Current Machine Learning Models Research Articles

Harnessing artificial intelligence for predicting and managing postoperative pain: a narrative literature review.

This review examines recent research on artificial intelligence focusing on machine learning (ML) models for predicting postoperative pain outcomes. We also identify technical, ethical, and practical hurdles that demand continued investigation and research. Current ML models leverage diverse datasets, algorithmic techniques, and validation methods to identify predictive biomarkers, risk factors, and phenotypic signatures associated with increased acute and chronic postoperative pain and persistent opioid use. ML models demonstrate satisfactory performance to predict pain outcomes and their prognostic trajectories, identify modifiable risk factors and at-risk patients who benefit from targeted pain management strategies, and show promise in pain prevention applications. However, further evidence is needed to evaluate the reliability, generalizability, effectiveness, and safety of ML-driven approaches before their integration into perioperative pain management practices. Artificial intelligence (AI) has the potential to enhance perioperative pain management by providing more accurate predictive models and personalized interventions. By leveraging ML algorithms, clinicians can better identify at-risk patients and tailor treatment strategies accordingly. However, successful implementation needs to address challenges in data quality, algorithmic complexity, and ethical and practical considerations. Future research should focus on validating AI-driven interventions in clinical practice and fostering interdisciplinary collaboration to advance perioperative care.

Higher-Order Convolutional Neural Networks for Essential Climate Variables Forecasting

Earth observation imaging technologies, particularly multispectral sensors, produce extensive high-dimensional data over time, thus offering a wealth of information on global dynamics. These data encapsulate crucial information in essential climate variables, such as varying levels of soil moisture and temperature. However, current cutting-edge machine learning models, including deep learning ones, often overlook the treasure trove of multidimensional data, thus analyzing each variable in isolation and losing critical interconnected information. In our study, we enhance conventional convolutional neural network models, specifically those based on the embedded temporal convolutional network framework, thus transforming them into models that inherently understand and interpret multidimensional correlations and dependencies. This transformation involves recasting the existing problem as a generalized case of N-dimensional observation analysis, which is followed by deriving essential forward and backward pass equations through tensor decompositions and compounded convolutions. Consequently, we adapt integral components of established embedded temporal convolutional network models, like encoder and decoder networks, thus enabling them to process 4D spatial time series data that encompass all essential climate variables concurrently. Through the rigorous exploration of diverse model architectures and an extensive evaluation of their forecasting prowess against top-tier methods, we utilize two new, long-term essential climate variables datasets with monthly intervals extending over four decades. Our empirical scrutiny, particularly focusing on soil temperature data, unveils that the innovative high-dimensional embedded temporal convolutional network model-centric approaches markedly excel in forecasting, thus surpassing their low-dimensional counterparts, even under the most challenging conditions characterized by a notable paucity of training data.

Developing a novel causal inference algorithm for personalized biomedical causal graph learning using meta machine learning

BackgroundModeling causality through graphs, referred to as causal graph learning, offers an appropriate description of the dynamics of causality. The majority of current machine learning models in clinical decision support systems only predict associations between variables, whereas causal graph learning models causality dynamics through graphs. However, building personalized causal graphs for each individual is challenging due to the limited amount of data available for each patient.MethodIn this study, we present a new algorithmic framework using meta-learning for learning personalized causal graphs in biomedicine. Our framework extracts common patterns from multiple patient graphs and applies this information to develop individualized graphs. In multi-task causal graph learning, the proposed optimized initial guess of shared commonality enables the rapid adoption of knowledge to new tasks for efficient causal graph learning.ResultsExperiments on one real-world biomedical causal graph learning benchmark data and four synthetic benchmarks show that our algorithm outperformed the baseline methods. Our algorithm can better understand the underlying patterns in the data, leading to more accurate predictions of the causal graph. Specifically, we reduce the structural hamming distance by 50-75%, indicating an improvement in graph prediction accuracy. Additionally, the false discovery rate is decreased by 20-30%, demonstrating that our algorithm made fewer incorrect predictions compared to the baseline algorithms.ConclusionTo the best of our knowledge, this is the first study to demonstrate the effectiveness of meta-learning in personalized causal graph learning and cause inference modeling for biomedicine. In addition, the proposed algorithm can also be generalized to transnational research areas where integrated analysis is necessary for various distributions of datasets, including different clinical institutions.

A multimodal Transformer Network for protein-small molecule interactions enhances predictions of kinase inhibition and enzyme-substrate relationships.

The activities of most enzymes and drugs depend on interactions between proteins and small molecules. Accurate prediction of these interactions could greatly accelerate pharmaceutical and biotechnological research. Current machine learning models designed for this task have a limited ability to generalize beyond the proteins used for training. This limitation is likely due to a lack of information exchange between the protein and the small molecule during the generation of the required numerical representations. Here, we introduce ProSmith, a machine learning framework that employs a multimodal Transformer Network to simultaneously process protein amino acid sequences and small molecule strings in the same input. This approach facilitates the exchange of all relevant information between the two molecule types during the computation of their numerical representations, allowing the model to account for their structural and functional interactions. Our final model combines gradient boosting predictions based on the resulting multimodal Transformer Network with independent predictions based on separate deep learning representations of the proteins and small molecules. The resulting predictions outperform recently published state-of-the-art models for predicting protein-small molecule interactions across three diverse tasks: predicting kinase inhibitions; inferring potential substrates for enzymes; and predicting Michaelis constants KM. The Python code provided can be used to easily implement and improve machine learning predictions involving arbitrary protein-small molecule interactions.

Machine learning techniques for predicting microvascular network hemodynamics

Existing techniques to image microvascular networks in vivo do not allow a direct and accurate measurement of several hemodynamic variables such as the wall shear stress (WSS). They cannot also provide a full 3D characterization of blood velocity and red blood cell (RBC) concentration profiles in every vessel in a microvascular network. Such physiological parameters however play critical roles in blood flow regulation, disease progression, angiogenesis, and hemostasis. Theoretical models of network blood flow, often used for hemodynamic predictions in experimentally acquired images of the microvascular network, cannot provide the full 3D blood velocity, RBC concentration and WSS profiles as these models assume one-dimensional flow governed by Poiseuille’s law. In contrast, high-fidelity computational models that simulate flow of blood as a suspension of deformable RBCs and model the 3D vessel geometry can readily predict such physiological variables with high accuracy. Such high-fidelity computational models have recently been applied to predict hemodynamic variables in microvascular networks in silico composed of many vessels and vascular junctions. Such models, however, tend to become computationally expensive with increasing size of the network; therefore, they are not feasible for use in large networks at organ scale. To overcome this limitation, we developed a machine learning (ML) approach that can predict hemodynamic variables in 3D and with high accuracy. The ML models are based on artificial neural network and convolution-based U-net models. Using experimentally acquired images of the microvascular networks, inputs at a few inlets, and a “bank” of database that we have developed, these ML models can predict flow and RBC distribution in all vessels in the network, 3D blood velocity and RBC concentration profiles over the cross-section of every vessel, 3D WSS profiles, and time-dependent variations in blood flow in any vessel. The prediction from the ML models compare very well against the true data. Also, the prediction time using ML is significantly less than that of the high-fidelity models. The current ML models, which to our knowledge are the first of their kind, are highly promising for image-based predictions of sub-cellular resolved capillary hemodynamics in organ-scale networks. This study paves the way for AI to make hemodynamic predictions in organ-scale capillary vessel networks while retaining the sub-cellular scale details and overcoming the limitations of the in vivo imaging techniques, with potential applications in hematological and microvascular disorders, angiogenesis, and vascular-mediated drug delivery. Funded by NIH (R01-EY033003) and NSF (CBET 2302212). This is the full abstract presented at the American Physiology Summit 2024 meeting and is only available in HTML format. There are no additional versions or additional content available for this abstract. Physiology was not involved in the peer review process.

Data-driven intensive care: alack of comprehensive datasets

Intensive care units provide adata-rich environment with the potential to generate datasets in the realm of big data, which could be utilized to train powerful machine learning (ML) models. However, the currently available datasets are too small and exhibit too little diversity due to their limitation to individual hospitals. This lack of extensive and varied datasets is aprimary reason for the limited generalizability and resulting low clinical utility of current ML models. Often, these models are based on data from single centers and suffer from poor external validity. There is an urgent need for the development of large-scale, multicentric, and multinational datasets. Ensuring data protection and minimizing re-identification risks pose central challenges in this process. The "Amsterdam University Medical Center database (AmsterdamUMCdb)" and the "Salzburg Intensive Care database (SICdb)" demonstrate that open access datasets are possible in Europe while complying with the data protection regulations of the General Data Protection Regulation (GDPR). Another challenge in building intensive care datasets is the absence of semantic definitions in the source data and the heterogeneity of data formats. Establishing binding industry standards for the semantic definition is crucial to ensure seamless semantic interoperability between datasets.

Do AI models produce better weather forecasts than physics-based models? A quantitative evaluation case study of Storm Ciarán

There has been huge recent interest in the potential of making operational weather forecasts using machine learning techniques. As they become a part of the weather forecasting toolbox, there is a pressing need to understand how well current machine learning models can simulate high-impact weather events. We compare short to medium-range forecasts of Storm Ciarán, a European windstorm that caused sixteen deaths and extensive damage in Northern Europe, made by machine learning and numerical weather prediction models. The four machine learning models considered (FourCastNet, Pangu-Weather, GraphCast and FourCastNet-v2) produce forecasts that accurately capture the synoptic-scale structure of the cyclone including the position of the cloud head, shape of the warm sector and location of the warm conveyor belt jet, and the large-scale dynamical drivers important for the rapid storm development such as the position of the storm relative to the upper-level jet exit. However, their ability to resolve the more detailed structures important for issuing weather warnings is more mixed. All of the machine learning models underestimate the peak amplitude of winds associated with the storm, only some machine learning models resolve the warm core seclusion and none of the machine learning models capture the sharp bent-back warm frontal gradient. Our study shows there is a great deal about the performance and properties of machine learning weather forecasts that can be derived from case studies of high-impact weather events such as Storm Ciarán.

DeFiTrust: A transformer-based framework for scam DeFi token detection using event logs and sentiment analysis

Decentralized finance (DeFi) token scams are rising, leading to significant losses for investors and users. Hence, it is vital to identify malicious DeFi tokens accurately and early to protect investor funds. Current Machine Learning models that detect scam tokens rely entirely on DeFi token transactional behavior and have overlooked the public opinion that can be derived from news and social media posts, which can provide helpful information for identifying scam tokens. Further, existing frameworks to detect scam tokens fail to explain why certain tokens are identified as scams, hindering stakeholder confidence in such systems. This paper proposes a new transformer-based framework called “DeFiTrust” to identify malicious DeFi tokens using transaction events and social media posts. DeFiTrust has two components: event log processing and sentiment analysis. The former creates a feature vector representing temporal variations in the event logs of a token. The latter analyzes social media posts to determine public opinion about the token. A fully-connected neural network then processes both pieces of information to determine whether the token is legitimate or a scam. We evaluated the proposed model against existing Machine Learning-based models using real-world data. The results demonstrate that DeFiTrust outperforms existing approaches and detects scam tokens earlier. Explanations generated using Integrated Gradients reveal that DeFiTrust identifies scam tokens based on valid reasons, enhancing stakeholder trust in the proposed model.

Effect of tokenization on transformers for biological sequences.

Deep-learning models are transforming biological research, including many bioinformatics and comparative genomics algorithms, such as sequence alignments, phylogenetic tree inference, and automatic classification of protein functions. Among these deep-learning algorithms, models for processing natural languages, developed in the natural language processing (NLP) community, were recently applied to biological sequences. However, biological sequences are different from natural languages, such as English, and French, in which segmentation of the text to separate words is relatively straightforward. Moreover, biological sequences are characterized by extremely long sentences, which hamper their processing by current machine-learning models, notably the transformer architecture. In NLP, one of the first processing steps is to transform the raw text to a list of tokens. Deep-learning applications to biological sequence data mostly segment proteins and DNA to single characters. In this work, we study the effect of alternative tokenization algorithms on eight different tasks in biology, from predicting the function of proteins and their stability, through nucleotide sequence alignment, to classifying proteins to specific families. We demonstrate that applying alternative tokenization algorithms can increase accuracy and at the same time, substantially reduce the input length compared to the trivial tokenizer in which each character is a token. Furthermore, applying these tokenization algorithms allows interpreting trained models, taking into account dependencies among positions. Finally, we trained these tokenizers on a large dataset of protein sequences containing more than 400 billion amino acids, which resulted in over a 3-fold decrease in the number of tokens. We then tested these tokenizers trained on large-scale data on the above specific tasks and showed that for some tasks it is highly beneficial to train database-specific tokenizers. Our study suggests that tokenizers are likely to be a critical component in future deep-network analysis of biological sequence data. Code, data, and trained tokenizers are available on https://github.com/technion-cs-nlp/BiologicalTokenizers.

Comparison of different machine learning models: Linear model, forest and SVM

Over the past century, machine learning has developed rapidly with the fast evolution of AI. As high-performance computing (HPC) has popularized over the past few decades, huge differences have been made in the field of machine learning, from expert systems to deep learning algorithms nowadays. This study summarizes some well-known models and compares three of the current popular machine learning models: the Linear Model, the Forest Model, and the Supporting Vector Machine (SVM), which aims to give and clearer view of those models in several aspects. Forest and SVM are more powerful in performance than the Linear Model due to their ability to deal with more types and more complex relationships and patterns. However, Forest and SVM have a higher complexity than the Linear Model and are less interpretable. Through the comparison of the three models, the basic model and algorithms will be better understood, the appropriate model will be selected faster when solving the problem, saving training time and space, and effectively saving resources and costs.

Applicable Machine Learning Model for Predicting Contrast-induced Nephropathy Based on Pre-catheterization Variables.

Objective Contrast agents used for radiological examinations are an important cause of acute kidney injury (AKI). We developed and validated a machine learning and clinical scoring prediction model to stratify the risk of contrast-induced nephropathy, considering the limitations of current classical and machine learning models. Methods This retrospective study included 38,481 percutaneous coronary intervention cases from 23,703 patients in a tertiary hospital. We divided the cases into development and internal test sets (8:2). Using the development set, we trained a gradient boosting machine prediction model (complex model). We then developed a simple model using seven variables based on variable importance. We validated the performance of the models using an internal test set and tested them externally in two other hospitals. Results The complex model had the best area under the receiver operating characteristic (AUROC) curve at 0.885 [95% confidence interval (CI) 0.876-0.894] in the internal test set and 0.837 (95% CI 0.819-0.854) and 0.850 (95% CI 0.781-0.918) in two different external validation sets. The simple model showed an AUROC of 0.795 (95% CI 0.781-0.808) in the internal test set and 0.766 (95% CI 0.744-0.789) and 0.782 (95% CI 0.687-0.877) in the two different external validation sets. This was higher than the value in the well-known scoring system (Mehran criteria, AUROC=0.67). The seven precatheterization variables selected for the simple model were age, known chronic kidney disease, hematocrit, troponin I, blood urea nitrogen, base excess, and N-terminal pro-brain natriuretic peptide. The simple model is available at http://52.78.230.235:8081/Conclusions We developed an AKI prediction machine learning model with reliable performance. This can aid in bedside clinical decision making.

Benchmarks for Pirá 2.0, a Reading Comprehension Dataset about the Ocean, the Brazilian Coast, and Climate Change

ABSTRACT Pirá is a reading comprehension dataset focused on the ocean, the Brazilian coast, and climate change, built from a collection of scientific abstracts and reports on these topics. This dataset represents a versatile language resource, particularly useful for testing the ability of current machine learning models to acquire expert scientific knowledge. Despite its potential, a detailed set of baselines has not yet been developed for Pirá. By creating these baselines, researchers can more easily utilize Pirá as a resource for testing machine learning models across a wide range of question answering tasks. In this paper, we define six benchmarks over the Pirá dataset, covering closed generative question answering, machine reading comprehension, information retrieval, open question answering, answer triggering, and multiple choice question answering. As part of this effort, we have also produced a curated version of the original dataset, where we fixed a number of grammar issues, repetitions, and other shortcomings. Furthermore, the dataset has been extended in several new directions, so as to face the aforementioned benchmarks: translation of supporting texts from English into Portuguese, classification labels for answerability, automatic paraphrases of questions and answers, and multiple choice candidates. The results described in this paper provide several points of reference for researchers interested in exploring the challenges provided by the Pirá dataset.

Nano1D: An accurate computer vision software for analysis and segmentation of low-dimensional nanostructures

Nanoparticles in microscopy images are usually analyzed qualitatively or manually and there is a need for autonomous quantitative analysis of these objects. In this paper, we present a physics-based computational model for accurate segmentation and geometrical analysis of one-dimensional deformable overlapping objects from microscopy images. This model, named Nano1D, has four steps of preprocessing, segmentation, separating overlapped objects and geometrical measurements. The model is tested on SEM images of Ag and Au nanowire taken from different microscopes, and thermally fragmented Ag nanowires transformed into nanoparticles with different lengths, diameters, and population densities. It successfully segments and analyzes their geometrical characteristics including lengths and average diameter. The function of the algorithm is not undermined by the size, number, density, orientation and overlapping of objects in images. The main strength of the model is shown to be its ability to segment and analyze overlapping objects successfully with more than 99 % accuracy, while current machine learning and computational models suffer from inaccuracy and inability to segment overlapping objects. Benefiting from a graphical user interface, Nano1D can analyze 1D nanoparticles including nanowires, nanotubes, nanorods in addition to other 1D features of microstructures like microcracks, dislocations etc.

Applying Machine Learning to Blood Count Data Predicts Sepsis with ICU Admission.

Timely diagnosis is crucial for sepsis treatment. Current machine learning (ML) models suffer from high complexity and limited applicability. We therefore created an ML model using only complete blood count (CBC) diagnostics. We collected non-intensive care unit (non-ICU) data from a German tertiary care centre (January 2014 to December 2021). Using patient age, sex, and CBC parameters (haemoglobin, platelets, mean corpuscular volume, white and red blood cells), we trained a boosted random forest, which predicts sepsis with ICU admission. Two external validations were conducted using data from another German tertiary care centre and the Medical Information Mart for Intensive Care IV database (MIMIC-IV). Using the subset of laboratory orders also including procalcitonin (PCT), an analogous model was trained with PCT as an additional feature. After exclusion, 1 381 358 laboratory requests (2016 from sepsis cases) were available. The CBC model shows an area under the receiver operating characteristic (AUROC) of 0.872 (95% CI, 0.857-0.887). External validations show AUROCs of 0.805 (95% CI, 0.787-0.824) for University Medicine Greifswald and 0.845 (95% CI, 0.837-0.852) for MIMIC-IV. The model including PCT revealed a significantly higher AUROC (0.857; 95% CI, 0.836-0.877) than PCT alone (0.790; 95% CI, 0.759-0.821; P < 0.001). Our results demonstrate that routine CBC results could significantly improve diagnosis of sepsis when combined with ML. The CBC model can facilitate early sepsis prediction in non-ICU patients with high robustness in external validations. Its implementation in clinical decision support systems has strong potential to provide an essential time advantage and increase patient safety.

POSEIDON: Peptidic Objects SEquence-based Interaction with cellular DOmaiNs: a new database and predictor

Cell-penetrating peptides (CPPs) are short chains of amino acids that have shown remarkable potential to cross the cell membrane and deliver coupled therapeutic cargoes into cells. Designing and testing different CPPs to target specific cells or tissues is crucial to ensure high delivery efficiency and reduced toxicity. However, in vivo/in vitro testing of various CPPs can be both time-consuming and costly, which has led to interest in computational methodologies, such as Machine Learning (ML) approaches, as faster and cheaper methods for CPP design and uptake prediction. However, most ML models developed to date focus on classification rather than regression techniques, because of the lack of informative quantitative uptake values. To address these challenges, we developed POSEIDON, an open-access and up-to-date curated database that provides experimental quantitative uptake values for over 2,300 entries and physicochemical properties of 1,315 peptides. POSEIDON also offers physicochemical properties, such as cell line, cargo, and sequence, among others. By leveraging this database along with cell line genomic features, we processed a dataset of over 1,200 entries to develop an ML regression CPP uptake predictor. Our results demonstrated that POSEIDON accurately predicted peptide cell line uptake, achieving a Pearson correlation of 0.87, Spearman correlation of 0.88, and r2 score of 0.76, on an independent test set. With its comprehensive and novel dataset, along with its potent predictive capabilities, the POSEIDON database and its associated ML predictor signify a significant leap forward in CPP research and development. The POSEIDON database and ML Predictor are available for free and with a user-friendly interface at https://moreiralab.com/resources/poseidon/, making them valuable resources for advancing research on CPP-related topics. Scientific Contribution Statement: Our research addresses the critical need for more efficient and cost-effective methodologies in Cell-Penetrating Peptide (CPP) research. We introduced POSEIDON, a comprehensive and freely accessible database that delivers quantitative uptake values for over 2,300 entries, along with detailed physicochemical profiles for 1,315 peptides. Recognizing the limitations of current Machine Learning (ML) models for CPP design, our work leveraged the rich dataset provided by POSEIDON to develop a highly accurate ML regression model for predicting CPP uptake.Graphical

An Improved Deep Learning Model for Detecting Rice Diseases

Early detection of rice plant diseases could help to quickly eradicate numerous diseases, such as fungi, viruses, and bacteria, consequently increasing rice yield. Traditional techniques for performing this task may not be the best because they take a long time, require experienced personnel, and are susceptible to a variety of infections. As a result, machine learning and deep learning approaches have recently been utilized to overcome these issues and present a more accurate model for detecting rice plant diseases. However, the current machine learning (ML) and deep learning (DL) models for this task produce unsatisfactory results due to many constraints, including high computational expenses and overfitting. To address these limitations and obtain more accurate disease detection for rice plants, we present a hybrid model of MobileNet and DNN (HMobileNetDNN). The small size of MobileNet minimizes computing costs, while the depth and complexity of DNN improve the model's capacity to capture complicated features, yielding satisfying results. Furthermore, the proposed HMobileNetDNN is also compared to four transfer learning-based DL models, namely ResNetV2, InceptionV3, MobilenetV2, and DensNet121, using the Paddy Doctor dataset. We employ several performance metrics to assess the effectiveness and efficiency of the models, like accuracy, precision, recall, F1 score, and area under the curve. The proposed model outperformed the comparing models, achieving values of 0.918, 0.918, 0.907, 0.912, and 0.949 for accuracy, precision, recall, F1 score, and area under the curve, respectively.

Real-time water quality prediction in water distribution networks using graph neural networks with sparse monitoring data

Ensuring the safety and reliability of drinking water supply requires accurate prediction of water quality in water distribution networks (WDNs). However, existing hydraulic model-based approaches for system state prediction face challenges in model calibration with limited sensor data and intensive computing requirements, while current machine learning models are lack of capacity to predict the system states at sites that are not monitored or included in model training. To address these gaps, this study proposes a novel gated graph neural network (GGNN) model for real-time water quality prediction in WDNs. The GGNN model integrates hydraulic flow directions and water quality data to represent the topology and system dynamics, and employs a masking operation for training to enhance prediction accuracy. Evaluation results from a real-world WDN demonstrate that the GGNN model is capable to achieve accurate water quality prediction across the entire WDN. Despite being trained with water quality data from a limited number of sensor sites, the model can achieve high predictive accuracies (Mean Absolute Error = 0.07 mg L−1 and Mean Absolute Percentage Error = 10.0 %) across the entire network including those unmonitored sites. Furthermore, water quality-based sensor placement significantly improves predictive accuracy, emphasizing the importance of careful sensor location selection. This research advances water quality prediction in WDNs by offering a practical and effective machine learning solution to address challenges related to limited sensor data and network complexity. This study provides a first step towards developing machine learning models to replace hydraulic models in WDN modelling.

Review of machine learning and deep learning models for toxicity prediction.

The ever-increasing number of chemicals has raised public concerns due to their adverse effects on human health and the environment. To protect public health and the environment, it is critical to assess the toxicity of these chemicals. Traditional in vitro and in vivo toxicity assays are complicated, costly, and time-consuming and may face ethical issues. These constraints raise the need for alternative methods for assessing the toxicity of chemicals. Recently, due to the advancement of machine learning algorithms and the increase in computational power, many toxicity prediction models have been developed using various machine learning and deep learning algorithms such as support vector machine, random forest, k-nearest neighbors, ensemble learning, and deep neural network. This review summarizes the machine learning- and deep learning-based toxicity prediction models developed in recent years. Support vector machine and random forest are the most popular machine learning algorithms, and hepatotoxicity, cardiotoxicity, and carcinogenicity are the frequently modeled toxicity endpoints in predictive toxicology. It is known that datasets impact model performance. The quality of datasets used in the development of toxicity prediction models using machine learning and deep learning is vital to the performance of the developed models. The different toxicity assignments for the same chemicals among different datasets of the same type of toxicity have been observed, indicating benchmarking datasets is needed for developing reliable toxicity prediction models using machine learning and deep learning algorithms. This review provides insights into current machine learning models in predictive toxicology, which are expected to promote the development and application of toxicity prediction models in the future.

Machine Learning Models for Blood Glucose Level Prediction in Patients With Diabetes Mellitus: Systematic Review and Network Meta-Analysis.

Machine learning (ML) models provide more choices to patients with diabetes mellitus (DM) to more properly manage blood glucose (BG) levels. However, because of numerous types of ML algorithms, choosing an appropriate model is vitally important. In a systematic review and network meta-analysis, this study aimed to comprehensively assess the performance of ML models in predicting BG levels. In addition, we assessed ML models used to detect and predict adverse BG (hypoglycemia) events by calculating pooled estimates of sensitivity and specificity. PubMed, Embase, Web of Science, and Institute of Electrical and Electronics Engineers Explore databases were systematically searched for studies on predicting BG levels and predicting or detecting adverse BG events using ML models, from inception to November 2022. Studies that assessed the performance of different ML models in predicting or detecting BG levels or adverse BG events of patients with DM were included. Studies with no derivation or performance metrics of ML models were excluded. The Quality Assessment of Diagnostic Accuracy Studies tool was applied to assess the quality of included studies. Primary outcomes were the relative ranking of ML models for predicting BG levels in different prediction horizons (PHs) and pooled estimates of the sensitivity and specificity of ML models in detecting or predicting adverse BG events. In total, 46 eligible studies were included for meta-analysis. Regarding ML models for predicting BG levels, the means of the absolute root mean square error (RMSE) in a PH of 15, 30, 45, and 60 minutes were 18.88 (SD 19.71), 21.40 (SD 12.56), 21.27 (SD 5.17), and 30.01 (SD 7.23) mg/dL, respectively. The neural network model (NNM) showed the highest relative performance in different PHs. Furthermore, the pooled estimates of the positive likelihood ratio and the negative likelihood ratio of ML models were 8.3 (95% CI 5.7-12.0) and 0.31 (95% CI 0.22-0.44), respectively, for predicting hypoglycemia and 2.4 (95% CI 1.6-3.7) and 0.37 (95% CI 0.29-0.46), respectively, for detecting hypoglycemia. Statistically significant high heterogeneity was detected in all subgroups, with different sources of heterogeneity. For predicting precise BG levels, the RMSE increases with a rise in the PH, and the NNM shows the highest relative performance among all the ML models. Meanwhile, current ML models have sufficient ability to predict adverse BG events, while their ability to detect adverse BG events needs to be enhanced. PROSPERO CRD42022375250; https://www.crd.york.ac.uk/prospero/display_record.php?RecordID=375250.

Prediction of multiclass surgical outcomes in glaucoma using multimodal deep learning based on free-text operative notes and structured EHR data.

Surgical outcome prediction is challenging but necessary for postoperative management. Current machine learning models utilize pre- and post-op data, excluding intraoperative information in surgical notes. Current models also usually predict binary outcomes even when surgeries have multiple outcomes that require different postoperative management. This study addresses these gaps by incorporating intraoperative information into multimodal models for multiclass glaucoma surgery outcome prediction. We developed and evaluated multimodal deep learning models for multiclass glaucoma trabeculectomy surgery outcomes using both structured EHR data and free-text operative notes. We compare those to baseline models that use structured EHR data exclusively, or neural network models that leverage only operative notes. The multimodal neural network had the highest performance with a macro AUROC of 0.750 and F1 score of 0.583. It outperformed the baseline machine learning model with structured EHR data alone (macro AUROC of 0.712 and F1 score of 0.486). Additionally, the multimodal model achieved the highest recall (0.692) for hypotony surgical failure, while the surgical success group had the highest precision (0.884) and F1 score (0.775). This study shows that operative notes are an important source of predictive information. The multimodal predictive model combining perioperative notes and structured pre- and post-op EHR data outperformed other models. Multiclass surgical outcome prediction can provide valuable insights for clinical decision-making. Our results show the potential of deep learning models to enhance clinical decision-making for postoperative management. They can be applied to other specialties to improve surgical outcome predictions.