nm, respectively. Unseen cases confirmed the prediction performance of our ML model. Finally, we applied the ML prediction model to enable rapid screening of organic fluorescent compound with desired Stokes shift. Our study presents a rapid and accurate method for predicting the Stokes shift of organic fluorescent compounds, which accelerate the design of organic fluorescent materials with desired Stokes shift. All source codes and dataset are freely available at https://github.com/Yihuan-Zhao93/Stocks-shiftsML.

FREDY: Federated Resilience Enhanced with Differential Privacy

Federated Learning is identified as a reliable technique for distributed training of ML models. Specifically, a set of dispersed nodes may collaborate through a federation in producing a jointly trained ML model without disclosing their data to each other. Each node performs local model training and then shares its trained model weights with a server node, usually called Aggregator in federated learning, as it aggregates the trained weights and then sends them back to its clients for another round of local training. Despite the data protection and security that FL provides to each client, there are still well-studied attacks such as membership inference attacks that can detect potential vulnerabilities of the FL system and thus expose sensitive data. In this paper, in order to prevent this kind of attack and address private data leakage, we introduce FREDY, a differential private federated learning framework that enables knowledge transfer from private data. Particularly, our approach has a teachers–student scheme. Each teacher model is trained on sensitive, disjoint data in a federated manner, and the student model is trained on the most voted predictions of the teachers on public unlabeled data which are noisy aggregated in order to guarantee the privacy of each teacher’s sensitive data. Only the student model is publicly accessible as the teacher models contain sensitive information. We show that our proposed approach guarantees the privacy of sensitive data against model inference attacks while it combines the federated learning settings for the model training procedures.

Fast Optimal Power Flow With Guarantees via an Unsupervised Generative Model

Climate change concerns are driving the widespread integration of renewable generation sources, storage systems, electric vehicles and diverse consumer loads. Inherent variabilities of these power entities introduce uncertainties that affect the economical operations and integrity of the electrical grid. Thus, optimal power flow (OPF) studies must be conducted at high granularity in order to account for these fluctuations. However, due to the non-convex nature of the OPF problem, solving it in real-time still remains an open challenge. In this paper, a novel data-driven approach that combines generative learning, information theory and domain knowledge is proposed to enable real-time OPF studies. In order to train this model, only <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">feasible</i> datapoints are necessary - not optimal values. Once trained, the time required to compute the solution is extremely fast and falls within the sub-second range. No information pertaining to the topology or power flow equations is required once the model is trained. Theoretical guarantees on optimality and feasibility of the solutions generated by the proposed ML model are established. Comprehensive practical and comparative studies conducted demonstrate the efficacy of the proposed data-driven approach for solving OPF in real-time.

KNN-Based ML Model for the Symbol Prediction in TCM Trellis Coded Modulation TCM Decoder

Machine Learning is a booming technology today. In a machine learning set of training, data is to be provided to the model for training and that model predicts the output. Machine Learning models are trained using a computer program known as ML algorithms.The new machine learning-based Transition Metric Unit (TMU) of 4D- 8PSK Trellis coded Modulation TCM Decoder is presented in this work. The classic Viterbi decoder's branch metric unit, or TMU, takes on a complex structure. Trellis coded Modulation (TCM) is a combination of 8 PSK modulations and Error Correcting Code (ECC). TMU is one of the complex units of the TCM decoder, which is essentially a Viterbi decoder. Similar to how the first Branch metric is determined in the straightforward Viterbi decoder, the TCM decoder performs this BM computation via the TMU unit. The TMU becomes challenging and uses more dynamic power as a result of the enormous constraint length and the vast number of encoder states.In the proposed algorithm innovative KNN (K nearest neighbours) based ML model is developed. It is a supervised learning model in which input and output both are provided to the model, training data also called the labels, when a new set of data will come the model will give output based on its previous set experience and data.Here we are using this ML model for the symbol prediction at the receiver end of the TCM decoder based on the previous learning. Using the proposed innovation, the paper perceives the optimization of the TCM Decoder which will further reduce the H/W requirements and low latency which results in less power consumption.

EHFT: An Ensembled Hyperopt Fine-Tuned Neural Network for Disease Detection in Tomato Plants

The identification of unhealthy plants in the crops at the early stage of cultivation helps for good farming. Unhealthy parts can be recognized using shape, color and texture, which are processed using feature extraction techniques. The feature extraction system stores the images in the matrix pixel format, which requires 3 channels for processing the images. Traditional neural networks utilize backpropagation techniques to adjust the random weights, which requires many resources while extracting a more significant number of features from a huge amount of data. These mechanisms also require more trainable parameters during the transformation of data from one layer to another. The proposed model implements the pre-trained model "RESNET152" (Residual Network), which is efficient for feature extraction and designs the last layer of the network as a "Tuned X-Gradient Boosting" ensemble algorithm for performing the binary classification of tomato leaves. RESNET can reduce computational resources because it implements residual blocks which fasten the learning rate by skipping a few connections in the network. The fine-tuned ensemble model helps the model identify the best parameters quickly. The learnable parameters are the essential elements of any ML model because they can easily identify the patterns associated with the different features. In the proposed model for feature extraction, pattern matching is the crucial step. Therefore, it is very necessary to tune the XGBOOST algorithm. Compared to the traditional approaches, the proposed model enhanced the accuracy performance in training and testing with 98.58% and 95.56%, correspondingly

“TimBre” Pilot Study Conducted Using Multi-Country Training and Validation Data for Screening of Pulmonary Tuberculosis Using Cough (Acoustic Sounds), Clinical &amp; Demographic Inputs

TimBre from Docturnal offers screening for multiple lung diseases – Pulmonary Tuberculosis, Pneumonia, Covid19 &amp; COPD. Detailed studies of TimBre in the past used a third-party Microphone Array that focused on a XY arrangement that provided high fidelity cough sounds with an average length of &gt;5 seconds and demographic data such as Height, Weight, BMI [1]. In the current study, cough sounds were collected from 7 different countries (India, Vietnam, Philippines, Uganda, Tanzania, Madagascar, and South Africa) using Mobile Phones from different manufacturers &amp; recorded solicited coughs in a clinic for a duration of 0.5 seconds. A plethora of demographic and clinical variables were provided of which a subset was used by TimBre algorithm. Most importantly, the .WAV files were recorded in a single channel at a sampling rate of 44.1kHz &amp; 16 bits. The study details two approaches wherein the first method was to concatenate all the 0.5 second WAV files based on a timestamp provided for each StudyID in the training &amp; scoring set while the second method involved using the 0.5 second snippets as-is in both training and validation sets without any concatenation. Both approaches used a combination of demographic, clinical and spectral variables. The first approach on the independent test set yielded a sensitivity and specificity of 68.6% and 71.7% respectively with an AUC of 0.75 while the second approach yielded a sensitivity &amp; specificity of 75.41% and 68.30% respectively with an AUC of 0.78. Thus, the ML model performed better in the second approach and we anticipate it to improve with additional training data.

(Invited) Machine Learning and Fast Experimental Screening-Assisted Development of Organic Solar Cell

Non-fullerene, a small molecular electron acceptor, has substantially improved the power conversion efficiency of organic photovoltaics (OPVs).[1] However, the large structural freedom of π-conjugated polymers and molecules makes it difficult to be explored with limited resources. Machine learning, which is based on the rapidly growing artificial intelligence technology, is a high-throughput method to accelerate the speed of material design and process optimization; however, it suffers from limitations in terms of prediction accuracy, interpretability, data collection, and available data (particularly, experimental data). This recognition motivates the present review, which focuses on utilizing the experimental dataset for ML to efficiently aid OPV research. The author discusses the trends in ML-OPV publications, the NFA category, and the effects of data size and explanatory variables (fingerprints or Mordred descriptors) on the prediction accuracy and explainability, which broadens the scope of ML and would be useful for the development of next-generation solar cell materials.[2]Despite the advance of ML, the predictive accuracy of ML currently remains insufficient for the design of OPV semiconductors that exhibit a complex connectivity between chemical structure and PCE. In this study, we examined the impact of data selection and the introduction of artificially generated failure data on ML predictions of NFA solar cells. The authors demonstrated that an ML model empowered by artificially generated failure data (~0% PCE by insoluble polymers based on an inappropriate choice of solubilizing side alkyl chains) led to improved predictions.[3] This approach was validated through the synthesis and characterization of twelve polymers (benzothiadiazole, thienothiophene, or tetrazine coupled with benzodithiophene; benzobisthiazole coupled with dioxo-benzodithiophene). Our work offers a facile approach to mitigate the difficulties of the ML-driven development of OPV materials that is also readily applicable to other material science fields.Reference[1] Kranthiraja, A. Saeki, Adv. Funct. Mater. 31 (2021) 2011168[2] Miyake, A. Saeki, J. Phys. Chem. Lett. 12 (2021) 12391.[3] Miyake, K. Kranthiraja, F. Ishiwari, A. Saeki, Chem. Mater. 34 (2022) 6912.

MRI-Based Surrogate Imaging Markers of Aggressiveness in Prostate Cancer: Development of a Machine Learning Model Based on Radiomic Features.

This retrospective study included 86 adult Hispanic men (60 ± 8.2 years, median prostate-specific antigen density (PSA-D) 0.15 ng/mL2) with PCa who underwent prebiopsy 3T MRI followed by targeted MRI-ultrasound fusion and systematic biopsy. Two observers performed 2D segmentation of lesions in T2WI/ADC images. We classified csPCa (GS ≥ 7) vs. non-csPCa (GS = 6). Univariate statistical tests were performed for different parameters, including prostate volume (PV), PSA-D, PI-RADS, and radiomic features. Multivariate models were built using the automatic feature selection algorithm Recursive Feature Elimination (RFE) and different classifiers. A stratified split separated the train/test (80%) and validation (20%) sets. Radiomic features derived from T2WI/ADC are associated with GS in patients with PCa. The best model found was multivariate, including image (T2WI/ADC) and clinical (PV and PSA-D) information. The validation area under the curve (AUC) was 0.80 for differentiating csPCa from non-csPCa, exhibiting better performance than PI-RADS (AUC: 0.71) and PSA-D (AUC: 0.78). Our multivariate ML model outperforms PI-RADS v2.1 and established clinical indicators like PSA-D in classifying csPCa accurately. This underscores MRI-derived radiomics' (T2WI/ADC) potential as a robust biomarker for assessing PCa aggressiveness in Hispanic patients.

Predicting aqueous sorption of organic pollutants on microplastics with machine learning

Microplastics (MPs) are ubiquitously distributed in freshwater systems and they can determine the environmental fate of organic pollutants (OPs) via sorption interaction. However, the diverse physicochemical properties of MPs and the wide range of OP species make a deeper understanding of sorption mechanisms challenging. Traditional isotherm-based sorption models are limited in their universality since they normally only consider the nature and characteristics of either sorbents or sorbates individually. Therefore, only specific equilibrium concentrations or specific sorption isotherms can be used to predict sorption. To systematically evaluate and predict OP sorption under the influence of both MPs and OPs properties, we collected 475 sorption data from peer-reviewed publications and developed a poly-parameter-linear-free-energy-relationship-embedded machine learning method to analyze the collected sorption datasets. Models of different algorithms were compared, and the genetic algorithm and support vector machine hybrid model displayed the best prediction performance (R2 of 0.93 and root-mean-square-error of 0.07). Finally, comparison results of three feature importance analysis tools (forward step wise method, Shapley method, and global sensitivity analysis) showed that chemical properties of MPs, excess molar refraction, and hydrogen-bonding interaction of OPs contribute the most to sorption, reflecting the dominant sorption mechanisms of hydrophobic partitioning, hydrogen bond formation, and π-π interaction, respectively. This study presents a novel sorbate-sorbent-based ML model with a wide applicability to expand our capacity in understanding the complicated process and mechanism of OP sorption on MPs.