Accurate Molecular Dynamics Enabled by Efficient Physically Constrained Machine Learning Approaches

Molecular dynamics (MD) simulations employing classical force fields constitute the cornerstone of contemporary atomistic modeling in chemistry, biology, and materials science. However, the predictive power of these simulations is only as good as the underlying interatomic potential. Classical potentials often fail to faithfully capture key quantum effects in molecules and materials. Here we enable the direct construction of flexible molecular force fields from high-level ab initio calculations by incorporating spatial and temporal physical symmetries into a gradient-domain machine learning (sGDML) model in an automatic data-driven way. The developed sGDML approach faithfully reproduces global force fields at quantum-chemical CCSD(T) level of accuracy and allows converged molecular dynamics simulations with fully quantized electrons and nuclei. We present MD simulations, for flexible molecules with up to a few dozen atoms and provide insights into the dynamical behavior of these molecules. Our approach provides the key missing ingredient for achieving spectroscopic accuracy in molecular simulations.

Classification and prediction tasks are common in health research. With the increasing availability of vast health data repositories (e.g. electronic medical record databases) and advances in computing power, traditional statistical approaches are being augmented or replaced with machine learning (ML) approaches to classify and predict health outcomes. ML describes the automated process of identifying ("learning") patterns in data to perform tasks. Developing an ML model includes selecting between many ML models (e.g. decision trees, support vector machines, neural networks); model specifications such as hyperparameter tuning; and evaluation of model performance. This process is conducted repeatedly to find the model and corresponding specifications that optimize some measure of model performance. ML models can make more accurate classifications and predictions than their statistical counterparts and confer greater flexibility when modelling unstructured data or interactions between covariates; however, many ML models require larger sample sizes to achieve good classification or predictive performance and have been criticized as "black box" for their poor transparency and interpretability. ML holds potential in family medicine for risk profiling of patients' disease risk and clinical decision support to present additional information at times of uncertainty or high demand. In the future, ML approaches are positioned to become commonplace in family medicine. As such, it is important to understand the objectives that can be addressed using ML approaches and the associated techniques and limitations. This article provides a brief introduction into the use of ML approaches for classification and prediction tasks in family medicine.

Summary A meticulous interpretation of steady-state or unsteady-state relative permeability (Kr) experimental data is required to determine a complete set of Kr curves. In this work, different machine learning (ML) models were developed to assist in a faster estimation of these curves from steady-state drainage coreflooding experimental runs. These ML algorithms include gradient boosting (GB), random forest (RF), extreme gradient boosting (XGB), and deep neural network (DNN) with a main focus on and comparison of the two latter algorithms (XGB and DNN). Based on existing mathematical models, a leading-edge framework was developed where a large database of Kr and capillary pressure (Pc) curves were generated. This database was used to perform thousands of coreflood simulation runs representing oil-water drainage steady-state experiments. The results obtained from these simulation runs, mainly pressure drop along with other conventional core analysis data, were used to estimate analytical Kr curves based on Darcy’s law. These analytically estimated Kr curves along with the previously generated Pc curves were fed as features into the ML model. The entire data set was split into 80% for training and 20% for testing. The k-fold cross-validation technique was applied to increase the model’s accuracy by splitting 80% of the training data into 10 folds. In this manner, for each of the 10 experiments, nine folds were used for training and the remaining fold was used for model validation. Once the model was trained and validated, it was subjected to blind testing on the remaining 20% of the data set. The ML model learns to capture fluid flow behavior inside the core from the training data set. In terms of applicability of these ML models, two sets of experimental data were needed as input; the first was the analytically estimated Kr curves from the steady-state drainage coreflooding experiments, while the other was the Pc curves estimated from centrifuge or mercury injection capillary pressure (MICP) measurements. The trained/tested model was then able to estimate Kr curves based on the experimental results fed as input. Furthermore, to test the performance of the ML model when only one set of experimental data is available to an end user, a recurrent neural network (RNN) algorithm was trained/tested to predict Kr curves in the absence of Pc curves as an input. The performance of the three developed models (XGB, DNN, and RNN) was assessed using the values of the coefficient of determination (R2) along with the loss calculated during training/validation of the model. The respective crossplots along with comparisons of ground truth vs. artificial intelligence (AI)-predicted curves indicated that the model is capable of making accurate predictions with an error percentage between 0.2% and 0.6% on history-matching experimental data for all three tested ML techniques. This implies that the AI-based model exhibits better efficiency and reliability in determining Kr curves when compared to conventional methods. The developed ML models by no means replace the need to conduct drainage coreflooding or centrifuge experiments but act as an alternative to existing commercial platforms that are used to interpret experimental data to predict Kr curves. The two main advantages of the developed ML models are their capability of predicting Kr curves within a matter of a few minutes as well as with limited intervention from the end user. The results also include a comparison between classical ML approaches, shallow neural networks, and DNNs in terms of accuracy in predicting the final Kr curves. The research presented here is an extension of the state-of-the-art framework proposed by Mathew et al. (2021). However, the two main aspects of the current study are the application of deep learning for the prediction of Kr curves and the application of feature engineering. The latter not only reduces the training/testing time for the ML models but also enables the end user to obtain the final predictions with the least set of experimental data. The various models discussed in this research work currently focus on the prediction of Kr curves for drainage steady-state experiments; however, the work can be extended to capture the imbibition cycle as well.

Preventing the expansion of perihematomal edema (PHE) represents a novel strategy for the improvement of neurological outcomes in intracerebral hemorrhage (ICH) patients. Our goal was to predict early and delayed PHE expansion using a machine learning approach. We enrolled 550 patients with spontaneous ICH to study early PHE expansion, and 389 patients to study delayed expansion. Two imaging researchers rated the shape and density of hematoma in non-contrast computed tomography (NCCT). We trained a radiological machine learning (ML) model, a radiomics ML model, and a combined ML model, using data from radiomics, traditional imaging, and clinical indicators. We then validated these models on an independent dataset by using a nested 4-fold cross-validation approach. We compared models with respect to their predictive performance, which was assessed using the receiver operating characteristic curve. For both early and delayed PHE expansion, the combined ML model was most predictive (early/delayed AUC values were 0.840/0.705), followed by the radiomics ML model (0.799/0.663), the radiological ML model (0.779/0.631), and the imaging readers (reader 1: 0.668/0.565, reader 2: 0.700/0.617). We validated a machine learning approach with high interpretability for the prediction of early and delayed PHE expansion. This new technique may assist clinical practice for the management of neurocritical patients with ICH. • This is the first study to use artificial intelligence technology for the prediction of perihematomal edema expansion. • A combined machine learning model, trained on data from radiomics, clinical indicators, and imaging features associated with hematoma expansion, outperformed all other methods.

Commentary: Machine learning and cardiac surgery risk prediction

BackgroundPhysical activity (PA) level is associated with multiple health benefits during early childhood. However, inconsistency in the methods for quantification of PA levels among preschoolers remains a problem.ObjectiveThis study aimed to develop PA intensity cut points for wrist-worn accelerometers by using machine learning (ML) approaches to assess PA in preschoolers.MethodsWrist- and hip-derived acceleration data were collected simultaneously from 34 preschoolers on 3 consecutive preschool days. Two supervised ML models, receiver operating characteristic curve (ROC) and ordinal logistic regression (OLR), and one unsupervised ML model, k-means cluster analysis, were applied to establish wrist-worn accelerometer vector magnitude (VM) cut points to classify accelerometer counts into sedentary behavior, light PA (LPA), moderate PA (MPA), and vigorous PA (VPA). Physical activity intensity levels identified by hip-worn accelerometer VM cut points were used as reference to train the supervised ML models. Vector magnitude counts were classified by intensity based on three newly established wrist methods and the hip reference to examine classification accuracy. Daily estimates of PA were compared to the hip-reference criterion.ResultsIn total, 3600 epochs with matched hip- and wrist-worn accelerometer VM counts were analyzed. All ML approaches performed differently on developing PA intensity cut points for wrist-worn accelerometers. Among the three ML models, k-means cluster analysis derived the following cut points: ≤2556 counts per minute (cpm) for sedentary behavior, 2557-7064 cpm for LPA, 7065-14532 cpm for MPA, and ≥14533 cpm for VPA; in addition, k-means cluster analysis had the highest classification accuracy, with more than 70% of the total epochs being classified into the correct PA categories, as examined by the hip reference. Additionally, k-means cut points exhibited the most accurate estimates on sedentary behavior, LPA, and VPA as the hip reference. None of the three wrist methods were able to accurately assess MPA.ConclusionsThis study demonstrates the potential of ML approaches in establishing cut points for wrist-worn accelerometers to assess PA in preschoolers. However, the findings from this study warrant additional validation studies.

Background: Acute coronary syndrome (ACS) remains one of the leading causes of death globally. Accurate and reliable mortality risk prediction of ACS patients is essential for developing targeted treatment strategies. Existing traditional models offer moderate discriminative value, and do not incorporate contemporary predictors of ACS prognosis. Machine learning (ML) models have emerged as an alternate method that may offer improved risk assessment. This study aims to compare machine learning models with traditional risk scores for predicting all-cause mortality in patients with ACS. Methods: PubMed, EMBASE, Web of Science and Cochrane databases were searched until 1st November 2023 for studies comparing ML models with traditional statistical methods for event prediction of ACS patients. The primary outcome was comparative discrimination measured by C-statistics with 95% confidence intervals between ML models and traditional methods in estimating the risk of all-cause mortality. Results: Ten studies were included (239627 patients). The summary C-statistic of all ML models across all endpoints was 0.89 (95% CI, 0.86-0.92), compared to traditional methods 0.82 (95% CI, 0.79-0.85). The difference in C-statistic between all ML models and traditional methods was 0.07 (p<0.05). Three models undertook external validation, and calibration was inconsistently reported. Conclusion: ML models demonstrated superior discrimination of all-cause mortality for ACS patients compared to traditional risk scores. Despite outperforming well-established prognostic tools such as the GRACE and TIMI scores, current clinical applications of ML approaches remain uncertain, particularly in view of the methodological limitations of existing studies and the need for greater model validation.

In-hospital cardiac arrest (IHCA) in the emergency department (ED) is not uncommon but often fatal. Using the machine learning (ML) approach, we sought to predict ED-based IHCA (EDCA) in patients presenting to the ED based on triage data. We retrieved 733,398 ED records from a tertiary teaching hospital over a 7year period (Jan. 1, 2009-Dec. 31, 2015). We included only adult patients (≥ 18y) and excluded cases presenting as out-of-hospital cardiac arrest. Primary outcome (EDCA) was identified via a resuscitation code. Patient demographics, triage data, and structured chief complaints (CCs), were extracted. Stratified split was used to divide the dataset into the training and testing cohort at a 3-to-1 ratio. Three supervised ML models were trained and performances were evaluated and compared to the National Early Warning Score 2 (NEWS2) and logistic regression (LR) model by the area under the receiver operating characteristic curve (AUC). We included 316,465 adult ED records for analysis. Of them, 636 (0.2%) developed EDCA. Of the constructed ML models, Random Forest outperformed the others with the best AUC result (0.931, 95% CI 0.911-0.949), followed by Gradient Boosting (0.930, 95% CI 0.909-0.948) and Extra Trees classifier (0.915, 95% CI 0.892-0.936). Although the differences between each of ML models and LR (AUC: 0.905, 95% CI 0.882-0.926) were not significant, all constructed ML models performed significantly better than using the NEWS2 scoring system (AUC 0.678, 95% CI 0.635-0.722). Our ML models showed excellent discriminatory performance to identify EDCA based only on the triage information. This ML approach has the potential to reduce unexpected resuscitation events if successfully implemented in the ED information system.

Chemical heuristics have been fundamental to the advancement of chemistry and materials science. These heuristics are typically established by scientists using knowledge and creativity to extract patterns from limited datasets. Machine learning offers opportunities to perfect this approach using computers and larger datasets. Here, we discuss the relationships between traditional heuristics and machine learning approaches. We show how traditional rules can be challenged by large-scale statistical assessment and how traditional concepts commonly used as features are feeding the machine learning techniques. We stress the waste involved in relearning chemical rules and the challenges in terms of data size requirements for purely data-driven approaches. Our view is that heuristic and machine learning approaches are at their best when they work together.

In this paper, we describe a machine learning (ML) approach for estimating interface orientation in multiphase flow using the volume of fluid (VOF) method on a uniform Cartesian mesh. By using complex shapes generated with the parametric radial star formula during training, we significantly improve prediction accuracy without increasing the network's structural complexity or processing cost. Our key contribution is the development of a robust ML model capable of reliably predicting interface orientation angles on uniform Cartesian grids. To enhance performance and robustness, we conducted a parametric study of the ML models' hyperparameters. The proposed method produced two models: a full augmented (9-cell) stencil and a compact (5-cell) stencil. Both were compared against popular finite difference/volume methods commonly used in VOF schemes. The results show that our approach is more accurate while remaining computationally efficient, particularly when employing a small stencil. Numerical studies, including challenging flow scenarios, demonstrate that the technique can predict interface orientation with an absolute mean error of less than 1 degree. Implemented in the OpenFOAM isoAdvector, our technique reliably produces accurate interface tracking with minimal deviation from the exact solution. These findings highlight the potential for incorporating machine learning approaches into classical numerical methods to improve the accuracy and reliability of the VOF method in a variety of challenging applications using uniform Cartesian meshes.

Introduction Spine trauma represents a major cause of disability and substantial healthcare costs. Despite classification systems and risk scores, outcome prediction following spine trauma remains challenging. Machine learning (ML) approaches offer multidimensionality but vary widely in methodology, performance, and clinical application. This systematic review aims to characterize the use of ML for outcome prediction and prognostication in spine trauma. Methods A systematic search of PubMed, Embase, Scopus, Web of Science, and Cochrane Library was performed for publications from January 2015 to May 2025. Studies reporting ML for predicting clinical outcomes in spine trauma patients were identified using pre-defined criteria. Primary outcomes were types and performance metrics of ML models for spine trauma prognostication. Secondary outcomes included cost-effectiveness, clinical utility and implementation feasibility. Results Twenty-seven studies were included, of which 26 were published from 2020 onwards. Only one study was prospective, while the rest were retrospective. Commonly used ML algorithms for spine trauma prognosis were logistic regression (LR), extreme gradient boosting (XGBoost), Random Forest (RF), light gradient boosting machine (LightGBM), and support vector machine (SVM). Outcomes predicted include mortality, neurological function or recovery, ambulation and walking, self-care and independence, complications, treatment failure, length of stay, and discharge destination. Most studies employed internal testing and validation; only four reported external validation. Seven studies reported direct comparison with non-ML models or risk scores, demonstrating comparable or superior performance of ML models. Performance metrics include area under the curve (AUC), accuracy, sensitivity, specificity, precision, F1-score, Brier score, and others. The highest AUC was reported by a nomogram constructed using consensus clustering to predict postoperative pulmonary infection in patients with acute cervical spinal cord injury (internal validation; AUC 0.993; 95% CI 0.981–1.000). Quality of evidence varied between studies, but risk of bias was a concern. Conclusions Existing evidence on ML for spine trauma prognosis is characterized by retrospective study designs, homogeneous cohorts, and lack of external validation, limiting their generalizability and real-world application. While ML approaches demonstrate encouraging predictive performance for spine trauma outcomes, their clinical adoption is limited by methodological heterogeneity. More robust validation, standardized reporting, and comparison with traditional tools are needed before ML models can reliably inform clinical decision-making.

Background This study aimed to establish and validate a machine learning (ML) model for predicting in-hospital mortality in critically ill patients with chronic kidney disease (CKD). Methods This study collected data on CKD patients from 2008 to 2019 using the Medical Information Mart for Intensive Care IV. Six ML approaches were used to build the model. Accuracy and area under the curve (AUC) were used to choose the best model. In addition, the best model was interpreted using SHapley Additive exPlanations (SHAP) values. Results There were 8527 CKD patients eligible for participation; the median age was 75.1 (interquartile range: 65.0–83.5) years, and 61.7% (5259/8527) were male. We developed six ML models with clinical variables as input factors. Among the six models developed, the eXtreme Gradient Boosting (XGBoost) model had the highest AUC, at 0.860. According to the SHAP values, the sequential organ failure assessment score, urine output, respiratory rate, and simplified acute physiology score II were the four most influential variables in the XGBoost model. Conclusions In conclusion, we successfully developed and validated ML models for predicting mortality in critically ill patients with CKD. Among all ML models, the XGBoost model is the most effective ML model that can help clinicians accurately manage and implement early interventions, which may reduce mortality in critically ill CKD patients with a high risk of death.

Efficient fertilizer application is vital for enhancing maize production and profitability in Sub-Saharan Africa, where soil fertility varies widely across regions. This study aimed to develop a machine learning approach for generating site-specific fertilizer recommendations for maize production in Ghana and to evaluate its performance against conventional and semi-mechanistic approaches. A random forest machine learning model was trained on 482 maize yield experiments, consisting of 3136 yield observations collected from 1991 to 2020, to predict maize yield response to different fertilizer rates. The model incorporated multiple explanatory variables, including soil properties, climate conditions, and management practices, to generate fertilizer response curves from which fertilizer recommendations were derived for 14 sites across three agro-ecological zones in Ghana where field validation experiments were conducted. On these sites, the recommendations were compared with recommendations derived from the Quantitative Evaluation of the Fertility of Tropical Soils (QUEFTS), Conventional Fertilizer Dose Response (CFDR), and Updated Conventional Fertilizer Dose Response (UCFDR) approaches and validated through field experiments. The machine learning approach generally recommended lower rates of phosphorus and potassium than the other approaches, while nitrogen recommendations were comparable. In the Guinea Savanna zone, the recommendations from the machine learning approach outperformed those from the other approaches, producing higher mean yields for three out of the four sites in the zone. In the Forest-Savanna Transition (FST) zone, the machine learning model recommendations led to higher mean yields at four sites, while the approaches based on QUEFTS and UCFDR performed best at two other sites. In the Semi-deciduous Forest zone, the recommendations of the QUEFTS approach resulted in the highest mean yields at three sites, and CFDR at one site. Despite high input prices during the period of experimentation, the machine learning approach-based recommendations demonstrated higher net profit margins in the FST zone, suggesting cost-effectiveness in this zone. These findings indicate that site-specific fertilizer recommendations are more efficient than blanket recommendations and that machine learning approaches offer a promising and innovative approach for generating cost-effective, site-specific fertilizer recommendations in tropical climates. • ML and QUEFTS approaches provided site-specific fertilizer recommendations. • ML and QUEFTS outperformed conventional fertilizer recommendation approaches. • CFDR and UCFDR were constrained by generalized, non-site-specific recommendations. • Machine learning supports sustainable fertilizer use.

There is limited understanding of the link between exposure to heavy metals and ischemic stroke (IS). This research aimed to develop efficient and interpretable machine learning (ML) models to associate the relationship between exposure to heavy metals and IS. The data of this research were obtained from the National Health and Nutrition Examination Survey (US NHANES, 2003-2018) database. Seven ML models were used to identify IS caused by exposure to heavy metals. To assess the strength of the models, we employed 10-fold cross-validation, the area under the curve (AUC), F1 scores, Brier scores, Matthews correlation coefficient (MCC), precision-recall (PR) curves, and decision curve analysis (DCA) curves. Following these tests, the best-performing model was selected. Finally, the DALEX package was used for feature explanation and decision-making visualization. A total of 15,575 participants were involved in this study. The best-performing ML models, which included logistic regression (LR) (AUC: 0.796) and XGBoost (AUC: 0.789), were selected. The DALEX package revealed that age, total mercury in blood, poverty-to-income ratio (PIR), and cadmium were the most significant contributors to IS in the logistic regression and XGBoost models. The logistic regression and XGBoost models showed high efficiency, accuracy, and robustness in identifying associations between heavy metal exposure and IS in NHANES 2003-2018 participants.

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.