Development Of Machine Learning Models Research Articles

Development of interpretable machine learning models for prediction of acute kidney injury after noncardiac surgery: a retrospective cohort study.

Early identification of patients at high-risk of postoperative acute kidney injury (AKI) can facilitate the development of preventive approaches. This study aimed to develop prediction models for postoperative AKI in noncardiac surgery using machine learning algorithms. The authors also evaluated the predictive performance of models that included only preoperative variables or only important predictors. Adult patients undergoing noncardiac surgery were retrospectively included in the study (76457 patients in the discovery cohort and 11910 patients in the validation cohort). AKI was determined using the KDIGO criteria. The prediction model was developed using 87 variables (56 preoperative variables and 31 intraoperative variables). A variety of machine learning algorithms were employed to develop the model, including logistic regression, random forest, extreme gradient boosting, and gradient boosting decision trees. The performance of different models was compared using the area under the receiver operating characteristic curve (AUROC). Shapley Additive Explanations (SHAP) analysis was employed for model interpretation. The patients in the discovery cohort had a median age of 52 years (IQR: 42-61 years), and 1179 patients (1.5%) developed AKI after surgery. The gradient boosting decision trees algorithm showed the best predictive performance using all available variables, or only preoperative variables. The AUROCs were 0.849 (95% CI: 0.835-0.863) and 0.828 (95% CI: 0.813-0.843), respectively. The SHAP analysis showed that age, surgical duration, preoperative serum creatinine, and gamma-glutamyltransferase, as well as American Society of Anesthesiologists physical status III were the most important five features. When gradually reducing the features, the AUROCs decreased from 0.852 (including the top 40 features) to 0.839 (including the top 10 features). In the validation cohort, the authors observed a similar pattern regarding the models' predictive performance. The machine learning models the authors developed had satisfactory predictive performance for identifying high-risk postoperative AKI patients. Furthermore, the authors found that model performance was only slightly affected when only preoperative variables or only the most important predictive features were included.

Studying the impact of marital status on diagnosis and survival prediction in pancreatic ductal carcinoma using machine learning methods

Pancreatic cancer is a commonly occurring malignant tumor, with pancreatic ductal carcinoma (PDAC) accounting for approximately 95% of cases. According of its poor prognosis, identifying prognostic factors of pancreatic ductal carcinoma can provide physicians with a reliable theoretical foundation when predicting patient survival. This study aimed to analyze the impact of marital status on survival outcomes of PDAC patients using propensity score matching and machine learning. The goal was to develop a prognosis prediction model specific to married patients with PDAC. We extracted a total of 206,968 patient records of pancreatic cancer from the SEER database. To ensure the baseline characteristics of married and unmarried individuals were balanced, we used a 1:1 propensity matching score. We then conducted Kaplan–Meier analysis and Cox proportional-hazards regression to examine the impact of marital status on PDAC survival before and after matching. Additionally, we developed machine learning models to predict 5-year CSS and OS for married patients with PDAC specifically. In total, 24,044 PDAC patients were included in this study. After 1:1 propensity matching, 8043 married patients and 8,043 unmarried patients were successfully enrolled. Multivariate analysis and the Kaplan–Meier curves demonstrated that unmarried individuals had a poorer survival rate than their married counterparts. Among the algorithms tested, the random forest performed the best, with 0.734 5-year CSS and 0.795 5-year OS AUC. This study found a significant association between marital status and survival in PDAC patients. Married patients had the best prognosis, while widowed patients had the worst. The random forest is a reliable model for predicting survival in married patients with PDAC.

An empirical investigation of challenges of specifying training data and runtime monitors for critical software with machine learning and their relation to architectural decisions

The development and operation of critical software that contains machine learning (ML) models requires diligence and established processes. Especially the training data used during the development of ML models have major influences on the later behaviour of the system. Runtime monitors are used to provide guarantees for that behaviour. Runtime monitors for example check that the data at runtime is compatible with the data used to train the model. In a first step towards identifying challenges when specifying requirements for training data and runtime monitors, we conducted and thematically analysed ten interviews with practitioners who develop ML models for critical applications in the automotive industry. We identified 17 themes describing the challenges and classified them in six challenge groups. In a second step, we found interconnection between the challenge themes through an additional semantic analysis of the interviews. We explored how the identified challenge themes and their interconnections can be mapped to different architecture views. This step involved identifying relevant architecture views such as data, context, hardware, AI model, and functional safety views that can address the identified challenges. The article presents a list of the identified underlying challenges, identified relations between the challenges and a mapping to architecture views. The intention of this work is to highlight once more that requirement specifications and system architecture are interlinked, even for AI-specific specification challenges such as specifying requirements for training data and runtime monitoring.

Variability and influence of fisheries acoustic echogram annotations on machine learning applications

High-frequency echosounders are the workhorse in fisheries and marine ecological surveys. Due to the inherent complexity of biological aggregations and ambiguity in interpreting echoes from species of similar size and anatomical compositions, echogram annotation typically requires combining spectral information referencing scattering physics, biological ground-truth from nearby net-trawls, and empirical school morphology of specific fish species. Here, we investigate the variability of echogram annotations and its influence on machine learning applications using data from the biennial Pacific hake acoustic-trawl survey. Compared to many other fish species, hake tend to possess less defined school boundaries with variable acoustic features and often form mixed-species aggregations in the mesopelagic. Nonnegative matrix factorization and hierarchical clustering of volume backscattering strength (Sv) distributions across the 18, 38, and 120 kHz channels revealed a spectrum of annotation region types that reflect differences in morphological and acoustic features as well as differences in annotator style. This variability likely contributes to the observed variable segmentation behavior of deep learning models trained using this dataset. These results highlight the importance of considering the diversity of echogram annotation, its connection to scattering physics and the underlying aggregation composition, and the incorporation of such information in developing machine learning models.

Leveraging Machine Learning to Develop Digital Engagement Phenotypes of Users in a Digital Diabetes Prevention Program: Evaluation Study.

Digital diabetes prevention programs (dDPPs) are effective "digital prescriptions" but have high attrition rates and program noncompletion. To address this, we developed a personalized automatic messaging system (PAMS) that leverages SMS text messaging and data integration into clinical workflows to increase dDPP engagement via enhanced patient-provider communication. Preliminary data showed positive results. However, further investigation is needed to determine how to optimize the tailoring of support technology such as PAMS based on a user's preferences to boost their dDPP engagement. This study evaluates leveraging machine learning (ML) to develop digital engagement phenotypes of dDPP users and assess ML's accuracy in predicting engagement with dDPP activities. This research will be used in a PAMS optimization process to improve PAMS personalization by incorporating engagement prediction and digital phenotyping. This study aims (1) to prove the feasibility of using dDPP user-collected data to build an ML model that predicts engagement and contributes to identifying digital engagement phenotypes, (2) to describe methods for developing ML models with dDPP data sets and present preliminary results, and (3) to present preliminary data on user profiling based on ML model outputs. Using the gradient-boosted forest model, we predicted engagement in 4 dDPP individual activities (physical activity, lessons, social activity, and weigh-ins) and general activity (engagement in any activity) based on previous short- and long-term activity in the app. The area under the receiver operating characteristic curve, the area under the precision-recall curve, and the Brier score metrics determined the performance of the model. Shapley values reflected the feature importance of the models and determined what variables informed user profiling through latent profile analysis. We developed 2 models using weekly and daily DPP data sets (328,821 and 704,242 records, respectively), which yielded predictive accuracies above 90%. Although both models were highly accurate, the daily model better fitted our research plan because it predicted daily changes in individual activities, which was crucial for creating the "digital phenotypes." To better understand the variables contributing to the model predictor, we calculated the Shapley values for both models to identify the features with the highest contribution to model fit; engagement with any activity in the dDPP in the last 7 days had the most predictive power. We profiled users with latent profile analysis after 2 weeks of engagement (Bayesian information criterion=-3222.46) with the dDPP and identified 6 profiles of users, including those with high engagement, minimal engagement, and attrition. Preliminary results demonstrate that applying ML methods with predicting power is an acceptable mechanism to tailor and optimize messaging interventions to support patient engagement and adherence to digital prescriptions. The results enable future optimization of our existing messaging platform and expansion of this methodology to other clinical domains. ClinicalTrials.gov NCT04773834; https://www.clinicaltrials.gov/ct2/show/NCT04773834. RR2-10.2196/26750.

Exploring the viability of AI-aided genetic algorithms in estimating the crack repair rate of self-healing concrete

Abstract As a potential replacement for traditional concrete, which has cracking and poor durability issues, self-healing concrete (SHC) has been the research subject. However, conducting lab trials can be expensive and time-consuming. Therefore, machine learning (ML)-based predictions can aid improved formulations of self-healing concrete. The aim of this work is to develop ML models that could analyze and forecast the rate of healing of the cracked area (CrA) of bacteria- and fiber-containing SHC. These models were constructed using gene expression programming (GEP) and multi-expression programming (MEP) tools. The discrepancy between expected and desired results, statistical tests, Taylor’s diagram, and R 2 values were additional metrics used to assess the constructed models. A SHapley Additive exPlanations (SHAP) approach was used to evaluate which input attributes were highly relevant. With R 2 = 0.93, MAE = 0.047, MAPE = 12.60%, and RMSE = 0.062, the GEP produced somewhat worse predictions than the MEP (R 2 = 0.93, MAE = 0.033, MAPE = 9.60%, and RMSE = 0.044). Bacteria had an indirect (negative) relationship with the CrA of SHC, while fiber had a direct (positive) association, according to the SHAP study. The SHAP study might help researchers and companies figure out how much of each raw material is needed for SHCs. Therefore, MEP and GEP models can be used to generate and test SHC compositions based on bacteria and polymeric fibers.

Individualized estimation of arterial carbon dioxide partial pressure using machine learning in children receiving mechanical ventilation

BackgroundMeasuring arterial partial pressure of carbon dioxide (PaCO2) is crucial for proper mechanical ventilation, but the current sampling method is invasive. End-tidal carbon dioxide (EtCO2) has been used as a surrogate, which can be measured non-invasively, but its limited accuracy is due to ventilation-perfusion mismatch. This study aimed to develop a non-invasive PaCO2 estimation model using machine learning.MethodsThis retrospective observational study included pediatric patients (< 18 years) admitted to the pediatric intensive care unit of a tertiary children’s hospital and received mechanical ventilation between January 2021 and June 2022. Clinical information, including mechanical ventilation parameters and laboratory test results, was used for machine learning. Linear regression, multilayer perceptron, and extreme gradient boosting were implemented. The dataset was divided into 7:3 ratios for training and testing. Model performance was assessed using the R2 value.ResultsWe analyzed total 2,427 measurements from 32 patients. The median (interquartile range) age was 16 (12−19.5) months, and 74.1% were female. The PaCO2 and EtCO2 were 63 (50−83) mmHg and 43 (35−54) mmHg, respectively. A significant discrepancy of 19 (12–31) mmHg existed between EtCO2 and the measured PaCO2. The R2 coefficient of determination for the developed models was 0.799 for the linear regression model, 0.851 for the multilayer perceptron model, and 0.877 for the extreme gradient boosting model. The correlations with PaCO2 were higher in all three models compared to EtCO2.ConclusionsWe developed machine learning models to non-invasively estimate PaCO2 in pediatric patients receiving mechanical ventilation, demonstrating acceptable performance. Further research is needed to improve reliability and external validation.

Systematic analysis of off-label and off-guideline cancer therapy usage in a real-world cohort of 165,912 US patients

Patients with cancer may be given treatments that are not officially approved (off-label) or recommended by guidelines (off-guideline). Here we present a data science framework to systematically characterize off-label and off-guideline usages using real-world data from de-identified electronic health records (EHR). We analyze treatment patterns in 165,912US patients with 14 common cancer types. We find that 18.6% and 4.4% of patients have received at least one line of off-label and off-guideline cancer drugs, respectively. Patients with worse performance status, in later lines, or treated at academic hospitals are significantly more likely to receive off-label and off-guideline drugs. To quantify how predictable off-guideline usage is, we developed machine learning models to predict which drug a patient is likely to receive based on their clinical characteristics and previous treatments. Finally, we demonstrate that our systematic analyses generate hypotheses about patients' response to treatments.

Descriptor-augmented machine learning for enzyme-chemical interaction predictions

Descriptors play a pivotal role in enzyme design for the greener synthesis of biochemicals, as they could characterize enzymes and chemicals from the physicochemical and evolutionary perspective. This study examined the effects of various descriptors on the performance of Random Forest model used for enzyme-chemical relationships prediction. We curated activity data of seven specific enzyme families from the literature and developed the pipeline for evaluation the machine learning model performance using 10-fold cross-validation. The influence of protein and chemical descriptors was assessed in three scenarios, which were predicting the activity of unknown relations between known enzymes and known chemicals (new relationship evaluation), predicting the activity of novel enzymes on known chemicals (new enzyme evaluation), and predicting the activity of new chemicals on known enzymes (new chemical evaluation). The results showed that protein descriptors significantly enhanced the classification performance of model on new enzyme evaluation in three out of the seven datasets with the greatest number of enzymes, whereas chemical descriptors appear no effect. A variety of sequence-based and structure-based protein descriptors were constructed, among which the esm-2 descriptor achieved the best results. Using enzyme families as labels showed that descriptors could cluster proteins well, which could explain the contributions of descriptors to the machine learning model. As a counterpart, in the new chemical evaluation, chemical descriptors made significant improvement in four out of the seven datasets, while protein descriptors appear no effect. We attempted to evaluate the generalization ability of the model by correlating the statistics of the datasets with the performance of the models. The results showed that datasets with higher sequence similarity were more likely to get better results in the new enzyme evaluation and datasets with more enzymes were more likely beneficial from the protein descriptor strategy. This work provides guidance for the development of machine learning models for specific enzyme families.

Using Convolutional Neural Networks for Blocking Prediction in Elastic Optical Networks

This paper presents a study on connection-blocking prediction in Elastic Optical Networks (EONs) using Convolutional Neural Networks (CNNs). In EONs, connections are established and torn down dynamically to fulfill the instantaneous requirements of the users. The dynamic allocation of the connections may cause spectrum fragmentation and lead to network performance degradation as connection blocking increases. Predicting potential blocking situations can be helpful during EON operations. For example, this prediction could be used in real networks to trigger proper spectrum defragmentation mechanisms at suitable moments, thereby enhancing network performance. Extensive simulations over the well-known NSFNET (National Science Foundation Network) backbone network topology were run by generating realistic traffic patterns. The obtained results are later used to train the developed machine learning models, which allow the prediction of connection-blocking events. Resource use was continuously monitored and recorded during the process. Two different Convolutional Neural Network models, a 1D CNN (One-Dimensional Convolutional Neural Network) and 2D CNN (Two-Dimensional Convolutional Neural Network), are proposed as the predicting methods, and their behavior is compared to other conventional models based on an SVM (Support Vector Machine) and KNN (K Nearest Neighbors). The results obtained show that the proposed 2D CNN predicts blocking with the best accuracy (92.17%), followed by the SVM, the proposed 1D CNN, and KNN. Results suggest that 2D CNN can be helpful in blocking prediction and might contribute to increasing the efficiency of future EON networks.

Machine learning-driven approach for predicting the condensation heat transfer coefficient (HTC) in the presence of non-condensable gases

This research focuses on developing a predictive model that can effectively estimate the condensation heat transfer coefficient (HTC) on vertical tube surfaces during free-fall. The model takes into consideration the presence of non-condensable gases. The main objective is to assemble an extensive and varied database that encompasses different geometric parameters and operational conditions. The motivation here stems from the urgent need to develop practical models and/or correlations that can accurately predict the condensation heat transfer efficiency for industrial heat exchangers affected by non-condensable gases, with a particular focus on passive containment cooling systems (PCCS) employed in nuclear power plants (NPPs). This innovative cooling system utilizes the force of gravity to condense water steam, effectively removing heat from the containment vessel in the event of an accident. For the development of the model, we employed the neural network toolbox in MATLAB to construct an Artificial Neural Network (ANN) known as a Multi-Layer Perceptron (MLP) network. This model was designed to predict the HTC during the condensation process involving two different types of Non-Condensable Gases (NCG): Air and Nitrogen, along with Helium as a representative light gas. To build the model, we utilized a dataset comprising 1,888 data points sourced from 21 experimental references. The input layer of our model receives a range of parameters, encompassing total pressure (Ptot), subcooling temperature (ΔTsub), mass fraction of non-condensable gases (wnc), wall length (L), hydraulic diameter (Dh), the ratio of the Helium to non-condensable gases (RHe/nc), liquid density (ρl), gas density (ρg), liquid viscosity (μl), gas viscosity (μg), saturation temperature (Tsat), latent heat (hfg), liquid thermal conductivity (kl), and gas specific heat (cp,g). The output data of our model corresponds to the condensation HTC (h). To ensure consistent scaling, the input data was normalized by scaling each feature within the range of 0 to 1. On the other hand, the output data underwent a transformation using the natural logarithm. By utilizing Pearson correlation coefficient analysis and importance analysis for parameter selection, a new correlation for heat transfer was developed and evaluated. The developed machine learning model exhibited outstanding effectiveness in precisely forecasting the condensation HTC with an accuracy range of ± 10 % and ± 20 %. The outcome of this investigation represents the notable progress in analyzing extensive datasets gathered through experiments. Specifically, it addresses the previously unexplored influence of Helium gas by harnessing the capabilities of machine learning.

Reduced representative methylome profiling of cell-free DNA for breast cancer detection.

Whole-genome methylation sequencing of cfDNA is not cost-effective for tumor detection. Here, we introduce reduced representative methylome profiling (RRMP), which employs restriction enzyme for depletion of AT-rich sequence to achieve enrichment and deep sequencing of CG-rich sequences. We first verified the ability of RRMP to enrich CG-rich sequences using tumor cell genomic DNA and analyzed differential methylation regions between tumor cells and normal whole blood cells. We then analyzed cfDNA from 29 breast cancer patients and 27 non-breast cancer individuals to detect breast cancer by building machine learning models. RRMP captured 81.9% CpG islands and 75.2% gene promoters when sequenced to 10 billion base pairs, with an enrichment efficiency being comparable to RRBS. RRMP allowed us to assess DNA methylation changes between tumor cells and whole blood cells. Applying our approach to cfDNA from 29 breast cancer patients and 27 non-breast cancer individuals, we developed machine learning models that could discriminate between breast cancer and non-breast cancer controls (AUC = 0.85), suggesting possibilities for truly non-invasive cancer detection. We developed a new method to achieve reduced representative methylome profiling of cell-free DNA for tumor detection.

Exploration and machine learning model development for T2 NSCLC with bronchus infiltration and obstructive pneumonia/atelectasis

In the 8th edition of the American Joint Committee on Cancer (AJCC) staging system for Non-Small Cell Lung Cancer (NSCLC), tumors exhibiting main bronchial infiltration (MBI) near the carina and those presenting with complete lung obstructive pneumonia/atelectasis (P/ATL) have been reclassified from T3 to T2. Our investigation into the Surveillance, Epidemiology, and End Results (SEER) database, spanning from 2007 to 2015 and adjusted via Propensity Score Matching (PSM) for additional variables, disclosed a notably inferior overall survival (OS) for patients afflicted with these conditions. Specifically, individuals with P/ATL experienced a median OS of 12 months compared to 15 months (p < 0.001). In contrast, MBI patients demonstrated a slightly worse prognosis with a median OS of 22 months versus 23 months (p = 0.037), with both conditions significantly correlated with lymph node metastasis (All p < 0.001). Upon evaluating different treatment approaches for these particular T2 NSCLC variants, while adjusting for other factors, surgery emerged as the optimal therapeutic strategy. We counted those who underwent surgery and found that compared to surgery alone, the MBI/(P/ATL) group experienced a much higher proportion of preoperative induction therapy or postoperative adjuvant therapy than the non-MBI/(P/ATL) group (41.3%/54.7% vs. 36.6%). However, for MBI patients, initial surgery followed by adjuvant treatment or induction therapy succeeded in significantly enhancing prognosis, a benefit that was not replicated for P/ATL patients. Leveraging the XGBoost model for a 5-year survival forecast and treatment determination for P/ATL and MBI patients yielded Area Under the Curve (AUC) scores of 0.853 for P/ATL and 0.814 for MBI, affirming the model's efficacy in prognostication and treatment allocation for these distinct T2 NSCLC categories.

Prediction of chemical reaction yields with large-scale multi-view pre-training

Developing machine learning models with high generalization capability for predicting chemical reaction yields is of significant interest and importance. The efficacy of such models depends heavily on the representation of chemical reactions, which has commonly been learned from SMILES or graphs of molecules using deep neural networks. However, the progression of chemical reactions is inherently determined by the molecular 3D geometric properties, which have been recently highlighted as crucial features in accurately predicting molecular properties and chemical reactions. Additionally, large-scale pre-training has been shown to be essential in enhancing the generalization capability of complex deep learning models. Based on these considerations, we propose the Reaction Multi-View Pre-training (ReaMVP) framework, which leverages self-supervised learning techniques and a two-stage pre-training strategy to predict chemical reaction yields. By incorporating multi-view learning with 3D geometric information, ReaMVP achieves state-of-the-art performance on two benchmark datasets. Notably, the experimental results indicate that ReaMVP has a significant advantage in predicting out-of-sample data, suggesting an enhanced generalization ability to predict new reactions. Scientific Contribution: This study presents the ReaMVP framework, which improves the generalization capability of machine learning models for predicting chemical reaction yields. By integrating sequential and geometric views and leveraging self-supervised learning techniques with a two-stage pre-training strategy, ReaMVP achieves state-of-the-art performance on benchmark datasets. The framework demonstrates superior predictive ability for out-of-sample data and enhances the prediction of new reactions.

Machine Learning Prediction of Treatment Response to Inhaled Corticosteroids in Asthma.

Although inhaled corticosteroids (ICS) are the first-line therapy for patients with persistent asthma, many patients continue to have exacerbations. We developed machine learning models to predict the ICS response in patients with asthma. The subjects included asthma patients of European ancestry (n = 1371; 448 children; 916 adults). A genome-wide association study was performed to identify the SNPs associated with ICS response. Using the SNPs identified, two machine learning models were developed to predict ICS response: (1) least absolute shrinkage and selection operator (LASSO) regression and (2) random forest. The LASSO regression model achieved an AUC of 0.71 (95% CI 0.67-0.76; sensitivity: 0.57; specificity: 0.75) in an independent test cohort, and the random forest model achieved an AUC of 0.74 (95% CI 0.70-0.78; sensitivity: 0.70; specificity: 0.68). The genes contributing to the prediction of ICS response included those associated with ICS responses in asthma (TPSAB1, FBXL16), asthma symptoms and severity (ABCA7, CNN2, PTRN3, and BSG/CD147), airway remodeling (ELANE, FSTL3), mucin production (GAL3ST), leukotriene synthesis (GPX4), allergic asthma (ZFPM1, SBNO2), and others. An accurate risk prediction of ICS response can be obtained using machine learning methods, with the potential to inform personalized treatment decisions. Further studies are needed to examine if the integration of richer phenotype data could improve risk prediction.

Prediction of survival in out-of-hospital cardiac arrest: the updated Swedish cardiac arrest risk score (SCARS) model.

Out-of-hospital cardiac arrest (OHCA) is a major health concern worldwide. Although one-third of all patients achieve a return of spontaneous circulation and may undergo a difficult period in the intensive care unit, only 1 in 10 survive. This study aims to improve our previously developed machine learning model for early prognostication of survival in OHCA. We studied all cases registered in the Swedish Cardiopulmonary Resuscitation Registry during 2010 and 2020 (n = 55 615). We compared the predictive performance of extreme gradient boosting (XGB), light gradient boosting machine (LightGBM), logistic regression, CatBoost, random forest, and TabNet. For each framework, we developed models that optimized (i) a weighted F1 score to penalize models that yielded more false negatives and (ii) a precision-recall area under the curve (PR AUC). LightGBM assigned higher importance values to a larger set of variables, while XGB made predictions using fewer predictors. The area under the curve receiver operating characteristic (AUC ROC) scores for LightGBM were 0.958 (optimized for weighted F1) and 0.961 (optimized for a PR AUC), while for XGB, the scores were 0.958 and 0.960, respectively. The calibration plots showed a subtle underestimation of survival for LightGBM, contrasting with a mild overestimation for XGB models. In the crucial range of 0-10% likelihood of survival, the XGB model, optimized with the PR AUC, emerged as a clinically safe model. We improved our previous prediction model by creating a parsimonious model with an AUC ROC at 0.96, with excellent calibration and no apparent risk of underestimating survival in the critical probability range (0-10%). The model is available at www.gocares.se.

“DompeKeys”: a set of novel substructure-based descriptors for efficient chemical space mapping, development and structural interpretation of machine learning models, and indexing of large databases

The conversion of chemical structures into computer-readable descriptors, able to capture key structural aspects, is of pivotal importance in the field of cheminformatics and computer-aided drug design. Molecular fingerprints represent a widely employed class of descriptors; however, their generation process is time-consuming for large databases and is susceptible to bias. Therefore, descriptors able to accurately detect predefined structural fragments and devoid of lengthy generation procedures would be highly desirable. To meet additional needs, such descriptors should also be interpretable by medicinal chemists, and suitable for indexing databases with trillions of compounds. To this end, we developed—as integral part of EXSCALATE, Dompé’s end-to-end drug discovery platform—the DompeKeys (DK), a new substructure-based descriptor set, which encodes the chemical features that characterize compounds of pharmaceutical interest. DK represent an exhaustive collection of curated SMARTS strings, defining chemical features at different levels of complexity, from specific functional groups and structural patterns to simpler pharmacophoric points, corresponding to a network of hierarchically interconnected substructures. Because of their extended and hierarchical structure, DK can be used, with good performance, in different kinds of applications. In particular, we demonstrate how they are very well suited for effective mapping of chemical space, as well as substructure search and virtual screening. Notably, the incorporation of DK yields highly performing machine learning models for the prediction of both compounds’ activity and metabolic reaction occurrence. The protocol to generate the DK is freely available at https://dompekeys.exscalate.eu and is fully integrated with the Molecular Anatomy protocol for the generation and analysis of hierarchically interconnected molecular scaffolds and frameworks, thus providing a comprehensive and flexible tool for drug design applications.

Leveraging explainable artificial intelligence to optimize clinical decision support.

To develop and evaluate a data-driven process to generate suggestions for improving alert criteria using explainable artificial intelligence (XAI) approaches. We extracted data on alerts generated from January 1, 2019 to December 31, 2020, at Vanderbilt University Medical Center. We developed machine learning models to predict user responses to alerts. We applied XAI techniques to generate global explanations and local explanations. We evaluated the generated suggestions by comparing with alert's historical change logs and stakeholder interviews. Suggestions that either matched (or partially matched) changes already made to the alert or were considered clinically correct were classified as helpful. The final dataset included 2 991 823 firings with 2689 features. Among the 5 machine learning models, the LightGBM model achieved the highest Area under the ROC Curve: 0.919 [0.918, 0.920]. We identified 96 helpful suggestions. A total of 278 807 firings (9.3%) could have been eliminated. Some of the suggestions also revealed workflow and education issues. We developed a data-driven process to generate suggestions for improving alert criteria using XAI techniques. Our approach could identify improvements regarding clinical decision support (CDS) that might be overlooked or delayed in manual reviews. It also unveils a secondary purpose for the XAI: to improve quality by discovering scenarios where CDS alerts are not accepted due to workflow, education, or staffing issues.

Clinical and neuroimaging predictors of benzodiazepine response in catatonia: A machine learning approach

Catatonia is a well characterized psychomotor syndrome combining motor, behavioural and neurovegetative signs. Benzodiazepines are the first-choice treatment, effective in 70 % of cases. Currently, the factors associated with benzodiazepine resistance remain unknown. We aimed to develop machine learning models using clinical and neuroimaging data to predict benzodiazepine response in catatonic patients. This study examined a cohort of catatonic patients who underwent standardized clinical evaluation, 3 T brain MRI, and benzodiazepine trial. Based on clinical response, patients were classified as benzodiazepine responders or non-responders. Cortical thickness and regional brain volumes were measured. Two machine learning models (linear model and gradient boosting tree model) were developed to identify predictors of treatment response using clinical, demographic, and neuroimaging data. The cohort included 65 catatonic patients, comprising 30 benzodiazepine responders and 35 non-responders. Using clinical data alone, the linear model achieved 63% precision, 51% recall, a specificity of 61%, and 58% AUC, while the gradient boosting tree (GBT) model attained 46% precision, 60% recall, a specificity of 62% and 64% AUC. Incorporating neuroimaging data improved model performance, with the linear model achieving 66% precision, 57% recall, a specificity of 67%, and 70% AUC, and the GBT model attaining 50% precision, 50% recall, a specificity of 62% and 70% AUC. The integration of imaging data with demographic and clinical information significantly enhanced the predictive performance of the models. The duration of the catatonic syndrome, along with the presence of mitgehen (passive obedience) and immobility/stupor, and the volume of the right medial orbito-frontal cortex emerged as important factors in predicting non-response to benzodiazepines.

Predicting failure pressure of corroded gas pipelines: A data-driven approach using machine learning

This paper presents a novel methodology utilizing machine learning (ML) techniques to accurately forecast the failure pressure of corroded pipelines, which are pivotal for oil, gas, and petroleum product transport. In contrast to prior approaches, this study overcomes their limitations by integrating physically significant factors tied to corroded pipeline failure mechanisms and delivering interpretable models. Robust ML models are constructed using dependable experimental data from pertinent literature sources, ensuring research rigor. To bolster model interpretability, the SHAP method is deployed, facilitating a comprehensive grasp of each feature's impact on predictions. The performance and efficacy of the developed ML models are rigorously assessed using real-world pipeline incidents reported by the PHMSA in the US, further validating their practical utility. The findings not only affirm the models' precision and dependability but also yield valuable insights, underscoring their importance. This study makes a substantial contribution to the enhancement of pipeline integrity management by enabling precise failure pressure predictions in corroded pipelines. The insights gleaned from this research hold practical implications, enabling well-informed choices in pipeline design, integrity management, and risk assessment. Thus, it stands as a valuable resource for pipeline operators aiming to optimize resource allocation and bolster overall pipeline safety and efficiency.